Attempt to build manylinux wheel

Allow to use vosk-server for transcription

.gitignore

@@ -10,6 +10,7 @@ gradlew
 gradlew.bat
 gradle
 local.properties
+gradle.properties
 
 # Android
 android/build

CMakeLists.txt

@@ -0,0 +1,20 @@
+cmake_minimum_required(VERSION 3.13)
+project(vosk-api CXX)
+
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_EXTENSIONS OFF)
+
+add_library(vosk
+  src/language_model.cc
+  src/model.cc
+  src/recognizer.cc
+  src/spk_model.cc
+  src/vosk_api.cc
+)
+
+find_package(kaldi REQUIRED)
+target_link_libraries(vosk PUBLIC kaldi-base kaldi-online2 kaldi-rnnlm fstngram)
+
+include(GNUInstallDirs)
+install(TARGETS vosk DESTINATION ${CMAKE_INSTALL_LIBDIR})
+install(FILES src/vosk_api.h DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})

README.md

@@ -1,17 +1,18 @@
-# About
+# Vosk Speech Recognition Toolkit
 
 Vosk is an offline open source speech recognition toolkit. It enables
-speech recognition models for 18 languages and dialects - English, Indian
+speech recognition for 20+ languages and dialects - English, Indian
 English, German, French, Spanish, Portuguese, Chinese, Russian, Turkish,
 Vietnamese, Italian, Dutch, Catalan, Arabic, Greek, Farsi, Filipino,
-Ukrainian.
+Ukrainian, Kazakh, Swedish, Japanese, Esperanto, Hindi, Czech, Polish.
+More to come.
 
 Vosk models are small (50 Mb) but provide continuous large vocabulary
 transcription, zero-latency response with streaming API, reconfigurable
 vocabulary and speaker identification.
 
 Speech recognition bindings implemented for various programming languages
-like Python, Java, Node.JS, C#, C++ and others.
+like Python, Java, Node.JS, C#, C++, Rust, Go and others.
 
 Vosk supplies speech recognition for chatbots, smart home appliances,
 virtual assistants. It can also create subtitles for movies,

android/build.gradle

@@ -1,25 +1,36 @@
 buildscript {
     repositories {
         google()
-        jcenter()
+        mavenCentral()
     }
     dependencies {
-        classpath 'com.android.tools.build:gradle:4.1.3'
+        classpath 'com.android.tools.build:gradle:4.2.0'
+        classpath 'com.vanniktech:gradle-maven-publish-plugin:0.18.0'
     }
 }
 
 allprojects {
-    version = '0.3.30'
+    version = '0.3.43'
 }
 
 subprojects {
 
     apply plugin: 'com.android.library'
     apply plugin: 'maven-publish'
+    apply plugin: 'com.vanniktech.maven.publish'
+
+    plugins.withId('com.vanniktech.maven.publish') {
+        mavenPublish {
+            group = 'com.alphacephei'
+            version = version
+            sonatypeHost = 's01'
+            androidVariantToPublish = 'release'
+        }
+    }
 
     repositories {
         google()
-        jcenter()
+        mavenCentral()
     }
 
     publishing {
@@ -29,8 +40,8 @@ subprojects {
                 version version
                 pom {
                     url = 'http://www.alphacephei.com.com/vosk/'
-                        licenses {
-                            license {
+                    licenses {
+                        license {
                             name = 'The Apache License, Version 2.0'
                             url = 'http://www.apache.org/licenses/LICENSE-2.0.txt'
                         }
@@ -49,10 +60,6 @@ subprojects {
                 }
             }
         }
-        repositories {
-            maven {
-                url = "$rootDir/repo"
-            }
-        }
     }
+
 }

android/lib/build-vosk.sh

@@ -29,7 +29,7 @@ set -x
 OS_NAME=`echo $(uname -s) | tr '[:upper:]' '[:lower:]'`
 ANDROID_TOOLCHAIN_PATH=$ANDROID_NDK_HOME/toolchains/llvm/prebuilt/${OS_NAME}-x86_64
 WORKDIR_BASE=`pwd`/build
-PATH=$PATH:$ANDROID_NDK_HOME/toolchains/llvm/prebuilt/${OS_NAME}-x86_64/bin
+PATH=$ANDROID_NDK_HOME/toolchains/llvm/prebuilt/${OS_NAME}-x86_64/bin:$PATH
 OPENFST_VERSION=1.8.0
 
 for arch in armeabi-v7a arm64-v8a x86_64 x86; do
@@ -40,7 +40,8 @@ case $arch in
     armeabi-v7a)
           BLAS_ARCH=ARMV7
           HOST=arm-linux-androideabi
-          AR=arm-linux-androideabi-ar
+          AR=llvm-ar
+          RANLIB=llvm-ranlib
           CC=armv7a-linux-androideabi21-clang
           CXX=armv7a-linux-androideabi21-clang++
           ARCHFLAGS="-mfloat-abi=softfp -mfpu=neon"
@@ -48,7 +49,8 @@ case $arch in
     arm64-v8a)
           BLAS_ARCH=ARMV8
           HOST=aarch64-linux-android
-          AR=aarch64-linux-android-ar
+          AR=llvm-ar
+          RANLIB=llvm-ranlib
           CC=aarch64-linux-android21-clang
           CXX=aarch64-linux-android21-clang++
           ARCHFLAGS=""
@@ -56,7 +58,8 @@ case $arch in
     x86_64)
           BLAS_ARCH=ATOM
           HOST=x86_64-linux-android
-          AR=x86_64-linux-android-ar
+          AR=llvm-ar
+          RANLIB=llvm-ranlib
           CC=x86_64-linux-android21-clang
           CXX=x86_64-linux-android21-clang++
           ARCHFLAGS=""
@@ -64,7 +67,8 @@ case $arch in
     x86)
           BLAS_ARCH=ATOM
           HOST=i686-linux-android
-          AR=i686-linux-android-ar
+          AR=llvm-ar
+          RANLIB=llvm-ranlib
           CC=i686-linux-android21-clang
           CXX=i686-linux-android21-clang++
           ARCHFLAGS=""
@@ -105,9 +109,9 @@ make install
 
 # Kaldi itself
 cd $WORKDIR
-git clone -b android-mix --single-branch https://github.com/alphacep/kaldi
+git clone -b vosk-android --single-branch https://github.com/alphacep/kaldi
 cd $WORKDIR/kaldi/src
-CXX=$CXX CXXFLAGS="$ARCHFLAGS -O3 -DFST_NO_DYNAMIC_LINKING" ./configure --use-cuda=no \
+CXX=$CXX AR=$AR RANLIB=$RANLIB CXXFLAGS="$ARCHFLAGS -O3 -DFST_NO_DYNAMIC_LINKING" ./configure --use-cuda=no \
     --mathlib=OPENBLAS_CLAPACK --shared \
     --android-incdir=${ANDROID_TOOLCHAIN_PATH}/sysroot/usr/include \
     --host=$HOST --openblas-root=${WORKDIR}/local \
@@ -118,12 +122,14 @@ make -j 8 online2 lm rnnlm
 
 # Vosk-api
 cd $WORKDIR
-#rm -rf vosk-api
-git clone -b master --single-branch https://github.com/alphacep/vosk-api
-cd vosk-api/src
-make -j 8 KALDI_ROOT=${WORKDIR}/kaldi OPENFST_ROOT=${WORKDIR}/local OPENBLAS_ROOT=${WORKDIR}/local CXX=$CXX EXTRA_LDFLAGS="-llog -static-libstdc++"
-
-# Copy JNI library to sources
-cp $WORKDIR/vosk-api/src/libvosk.so $WORKDIR/../../src/main/jniLibs/$arch/libvosk.so
+mkdir -p $WORKDIR/vosk
+make -j 8 -C ${WORKDIR_BASE}/../../../src \
+    OUTDIR=$WORKDIR/vosk \
+    KALDI_ROOT=${WORKDIR}/kaldi \
+    OPENFST_ROOT=${WORKDIR}/local \
+    OPENBLAS_ROOT=${WORKDIR}/local \
+    CXX=$CXX \
+    EXTRA_LDFLAGS="-llog -static-libstdc++ -Wl,-soname,libvosk.so"
+cp $WORKDIR/vosk/libvosk.so $WORKDIR/../../src/main/jniLibs/$arch/libvosk.so
 
 done

android/lib/build.gradle

@@ -10,6 +10,7 @@ android {
         versionCode 6
         versionName = version
         archivesBaseName = archiveName
+        ndkVersion = "22.1.7171670"
     }
     compileOptions {
         sourceCompatibility JavaVersion.VERSION_1_8
@@ -19,11 +20,11 @@ android {
 
 task buildVosk(type: Exec) {
     commandLine './build-vosk.sh'
-    environment ANDROID_NDK_HOME: android.getSdkDirectory()
+    environment ANDROID_NDK_HOME: android.getNdkDirectory()
 }
 
 dependencies {
-    implementation 'net.java.dev.jna:jna:4.4.0@aar'
+    api 'net.java.dev.jna:jna:4.4.0@aar'
 }
 
 //preBuild.dependsOn buildVosk
@@ -37,6 +38,18 @@ publishing {
                 name = pomName
                 description = pomDescription
             }
+            //generate pom nodes for dependencies
+            pom.withXml {
+                def dependenciesNode = asNode().appendNode('dependencies')
+                configurations.implementation.allDependencies.each { dependency ->
+                    if (dependency.name != 'unspecified') {
+                        def dependencyNode = dependenciesNode.appendNode('dependency')
+                        dependencyNode.appendNode('groupId', dependency.group)
+                        dependencyNode.appendNode('artifactId', dependency.name)
+                        dependencyNode.appendNode('version', dependency.version)
+                    }
+                }
+            }
         }
     }
 }

android/lib/settings.gradle

@@ -1 +0,0 @@
-include 'model-en'

android/lib/src/main/java/org/vosk/LibVosk.java

@@ -36,6 +36,8 @@ public class LibVosk {
 
     public static native void vosk_recognizer_set_words(Pointer recognizer, boolean words);
 
+    public static native void vosk_recognizer_set_partial_words(Pointer recognizer, boolean partial_words);
+
     public static native void vosk_recognizer_set_spk_model(Pointer recognizer, Pointer spk_model);
 
     public static native boolean vosk_recognizer_accept_waveform(Pointer recognizer, byte[] data, int len);

android/lib/src/main/java/org/vosk/Recognizer.java

@@ -23,6 +23,10 @@ public class Recognizer extends PointerType implements AutoCloseable {
         LibVosk.vosk_recognizer_set_words(this.getPointer(), words);
     }
 
+    public void setPartialWords(boolean partial_words) {
+        LibVosk.vosk_recognizer_set_partial_words(this.getPointer(), partial_words);
+    }
+
     public void setSpeakerModel(SpeakerModel spkModel) {
         LibVosk.vosk_recognizer_set_spk_model(this.getPointer(), spkModel.getPointer());
     }

c/Makefile

@@ -1,16 +1,16 @@
 CFLAGS=-I../src
-LDFLAGS=-L../src -lvosk -ldl -lpthread -Wl,-rpath=../src
+LDFLAGS=-L../src -lvosk -ldl -lpthread -Wl,-rpath,../src
 
 all: test_vosk test_vosk_speaker
 
 test_vosk: test_vosk.o
-	g++ $^ -o $@ $(LDFLAGS)
+	gcc $^ -o $@ $(LDFLAGS)
 
 test_vosk_speaker: test_vosk_speaker.o
-	g++ $^ -o $@ $(LDFLAGS)
+	gcc $^ -o $@ $(LDFLAGS)
 
 %.o: %.c
-	g++ $(CFLAGS) -c -o $@ $<
+	gcc $(CFLAGS) -c -o $@ $<
 
 clean:
 	rm -f *.o *.a test_vosk test_vosk_speaker

csharp/demo/VoskDemo.cs

@@ -34,7 +34,7 @@ public class VoskDemo
             while((bytesRead = source.Read(buffer, 0, buffer.Length)) > 0) {
                 float[] fbuffer = new float[bytesRead / 2];
                 for (int i = 0, n = 0; i < fbuffer.Length; i++, n+=2) {
-                    fbuffer[i] = (short)(buffer[n] | buffer[n+1] << 8);
+                    fbuffer[i] = BitConverter.ToInt16(buffer, n);
                 }
                 if (rec.AcceptWaveform(fbuffer, fbuffer.Length)) {
                     Console.WriteLine(rec.Result());

csharp/demo/VoskDemo.csproj

@@ -11,7 +11,7 @@
   </PropertyGroup>
 
   <ItemGroup>
-    <PackageReference Include="Vosk" Version="0.3.30" />
+    <PackageReference Include="Vosk" Version="0.3.43" />
   </ItemGroup>
 
 </Project>

csharp/nuget/Vosk.nuspec

@@ -2,13 +2,14 @@
 <package>
   <metadata>
     <id>Vosk</id>
-    <version>0.3.30</version>
+    <version>0.3.43</version>
     <authors>Alpha Cephei Inc</authors>
     <owners>Alpha Cephei Inc</owners>
     <license type="expression">Apache-2.0</license>
     <projectUrl>https://alphacephei.com/vosk/</projectUrl>
     <requireLicenseAcceptance>false</requireLicenseAcceptance>
-    <description>Vosk is an offline open source speech recognition toolkit. It enables speech recognition models for 16 languages and dialects - English, Indian English, German, French, Spanish, Portuguese, Chinese, Russian, Turkish, Vietnamese, Italian, Dutch, Catalan, Arabic, Greek, Farsi.
+    <title>Vosk Speech Recognition Toolkit</title>
+    <description>Vosk is an offline open source speech recognition toolkit. It enables speech recognition models for 20+ languages and dialects - English, Indian English, German, French, Spanish, Portuguese, Chinese, Russian, Turkish, Vietnamese, Italian, Dutch, Catalan, Arabic, Greek, Farsi, Filipino, Ukrainian, Kazakh, Swedish, Japanese, Esperanto, Hindi, Czech. More to come.
 
 Vosk models are small (50 Mb) but provide continuous large vocabulary transcription, zero-latency response with streaming API, reconfigurable vocabulary and speaker identification.
 
@@ -18,7 +19,8 @@ Vosk supplies speech recognition for chatbots, smart home appliances, virtual as
 
 Vosk scales from small devices like Raspberry Pi or Android smartphone to big clusters.</description>
     <releaseNotes>See for details https://github.com/alphacep/vosk-api/releases</releaseNotes>
-    <copyright>Copyright 2020 Alpha Cephei Inc</copyright>
+    <repository type="git" url="https://github.com/alphacep/vosk-api.git" branch="master"/>
+    <copyright>Copyright 2020-2050 Alpha Cephei Inc</copyright>
     <tags>speech recognition voice stt asr speech-to-text ai offline privacy</tags>
     <dependencies>
       <group targetFramework=".NETStandard2.0"/>

csharp/nuget/build/Vosk.targets

@@ -2,7 +2,7 @@
   <ItemGroup>
     <NativeLibs Include="$(MSBuildThisFileDirectory)\lib\linux-x64\*.so" Condition="'$([MSBuild]::IsOsPlatform(Linux))'" />
     <NativeLibs Include="$(MSBuildThisFileDirectory)\lib\win-x64\*.dll" Condition="'$([MSBuild]::IsOsPlatform(Windows))'" />
-    <NativeLibs Include="$(MSBuildThisFileDirectory)\lib\osx-x64\*.dylib" Condition="'$([MSBuild]::IsOsPlatform(OSX))'" />
+    <NativeLibs Include="$(MSBuildThisFileDirectory)\lib\osx-universal\*.dylib" Condition="'$([MSBuild]::IsOsPlatform(OSX))'" />
     <None Include="@(NativeLibs)">
       <Link>%(FileName)%(Extension)</Link>
       <CopyToOutputDirectory>PreserveNewest</CopyToOutputDirectory>

csharp/nuget/src/Model.cs

@@ -32,7 +32,7 @@ public class Model : global::System.IDisposable {
   public Model(string model_path) : this(VoskPINVOKE.new_Model(model_path)) {
   }
 
-  public int vosk_model_find_word(string word) {
+  public int FindWord(string word) {
     return VoskPINVOKE.Model_vosk_model_find_word(handle, word);
   }
 

csharp/nuget/src/VoskPINVOKE.cs

@@ -38,6 +38,9 @@ class VoskPINVOKE {
   [global::System.Runtime.InteropServices.DllImport("libvosk", EntryPoint="vosk_recognizer_set_words")]
   public static extern void VoskRecognizer_SetWords(global::System.Runtime.InteropServices.HandleRef jarg1, int jarg2);
 
+  [global::System.Runtime.InteropServices.DllImport("libvosk", EntryPoint="vosk_recognizer_set_partial_words")]
+  public static extern void VoskRecognizer_SetPartialWords(global::System.Runtime.InteropServices.HandleRef jarg1, int jarg2);
+
   [global::System.Runtime.InteropServices.DllImport("libvosk", EntryPoint="vosk_recognizer_set_spk_model")]
   public static extern void VoskRecognizer_SetSpkModel(global::System.Runtime.InteropServices.HandleRef jarg1, global::System.Runtime.InteropServices.HandleRef jarg2);
 

csharp/nuget/src/VoskRecognizer.cs

@@ -46,6 +46,10 @@ public class VoskRecognizer : System.IDisposable {
     VoskPINVOKE.VoskRecognizer_SetWords(handle, words ? 1 : 0);
   }
 
+  public void SetPartialWords(bool partial_words) {
+    VoskPINVOKE.VoskRecognizer_SetPartialWords(handle, partial_words ? 1 : 0);
+  }
+
   public void SetSpkModel(SpkModel spk_model) {
     VoskPINVOKE.VoskRecognizer_SetSpkModel(handle, SpkModel.getCPtr(spk_model));
   }

go/COPYING

@@ -0,0 +1,176 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. this License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS

go/README.md

@@ -0,0 +1 @@
+See example subfolder for instructions how to use the module

go/doc.go

@@ -0,0 +1,7 @@
+// Go bindings for Vosk speech recognition toolkit. Vosk is an offline
+// open source speech to text API for Android, iOS, Raspberry Pi and
+// servers. It enables speech recognition models for 18 languages and
+// dialects - English, Indian English, German, French, Spanish, Portuguese,
+// Chinese, Russian, Turkish, Vietnamese, Italian, Dutch, Catalan, Arabic,
+// Greek, Farsi, Filipino, Ukrainian.
+package vosk

go/example/README.md

@@ -0,0 +1,31 @@
+To try this package do the following steps:
+
+On Linux (we download library and set LD_LIBRARY_PATH)
+
+```
+git clone https://github.com/alphacep/vosk-api
+cd vosk-api/go/example
+wget https://github.com/alphacep/vosk-api/releases/download/v0.3.42/vosk-linux-x86_64-0.3.42.zip
+unzip vosk-linux-x86_64-0.3.42.zip
+wget https://alphacephei.com/vosk/models/vosk-model-small-en-us-0.15.zip
+unzip vosk-model-small-en-us-0.15.zip
+mv vosk-model-small-en-us-0.15 model
+cp ../../python/example/test.wav .
+VOSK_PATH=`pwd`/vosk-linux-x86_64-0.3.42 LD_LIBRARY_PATH=$VOSK_PATH CGO_CPPFLAGS="-I $VOSK_PATH" CGO_LDFLAGS="-L $VOSK_PATH" go run . -f test.wav
+```
+
+for Windows (we place DLLs in current folder where linker finds them):
+
+```
+git clone https://github.com/alphacep/vosk-api
+cd vosk-api/go/example
+wget https://github.com/alphacep/vosk-api/releases/download/v0.3.42/vosk-linux-x86_64-0.3.42.zip
+unzip vosk-linux-x86_64-0.3.42.zip
+cp vosk-linux-x86_64-0.3.42/*.dll .
+cp vosk-linux-x86_64-0.3.42/*.h .
+wget https://alphacephei.com/vosk/models/vosk-model-small-en-us-0.15.zip
+unzip vosk-model-small-en-us-0.15.zip
+mv vosk-model-small-en-us-0.15 model
+cp ../../python/example/test.wav .
+VOSK_PATH=`pwd` LD_LIBRARY_PATH=$VOSK_PATH CGO_CPPFLAGS="-I $VOSK_PATH" CGO_LDFLAGS="-L $VOSK_PATH -lvosk -lpthread -dl" go run . -f test.wav
+```

go/example/doc.go

@@ -0,0 +1,2 @@
+// Example package for Vosk Go bindings.
+package main

go/example/main.go

@@ -1,36 +0,0 @@
-package main
-
-import (
-    "flag"
-    "os"
-    ".."
-)
-
-func main() {
-    var filename string
-    flag.StringVar(&filename, "f", "", "file to transcribe")
-    flag.Parse()
-    model, err := vosk.NewModel("model")
-    rec, err := vosk.NewRecognizer(model)
-
-    file, err := os.Open(filename)
-    if err != nil {
-        panic(err)
-    }
-    defer file.Close()
-
-    fileinfo, err := file.Stat()
-    if err != nil {
-        panic(err)
-    }
-
-    filesize := fileinfo.Size()
-    buffer := make([]byte, filesize)
-
-    _, err = file.Read(buffer)
-    if err != nil {
-        panic(err)
-    }
-
-    println(vosk.VoskFinalResult(rec, buffer))
-}

go/example/test_simple.go

@@ -0,0 +1,56 @@
+package main
+
+import (
+	"bufio"
+	"flag"
+	"fmt"
+	"io"
+	"log"
+	"os"
+
+	vosk "github.com/alphacep/vosk-api/go"
+)
+
+func main() {
+	var filename string
+	flag.StringVar(&filename, "f", "", "file to transcribe")
+	flag.Parse()
+
+	model, err := vosk.NewModel("model")
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	sampleRate := 16000.0
+	rec, err := vosk.NewRecognizer(model, sampleRate)
+	if err != nil {
+		log.Fatal(err)
+	}
+	rec.SetWords(1)
+
+	file, err := os.Open(filename)
+	if err != nil {
+		panic(err)
+	}
+	defer file.Close()
+
+	reader := bufio.NewReader(file)
+	buf := make([]byte, 4096)
+
+	for {
+		_, err := reader.Read(buf)
+		if err != nil {
+			if err != io.EOF {
+				log.Fatal(err)
+			}
+
+			break
+		}
+
+		if rec.AcceptWaveform(buf) != 0 {
+			fmt.Println(string(rec.Result()))
+		}
+	}
+
+	fmt.Println(string(rec.FinalResult()))
+}

go/go.mod

@@ -0,0 +1,7 @@
+module github.com/alphacep/vosk-api/go
+
+go 1.16
+
+replace (
+    github.com/alphacep/vosk-api/go => ./
+)

go/vosk.go

@@ -8,55 +8,145 @@ import "C"
 
 // VoskModel contains a reference to the C VoskModel
 type VoskModel struct {
-    model *C.struct_VoskModel
-}
-
-// VoskSpkModel contains a reference to the C VoskSpkModel
-type VoskSpkModel struct {
-    spkModel *C.struct_VoskSpkModel
-}
-
-// VoskRecognizer contains a reference to the C VoskRecognizer
-type VoskRecognizer struct {
-    rec *C.struct_VoskRecognizer
-}
-
-func VoskFinalResult(recognizer *VoskRecognizer, buffer []byte) string {
-    cbuf := C.CBytes(buffer)
-    defer C.free(cbuf)
-    _ = C.vosk_recognizer_accept_waveform(recognizer.rec, (*C.char)(cbuf), C.int(len(buffer)))
-    result := C.GoString(C.vosk_recognizer_final_result(recognizer.rec))
-    return result
+	model *C.struct_VoskModel
 }
 
 // NewModel creates a new VoskModel instance
 func NewModel(modelPath string) (*VoskModel, error) {
-    var internal *C.struct_VoskModel
-    internal = C.vosk_model_new(C.CString(modelPath))
-    model := &VoskModel{model: internal}
-    return model, nil
+	internal := C.vosk_model_new(C.CString(modelPath))
+	model := &VoskModel{model: internal}
+	return model, nil
 }
 
-// NewRecognizer creates a new VoskRecognizer instance
-func NewRecognizer(model *VoskModel) (*VoskRecognizer, error) {
-    var internal *C.struct_VoskRecognizer
-    internal = C.vosk_recognizer_new(model.model, 16000.0)
-    rec := &VoskRecognizer{rec: internal}
-    return rec, nil
+func (m *VoskModel) Free() {
+	C.vosk_model_free(m.model)
 }
 
 func freeModel(model *VoskModel) {
-    C.vosk_model_free(model.model)
+	C.vosk_model_free(model.model)
 }
 
-func freeRecognizer(recognizer *VoskRecognizer) {
-    C.vosk_recognizer_free(recognizer.rec)
+// FindWord checks if a word can be recognized by the model.
+// Returns the word symbol if the word exists inside the model or
+// -1 otherwise.
+func (m *VoskModel) FindWord(word []byte) int {
+	cbuf := C.CBytes(word)
+	defer C.free(cbuf)
+	i := C.vosk_model_find_word(m.model, (*C.char)(cbuf))
+	return int(i)
+}
+
+// VoskSpkModel contains a reference to the C VoskSpkModel
+type VoskSpkModel struct {
+	spkModel *C.struct_VoskSpkModel
 }
 
 // NewSpkModel creates a new VoskSpkModel instance
 func NewSpkModel(spkModelPath string) (*VoskSpkModel, error) {
-    var internal *C.struct_VoskSpkModel
-    internal = C.vosk_spk_model_new(C.CString(spkModelPath))
-    spkModel := &VoskSpkModel{spkModel: internal}
-    return spkModel, nil
+	internal := C.vosk_spk_model_new(C.CString(spkModelPath))
+	spkModel := &VoskSpkModel{spkModel: internal}
+	return spkModel, nil
+}
+
+func freeSpkModel(model *VoskSpkModel) {
+	C.vosk_spk_model_free(model.spkModel)
+}
+
+func(s *VoskSpkModel) Free() {
+	C.vosk_spk_model_free(s.spkModel)
+}
+
+// VoskRecognizer contains a reference to the C VoskRecognizer
+type VoskRecognizer struct {
+	rec *C.struct_VoskRecognizer
+}
+
+func freeRecognizer(recognizer *VoskRecognizer) {
+	C.vosk_recognizer_free(recognizer.rec)
+}
+
+func (r *VoskRecognizer) Free() {
+	C.vosk_recognizer_free(r.rec)
+}
+
+// NewRecognizer creates a new VoskRecognizer instance
+func NewRecognizer(model *VoskModel, sampleRate float64) (*VoskRecognizer, error) {
+	internal := C.vosk_recognizer_new(model.model, C.float(sampleRate))
+	rec := &VoskRecognizer{rec: internal}
+	return rec, nil
+}
+
+// NewRecognizerSpk creates a new VoskRecognizer instance with a speaker model.
+func NewRecognizerSpk(model *VoskModel, sampleRate float64, spkModel *VoskSpkModel) (*VoskRecognizer, error) {
+	internal := C.vosk_recognizer_new_spk(model.model, C.float(sampleRate), spkModel.spkModel)
+	rec := &VoskRecognizer{rec: internal}
+	return rec, nil
+}
+
+// NewRecognizerGrm creates a new VoskRecognizer instance with the phrase list.
+func NewRecognizerGrm(model *VoskModel, sampleRate float64, grammer []byte) (*VoskRecognizer, error) {
+	cbuf := C.CBytes(grammer)
+	defer C.free(cbuf)
+	internal := C.vosk_recognizer_new_grm(model.model, C.float(sampleRate), (*C.char)(cbuf))
+	rec := &VoskRecognizer{rec: internal}
+	return rec, nil
+}
+
+// SetSpkModel adds a speaker model to an already initialized recognizer.
+func (r *VoskRecognizer) SetSpkModel(spkModel *VoskSpkModel) {
+	C.vosk_recognizer_set_spk_model(r.rec, spkModel.spkModel)
+}
+
+// SetMaxAlternatives configures the recognizer to output n-best results.
+func (r *VoskRecognizer) SetMaxAlternatives(maxAlternatives int) {
+	C.vosk_recognizer_set_max_alternatives(r.rec, C.int(maxAlternatives))
+}
+
+// SetWords enables words with times in the ouput.
+func (r *VoskRecognizer) SetWords(words int) {
+	C.vosk_recognizer_set_words(r.rec, C.int(words))
+}
+
+// AcceptWaveform accepts and processes a new chunk of the voice data.
+func (r *VoskRecognizer) AcceptWaveform(buffer []byte) int {
+	cbuf := C.CBytes(buffer)
+	defer C.free(cbuf)
+	i := C.vosk_recognizer_accept_waveform(r.rec, (*C.char)(cbuf), C.int(len(buffer)))
+	return int(i)
+}
+
+// Result returns a speech recognition result.
+func (r *VoskRecognizer) Result() []byte {
+	return []byte(C.GoString(C.vosk_recognizer_result(r.rec)))
+}
+
+// PartialResult returns a partial speech recognition result.
+func (r *VoskRecognizer) PartialResult() []byte {
+	return []byte(C.GoString(C.vosk_recognizer_partial_result(r.rec)))
+}
+
+// FinalResult returns a speech recognition result. Same as result, but doesn't wait
+// for silence.
+func (r *VoskRecognizer) FinalResult() []byte {
+	return []byte(C.GoString(C.vosk_recognizer_final_result(r.rec)))
+}
+
+// Reset resets the recognizer.
+func (r *VoskRecognizer) Reset() {
+	C.vosk_recognizer_reset(r.rec)
+}
+
+// SetLogLevel sets the log level for Kaldi messages.
+func SetLogLevel(logLevel int) {
+	C.vosk_set_log_level(C.int(logLevel))
+}
+
+// GPUInit automatically selects a CUDA device and allows multithreading.
+func GPUInit() {
+	C.vosk_gpu_init()
+}
+
+// GPUThreadInit inits CUDA device in a multi-threaded environment.
+func GPUThreadInit() {
+	C.vosk_gpu_thread_init()
 }

ios/VoskApiTest.xcodeproj/project.pbxproj

@@ -13,10 +13,10 @@
 		92375229240C550B00DD6076 /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 92375228240C550B00DD6076 /* Assets.xcassets */; };
 		9237522C240C550B00DD6076 /* LaunchScreen.storyboard in Resources */ = {isa = PBXBuildFile; fileRef = 9237522A240C550B00DD6076 /* LaunchScreen.storyboard */; };
 		92375234240C558900DD6076 /* Vosk.swift in Sources */ = {isa = PBXBuildFile; fileRef = 92375233240C558900DD6076 /* Vosk.swift */; };
-		9237523C240C642000DD6076 /* libkaldiwrap.a in Frameworks */ = {isa = PBXBuildFile; fileRef = 9237523A240C642000DD6076 /* libkaldiwrap.a */; };
 		92375244240C6DAF00DD6076 /* Accelerate.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 92375243240C6DAF00DD6076 /* Accelerate.framework */; };
-		92375246240C6DC900DD6076 /* libstdc++.tbd in Frameworks */ = {isa = PBXBuildFile; fileRef = 92375245240C6DC900DD6076 /* libstdc++.tbd */; };
 		92375274240C6F1E00DD6076 /* 10001-90210-01803.wav in Resources */ = {isa = PBXBuildFile; fileRef = 92375256240C6E3D00DD6076 /* 10001-90210-01803.wav */; };
+		925527A9273C492C00FFD9CC /* libvosk.a in Frameworks */ = {isa = PBXBuildFile; fileRef = 925527A8273C492C00FFD9CC /* libvosk.a */; };
+		92833003273C466E00058B52 /* libc++.tbd in Frameworks */ = {isa = PBXBuildFile; fileRef = 92833002273C466E00058B52 /* libc++.tbd */; };
 		92BACED125BE125A00B5CC93 /* vosk-model-small-en-us-0.15 in Resources */ = {isa = PBXBuildFile; fileRef = 928CC50C25BE124400490481 /* vosk-model-small-en-us-0.15 */; };
 		92D6B8D325BDFEAC007FF08D /* VoskModel.swift in Sources */ = {isa = PBXBuildFile; fileRef = 92D6B8D225BDFEAC007FF08D /* VoskModel.swift */; };
 		92D86BD6253F823F0040D53F /* vosk-model-spk-0.4 in Resources */ = {isa = PBXBuildFile; fileRef = 92D86BD4253F823F0040D53F /* vosk-model-spk-0.4 */; };
@@ -31,10 +31,10 @@
 		9237522B240C550B00DD6076 /* Base */ = {isa = PBXFileReference; lastKnownFileType = file.storyboard; name = Base; path = Base.lproj/LaunchScreen.storyboard; sourceTree = "<group>"; };
 		9237522D240C550B00DD6076 /* Info.plist */ = {isa = PBXFileReference; lastKnownFileType = text.plist.xml; path = Info.plist; sourceTree = "<group>"; };
 		92375233240C558900DD6076 /* Vosk.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = Vosk.swift; sourceTree = "<group>"; };
-		9237523A240C642000DD6076 /* libkaldiwrap.a */ = {isa = PBXFileReference; lastKnownFileType = archive.ar; path = libkaldiwrap.a; sourceTree = "<group>"; };
 		92375243240C6DAF00DD6076 /* Accelerate.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Accelerate.framework; path = System/Library/Frameworks/Accelerate.framework; sourceTree = SDKROOT; };
-		92375245240C6DC900DD6076 /* libstdc++.tbd */ = {isa = PBXFileReference; lastKnownFileType = "sourcecode.text-based-dylib-definition"; name = "libstdc++.tbd"; path = "usr/lib/libstdc++.tbd"; sourceTree = SDKROOT; };
 		92375256240C6E3D00DD6076 /* 10001-90210-01803.wav */ = {isa = PBXFileReference; lastKnownFileType = audio.wav; path = "10001-90210-01803.wav"; sourceTree = "<group>"; };
+		925527A8273C492C00FFD9CC /* libvosk.a */ = {isa = PBXFileReference; lastKnownFileType = archive.ar; path = libvosk.a; sourceTree = "<group>"; };
+		92833002273C466E00058B52 /* libc++.tbd */ = {isa = PBXFileReference; lastKnownFileType = "sourcecode.text-based-dylib-definition"; name = "libc++.tbd"; path = "usr/lib/libc++.tbd"; sourceTree = SDKROOT; };
 		928CC50C25BE124400490481 /* vosk-model-small-en-us-0.15 */ = {isa = PBXFileReference; lastKnownFileType = folder; name = "vosk-model-small-en-us-0.15"; path = "/Users/shmyrev/Documents/IOS/VoskApiTest/VoskApiTest/Vosk/vosk-model-small-en-us-0.15"; sourceTree = "<absolute>"; };
 		92AA22AD244CDD1200DA464B /* vosk_api.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = vosk_api.h; sourceTree = "<group>"; };
 		92AA22AE244CDD5200DA464B /* bridging.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = bridging.h; sourceTree = "<group>"; };
@@ -47,9 +47,9 @@
 			isa = PBXFrameworksBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
-				92375246240C6DC900DD6076 /* libstdc++.tbd in Frameworks */,
+				92833003273C466E00058B52 /* libc++.tbd in Frameworks */,
 				92375244240C6DAF00DD6076 /* Accelerate.framework in Frameworks */,
-				9237523C240C642000DD6076 /* libkaldiwrap.a in Frameworks */,
+				925527A9273C492C00FFD9CC /* libvosk.a in Frameworks */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
 		};
@@ -93,11 +93,11 @@
 		92375239240C642000DD6076 /* Vosk */ = {
 			isa = PBXGroup;
 			children = (
-				928CC50C25BE124400490481 /* vosk-model-small-en-us-0.15 */,
 				92D86BD4253F823F0040D53F /* vosk-model-spk-0.4 */,
-				92375256240C6E3D00DD6076 /* 10001-90210-01803.wav */,
+				928CC50C25BE124400490481 /* vosk-model-small-en-us-0.15 */,
+				925527A8273C492C00FFD9CC /* libvosk.a */,
 				92AA22AD244CDD1200DA464B /* vosk_api.h */,
-				9237523A240C642000DD6076 /* libkaldiwrap.a */,
+				92375256240C6E3D00DD6076 /* 10001-90210-01803.wav */,
 			);
 			name = Vosk;
 			path = VoskApiTest/Vosk;
@@ -106,7 +106,7 @@
 		92375242240C6DAF00DD6076 /* Frameworks */ = {
 			isa = PBXGroup;
 			children = (
-				92375245240C6DC900DD6076 /* libstdc++.tbd */,
+				92833002273C466E00058B52 /* libc++.tbd */,
 				92375243240C6DAF00DD6076 /* Accelerate.framework */,
 			);
 			name = Frameworks;
@@ -145,7 +145,7 @@
 					9237521D240C550B00DD6076 = {
 						CreatedOnToolsVersion = 8.3.2;
 						LastSwiftMigration = 0920;
-						ProvisioningStyle = Automatic;
+						ProvisioningStyle = Manual;
 					};
 				};
 			};
@@ -223,7 +223,6 @@
 				ALWAYS_SEARCH_USER_PATHS = NO;
 				CLANG_ANALYZER_NONNULL = YES;
 				CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE;
-				CLANG_CXX_LANGUAGE_STANDARD = "gnu++0x";
 				CLANG_CXX_LIBRARY = "libc++";
 				CLANG_ENABLE_MODULES = YES;
 				CLANG_ENABLE_OBJC_ARC = YES;
@@ -245,7 +244,7 @@
 				CLANG_WARN_SUSPICIOUS_MOVE = YES;
 				CLANG_WARN_UNREACHABLE_CODE = YES;
 				CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
-				"CODE_SIGN_IDENTITY[sdk=iphoneos*]" = "iPhone Developer";
+				"CODE_SIGN_IDENTITY[sdk=iphoneos*]" = "";
 				COPY_PHASE_STRIP = NO;
 				DEBUG_INFORMATION_FORMAT = dwarf;
 				ENABLE_STRICT_OBJC_MSGSEND = YES;
@@ -282,7 +281,6 @@
 				ALWAYS_SEARCH_USER_PATHS = NO;
 				CLANG_ANALYZER_NONNULL = YES;
 				CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE;
-				CLANG_CXX_LANGUAGE_STANDARD = "gnu++0x";
 				CLANG_CXX_LIBRARY = "libc++";
 				CLANG_ENABLE_MODULES = YES;
 				CLANG_ENABLE_OBJC_ARC = YES;
@@ -304,7 +302,7 @@
 				CLANG_WARN_SUSPICIOUS_MOVE = YES;
 				CLANG_WARN_UNREACHABLE_CODE = YES;
 				CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
-				"CODE_SIGN_IDENTITY[sdk=iphoneos*]" = "iPhone Developer";
+				"CODE_SIGN_IDENTITY[sdk=iphoneos*]" = "";
 				COPY_PHASE_STRIP = NO;
 				DEBUG_INFORMATION_FORMAT = "dwarf-with-dsym";
 				ENABLE_NS_ASSERTIONS = NO;
@@ -319,6 +317,7 @@
 				GCC_WARN_UNUSED_VARIABLE = YES;
 				IPHONEOS_DEPLOYMENT_TARGET = 10.3;
 				MTL_ENABLE_DEBUG_INFO = NO;
+				ONLY_ACTIVE_ARCH = YES;
 				SDKROOT = iphoneos;
 				SWIFT_OBJC_BRIDGING_HEADER = bridging.h;
 				SWIFT_OPTIMIZATION_LEVEL = "-Owholemodule";
@@ -333,7 +332,10 @@
 			buildSettings = {
 				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
 				CLANG_ENABLE_MODULES = YES;
-				ENABLE_BITCODE = YES;
+				"CODE_SIGN_IDENTITY[sdk=iphoneos*]" = "iPhone Developer";
+				CODE_SIGN_STYLE = Manual;
+				DEVELOPMENT_TEAM = "";
+				ENABLE_BITCODE = NO;
 				INFOPLIST_FILE = VoskApiTest/Info.plist;
 				LD_RUNPATH_SEARCH_PATHS = "$(inherited) @executable_path/Frameworks";
 				LIBRARY_SEARCH_PATHS = (
@@ -342,11 +344,13 @@
 				);
 				PRODUCT_BUNDLE_IDENTIFIER = com.alphacephei.VoskApiTest;
 				PRODUCT_NAME = "$(TARGET_NAME)";
+				PROVISIONING_PROFILE_SPECIFIER = "";
 				SWIFT_INSTALL_OBJC_HEADER = YES;
 				SWIFT_OBJC_BRIDGING_HEADER = VoskApiTest/bridging.h;
 				SWIFT_OPTIMIZATION_LEVEL = "-Onone";
 				SWIFT_SWIFT3_OBJC_INFERENCE = Default;
 				SWIFT_VERSION = 4.0;
+				TARGETED_DEVICE_FAMILY = "1,2";
 			};
 			name = Debug;
 		};
@@ -355,7 +359,10 @@
 			buildSettings = {
 				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
 				CLANG_ENABLE_MODULES = YES;
-				ENABLE_BITCODE = YES;
+				"CODE_SIGN_IDENTITY[sdk=iphoneos*]" = "iPhone Developer";
+				CODE_SIGN_STYLE = Manual;
+				DEVELOPMENT_TEAM = "";
+				ENABLE_BITCODE = NO;
 				INFOPLIST_FILE = VoskApiTest/Info.plist;
 				LD_RUNPATH_SEARCH_PATHS = "$(inherited) @executable_path/Frameworks";
 				LIBRARY_SEARCH_PATHS = (
@@ -364,10 +371,12 @@
 				);
 				PRODUCT_BUNDLE_IDENTIFIER = com.alphacephei.VoskApiTest;
 				PRODUCT_NAME = "$(TARGET_NAME)";
+				PROVISIONING_PROFILE_SPECIFIER = "";
 				SWIFT_INSTALL_OBJC_HEADER = YES;
 				SWIFT_OBJC_BRIDGING_HEADER = VoskApiTest/bridging.h;
 				SWIFT_SWIFT3_OBJC_INFERENCE = Default;
 				SWIFT_VERSION = 4.0;
+				TARGETED_DEVICE_FAMILY = "1,2";
 			};
 			name = Release;
 		};

ios/VoskApiTest.xcodeproj/project.xcworkspace/contents.xcworkspacedata

@@ -1,9 +1,6 @@
 <?xml version="1.0" encoding="UTF-8"?>
 <Workspace
    version = "1.0">
-   <FileRef
-      location = "group:/Users/shmyrev/Documents/IOS/VoskApiTest/VoskApiTest/Vosk/vosk-model-small-en-us-0.15">
-   </FileRef>
    <FileRef
       location = "self:">
    </FileRef>

ios/VoskApiTest/Vosk.swift

@@ -14,7 +14,7 @@ public final class Vosk {
     var recognizer : OpaquePointer!
 
     init(model: VoskModel, sampleRate: Float) {
-        recognizer = vosk_recognizer_new_spk(model.model, model.spkModel, sampleRate)
+        recognizer = vosk_recognizer_new_spk(model.model, sampleRate, model.spkModel)
     }
     
     deinit {

ios/VoskApiTest/Vosk/vosk_api.h

@@ -1,4 +1,4 @@
-// Copyright 2020 Alpha Cephei Inc.
+// Copyright 2020-2021 Alpha Cephei Inc.
 //
 // Licensed under the Apache License, Version 2.0 (the "License");
 // you may not use this file except in compliance with the License.
@@ -43,7 +43,7 @@ typedef struct VoskRecognizer VoskRecognizer;
 /** Loads model data from the file and returns the model object
  *
  * @param model_path: the path of the model on the filesystem
- @ @returns model object */
+ * @returns model object or NULL if problem occured */
 VoskModel *vosk_model_new(const char *model_path);
 
 
@@ -55,10 +55,18 @@ VoskModel *vosk_model_new(const char *model_path);
 void vosk_model_free(VoskModel *model);
 
 
+/** Check if a word can be recognized by the model
+ * @param word: the word
+ * @returns the word symbol if @param word exists inside the model
+ * or -1 otherwise.
+ * Reminding that word symbol 0 is for <epsilon> */
+int vosk_model_find_word(VoskModel *model, const char *word);
+
+
 /** Loads speaker model data from the file and returns the model object
  *
  * @param model_path: the path of the model on the filesystem
- * @returns model object */
+ * @returns model object or NULL if problem occured */
 VoskSpkModel *vosk_spk_model_new(const char *model_path);
 
 
@@ -71,9 +79,13 @@ void vosk_spk_model_free(VoskSpkModel *model);
 
 /** Creates the recognizer object
  *
- *  The recognizers process the speech and return text using shared model data 
- *  @param sample_rate The sample rate of the audio you going to feed into the recognizer
- *  @returns recognizer object */
+ *  The recognizers process the speech and return text using shared model data
+ *  @param model       VoskModel containing static data for recognizer. Model can be
+ *                     shared across recognizers, even running in different threads.
+ *  @param sample_rate The sample rate of the audio you going to feed into the recognizer.
+ *                     Make sure this rate matches the audio content, it is a common
+ *                     issue causing accuracy problems.
+ *  @returns recognizer object or NULL if problem occured */
 VoskRecognizer *vosk_recognizer_new(VoskModel *model, float sample_rate);
 
 
@@ -82,10 +94,14 @@ VoskRecognizer *vosk_recognizer_new(VoskModel *model, float sample_rate);
  *  With the speaker recognition mode the recognizer not just recognize
  *  text but also return speaker vectors one can use for speaker identification
  *
+ *  @param model       VoskModel containing static data for recognizer. Model can be
+ *                     shared across recognizers, even running in different threads.
+ *  @param sample_rate The sample rate of the audio you going to feed into the recognizer.
+ *                     Make sure this rate matches the audio content, it is a common
+ *                     issue causing accuracy problems.
  *  @param spk_model speaker model for speaker identification
- *  @param sample_rate The sample rate of the audio you going to feed into the recognizer
- *  @returns recognizer object */
-VoskRecognizer *vosk_recognizer_new_spk(VoskModel *model, VoskSpkModel *spk_model, float sample_rate);
+ *  @returns recognizer object or NULL if problem occured */
+VoskRecognizer *vosk_recognizer_new_spk(VoskModel *model, float sample_rate, VoskSpkModel *spk_model);
 
 
 /** Creates the recognizer object with the phrase list
@@ -98,42 +114,46 @@ VoskRecognizer *vosk_recognizer_new_spk(VoskModel *model, VoskSpkModel *spk_mode
  *  Only recognizers with lookahead models support this type of quick configuration.
  *  Precompiled HCLG graph models are not supported.
  *
- *  @param sample_rate The sample rate of the audio you going to feed into the recognizer
+ *  @param model       VoskModel containing static data for recognizer. Model can be
+ *                     shared across recognizers, even running in different threads.
+ *  @param sample_rate The sample rate of the audio you going to feed into the recognizer.
+ *                     Make sure this rate matches the audio content, it is a common
+ *                     issue causing accuracy problems.
  *  @param grammar The string with the list of phrases to recognize as JSON array of strings,
  *                 for example "["one two three four five", "[unk]"]".
  *
- *  @returns recognizer object */
+ *  @returns recognizer object or NULL if problem occured */
 VoskRecognizer *vosk_recognizer_new_grm(VoskModel *model, float sample_rate, const char *grammar);
 
 
-/** Accept voice data
+/** Adds speaker model to already initialized recognizer
  *
- *  accept and process new chunk of voice data
+ * Can add speaker recognition model to already created recognizer. Helps to initialize
+ * speaker recognition for grammar-based recognizer.
  *
- *  @param data - audio data in PCM 16-bit mono format
- *  @param length - length of the audio data
- *  @returns true if silence is occured and you can retrieve a new utterance with result method */
-int vosk_recognizer_accept_waveform(VoskRecognizer *recognizer, const char *data, int length);
+ * @param spk_model Speaker recognition model */
+void vosk_recognizer_set_spk_model(VoskRecognizer *recognizer, VoskSpkModel *spk_model);
 
 
-/** Same as above but the version with the short data for language bindings where you have
- *  audio as array of shorts */
-int vosk_recognizer_accept_waveform_s(VoskRecognizer *recognizer, const short *data, int length);
-
-
-/** Same as above but the version with the float data for language bindings where you have
- *  audio as array of floats */
-int vosk_recognizer_accept_waveform_f(VoskRecognizer *recognizer, const float *data, int length);
-
-
-/** Returns speech recognition result
- *
- * @returns the result in JSON format which contains decoded line, decoded
- *          words, times in seconds and confidences. You can parse this result
- *          with any json parser
+/** Configures recognizer to output n-best results
+ *
+ * <pre>
+ *   {
+ *      "alternatives": [
+ *          { "text": "one two three four five", "confidence": 0.97 },
+ *          { "text": "one two three for five", "confidence": 0.03 },
+ *      ]
+ *   }
+ * </pre>
+ *
+ * @param max_alternatives - maximum alternatives to return from recognition results
+ */
+void vosk_recognizer_set_max_alternatives(VoskRecognizer *recognizer, int max_alternatives);
+
+
+/** Enables words with times in the output
  *
  * <pre>
- * {
  *   "result" : [{
  *       "conf" : 1.000000,
  *       "end" : 1.110000,
@@ -156,13 +176,54 @@ int vosk_recognizer_accept_waveform_f(VoskRecognizer *recognizer, const float *d
  *       "word" : "zero"
  *     }, {
  *       "conf" : 1.000000,
- *      "end" : 2.610000,
+ *       "end" : 2.610000,
  *       "start" : 2.340000,
  *       "word" : "one"
  *     }],
- *   "text" : "what zero zero zero one"
+ * </pre>
+ *
+ * @param words - boolean value
+ */
+void vosk_recognizer_set_words(VoskRecognizer *recognizer, int words);
+
+
+/** Accept voice data
+ *
+ *  accept and process new chunk of voice data
+ *
+ *  @param data - audio data in PCM 16-bit mono format
+ *  @param length - length of the audio data
+ *  @returns 1 if silence is occured and you can retrieve a new utterance with result method 
+ *           0 if decoding continues
+ *           -1 if exception occured */
+int vosk_recognizer_accept_waveform(VoskRecognizer *recognizer, const char *data, int length);
+
+
+/** Same as above but the version with the short data for language bindings where you have
+ *  audio as array of shorts */
+int vosk_recognizer_accept_waveform_s(VoskRecognizer *recognizer, const short *data, int length);
+
+
+/** Same as above but the version with the float data for language bindings where you have
+ *  audio as array of floats */
+int vosk_recognizer_accept_waveform_f(VoskRecognizer *recognizer, const float *data, int length);
+
+
+/** Returns speech recognition result
+ *
+ * @returns the result in JSON format which contains decoded line, decoded
+ *          words, times in seconds and confidences. You can parse this result
+ *          with any json parser
+ *
+ * <pre>
+ *  {
+ *    "text" : "what zero zero zero one"
  *  }
  * </pre>
+ *
+ * If alternatives enabled it returns result with alternatives, see also vosk_recognizer_set_alternatives().
+ *
+ * If word times enabled returns word time, see also vosk_recognizer_set_word_times().
  */
 const char *vosk_recognizer_result(VoskRecognizer *recognizer);
 
@@ -174,7 +235,7 @@ const char *vosk_recognizer_result(VoskRecognizer *recognizer);
  *
  * <pre>
  * {
- *  "partial" : "cyril one eight zero"
+ *    "partial" : "cyril one eight zero"
  * }
  * </pre>
  */
@@ -190,6 +251,12 @@ const char *vosk_recognizer_partial_result(VoskRecognizer *recognizer);
 const char *vosk_recognizer_final_result(VoskRecognizer *recognizer);
 
 
+/** Resets the recognizer
+ *
+ *  Resets current results so the recognition can continue from scratch */
+void vosk_recognizer_reset(VoskRecognizer *recognizer);
+
+
 /** Releases recognizer object
  *
  *  Underlying model is also unreferenced and if needed released */
@@ -204,6 +271,20 @@ void vosk_recognizer_free(VoskRecognizer *recognizer);
  */
 void vosk_set_log_level(int log_level);
 
+/**
+ *  Init, automatically select a CUDA device and allow multithreading.
+ *  Must be called once from the main thread.
+ *  Has no effect if HAVE_CUDA flag is not set.
+ */
+void vosk_gpu_init();
+
+/**
+ *  Init CUDA device in a multi-threaded environment.
+ *  Must be called for each thread.
+ *  Has no effect if HAVE_CUDA flag is not set.
+ */
+void vosk_gpu_thread_init();
+
 #ifdef __cplusplus
 }
 #endif

java/demo/build.gradle

@@ -8,12 +8,9 @@ application {
 
 repositories {
     mavenCentral()
-    maven {
-        url 'https://alphacephei.com/maven/'
-    }
 }
 
 dependencies {
     implementation group: 'net.java.dev.jna', name: 'jna', version: '5.7.0'
-    implementation group: 'com.alphacephei', name: 'vosk', version: '0.3.30+'
+    implementation group: 'com.alphacephei', name: 'vosk', version: '0.3.43+'
 }

java/lib/build.gradle

@@ -1,18 +1,31 @@
+buildscript {
+  repositories {
+    mavenCentral()
+  }
+}
+
 plugins {
     id 'java-library'
     id 'maven-publish'
+    id 'com.vanniktech.maven.publish' version '0.18.0'
 }
 
-archivesBaseName = 'vosk'
-group = 'com.alphacephei'
-version = '0.3.30'
-
 repositories {
     mavenCentral()
 }
 
+archivesBaseName = 'vosk'
+group = 'com.alphacephei'
+version = '0.3.43'
+
+mavenPublish {
+    group = 'com.alphacephei'
+    version = version
+    sonatypeHost = 's01'
+}
+
 dependencies {
-    implementation group: 'net.java.dev.jna', name: 'jna', version: '5.7.0'
+    api group: 'net.java.dev.jna', name: 'jna', version: '5.7.0'
     testImplementation 'junit:junit:4.13'
 }
 
@@ -45,14 +58,14 @@ publishing {
             }
         }
     }
-    repositories {
-        maven {
-            url = "repo"
-        }
-    }
 }
 
 test {
     dependsOn cleanTest
     testLogging.showStandardStreams = true
 }
+
+java {
+    withSourcesJar()
+    withJavadocJar()
+}

java/lib/src/main/java/org/vosk/LibVosk.java

@@ -62,6 +62,8 @@ public class LibVosk {
 
     public static native void vosk_recognizer_set_words(Pointer recognizer, boolean words);
 
+    public static native void vosk_recognizer_set_partial_words(Pointer recognizer, boolean partial_words);
+
     public static native void vosk_recognizer_set_spk_model(Pointer recognizer, Pointer spk_model);
 
     public static native boolean vosk_recognizer_accept_waveform(Pointer recognizer, byte[] data, int len);

java/lib/src/main/java/org/vosk/Model.java

@@ -1,13 +1,18 @@
 package org.vosk;
 
+import java.io.IOException;
 import com.sun.jna.PointerType;
 
 public class Model extends PointerType implements AutoCloseable {
     public Model() {
     }
 
-    public Model(String path) {
+    public Model(String path) throws IOException {
         super(LibVosk.vosk_model_new(path));
+
+        if (getPointer() == null) {
+            throw new IOException("Failed to create a model");
+        }
     }
 
     @Override

java/lib/src/main/java/org/vosk/Recognizer.java

@@ -1,10 +1,15 @@
 package org.vosk;
 
 import com.sun.jna.PointerType;
+import java.io.IOException;
 
 public class Recognizer extends PointerType implements AutoCloseable {
-    public Recognizer(Model model, float sampleRate) {
+    public Recognizer(Model model, float sampleRate) throws IOException {
         super(LibVosk.vosk_recognizer_new(model, sampleRate));
+
+        if (getPointer() == null) {
+            throw new IOException("Failed to create a recognizer");
+        }
     }
 
     public Recognizer(Model model, float sampleRate, SpeakerModel spkModel) {
@@ -23,6 +28,10 @@ public class Recognizer extends PointerType implements AutoCloseable {
         LibVosk.vosk_recognizer_set_words(this.getPointer(), words);
     }
 
+    public void setPartialWords(boolean partial_words) {
+        LibVosk.vosk_recognizer_set_partial_words(this.getPointer(), partial_words);
+    }
+
     public void setSpeakerModel(SpeakerModel spkModel) {
         LibVosk.vosk_recognizer_set_spk_model(this.getPointer(), spkModel.getPointer());
     }

java/lib/src/main/java/org/vosk/SpeakerModel.java

@@ -1,13 +1,18 @@
 package org.vosk;
 
 import com.sun.jna.PointerType;
+import java.io.IOException;
 
 public class SpeakerModel extends PointerType implements AutoCloseable {
     public SpeakerModel() {
     }
 
-    public SpeakerModel(String path) {
+    public SpeakerModel(String path) throws IOException {
         super(LibVosk.vosk_spk_model_new(path));
+
+        if (getPointer() == null) {
+            throw new IOException("Failed to create a speaker model");
+        }
     }
 
     @Override

java/lib/src/test/java/org/vosk/test/DecoderTest.java

@@ -30,6 +30,7 @@ public class DecoderTest {
 
             recognizer.setMaxAlternatives(10);
             recognizer.setWords(true);
+            recognizer.setPartialWords(true);
 
             int nbytes;
             byte[] b = new byte[4096];
@@ -93,4 +94,10 @@ public class DecoderTest {
         }
         Assert.assertTrue(true);
     }
+
+
+    @Test(expected = IOException.class)
+    public void decoderTestException() throws IOException {
+        Model model = new Model("model_missing");
+    }
 }

nodejs/README.md

@@ -2,18 +2,19 @@ This is an FFI-NAPI wrapper for the Vosk library.
 
 ## Usage
 
-It mostly follows Vosk interface, some methods are not yet fully implemented.
+Bindings mostly follow Vosk interface, some methods are not yet fully implemented.
 
-To use it you need to compile libvosk library, see Python module build
-instructions for details. You can find prebuilt library inside python
-wheel.
+See [demo folder](https://github.com/alphacep/vosk-api/tree/master/nodejs/demo) for
+details.
 
 ## About
 
 Vosk is an offline open source speech recognition toolkit. It enables
-speech recognition models for 17 languages and dialects - English, Indian
+speech recognition for 20+ languages and dialects - English, Indian
 English, German, French, Spanish, Portuguese, Chinese, Russian, Turkish,
-Vietnamese, Italian, Dutch, Catalan, Arabic, Greek, Farsi, Filipino.
+Vietnamese, Italian, Dutch, Catalan, Arabic, Greek, Farsi, Filipino,
+Ukrainian, Kazakh, Swedish, Japanese, Esperanto, Hindi, Czech, Polish.
+More to come.
 
 Vosk models are small (50 Mb) but provide continuous large vocabulary
 transcription, zero-latency response with streaming API, reconfigurable

nodejs/demo/test_microphone.js

@@ -18,11 +18,11 @@ const rec = new vosk.Recognizer({model: model, sampleRate: SAMPLE_RATE});
 var micInstance = mic({
     rate: String(SAMPLE_RATE),
     channels: '1',
-    debug: false
+    debug: false,
+    device: 'default',    
 });
 
 var micInputStream = micInstance.getAudioStream();
-micInstance.start();
 
 micInputStream.on('data', data => {
     if (rec.acceptWaveform(data))
@@ -31,9 +31,16 @@ micInputStream.on('data', data => {
         console.log(rec.partialResult());
 });
 
-process.on('SIGINT', function() {
+micInputStream.on('audioProcessExitComplete', function() {
+    console.log("Cleaning up");
     console.log(rec.finalResult());
-    console.log("\nDone");
     rec.free();
     model.free();
 });
+
+process.on('SIGINT', function() {
+    console.log("\nStopping");
+    micInstance.stop();
+});
+
+micInstance.start();

nodejs/demo/test_simple.js

@@ -29,10 +29,13 @@ wfReader.on('format', async ({ audioFormat, sampleRate, channels }) => {
     const rec = new vosk.Recognizer({model: model, sampleRate: sampleRate});
     rec.setMaxAlternatives(10);
     rec.setWords(true);
+    rec.setPartialWords(true);
     for await (const data of wfReadable) {
         const end_of_speech = rec.acceptWaveform(data);
         if (end_of_speech) {
               console.log(JSON.stringify(rec.result(), null, 4));
+        } else {
+              console.log(JSON.stringify(rec.partialResult(), null, 4));
         }
     }
     console.log(JSON.stringify(rec.finalResult(rec), null, 4));

nodejs/demo/test_srt.js

@@ -46,8 +46,8 @@ ffmpeg_run.on('exit', code => {
             subs.push({
                 type: 'cue',
                 data: {
-                start: words[0].start,
-                end: words[0].end,
+                start: words[0].start * 1000,
+                end: words[0].end * 1000,
                 text: words[0].word
                 }
             });
@@ -61,8 +61,8 @@ ffmpeg_run.on('exit', code => {
                 subs.push({
                     type: 'cue',
                     data: {
-                      start: words[start_index].start,
-                      end: words[i].end,
+                      start: words[start_index].start * 1000,
+                      end: words[i].end * 1000,
                       text: text.slice(0, text.length-1)
                     }
                 });
@@ -74,8 +74,8 @@ ffmpeg_run.on('exit', code => {
             subs.push({
                 type: 'cue',
                 data: {
-                  start: words[start_index].start,
-                  end: words[words.length-1].end,
+                  start: words[start_index].start * 1000,
+                  end: words[words.length-1].end * 1000,
                   text: text
                 }
             });

nodejs/index.js

@@ -76,7 +76,7 @@ if (os.platform() == 'win32') {
 
     soname = path.join(__dirname, "lib", "win-x86_64", "libvosk.dll")
 } else if (os.platform() == 'darwin') {
-    soname = path.join(__dirname, "lib", "osx-x86_64", "libvosk.dylib")
+    soname = path.join(__dirname, "lib", "osx-universal", "libvosk.dylib")
 } else {
     soname = path.join(__dirname, "lib", "linux-x86_64", "libvosk.so")
 }
@@ -93,6 +93,7 @@ const libvosk = ffi.Library(soname, {
     'vosk_recognizer_free': ['void', [vosk_recognizer_ptr]],
     'vosk_recognizer_set_max_alternatives': ['void', [vosk_recognizer_ptr, 'int']],
     'vosk_recognizer_set_words': ['void', [vosk_recognizer_ptr, 'bool']],
+    'vosk_recognizer_set_partial_words': ['void', [vosk_recognizer_ptr, 'bool']],
     'vosk_recognizer_set_spk_model': ['void', [vosk_recognizer_ptr, vosk_spk_model_ptr]],
     'vosk_recognizer_accept_waveform': ['bool', [vosk_recognizer_ptr, 'pointer', 'int']],
     'vosk_recognizer_result': ['string', [vosk_recognizer_ptr]],
@@ -301,6 +302,11 @@ class Recognizer {
         libvosk.vosk_recognizer_set_words(this.handle, words);
     }
 
+    /** Same as above, but for partial results*/
+    setPartialWords(partial_words) {
+        libvosk.vosk_recognizer_set_partial_words(this.handle, partial_words);
+    }
+
     /** Adds speaker recognition model to already created recognizer. Helps to initialize
      * speaker recognition for grammar-based recognizer.
      *

nodejs/package.json

@@ -1,6 +1,6 @@
 {
   "name": "vosk",
-  "version": "0.3.30",
+  "version": "0.3.43",
   "description": "Node binding for continuous offline voice recoginition with Vosk library.",
   "repository": {
     "type": "git",

python/README.md

@@ -1,9 +1,11 @@
 This is a Python module for Vosk.
 
 Vosk is an offline open source speech recognition toolkit. It enables
-speech recognition models for 17 languages and dialects - English, Indian
+speech recognition for 20+ languages and dialects - English, Indian
 English, German, French, Spanish, Portuguese, Chinese, Russian, Turkish,
-Vietnamese, Italian, Dutch, Catalan, Arabic, Greek, Farsi, Filipino.
+Vietnamese, Italian, Dutch, Catalan, Arabic, Greek, Farsi, Filipino,
+Ukrainian, Kazakh, Swedish, Japanese, Esperanto, Hindi, Czech, Polish.
+More to come.
 
 Vosk models are small (50 Mb) but provide continuous large vocabulary
 transcription, zero-latency response with streaming API, reconfigurable

python/example/test_alternatives.py

@@ -8,16 +8,12 @@ import json
 
 SetLogLevel(0)
 
-if not os.path.exists("model"):
-    print ("Please download the model from https://alphacephei.com/vosk/models and unpack as 'model' in the current folder.")
-    exit (1)
-
 wf = wave.open(sys.argv[1], "rb")
 if wf.getnchannels() != 1 or wf.getsampwidth() != 2 or wf.getcomptype() != "NONE":
     print ("Audio file must be WAV format mono PCM.")
     exit (1)
 
-model = Model("model")
+model = Model(lang="en-us")
 rec = KaldiRecognizer(model, wf.getframerate())
 rec.SetMaxAlternatives(10)
 rec.SetWords(True)

python/example/test_empty.py

@@ -4,7 +4,7 @@ from vosk import Model, KaldiRecognizer
 import sys
 import json
 
-model = Model("model")
+model = Model(lang="en-us")
 rec = KaldiRecognizer(model, 8000)
 
 res = json.loads(rec.FinalResult())

python/example/test_ffmpeg.py

@@ -8,12 +8,8 @@ import subprocess
 
 SetLogLevel(0)
 
-if not os.path.exists("model"):
-    print ("Please download the model from https://alphacephei.com/vosk/models and unpack as 'model' in the current folder.")
-    exit (1)
-
 sample_rate=16000
-model = Model("model")
+model = Model(lang="en-us")
 rec = KaldiRecognizer(model, sample_rate)
 
 process = subprocess.Popen(['ffmpeg', '-loglevel', 'quiet', '-i',

python/example/test_gpu_batch.py

@@ -0,0 +1,59 @@
+#!/usr/bin/env python3
+
+import sys
+import os
+import wave
+from time import sleep
+import json
+from timeit import default_timer as timer
+
+
+from vosk import BatchModel, BatchRecognizer, GpuInit
+
+GpuInit()
+
+model = BatchModel()
+
+fnames = open(sys.argv[1]).readlines()
+fds = [open(x.strip(), "rb") for x in fnames]
+uids = [fname.strip().split('/')[-1][:-4] for fname in fnames]
+recs = [BatchRecognizer(model, 16000) for x in fnames]
+results = [""] * len(fnames)
+ended = set()
+tot_samples = 0
+
+start_time = timer()
+
+while True:
+
+    # Feed in the data
+    for i, fd in enumerate(fds):
+        if i in ended:
+            continue
+        data = fd.read(8000)
+        if len(data) == 0:
+            recs[i].FinishStream()
+            ended.add(i)
+            continue
+        recs[i].AcceptWaveform(data)
+        tot_samples += len(data)
+
+    # Wait for results from CUDA
+    model.Wait()
+
+    # Retrieve and add results
+    for i, fd in enumerate(fds):
+       res = recs[i].Result()
+       if len(res) != 0:
+           results[i] = results[i] + " " + json.loads(res)['text']
+
+    if len(ended) == len(fds):
+        break
+
+end_time = timer()
+
+for i in range(len(results)):
+    print (uids[i], results[i].strip())
+
+print ("Processed %.3f seconds of audio in %.3f seconds (%.3f xRT)" % (tot_samples / 16000.0 / 2, end_time - start_time, 
+    (tot_samples / 16000.0 / 2 / (end_time - start_time))), file=sys.stderr)

python/example/test_microphone.py

@@ -37,9 +37,6 @@ parser = argparse.ArgumentParser(
 parser.add_argument(
     '-f', '--filename', type=str, metavar='FILENAME',
     help='audio file to store recording to')
-parser.add_argument(
-    '-m', '--model', type=str, metavar='MODEL_PATH',
-    help='Path to the model')
 parser.add_argument(
     '-d', '--device', type=int_or_str,
     help='input device (numeric ID or substring)')
@@ -48,18 +45,12 @@ parser.add_argument(
 args = parser.parse_args(remaining)
 
 try:
-    if args.model is None:
-        args.model = "model"
-    if not os.path.exists(args.model):
-        print ("Please download a model for your language from https://alphacephei.com/vosk/models")
-        print ("and unpack as 'model' in the current folder.")
-        parser.exit(0)
     if args.samplerate is None:
         device_info = sd.query_devices(args.device, 'input')
         # soundfile expects an int, sounddevice provides a float:
         args.samplerate = int(device_info['default_samplerate'])
 
-    model = vosk.Model(args.model)
+    model = vosk.Model(lang="en-us")
 
     if args.filename:
         dump_fn = open(args.filename, "wb")

python/example/test_nlsml.py

@@ -0,0 +1,27 @@
+#!/usr/bin/env python3
+
+from vosk import Model, KaldiRecognizer, SetLogLevel
+import sys
+import os
+import wave
+
+SetLogLevel(0)
+
+wf = wave.open(sys.argv[1], "rb")
+if wf.getnchannels() != 1 or wf.getsampwidth() != 2 or wf.getcomptype() != "NONE":
+    print ("Audio file must be WAV format mono PCM.")
+    exit (1)
+
+model = Model(lang="en-us")
+rec = KaldiRecognizer(model, wf.getframerate())
+rec.SetMaxAlternatives(10)
+rec.SetNLSML(True)
+
+while True:
+    data = wf.readframes(4000)
+    if len(data) == 0:
+        break
+    if rec.AcceptWaveform(data):
+        print(rec.Result())
+
+print(rec.FinalResult())

python/example/test_reset.py

@@ -8,16 +8,12 @@ import json
 
 SetLogLevel(0)
 
-if not os.path.exists("model"):
-    print ("Please download the model from https://alphacephei.com/vosk/models and unpack as 'model' in the current folder.")
-    exit (1)
-
 wf = wave.open(sys.argv[1], "rb")
 if wf.getnchannels() != 1 or wf.getsampwidth() != 2 or wf.getcomptype() != "NONE":
     print ("Audio file must be WAV format mono PCM.")
     exit (1)
 
-model = Model("model")
+model = Model(lang="en-us")
 rec = KaldiRecognizer(model, wf.getframerate())
 
 while True:

python/example/test_simple.py

@@ -5,20 +5,23 @@ import sys
 import os
 import wave
 
+# You can set log level to -1 to disable debug messages
 SetLogLevel(0)
 
-if not os.path.exists("model"):
-    print ("Please download the model from https://alphacephei.com/vosk/models and unpack as 'model' in the current folder.")
-    exit (1)
-
 wf = wave.open(sys.argv[1], "rb")
 if wf.getnchannels() != 1 or wf.getsampwidth() != 2 or wf.getcomptype() != "NONE":
     print ("Audio file must be WAV format mono PCM.")
     exit (1)
 
-model = Model("model")
+model = Model(lang="en-us")
+
+# You can also init model by name or with a folder path
+# model = Model(model_name="vosk-model-en-us-0.21")
+# model = Model("models/en")
+
 rec = KaldiRecognizer(model, wf.getframerate())
 rec.SetWords(True)
+rec.SetPartialWords(True)
 
 while True:
     data = wf.readframes(4000)

python/example/test_speaker.py

@@ -7,13 +7,8 @@ import json
 import os
 import numpy as np
 
-model_path = "model"
 spk_model_path = "model-spk"
 
-if not os.path.exists(model_path):
-    print ("Please download the model from https://alphacephei.com/vosk/models and unpack as {} in the current folder.".format(model_path))
-    exit (1)
-
 if not os.path.exists(spk_model_path):
     print ("Please download the speaker model from https://alphacephei.com/vosk/models and unpack as {} in the current folder.".format(spk_model_path))
     exit (1)
@@ -24,7 +19,7 @@ if wf.getnchannels() != 1 or wf.getsampwidth() != 2 or wf.getcomptype() != "NONE
     exit (1)
 
 # Large vocabulary free form recognition
-model = Model(model_path)
+model = Model(lang="en-us")
 spk_model = SpkModel(spk_model_path)
 #rec = KaldiRecognizer(model, wf.getframerate(), spk_model)
 rec = KaldiRecognizer(model, wf.getframerate())

python/example/test_srt.py

@@ -11,12 +11,8 @@ import datetime
 
 SetLogLevel(-1)
 
-if not os.path.exists("model"):
-    print ("Please download the model from https://alphacephei.com/vosk/models and unpack as 'model' in the current folder.")
-    exit (1)
-
 sample_rate=16000
-model = Model("model")
+model = Model(lang="en-us")
 rec = KaldiRecognizer(model, sample_rate)
 rec.SetWords(True)
 

python/example/test_text.py

@@ -5,12 +5,7 @@ import sys
 import json
 import os
 
-if not os.path.exists("model"):
-    print ("Please download the model from https://alphacephei.com/vosk/models and unpack as 'model' in the current folder.")
-    exit (1)
-
-
-model = Model("model")
+model = Model(lang="en-us")
 
 # Large vocabulary free form recognition
 rec = KaldiRecognizer(model, 16000)

python/example/test_webvtt.py

@@ -10,13 +10,8 @@ import textwrap
 
 SetLogLevel(-1)
 
-if not os.path.exists('model'):
-    print('Please download the model from https://alphacephei.com/vosk/models'
-          ' and unpack as `model` in the current folder.')
-    exit(1)
-
 sample_rate = 16000
-model = Model('model')
+model = Model(lang="en-us")
 rec = KaldiRecognizer(model, sample_rate)
 rec.SetWords(True)
 
@@ -67,6 +62,6 @@ def transcribe():
 
 if __name__ == '__main__':
     if not (1 < len(sys.argv) < 4):
-        print(f'Usage: {sys.argv[0]} audiofile [output file]')
+        print('Usage: {} audiofile [output file]'.format(sys.argv[0]))
         exit(1)
     transcribe()

python/example/test_words.py

@@ -5,16 +5,12 @@ import sys
 import os
 import wave
 
-if not os.path.exists("model"):
-    print ("Please download the model from https://alphacephei.com/vosk/models and unpack as 'model' in the current folder.")
-    exit (1)
-
 wf = wave.open(sys.argv[1], "rb")
 if wf.getnchannels() != 1 or wf.getsampwidth() != 2 or wf.getcomptype() != "NONE":
     print ("Audio file must be WAV format mono PCM.")
     exit (1)
 
-model = Model("model")
+model = Model(lang="en-us")
 
 # You can also specify the possible word or phrase list as JSON list, the order doesn't have to be strict
 rec = KaldiRecognizer(model, wf.getframerate(), '["oh one two three four five six seven eight nine zero", "[unk]"]')

python/setup.py

@@ -25,13 +25,15 @@ else:
         def get_tag(self):
             abi = 'none'
             if system == 'Darwin':
-                oses = 'macosx_10_6_x86_64'
+                oses = 'macosx_10_6_universal2'
             elif system == 'Windows' and architecture == '32bit':
                 oses = 'win32'
             elif system == 'Windows' and architecture == '64bit':
                 oses = 'win_amd64'
             elif system == 'Linux' and architecture == '64bit':
                 oses = 'linux_x86_64'
+            elif system == 'Linux' and architecture == 'aarch64':
+                oses = 'manylinux2014_aarch64'
             elif system == 'Linux':
                 oses = 'linux_' + architecture
             else:
@@ -44,7 +46,7 @@ with open("README.md", "r") as fh:
 
 setuptools.setup(
     name="vosk",
-    version="0.3.30",
+    version="0.3.43",
     author="Alpha Cephei Inc",
     author_email="contact@alphacephei.com",
     description="Offline open source speech recognition API based on Kaldi and Vosk",
@@ -53,6 +55,9 @@ setuptools.setup(
     url="https://github.com/alphacep/vosk-api",
     packages=setuptools.find_packages(),
     package_data = {'vosk': ['*.so', '*.dll', '*.dyld']},
+    entry_points = {
+        'console_scripts': ['vosk-transcriber=vosk.transcriber.cli:main'],
+    },
     include_package_data=True,
     classifiers=[
         'Programming Language :: Python :: 3',
@@ -65,7 +70,7 @@ setuptools.setup(
     cmdclass=cmdclass,
     python_requires='>=3',
     zip_safe=False, # Since we load so file from the filesystem, we can not run from zip file
-    setup_requires=['cffi>=1.0'],
-    install_requires=['cffi>=1.0'],
+    setup_requires=['cffi>=1.0', 'requests', 'tqdm', 'srt'],
+    install_requires=['cffi>=1.0', 'requests', 'tqdm', 'srt'],
     cffi_modules=['vosk_builder.py:ffibuilder'],
 )

python/vosk/__init__.py

@@ -1,7 +1,18 @@
 import os
 import sys
 
+import requests
+from urllib.request import urlretrieve
+from zipfile import ZipFile
+from re import match
+from pathlib import Path
 from .vosk_cffi import ffi as _ffi
+from tqdm import tqdm
+
+# Remote location of the models and local folders
+MODEL_PRE_URL = 'https://alphacephei.com/vosk/models/'
+MODEL_LIST_URL = MODEL_PRE_URL + 'model-list.json'
+MODEL_DIRS = [os.getenv('VOSK_MODEL_PATH'), Path('/usr/share/vosk'), Path.home() / 'AppData/Local/vosk', Path.home() / '.cache/vosk']
 
 def open_dll():
     dlldir = os.path.abspath(os.path.dirname(__file__))
@@ -20,10 +31,26 @@ def open_dll():
 
 _c = open_dll()
 
-class Model(object):
+def list_models():
+    response = requests.get(MODEL_LIST_URL)
+    for model in response.json():
+        print(model['name']) 
 
-    def __init__(self, model_path):
-        self._handle = _c.vosk_model_new(model_path.encode('utf-8'))
+def list_languages():
+    response = requests.get(MODEL_LIST_URL)
+    languages = set([m['lang'] for m in response.json()])
+    for lang in languages:
+        print (lang)
+
+class Model(object):
+    def __init__(self, model_path=None, model_name=None, lang=None):
+        if model_path != None:
+            self._handle = _c.vosk_model_new(model_path.encode('utf-8'))
+        else:
+            model_path = self.get_model_path(model_name, lang)
+            self._handle = _c.vosk_model_new(model_path.encode('utf-8'))
+        if self._handle == _ffi.NULL:
+            raise Exception("Failed to create a model")
 
     def __del__(self):
         _c.vosk_model_free(self._handle)
@@ -31,11 +58,76 @@ class Model(object):
     def vosk_model_find_word(self, word):
         return _c.vosk_model_find_word(self._handle, word.encode('utf-8'))
 
+    def get_model_path(self, model_name, lang):
+        if model_name is None:
+            model_path = self.get_model_by_lang(lang)
+        else:
+            model_path = self.get_model_by_name(model_name)
+        return str(model_path)
+
+    def get_model_by_name(self, model_name):
+        for directory in MODEL_DIRS:
+            if directory is None or not Path(directory).exists():
+                continue
+            model_file_list = os.listdir(directory)
+            model_file = [model for model in model_file_list if model == model_name]
+            if model_file != []:
+                return Path(directory, model_file[0])
+        response = requests.get(MODEL_LIST_URL)
+        result_model = [model['name'] for model in response.json() if model['name'] == model_name]
+        if result_model == []:
+            raise Exception("model name %s does not exist" % (model_name))
+        else:
+            self.download_model(Path(directory, result_model[0]))
+            return Path(directory, result_model[0])
+
+    def get_model_by_lang(self, lang):
+        for directory in MODEL_DIRS:
+            if directory is None or not Path(directory).exists():
+                continue
+            model_file_list = os.listdir(directory)
+            model_file = [model for model in model_file_list if match(r'vosk-model(-small)?-{}'.format(lang), model)]
+            if model_file != []:
+                return Path(directory, model_file[0])
+        response = requests.get(MODEL_LIST_URL)
+        result_model = [model['name'] for model in response.json() if model['lang'] == lang and model['type'] == 'small' and model['obsolete'] == 'false']
+        if result_model == []:
+            raise Exception("lang %s does not exist" % (lang))
+        else:
+            self.download_model(Path(directory, result_model[0]))
+            return Path(directory, result_model[0])
+
+    def download_model(self, model_name):
+        if not (model_name.parent).exists():
+            (model_name.parent).mkdir(parents=True)
+        with tqdm(unit='B', unit_scale=True, unit_divisor=1024, miniters=1,
+                desc=(MODEL_PRE_URL + str(model_name.name) + '.zip').split('/')[-1]) as t:
+            reporthook = self.download_progress_hook(t)
+            urlretrieve(MODEL_PRE_URL + str(model_name.name) + '.zip', str(model_name) + '.zip', 
+                reporthook=reporthook, data=None)
+            t.total = t.n
+            with ZipFile(str(model_name) + '.zip', 'r') as model_ref:
+                model_ref.extractall(model_name.parent)
+            Path(str(model_name) + '.zip').unlink()
+
+    def download_progress_hook(self, t):
+        last_b = [0]
+        def update_to(b=1, bsize=1, tsize=None):
+            if tsize not in (None, -1):
+                t.total = tsize
+            displayed = t.update((b - last_b[0]) * bsize)
+            last_b[0] = b
+            return displayed
+        return update_to
+
 class SpkModel(object):
 
     def __init__(self, model_path):
         self._handle = _c.vosk_spk_model_new(model_path.encode('utf-8'))
 
+        if self._handle == _ffi.NULL:
+            raise Exception("Failed to create a speaker model")
+
     def __del__(self):
         _c.vosk_spk_model_free(self._handle)
 
@@ -51,6 +143,9 @@ class KaldiRecognizer(object):
         else:
             raise TypeError("Unknown arguments")
 
+        if self._handle == _ffi.NULL:
+            raise Exception("Failed to create a recognizer")
+
     def __del__(self):
         _c.vosk_recognizer_free(self._handle)
 
@@ -60,11 +155,20 @@ class KaldiRecognizer(object):
     def SetWords(self, enable_words):
         _c.vosk_recognizer_set_words(self._handle, 1 if enable_words else 0)
 
+    def SetPartialWords(self, enable_partial_words):
+        _c.vosk_recognizer_set_partial_words(self._handle, 1 if enable_partial_words else 0)
+
+    def SetNLSML(self, enable_nlsml):
+        _c.vosk_recognizer_set_nlsml(self._handle, 1 if enable_nlsml else 0)
+
     def SetSpkModel(self, spk_model):
         _c.vosk_recognizer_set_spk_model(self._handle, spk_model._handle)
 
     def AcceptWaveform(self, data):
-        return _c.vosk_recognizer_accept_waveform(self._handle, data, len(data))
+        res = _c.vosk_recognizer_accept_waveform(self._handle, data, len(data))
+        if res < 0:
+            raise Exception("Failed to process waveform")
+        return res
 
     def Result(self):
         return _ffi.string(_c.vosk_recognizer_result(self._handle)).decode('utf-8')
@@ -89,3 +193,43 @@ def GpuInit():
 
 def GpuThreadInit():
     _c.vosk_gpu_thread_init()
+
+class BatchModel(object):
+
+    def __init__(self, *args):
+        self._handle = _c.vosk_batch_model_new()
+
+        if self._handle == _ffi.NULL:
+            raise Exception("Failed to create a model")
+
+    def __del__(self):
+        _c.vosk_batch_model_free(self._handle)
+
+    def Wait(self):
+        _c.vosk_batch_model_wait(self._handle)
+
+class BatchRecognizer(object):
+
+    def __init__(self, *args):
+        self._handle = _c.vosk_batch_recognizer_new(args[0]._handle, args[1])
+
+        if self._handle == _ffi.NULL:
+            raise Exception("Failed to create a recognizer")
+
+    def __del__(self):
+        _c.vosk_batch_recognizer_free(self._handle)
+
+    def AcceptWaveform(self, data):
+        res = _c.vosk_batch_recognizer_accept_waveform(self._handle, data, len(data))
+
+    def Result(self):
+        ptr = _c.vosk_batch_recognizer_front_result(self._handle)
+        res = _ffi.string(ptr).decode('utf-8')
+        _c.vosk_batch_recognizer_pop(self._handle)
+        return res
+
+    def FinishStream(self):
+        _c.vosk_batch_recognizer_finish_stream(self._handle)
+
+    def GetPendingChunks(self):
+        return _c.vosk_batch_recognizer_get_pending_chunks(self._handle)

python/vosk/transcriber/cli.py

@@ -0,0 +1,85 @@
+#!/usr/bin/env python3
+
+import logging
+import argparse
+import os
+
+from pathlib import Path
+from vosk import list_models, list_languages
+from vosk.transcriber.transcriber import Transcriber
+
+parser = argparse.ArgumentParser(
+        description = 'Transcribe audio file and save result in selected format')
+parser.add_argument(
+        '--model', '-m', type=str,
+        help='model path')
+parser.add_argument(
+        '--server', '-s', const='ws://localhost:2700', action='store_const',  
+        help='use server for recognition')
+parser.add_argument(
+        '--list-models', default=False, action='store_true', 
+        help='list available models')
+parser.add_argument(
+        '--list-languages', default=False, action='store_true',
+        help='list available languages')
+parser.add_argument(
+        '--model-name', '-n', type=str,
+        help='select model by name')
+parser.add_argument(
+        '--lang', '-l', default='en-us', type=str,
+        help='select model by language')
+parser.add_argument(
+        '--input', '-i', type=str,
+        help='audiofile')
+parser.add_argument(
+        '--output', '-o', default='', type=str,
+        help='optional output filename path')
+parser.add_argument(
+        '--output-type', '-t', default='txt', type=str,
+        help='optional arg output data type')
+parser.add_argument(
+        '--tasks', '-ts', default=10, type=int,
+        help='number of parallel recognition tasks')
+parser.add_argument(
+        '--log-level', default='INFO',
+        help='logging level')
+
+def main():
+
+    args = parser.parse_args()
+    log_level = args.log_level.upper()
+    logging.getLogger().setLevel(log_level)
+
+    if args.list_models == True:
+        list_models()
+        return
+
+    if args.list_languages == True:
+        list_languages()
+        return
+
+    if not args.input:
+        logging.info("Please specify input file or directory")
+        exit(1)
+
+    if not Path(args.input).exists():
+        logging.info("File/folder '%s' does not exist, please specify an existing file/directory" % (args.input))
+        exit(1)
+
+    transcriber = Transcriber(args)
+
+    if Path(args.input).is_dir():
+        task_list = [(Path(args.input, fn), Path(args.output, Path(fn).stem).with_suffix('.' + args.output_type)) for fn in os.listdir(args.input)]
+    elif Path(args.input).is_file():
+        if args.output == '':
+            task_list = [(Path(args.input), args.output)]
+        else:
+            task_list = [(Path(args.input), Path(args.output))]
+    else:
+        logging.info("Wrong arguments")
+        exit(1)
+
+    transcriber.process_task_list(args, task_list)    
+
+if __name__ == "__main__":
+    main()

python/vosk/transcriber/transcriber.py

@@ -0,0 +1,172 @@
+import json
+import subprocess
+import srt
+import datetime
+import os
+import logging
+import asyncio
+import websockets
+
+from queue import Queue
+from pathlib import Path
+from timeit import default_timer as timer
+from vosk import KaldiRecognizer, Model
+from multiprocessing.dummy import Pool
+
+CHUNK_SIZE = 4000
+SAMPLE_RATE = 16000.0
+
+class Transcriber:
+
+    def __init__(self, args):
+        self.model = Model(model_path=args.model, model_name=args.model_name, lang=args.lang)
+        self.args = args
+
+    def recognize_stream(self, rec, stream):
+        tot_samples = 0
+        result = []
+
+        while True:
+            data = stream.stdout.read(CHUNK_SIZE)
+
+            if len(data) == 0:
+                break
+
+            tot_samples += len(data)
+            if rec.AcceptWaveform(data):
+                jres = json.loads(rec.Result())
+                logging.info(jres)
+                result.append(jres)
+            else:
+                jres = json.loads(rec.PartialResult())
+                logging.info(jres)
+
+        jres = json.loads(rec.FinalResult())
+        logging.info(jres)
+        result.append(jres)
+
+        return result, tot_samples
+
+    async def recognize_stream_server(self, proc):
+        async with websockets.connect(self.args.server) as websocket:
+            tot_samples = 0
+            result = []
+
+            await websocket.send('{ "config" : { "sample_rate" : %f } }' % (SAMPLE_RATE))
+            while True:
+                data = await proc.stdout.read(CHUNK_SIZE)
+                tot_samples += len(data)
+                if len(data) == 0:
+                    break
+                await websocket.send(data)
+                jres = json.loads(await websocket.recv())
+                logging.info(jres)
+                if not 'partial' in jres:
+                    result.append(jres)
+            await websocket.send('{"eof" : 1}')
+            jres = json.loads(await websocket.recv())
+            logging.info(jres)
+            result.append(jres)
+
+            return result, tot_samples
+
+
+    def format_result(self, result, words_per_line=7):
+        final_result = ''
+        if self.args.output_type == 'srt':
+            subs = []
+
+            for i, res in enumerate(result):
+                if not 'result' in res:
+                    continue
+                words = res['result']
+
+                for j in range(0, len(words), words_per_line):
+                    line = words[j : j + words_per_line]
+                    s = srt.Subtitle(index=len(subs),
+                            content = ' '.join([l['word'] for l in line]),
+                            start=datetime.timedelta(seconds=line[0]['start']),
+                            end=datetime.timedelta(seconds=line[-1]['end']))
+                    subs.append(s)
+            final_result = srt.compose(subs)
+
+        elif self.args.output_type == 'txt':
+            for part in result:
+                final_result += part['text'] + ' '
+        return final_result
+
+    def resample_ffmpeg(self, infile):
+        cmd = "ffmpeg -nostdin -loglevel quiet -i {} -ar {} -ac 1 -f s16le -".format(str(infile), SAMPLE_RATE)
+        stream = subprocess.Popen(cmd.split(), stdout=subprocess.PIPE)
+        return stream
+
+    async def resample_ffmpeg_async(self, infile):
+        cmd = "ffmpeg -nostdin -loglevel quiet -i {} -ar {} -ac 1 -f s16le -".format(str(infile), SAMPLE_RATE)
+        return await asyncio.create_subprocess_shell(cmd, stdout=subprocess.PIPE)
+
+    async def server_worker(self):
+        while True:
+            try:
+                input_file, output_file = self.queue.get_nowait()
+            except:
+                break
+
+            logging.info('Recognizing {}'.format(input_file))
+            start_time = timer()
+            proc = await self.resample_ffmpeg_async(input_file)
+            result, tot_samples = await self.recognize_stream_server(proc)
+
+            final_result = self.format_result(result)
+            if output_file != '':
+                logging.info('File {} processing complete'.format(output_file))
+                with open(output_file, 'w', encoding='utf-8') as fh:
+                    fh.write(final_result)
+            else:
+                print(final_result)
+
+            await proc.wait()
+
+            elapsed = timer() - start_time
+            logging.info('Execution time: {:.3f} sec; xRT {:.3f}'.format(elapsed, float(elapsed) * (2 * SAMPLE_RATE) / tot_samples))
+            self.queue.task_done()
+
+    def pool_worker(self, inputdata):
+        logging.info('Recognizing {}'.format(inputdata[0]))
+        start_time = timer()
+
+        try:
+            stream = self.resample_ffmpeg(inputdata[0])
+        except Exception:
+            logging.info('Missing ffmpeg, please install and try again')
+            return
+
+        rec = KaldiRecognizer(self.model, SAMPLE_RATE)
+        rec.SetWords(True)
+        result, tot_samples = self.recognize_stream(rec, stream)
+        final_result = self.format_result(result)
+
+        if inputdata[1] != '':
+            logging.info('File {} processing complete'.format(inputdata[1]))
+            with open(inputdata[1], 'w', encoding='utf-8') as fh:
+                fh.write(final_result)
+        else:
+            print(final_result)
+
+        elapsed = timer() - start_time
+        logging.info('Execution time: {:.3f} sec; xRT {:.3f}'.format(elapsed, float(elapsed) * (2 * SAMPLE_RATE) / tot_samples))
+
+    async def process_task_list_server(self, task_list):
+        self.queue = Queue()
+        [self.queue.put(x) for x in task_list]
+        workers = [asyncio.create_task(self.server_worker()) for i in range(self.args.tasks)]
+        await asyncio.gather(*workers)
+
+    def process_task_list_pool(self, task_list):
+        with Pool() as pool:
+            pool.map(self.pool_worker, task_list)
+
+    def process_task_list(self, args, task_list):
+        if self.args.server is None:
+            self.process_task_list_pool(task_list)
+        else:
+            asyncio.run(self.process_task_list_server(task_list))

ruby/lib/vosk.rb

@@ -0,0 +1,5 @@
+class Vosk
+  def self.hi
+    puts "Hello world!"
+  end
+end

ruby/vosk.gemspec

@@ -0,0 +1,12 @@
+Gem::Specification.new do |s|
+  s.name        = "vosk"
+  s.version     = "0.3.43"
+  s.summary     = "Offline speech recognition API"
+  s.description = "Vosk is an offline open source speech recognition toolkit. It enables speech recognition for 20+ languages and dialects - English, Indian English, German, French, Spanish, Portuguese, Chinese, Russian, Turkish, Vietnamese, Italian, Dutch, Catalan, Arabic, Greek, Farsi, Filipino, Ukrainian, Kazakh, Swedish, Japanese, Esperanto, Hindi, Czech, Polish. More to come."
+  s.authors     = ["Alpha Cephei Inc"]
+  s.email       = "contact@alphacphei.com"
+  s.files       = ["lib/vosk.rb"]
+  s.homepage    =
+    "https://rubygems.org/gems/vosk"
+  s.license       = "Apache 2.0"
+end

rust/README.md

@@ -0,0 +1,5 @@
+See
+
+https://github.com/Bear-03/vosk-rs
+
+https://crates.io/crates/vosk

src/Makefile

@@ -1,71 +1,114 @@
+# Locations of the dependencies
 KALDI_ROOT?=$(HOME)/travis/kaldi
 OPENFST_ROOT?=$(KALDI_ROOT)/tools/openfst
 OPENBLAS_ROOT?=$(KALDI_ROOT)/tools/OpenBLAS/install
-HAVE_CUDA?=0
+MKL_ROOT?=/opt/intel/mkl
 CUDA_ROOT?=/usr/local/cuda
-EXT?=so
-CXX?=g++
-HAVE_OPENBLAS_CLAPACK=1
+USE_SHARED?=0
+# Math libraries
+HAVE_OPENBLAS_CLAPACK?=1
+HAVE_MKL?=0
 HAVE_ACCELERATE=0
-EXTRA_CFLGAS?=
+HAVE_CUDA?=0
+# Compiler
+CXX?=g++
+EXT?=so
+# Extra
+EXTRA_CFLAGS?=
 EXTRA_LDFLAGS?=
+OUTDIR?=.
 
 VOSK_SOURCES= \
-	kaldi_recognizer.cc \
+	recognizer.cc \
 	language_model.cc \
 	model.cc \
 	spk_model.cc \
 	vosk_api.cc
 
-CFLAGS=-g -O2 -std=c++17 -fPIC -DFST_NO_DYNAMIC_LINKING $(EXTRA_CFLAGS) \
-	-I. -I$(KALDI_ROOT)/src -I$(OPENFST_ROOT)/include -I$(OPENBLAS_ROOT)/include
+VOSK_HEADERS= \
+	recognizer.h \
+	language_model.h \
+	model.h \
+	spk_model.h \
+	vosk_api.h
 
-LIBS= \
-	$(KALDI_ROOT)/src/online2/kaldi-online2.a \
-	$(KALDI_ROOT)/src/decoder/kaldi-decoder.a \
-	$(KALDI_ROOT)/src/ivector/kaldi-ivector.a \
-	$(KALDI_ROOT)/src/gmm/kaldi-gmm.a \
-	$(KALDI_ROOT)/src/nnet3/kaldi-nnet3.a \
-	$(KALDI_ROOT)/src/tree/kaldi-tree.a \
-	$(KALDI_ROOT)/src/feat/kaldi-feat.a \
-	$(KALDI_ROOT)/src/lat/kaldi-lat.a \
-	$(KALDI_ROOT)/src/lm/kaldi-lm.a \
-	$(KALDI_ROOT)/src/rnnlm/kaldi-rnnlm.a \
-	$(KALDI_ROOT)/src/hmm/kaldi-hmm.a \
-	$(KALDI_ROOT)/src/transform/kaldi-transform.a \
-	$(KALDI_ROOT)/src/cudamatrix/kaldi-cudamatrix.a \
-	$(KALDI_ROOT)/src/matrix/kaldi-matrix.a \
-	$(KALDI_ROOT)/src/fstext/kaldi-fstext.a \
-	$(KALDI_ROOT)/src/util/kaldi-util.a \
-	$(KALDI_ROOT)/src/base/kaldi-base.a \
-	$(OPENFST_ROOT)/lib/libfst.a \
-	$(OPENFST_ROOT)/lib/libfstngram.a
+CFLAGS=-g -O3 -std=c++17 -Wno-deprecated-declarations -fPIC -DFST_NO_DYNAMIC_LINKING \
+       -I. -I$(KALDI_ROOT)/src -I$(OPENFST_ROOT)/include $(EXTRA_CFLAGS)
 
+LDFLAGS=
+
+ifeq ($(USE_SHARED), 0)
+    LIBS = \
+        $(KALDI_ROOT)/src/online2/kaldi-online2.a \
+        $(KALDI_ROOT)/src/decoder/kaldi-decoder.a \
+        $(KALDI_ROOT)/src/ivector/kaldi-ivector.a \
+        $(KALDI_ROOT)/src/gmm/kaldi-gmm.a \
+        $(KALDI_ROOT)/src/tree/kaldi-tree.a \
+        $(KALDI_ROOT)/src/feat/kaldi-feat.a \
+        $(KALDI_ROOT)/src/lat/kaldi-lat.a \
+        $(KALDI_ROOT)/src/lm/kaldi-lm.a \
+        $(KALDI_ROOT)/src/rnnlm/kaldi-rnnlm.a \
+        $(KALDI_ROOT)/src/hmm/kaldi-hmm.a \
+        $(KALDI_ROOT)/src/nnet3/kaldi-nnet3.a \
+        $(KALDI_ROOT)/src/transform/kaldi-transform.a \
+        $(KALDI_ROOT)/src/cudamatrix/kaldi-cudamatrix.a \
+        $(KALDI_ROOT)/src/matrix/kaldi-matrix.a \
+        $(KALDI_ROOT)/src/fstext/kaldi-fstext.a \
+        $(KALDI_ROOT)/src/util/kaldi-util.a \
+        $(KALDI_ROOT)/src/base/kaldi-base.a \
+        $(OPENFST_ROOT)/lib/libfst.a \
+        $(OPENFST_ROOT)/lib/libfstngram.a
+else
+    LDFLAGS += \
+        -L$(KALDI_ROOT)/libs \
+        -lkaldi-online2 -lkaldi-decoder -lkaldi-ivector -lkaldi-gmm -lkaldi-tree \
+        -lkaldi-feat -lkaldi-lat -lkaldi-lm -lkaldi-rnnlm -lkaldi-hmm -lkaldi-nnet3 \
+        -lkaldi-transform -lkaldi-cudamatrix -lkaldi-matrix -lkaldi-fstext \
+        -lkaldi-util -lkaldi-base -lfst -lfstngram
+endif
 
 ifeq ($(HAVE_OPENBLAS_CLAPACK), 1)
-LIBS += \
-	$(OPENBLAS_ROOT)/lib/libopenblas.a \
-	$(OPENBLAS_ROOT)/lib/liblapack.a \
-	$(OPENBLAS_ROOT)/lib/libblas.a \
-	$(OPENBLAS_ROOT)/lib/libf2c.a
+    CFLAGS += -I$(OPENBLAS_ROOT)/include
+    ifeq ($(USE_SHARED), 0)
+        LIBS += \
+            $(OPENBLAS_ROOT)/lib/libopenblas.a \
+            $(OPENBLAS_ROOT)/lib/liblapack.a \
+            $(OPENBLAS_ROOT)/lib/libblas.a \
+            $(OPENBLAS_ROOT)/lib/libf2c.a
+    else
+        LDFLAGS += -lopenblas -llapack -lblas -lf2c
+    endif
+endif
+
+ifeq ($(HAVE_MKL), 1)
+    CFLAGS += -DHAVE_MKL=1 -I$(MKL_ROOT)/include
+    LDFLAGS += -L$(MKL_ROOT)/lib/intel64 -Wl,-rpath=$(MKL_ROOT)/lib/intel64 -lmkl_rt -lmkl_intel_lp64 -lmkl_core -lmkl_sequential
 endif
 
 ifeq ($(HAVE_ACCELERATE), 1)
-LIBS += \
-	-framework Accelerate
+    LDFLAGS += -framework Accelerate
 endif
 
 ifeq ($(HAVE_CUDA), 1)
-CFLAGS+=-DHAVE_CUDA=1 -I$(CUDA_ROOT)/include
-LIBS+=-L$(CUDA_ROOT)/lib64 -lcublas -lcusparse -lcudart -lcurand -lcufft -lcusolver -lnvToolsExt
+    VOSK_SOURCES += batch_recognizer.cc batch_model.cc
+    VOSK_HEADERS += batch_recognizer.h batch_model.h
+
+    CFLAGS+=-DHAVE_CUDA=1 -I$(CUDA_ROOT)/include
+
+    LIBS := \
+        $(KALDI_ROOT)/src/cudadecoder/kaldi-cudadecoder.a \
+        $(KALDI_ROOT)/src/cudafeat/kaldi-cudafeat.a \
+        $(LIBS)
+
+    LDFLAGS += -L$(CUDA_ROOT)/lib64 -lcuda -lcublas -lcusparse -lcudart -lcurand -lcufft -lcusolver -lnvToolsExt
 endif
 
-all: libvosk.$(EXT)
+all: $(OUTDIR)/libvosk.$(EXT)
 
-libvosk.$(EXT): $(VOSK_SOURCES:.cc=.o)
-	$(CXX) --shared -s -o $@ $^ $(LIBS) -lm -latomic $(EXTRA_LDFLAGS)
+$(OUTDIR)/libvosk.$(EXT): $(VOSK_SOURCES:%.cc=$(OUTDIR)/%.o) $(LIBS)
+	$(CXX) --shared -s -o $@ $^ $(LDFLAGS) $(EXTRA_LDFLAGS)
 
-%.o: %.cc
+$(OUTDIR)/%.o: %.cc $(VOSK_HEADERS)
 	$(CXX) $(CFLAGS) -c -o $@ $<
 
 clean:

src/batch_model.cc

@@ -0,0 +1,121 @@
+// Copyright 2019-2020 Alpha Cephei Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//       http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "batch_model.h"
+
+#include <sys/stat.h>
+
+using namespace fst;
+using namespace kaldi::nnet3;
+using CorrelationID = CudaOnlinePipelineDynamicBatcher::CorrelationID;
+
+BatchModel::BatchModel() {
+    BatchedThreadedNnet3CudaOnlinePipelineConfig batched_decoder_config;
+
+    kaldi::ParseOptions po("something");
+    batched_decoder_config.Register(&po);
+    po.ReadConfigFile("model/conf/model.conf");
+
+    struct stat buffer;
+
+    string nnet3_rxfilename_ = "model/am/final.mdl";
+    string hclg_fst_rxfilename_ = "model/graph/HCLG.fst";
+    string word_syms_rxfilename_ = "model/graph/words.txt";
+    string winfo_rxfilename_ = "model/graph/phones/word_boundary.int";
+    string std_fst_rxfilename_ = "model/rescore/G.fst";
+    string carpa_rxfilename_ = "model/rescore/G.carpa";
+
+    trans_model_ = new kaldi::TransitionModel();
+    nnet_ = new kaldi::nnet3::AmNnetSimple();
+    {
+        bool binary;
+        kaldi::Input ki(nnet3_rxfilename_, &binary);
+        trans_model_->Read(ki.Stream(), binary);
+        nnet_->Read(ki.Stream(), binary);
+        SetBatchnormTestMode(true, &(nnet_->GetNnet()));
+        SetDropoutTestMode(true, &(nnet_->GetNnet()));
+        nnet3::CollapseModel(nnet3::CollapseModelConfig(), &(nnet_->GetNnet()));
+    }
+
+    if (stat(hclg_fst_rxfilename_.c_str(), &buffer) == 0) {
+        KALDI_LOG << "Loading HCLG from " << hclg_fst_rxfilename_;
+        hclg_fst_ = fst::ReadFstKaldiGeneric(hclg_fst_rxfilename_);
+    }
+
+    KALDI_LOG << "Loading words from " << word_syms_rxfilename_;
+    if (!(word_syms_ = fst::SymbolTable::ReadText(word_syms_rxfilename_))) {
+        KALDI_ERR << "Could not read symbol table from file "
+                  << word_syms_rxfilename_;
+    }
+    KALDI_ASSERT(word_syms_);
+
+    if (stat(winfo_rxfilename_.c_str(), &buffer) == 0) {
+        KALDI_LOG << "Loading winfo " << winfo_rxfilename_;
+        kaldi::WordBoundaryInfoNewOpts opts;
+        winfo_ = new kaldi::WordBoundaryInfo(opts, winfo_rxfilename_);
+    }
+
+    batched_decoder_config.num_worker_threads = -1;
+    batched_decoder_config.max_batch_size = 32;
+    batched_decoder_config.num_channels = 600;
+    batched_decoder_config.reset_on_endpoint = true;
+    batched_decoder_config.use_gpu_feature_extraction = true;
+
+    batched_decoder_config.feature_opts.feature_type = "mfcc";
+    batched_decoder_config.feature_opts.mfcc_config = "model/conf/mfcc.conf";
+    batched_decoder_config.feature_opts.ivector_extraction_config = "model/conf/ivector.conf";
+    batched_decoder_config.decoder_opts.max_active = 7000;
+    batched_decoder_config.decoder_opts.default_beam = 13.0;
+    batched_decoder_config.decoder_opts.lattice_beam = 6.0;
+    batched_decoder_config.compute_opts.acoustic_scale = 1.0;
+    batched_decoder_config.compute_opts.frame_subsampling_factor = 3;
+
+    int32 nnet_left_context, nnet_right_context;
+    nnet3::ComputeSimpleNnetContext(nnet_->GetNnet(), &nnet_left_context,
+                                    &nnet_right_context);
+
+    batched_decoder_config.compute_opts.frames_per_chunk = std::max(51, (nnet_right_context + 3 - nnet_right_context % 3));
+
+    cuda_pipeline_ = new BatchedThreadedNnet3CudaOnlinePipeline 
+         (batched_decoder_config, *hclg_fst_, *nnet_, *trans_model_);
+    cuda_pipeline_->SetSymbolTable(*word_syms_);
+
+    CudaOnlinePipelineDynamicBatcherConfig dynamic_batcher_config;
+    dynamic_batcher_ = new CudaOnlinePipelineDynamicBatcher(dynamic_batcher_config,
+                                                            *cuda_pipeline_);
+
+    samples_per_chunk_ = cuda_pipeline_->GetNSampsPerChunk();
+    last_id_ = 0;
+}
+
+uint64_t BatchModel::GetID(BatchRecognizer *recognizer) {
+    return last_id_++;
+}
+
+BatchModel::~BatchModel() {
+
+    delete trans_model_;
+    delete nnet_;
+    delete word_syms_;
+    delete winfo_;
+    delete hclg_fst_;
+
+    delete cuda_pipeline_;
+    delete dynamic_batcher_;
+}
+
+void BatchModel::WaitForCompletion()
+{
+    dynamic_batcher_->WaitForCompletion();
+}

src/batch_model.h

@@ -0,0 +1,68 @@
+// Copyright 2019 Alpha Cephei Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//       http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef VOSK_BATCH_MODEL_H
+#define VOSK_BATCH_MODEL_H
+
+#include "base/kaldi-common.h"
+#include "util/common-utils.h"
+#include "fstext/fstext-lib.h"
+#include "fstext/fstext-utils.h"
+#include "decoder/lattice-faster-decoder.h"
+#include "feat/feature-mfcc.h"
+#include "lat/kaldi-lattice.h"
+#include "lat/word-align-lattice.h"
+#include "lat/compose-lattice-pruned.h"
+#include "nnet3/am-nnet-simple.h"
+#include "nnet3/nnet-am-decodable-simple.h"
+#include "nnet3/nnet-utils.h"
+
+#include "cudadecoder/cuda-online-pipeline-dynamic-batcher.h"
+#include "cudadecoder/batched-threaded-nnet3-cuda-online-pipeline.h"
+#include "cudadecoder/batched-threaded-nnet3-cuda-pipeline2.h"
+#include "cudadecoder/cuda-pipeline-common.h"
+
+#include "model.h"
+
+using namespace kaldi;
+using namespace kaldi::cuda_decoder;
+
+class BatchRecognizer;
+
+class BatchModel {
+    public:
+        BatchModel();
+        ~BatchModel();
+
+        uint64_t GetID(BatchRecognizer *recognizer);
+        void WaitForCompletion();
+
+    private:
+        friend class BatchRecognizer;
+
+        kaldi::TransitionModel *trans_model_ = nullptr;
+        kaldi::nnet3::AmNnetSimple *nnet_ = nullptr;
+        const fst::SymbolTable *word_syms_ = nullptr;
+
+        fst::Fst<fst::StdArc> *hclg_fst_ = nullptr;
+        kaldi::WordBoundaryInfo *winfo_ = nullptr;
+
+        BatchedThreadedNnet3CudaOnlinePipeline *cuda_pipeline_ = nullptr;
+        CudaOnlinePipelineDynamicBatcher *dynamic_batcher_ = nullptr;
+
+        int32 samples_per_chunk_;
+        uint64_t last_id_;
+};
+
+#endif /* VOSK_BATCH_MODEL_H */

src/batch_recognizer.cc

@@ -0,0 +1,202 @@
+// Copyright 2019-2020 Alpha Cephei Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//       http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "batch_recognizer.h"
+
+#include "fstext/fstext-utils.h"
+#include "lat/sausages.h"
+#include "json.h"
+
+BatchRecognizer::BatchRecognizer(BatchModel *model, float
+                                 sample_frequency) : model_(model), sample_frequency_(sample_frequency),
+                                 initialized_(false), callbacks_set_(false), nlsml_(false) {
+    id_ = model->GetID(this);
+
+
+    resampler_ = new LinearResample(
+        sample_frequency, 16000.0f,
+        std::min(sample_frequency / 2, 16000.0f / 2), 6);
+}
+
+BatchRecognizer::~BatchRecognizer() {
+    delete resampler_;
+    // Drop the ID
+}
+
+void BatchRecognizer::FinishStream()
+{
+    SubVector<BaseFloat> chunk = buffer_.Range(0, buffer_.Dim());
+    model_->dynamic_batcher_->Push(id_, !initialized_, true, chunk);
+}
+
+void BatchRecognizer::PushLattice(CompactLattice &clat, BaseFloat offset)
+{
+    fst::ScaleLattice(fst::GraphLatticeScale(0.9), &clat);
+
+    CompactLattice aligned_lat;
+    WordAlignLattice(clat, *model_->trans_model_, *model_->winfo_, 0, &aligned_lat);
+
+    MinimumBayesRisk mbr(aligned_lat);
+    const vector<BaseFloat> &conf = mbr.GetOneBestConfidences();
+    const vector<int32> &words = mbr.GetOneBest();
+    const vector<pair<BaseFloat, BaseFloat> > &times =
+          mbr.GetOneBestTimes();
+
+    int size = words.size();
+
+    if (nlsml_) {
+
+        std::stringstream ss;
+        std::stringstream text;
+        ss << "<?xml version=\"1.0\"?>\n";
+        ss << "<result grammar=\"default\">\n";
+        BaseFloat confidence = 0.0;
+        for (int i = 0; i < size; i++) {
+            if (i) {
+                text << " ";
+            }
+            confidence += conf[i];
+            text << model_->word_syms_->Find(words[i]);
+        }
+        confidence /= size;
+
+        ss << "<interpretation grammar=\"default\" confidence=\"" << confidence << "\">\n";
+        ss << "<input mode=\"speech\">" << text.str() << "</input>\n";
+        ss << "<instance>" << text.str() << "</instance>\n";
+        ss << "</interpretation>\n";
+        ss << "</result>\n";
+
+        results_.push(ss.str());
+
+    } else {
+        json::JSON obj;
+        stringstream text;
+
+        // Create JSON object
+        for (int i = 0; i < size; i++) {
+            json::JSON word;
+
+            word["word"] = model_->word_syms_->Find(words[i]);
+            word["start"] = round(times[i].first) * 0.03 + offset;
+            word["end"] = round(times[i].second) * 0.03 + offset;
+            word["conf"] = conf[i];
+            obj["result"].append(word);
+
+            if (i) {
+                text << " ";
+            }
+            text << model_->word_syms_->Find(words[i]);
+        }
+        obj["text"] = text.str();
+
+//      KALDI_LOG << "Result " << id << " " << obj.dump();
+
+        results_.push(obj.dump());
+    }
+}
+
+void BatchRecognizer::SetNLSML(bool nlsml)
+{
+    nlsml_ = nlsml;
+}
+
+
+void BatchRecognizer::AcceptWaveform(const char *data, int len)
+{
+    uint64_t id = id_;
+    if (!callbacks_set_) {
+        // Define the callback for results.
+#if 0
+         model_->cuda_pipeline_->SetBestPathCallback(
+          id,
+          [&, id](const std::string &str, bool partial,
+                       bool endpoint_detected) {
+              if (partial) {
+                  KALDI_LOG << "id #" << id << " [partial] : " << str << ":";
+              }
+
+              if (endpoint_detected) {
+                  KALDI_LOG << "id #" << id << " [endpoint detected]";
+              }
+
+              if (!partial) {
+                  KALDI_LOG << "id #" << id << " : " << str;
+              }
+            });
+#endif
+        model_->cuda_pipeline_->SetLatticeCallback(
+          id,
+          [&, id](SegmentedLatticeCallbackParams& params) {
+              if (params.results.empty()) {
+                  KALDI_WARN << "Empty result for callback";
+                  return;
+              }
+              CompactLattice *clat = params.results[0].GetLatticeResult();
+              BaseFloat offset = params.results[0].GetTimeOffsetSeconds();
+              PushLattice(*clat, offset);
+          },
+          CudaPipelineResult::RESULT_TYPE_LATTICE);
+        callbacks_set_ = true;
+    }
+
+    Vector<BaseFloat> input_wave(len / 2);
+    for (int i = 0; i < len / 2; i++)
+        input_wave(i) = *(((short *)data) + i);
+
+    Vector<BaseFloat> resampled_wave;
+    resampler_->Resample(input_wave, true, &resampled_wave);
+
+    int32 end = buffer_.Dim();
+    buffer_.Resize(end + resampled_wave.Dim(), kCopyData);
+    buffer_.Range(end, resampled_wave.Dim()).CopyFromVec(resampled_wave);
+
+    // Pick chunks and submit them to the batcher
+    int32 i = 0;
+    while (i + model_->samples_per_chunk_ <= buffer_.Dim()) {
+        model_->dynamic_batcher_->Push(id_, !initialized_, false,
+                                       buffer_.Range(i, model_->samples_per_chunk_));
+        initialized_ = true;
+        i += model_->samples_per_chunk_;
+    }
+
+    // Keep remaining data
+    if (i > 0) {
+        int32 tail = buffer_.Dim() - i;
+        for (int j = 0; j < tail; j++) {
+            buffer_(j) = buffer_(i + j);
+        }
+        buffer_.Resize(tail, kCopyData);
+    }
+}
+
+const char* BatchRecognizer::FrontResult()
+{
+    if (results_.empty()) {
+        return "";
+    }
+    return results_.front().c_str();
+}
+
+void BatchRecognizer::Pop()
+{
+    if (results_.empty()) {
+        return;
+    }
+    results_.pop();
+}
+
+int BatchRecognizer::GetNumPendingChunks()
+{
+    return model_->dynamic_batcher_->GetNumPendingChunks(id_);
+}

src/batch_recognizer.h

@@ -0,0 +1,55 @@
+// Copyright 2019 Alpha Cephei Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//       http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef VOSK_BATCH_RECOGNIZER_H
+#define VOSK_BATCH_RECOGNIZER_H
+
+#include "base/kaldi-common.h"
+#include "util/common-utils.h"
+#include "feat/resample.h"
+
+#include <queue>
+
+#include "batch_model.h"
+
+using namespace kaldi;
+
+class BatchRecognizer {
+    public:
+        BatchRecognizer(BatchModel *model, float sample_frequency);
+        ~BatchRecognizer();
+
+        void AcceptWaveform(const char *data, int len);
+        int GetNumPendingChunks();
+        const char *FrontResult();
+        void Pop();
+        void FinishStream();
+        void SetNLSML(bool nlsml);
+
+    private:
+
+        void PushLattice(CompactLattice &clat, BaseFloat offset);
+
+        BatchModel *model_;
+        uint64_t id_;
+        bool initialized_;
+        bool callbacks_set_;
+        bool nlsml_;
+        float sample_frequency_;
+        std::queue<std::string> results_;
+        LinearResample *resampler_;
+        kaldi::Vector<BaseFloat> buffer_;
+};
+
+#endif /* VOSK_BATCH_RECOGNIZER_H */

src/json.h

@@ -424,7 +424,7 @@ class JSON
         Class Type = Class::Null;
 };
 
-JSON Array() {
+inline JSON Array() {
     return JSON::Make( JSON::Class::Array );
 }
 
@@ -435,11 +435,11 @@ JSON Array( T... args ) {
     return arr;
 }
 
-JSON Object() {
+inline JSON Object() {
     return JSON::Make( JSON::Class::Object );
 }
 
-std::ostream& operator<<( std::ostream &os, const JSON &json ) {
+inline std::ostream& operator<<( std::ostream &os, const JSON &json ) {
     os << json.dump();
     return os;
 }
@@ -647,7 +647,7 @@ namespace {
     }
 }
 
-JSON JSON::Load( const string &str ) {
+inline JSON JSON::Load( const string &str ) {
     size_t offset = 0;
     return parse_next( str, offset );
 }

src/model.cc

@@ -109,10 +109,10 @@ Model::Model(const char *model_path) : model_path_str_(model_path) {
 
     struct stat buffer;
     string am_v2_path = model_path_str_ + "/am/final.mdl";
-    string mfcc_v2_path = model_path_str_ + "/conf/mfcc.conf";
+    string model_conf_v2_path = model_path_str_ + "/conf/model.conf";
     string am_v1_path = model_path_str_ + "/final.mdl";
     string mfcc_v1_path = model_path_str_ + "/mfcc.conf";
-    if (stat(am_v2_path.c_str(), &buffer) == 0 && stat(mfcc_v2_path.c_str(), &buffer) == 0) {
+    if (stat(am_v2_path.c_str(), &buffer) == 0 && stat(model_conf_v2_path.c_str(), &buffer) == 0) {
         ConfigureV2();
         ReadDataFiles();
     } else if (stat(am_v1_path.c_str(), &buffer) == 0 && stat(mfcc_v1_path.c_str(), &buffer) == 0) {
@@ -168,13 +168,13 @@ void Model::ConfigureV1()
     std_fst_rxfilename_ = model_path_str_ + "/rescore/G.fst";
     final_ie_rxfilename_ = model_path_str_ + "/ivector/final.ie";
     mfcc_conf_rxfilename_ = model_path_str_ + "/mfcc.conf";
+    fbank_conf_rxfilename_ = model_path_str_ + "/fbank.conf";
     global_cmvn_stats_rxfilename_ = model_path_str_ + "/global_cmvn.stats";
     pitch_conf_rxfilename_ = model_path_str_ + "/pitch.conf";
     rnnlm_word_feats_rxfilename_ = model_path_str_ + "/rnnlm/word_feats.txt";
     rnnlm_feat_embedding_rxfilename_ = model_path_str_ + "/rnnlm/feat_embedding.final.mat";
     rnnlm_config_rxfilename_ = model_path_str_ + "/rnnlm/special_symbol_opts.conf";
     rnnlm_lm_rxfilename_ = model_path_str_ + "/rnnlm/final.raw";
-    rnnlm_lm_fst_rxfilename_ = model_path_str_ + "/rescore/G.fst";
 }
 
 void Model::ConfigureV2()
@@ -197,13 +197,13 @@ void Model::ConfigureV2()
     std_fst_rxfilename_ = model_path_str_ + "/rescore/G.fst";
     final_ie_rxfilename_ = model_path_str_ + "/ivector/final.ie";
     mfcc_conf_rxfilename_ = model_path_str_ + "/conf/mfcc.conf";
+    fbank_conf_rxfilename_ = model_path_str_ + "/conf/fbank.conf";
     global_cmvn_stats_rxfilename_ = model_path_str_ + "/am/global_cmvn.stats";
     pitch_conf_rxfilename_ = model_path_str_ + "/conf/pitch.conf";
     rnnlm_word_feats_rxfilename_ = model_path_str_ + "/rnnlm/word_feats.txt";
     rnnlm_feat_embedding_rxfilename_ = model_path_str_ + "/rnnlm/feat_embedding.final.mat";
     rnnlm_config_rxfilename_ = model_path_str_ + "/rnnlm/special_symbol_opts.conf";
     rnnlm_lm_rxfilename_ = model_path_str_ + "/rnnlm/final.raw";
-    rnnlm_lm_fst_rxfilename_ = model_path_str_ + "/rescore/G.fst";
 }
 
 void Model::ReadDataFiles()
@@ -215,9 +215,17 @@ void Model::ReadDataFiles()
          " lattice-beam=" << nnet3_decoding_config_.lattice_beam;
     KALDI_LOG << "Silence phones " << endpoint_config_.silence_phones;
 
-    feature_info_.feature_type = "mfcc";
-    ReadConfigFromFile(mfcc_conf_rxfilename_, &feature_info_.mfcc_opts);
-    feature_info_.mfcc_opts.frame_opts.allow_downsample = true; // It is safe to downsample
+    if (stat(mfcc_conf_rxfilename_.c_str(), &buffer) == 0) {
+        feature_info_.feature_type = "mfcc";
+        ReadConfigFromFile(mfcc_conf_rxfilename_, &feature_info_.mfcc_opts);
+        feature_info_.mfcc_opts.frame_opts.allow_downsample = true; // It is safe to downsample
+    } else if (stat(fbank_conf_rxfilename_.c_str(), &buffer) == 0) {
+        feature_info_.feature_type = "fbank";
+        ReadConfigFromFile(fbank_conf_rxfilename_, &feature_info_.fbank_opts);
+        feature_info_.fbank_opts.frame_opts.allow_downsample = true; // It is safe to downsample
+    } else {
+        KALDI_ERR << "Failed to find feature config file";
+    }
 
     feature_info_.silence_weighting_config.silence_weight = 1e-3;
     feature_info_.silence_weighting_config.silence_phones_str = endpoint_config_.silence_phones;
@@ -233,9 +241,9 @@ void Model::ReadDataFiles()
         SetDropoutTestMode(true, &(nnet_->GetNnet()));
         nnet3::CollapseModel(nnet3::CollapseModelConfig(), &(nnet_->GetNnet()));
     }
+
     decodable_info_ = new nnet3::DecodableNnetSimpleLoopedInfo(decodable_opts_,
                                                                nnet_);
-
     if (stat(final_ie_rxfilename_.c_str(), &buffer) == 0) {
         KALDI_LOG << "Loading i-vector extractor from " << final_ie_rxfilename_;
 
@@ -263,7 +271,8 @@ void Model::ReadDataFiles()
     if (stat(pitch_conf_rxfilename_.c_str(), &buffer) == 0) {
         KALDI_LOG << "Using pitch in feature pipeline";
         feature_info_.add_pitch = true;
-        ReadConfigFromFile(pitch_conf_rxfilename_, &feature_info_.pitch_opts);
+        ReadConfigsFromFile(pitch_conf_rxfilename_,
+                            &feature_info_.pitch_opts, &feature_info_.pitch_process_opts);
     }
 
     if (stat(hclg_fst_rxfilename_.c_str(), &buffer) == 0) {
@@ -296,12 +305,19 @@ void Model::ReadDataFiles()
         winfo_ = new kaldi::WordBoundaryInfo(opts, winfo_rxfilename_);
     }
 
+    if (stat(carpa_rxfilename_.c_str(), &buffer) == 0) {
+
+        KALDI_LOG << "Loading subtract G.fst model from " << std_fst_rxfilename_;
+        graph_lm_fst_ = fst::ReadAndPrepareLmFst(std_fst_rxfilename_);
+        KALDI_LOG << "Loading CARPA model from " << carpa_rxfilename_;
+        ReadKaldiObject(carpa_rxfilename_, &const_arpa_);
+    }
+
     // RNNLM Rescoring
     if (stat(rnnlm_lm_rxfilename_.c_str(), &buffer) == 0) {
         KALDI_LOG << "Loading RNNLM model from " << rnnlm_lm_rxfilename_;
 
         ReadKaldiObject(rnnlm_lm_rxfilename_, &rnnlm);
-        rnnlm_lm_fst_ = fst::ReadAndPrepareLmFst(rnnlm_lm_fst_rxfilename_);
         Matrix<BaseFloat> feature_embedding_mat;
         ReadKaldiObject(rnnlm_feat_embedding_rxfilename_, &feature_embedding_mat);
         SparseMatrix<BaseFloat> word_feature_mat;
@@ -319,17 +335,9 @@ void Model::ReadDataFiles()
 
         ReadConfigFromFile(rnnlm_config_rxfilename_, &rnnlm_compute_opts);
 
-    } else if (stat(carpa_rxfilename_.c_str(), &buffer) == 0) {
-
-        KALDI_LOG << "Loading CARPA model from " << carpa_rxfilename_;
-        std_lm_fst_ = fst::ReadFstKaldi(std_fst_rxfilename_);
-        fst::Project(std_lm_fst_, fst::ProjectType::OUTPUT);
-        if (std_lm_fst_->Properties(fst::kILabelSorted, true) == 0) {
-            fst::ILabelCompare<fst::StdArc> ilabel_comp;
-            fst::ArcSort(std_lm_fst_, ilabel_comp);
-        }
-        ReadKaldiObject(carpa_rxfilename_, &const_arpa_);
+        rnnlm_enabled_ = true;
     }
+
 }
 
 void Model::Ref() 
@@ -363,5 +371,5 @@ Model::~Model() {
     delete hclg_fst_;
     delete hcl_fst_;
     delete g_fst_;
-    delete std_lm_fst_;
+    delete graph_lm_fst_;
 }

src/model.h

@@ -21,7 +21,7 @@
 #include "online2/onlinebin-util.h"
 #include "online2/online-timing.h"
 #include "online2/online-endpoint.h"
-#include "online2/online-nnet3-decoding.h"
+#include "online2/online-nnet3-incremental-decoding.h"
 #include "online2/online-feature-pipeline.h"
 #include "lat/lattice-functions.h"
 #include "lat/sausages.h"
@@ -36,7 +36,7 @@
 using namespace kaldi;
 using namespace std;
 
-class KaldiRecognizer;
+class Recognizer;
 
 class Model {
 
@@ -52,7 +52,7 @@ protected:
     void ConfigureV2();
     void ReadDataFiles();
 
-    friend class KaldiRecognizer;
+    friend class Recognizer;
 
     string model_path_str_;
     string nnet3_rxfilename_;
@@ -66,17 +66,17 @@ protected:
     string std_fst_rxfilename_;
     string final_ie_rxfilename_;
     string mfcc_conf_rxfilename_;
+    string fbank_conf_rxfilename_;
     string global_cmvn_stats_rxfilename_;
     string pitch_conf_rxfilename_;
 
     string rnnlm_word_feats_rxfilename_;
     string rnnlm_feat_embedding_rxfilename_;
     string rnnlm_config_rxfilename_;
-    string rnnlm_lm_fst_rxfilename_;
     string rnnlm_lm_rxfilename_;
 
     kaldi::OnlineEndpointConfig endpoint_config_;
-    kaldi::LatticeFasterDecoderConfig nnet3_decoding_config_;
+    kaldi::LatticeIncrementalDecoderConfig nnet3_decoding_config_;
     kaldi::nnet3::NnetSimpleLoopedComputationOptions decodable_opts_;
     kaldi::OnlineNnet2FeaturePipelineInfo feature_info_;
 
@@ -92,13 +92,13 @@ protected:
     fst::Fst<fst::StdArc> *hcl_fst_ = nullptr;
     fst::Fst<fst::StdArc> *g_fst_ = nullptr;
 
-    fst::VectorFst<fst::StdArc> *std_lm_fst_ = nullptr;
+    fst::VectorFst<fst::StdArc> *graph_lm_fst_ = nullptr;
     kaldi::ConstArpaLm const_arpa_;
 
     kaldi::rnnlm::RnnlmComputeStateComputationOptions rnnlm_compute_opts;
     CuMatrix<BaseFloat> word_embedding_mat;
-    fst::VectorFst<fst::StdArc> *rnnlm_lm_fst_ = NULL;
     kaldi::nnet3::Nnet rnnlm;
+    bool rnnlm_enabled_ = false;
 
     std::atomic<int> ref_cnt_;
 };

src/recognizer.cc

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "kaldi_recognizer.h"
+#include "recognizer.h"
 #include "json.h"
 #include "fstext/fstext-utils.h"
 #include "lat/sausages.h"
@@ -21,7 +21,7 @@
 using namespace fst;
 using namespace kaldi::nnet3;
 
-KaldiRecognizer::KaldiRecognizer(Model *model, float sample_frequency) : model_(model), spk_model_(0), sample_frequency_(sample_frequency) {
+Recognizer::Recognizer(Model *model, float sample_frequency) : model_(model), spk_model_(0), sample_frequency_(sample_frequency) {
 
     model_->Ref();
 
@@ -36,7 +36,7 @@ KaldiRecognizer::KaldiRecognizer(Model *model, float sample_frequency) : model_(
         }
     }
 
-    decoder_ = new kaldi::SingleUtteranceNnet3Decoder(model_->nnet3_decoding_config_,
+    decoder_ = new kaldi::SingleUtteranceNnet3IncrementalDecoder(model_->nnet3_decoding_config_,
             *model_->trans_model_,
             *model_->decodable_info_,
             model_->hclg_fst_ ? *model_->hclg_fst_ : *decode_fst_,
@@ -46,7 +46,7 @@ KaldiRecognizer::KaldiRecognizer(Model *model, float sample_frequency) : model_(
     InitRescoring();
 }
 
-KaldiRecognizer::KaldiRecognizer(Model *model, float sample_frequency, char const *grammar) : model_(model), spk_model_(0), sample_frequency_(sample_frequency)
+Recognizer::Recognizer(Model *model, float sample_frequency, char const *grammar) : model_(model), spk_model_(0), sample_frequency_(sample_frequency)
 {
     model_->Ref();
 
@@ -97,7 +97,7 @@ KaldiRecognizer::KaldiRecognizer(Model *model, float sample_frequency, char cons
         KALDI_WARN << "Runtime graphs are not supported by this model";
     }
 
-    decoder_ = new kaldi::SingleUtteranceNnet3Decoder(model_->nnet3_decoding_config_,
+    decoder_ = new kaldi::SingleUtteranceNnet3IncrementalDecoder(model_->nnet3_decoding_config_,
             *model_->trans_model_,
             *model_->decodable_info_,
             model_->hclg_fst_ ? *model_->hclg_fst_ : *decode_fst_,
@@ -107,7 +107,7 @@ KaldiRecognizer::KaldiRecognizer(Model *model, float sample_frequency, char cons
     InitRescoring();
 }
 
-KaldiRecognizer::KaldiRecognizer(Model *model, float sample_frequency, SpkModel *spk_model) : model_(model), spk_model_(spk_model), sample_frequency_(sample_frequency) {
+Recognizer::Recognizer(Model *model, float sample_frequency, SpkModel *spk_model) : model_(model), spk_model_(spk_model), sample_frequency_(sample_frequency) {
 
     model_->Ref();
     spk_model->Ref();
@@ -123,7 +123,7 @@ KaldiRecognizer::KaldiRecognizer(Model *model, float sample_frequency, SpkModel
         }
     }
 
-    decoder_ = new kaldi::SingleUtteranceNnet3Decoder(model_->nnet3_decoding_config_,
+    decoder_ = new kaldi::SingleUtteranceNnet3IncrementalDecoder(model_->nnet3_decoding_config_,
             *model_->trans_model_,
             *model_->decodable_info_,
             model_->hclg_fst_ ? *model_->hclg_fst_ : *decode_fst_,
@@ -135,27 +135,27 @@ KaldiRecognizer::KaldiRecognizer(Model *model, float sample_frequency, SpkModel
     InitRescoring();
 }
 
-KaldiRecognizer::~KaldiRecognizer() {
+Recognizer::~Recognizer() {
     delete decoder_;
     delete feature_pipeline_;
     delete silence_weighting_;
     delete g_fst_;
     delete decode_fst_;
     delete spk_feature_;
-    delete lm_fst_;
 
-    delete info;
-    delete lm_to_subtract_det_backoff;
-    delete lm_to_subtract_det_scale;
-    delete lm_to_add_orig;
-    delete lm_to_add;
+    delete lm_to_subtract_;
+    delete carpa_to_add_;
+    delete carpa_to_add_scale_;
+    delete rnnlm_info_;
+    delete rnnlm_to_add_;
+    delete rnnlm_to_add_scale_;
 
     model_->Unref();
     if (spk_model_)
          spk_model_->Unref();
 }
 
-void KaldiRecognizer::InitState()
+void Recognizer::InitState()
 {
     frame_offset_ = 0;
     samples_processed_ = 0;
@@ -164,27 +164,28 @@ void KaldiRecognizer::InitState()
     state_ = RECOGNIZER_INITIALIZED;
 }
 
-void KaldiRecognizer::InitRescoring()
+void Recognizer::InitRescoring()
 {
-    if (model_->rnnlm_lm_fst_) {
-        float lm_scale = 0.5;
-        int lm_order = 4;
+    if (model_->graph_lm_fst_) {
 
-        info = new kaldi::rnnlm::RnnlmComputeStateInfo(model_->rnnlm_compute_opts, model_->rnnlm, model_->word_embedding_mat);
-        lm_to_subtract_det_backoff = new fst::BackoffDeterministicOnDemandFst<fst::StdArc>(*model_->rnnlm_lm_fst_);
-        lm_to_subtract_det_scale = new fst::ScaleDeterministicOnDemandFst(-lm_scale, lm_to_subtract_det_backoff);
-        lm_to_add_orig = new kaldi::rnnlm::KaldiRnnlmDeterministicFst(lm_order, *info);
-        lm_to_add = new fst::ScaleDeterministicOnDemandFst(lm_scale, lm_to_add_orig);
-
-    } else if (model_->std_lm_fst_) {
-        fst::CacheOptions cache_opts(true, 50000);
+        fst::CacheOptions cache_opts(true, -1);
         fst::ArcMapFstOptions mapfst_opts(cache_opts);
-        fst::StdToLatticeMapper<kaldi::BaseFloat> mapper;
-        lm_fst_ = new fst::ArcMapFst<fst::StdArc, kaldi::LatticeArc, fst::StdToLatticeMapper<kaldi::BaseFloat> >(*model_->std_lm_fst_, mapper, mapfst_opts);
+        fst::StdToLatticeMapper<BaseFloat> mapper;
+
+        lm_to_subtract_ = new fst::ArcMapFst<fst::StdArc, LatticeArc, fst::StdToLatticeMapper<BaseFloat> >(*model_->graph_lm_fst_, mapper, mapfst_opts);
+        carpa_to_add_ = new ConstArpaLmDeterministicFst(model_->const_arpa_);
+
+        if (model_->rnnlm_enabled_) {
+           int lm_order = 4;
+           rnnlm_info_ = new kaldi::rnnlm::RnnlmComputeStateInfo(model_->rnnlm_compute_opts, model_->rnnlm, model_->word_embedding_mat);
+           rnnlm_to_add_ = new kaldi::rnnlm::KaldiRnnlmDeterministicFst(lm_order, *rnnlm_info_);
+           rnnlm_to_add_scale_ = new fst::ScaleDeterministicOnDemandFst(0.5, rnnlm_to_add_);
+           carpa_to_add_scale_ = new fst::ScaleDeterministicOnDemandFst(-0.5, carpa_to_add_);
+        }
     }
 }
 
-void KaldiRecognizer::CleanUp()
+void Recognizer::CleanUp()
 {
     delete silence_weighting_;
     silence_weighting_ = new kaldi::OnlineSilenceWeighting(*model_->trans_model_, model_->feature_info_.silence_weighting_config, 3);
@@ -207,7 +208,7 @@ void KaldiRecognizer::CleanUp()
         delete feature_pipeline_;
 
         feature_pipeline_ = new kaldi::OnlineNnet2FeaturePipeline (model_->feature_info_);
-        decoder_ = new kaldi::SingleUtteranceNnet3Decoder(model_->nnet3_decoding_config_,
+        decoder_ = new kaldi::SingleUtteranceNnet3IncrementalDecoder(model_->nnet3_decoding_config_,
             *model_->trans_model_,
             *model_->decodable_info_,
             model_->hclg_fst_ ? *model_->hclg_fst_ : *decode_fst_,
@@ -222,7 +223,7 @@ void KaldiRecognizer::CleanUp()
     }
 }
 
-void KaldiRecognizer::UpdateSilenceWeights()
+void Recognizer::UpdateSilenceWeights()
 {
     if (silence_weighting_->Active() && feature_pipeline_->NumFramesReady() > 0 &&
         feature_pipeline_->IvectorFeature() != nullptr) {
@@ -235,17 +236,27 @@ void KaldiRecognizer::UpdateSilenceWeights()
     }
 }
 
-void KaldiRecognizer::SetMaxAlternatives(int max_alternatives)
+void Recognizer::SetMaxAlternatives(int max_alternatives)
 {
     max_alternatives_ = max_alternatives;
 }
 
-void KaldiRecognizer::SetWords(bool words)
+void Recognizer::SetWords(bool words)
 {
     words_ = words;
 }
 
-void KaldiRecognizer::SetSpkModel(SpkModel *spk_model)
+void Recognizer::SetPartialWords(bool partial_words)
+{
+    partial_words_ = partial_words;
+}
+
+void Recognizer::SetNLSML(bool nlsml)
+{
+    nlsml_ = nlsml;
+}
+
+void Recognizer::SetSpkModel(SpkModel *spk_model)
 {
     if (state_ == RECOGNIZER_RUNNING) {
         KALDI_ERR << "Can't add speaker model to already running recognizer";
@@ -256,7 +267,7 @@ void KaldiRecognizer::SetSpkModel(SpkModel *spk_model)
     spk_feature_ = new OnlineMfcc(spk_model_->spkvector_mfcc_opts);
 }
 
-bool KaldiRecognizer::AcceptWaveform(const char *data, int len)
+bool Recognizer::AcceptWaveform(const char *data, int len)
 {
     Vector<BaseFloat> wave;
     wave.Resize(len / 2, kUndefined);
@@ -265,7 +276,7 @@ bool KaldiRecognizer::AcceptWaveform(const char *data, int len)
     return AcceptWaveform(wave);
 }
 
-bool KaldiRecognizer::AcceptWaveform(const short *sdata, int len)
+bool Recognizer::AcceptWaveform(const short *sdata, int len)
 {
     Vector<BaseFloat> wave;
     wave.Resize(len, kUndefined);
@@ -274,7 +285,7 @@ bool KaldiRecognizer::AcceptWaveform(const short *sdata, int len)
     return AcceptWaveform(wave);
 }
 
-bool KaldiRecognizer::AcceptWaveform(const float *fdata, int len)
+bool Recognizer::AcceptWaveform(const float *fdata, int len)
 {
     Vector<BaseFloat> wave;
     wave.Resize(len, kUndefined);
@@ -283,7 +294,7 @@ bool KaldiRecognizer::AcceptWaveform(const float *fdata, int len)
     return AcceptWaveform(wave);
 }
 
-bool KaldiRecognizer::AcceptWaveform(Vector<BaseFloat> &wdata)
+bool Recognizer::AcceptWaveform(Vector<BaseFloat> &wdata)
 {
     // Cleanup if we finalized previous utterance or the whole feature pipeline
     if (!(state_ == RECOGNIZER_RUNNING || state_ == RECOGNIZER_INITIALIZED)) {
@@ -342,7 +353,7 @@ static void RunNnetComputation(const MatrixBase<BaseFloat> &features,
 
 #define MIN_SPK_FEATS 50
 
-bool KaldiRecognizer::GetSpkVector(Vector<BaseFloat> &out_xvector, int *num_spk_frames)
+bool Recognizer::GetSpkVector(Vector<BaseFloat> &out_xvector, int *num_spk_frames)
 {
     vector<int32> nonsilence_frames;
     if (silence_weighting_->Active() && feature_pipeline_->NumFramesReady() > 0) {
@@ -407,14 +418,23 @@ bool KaldiRecognizer::GetSpkVector(Vector<BaseFloat> &out_xvector, int *num_spk_
     return true;
 }
 
-
-const char *KaldiRecognizer::MbrResult(CompactLattice &rlat)
+// If we can't align, we still need to prepare for MBR
+static void CopyLatticeForMbr(CompactLattice &lat, CompactLattice *lat_out)
 {
+    *lat_out = lat;
+    RmEpsilon(lat_out, true);
+    fst::CreateSuperFinal(lat_out);
+    TopSortCompactLatticeIfNeeded(lat_out);
+}
+
+const char *Recognizer::MbrResult(CompactLattice &rlat)
+{
+
     CompactLattice aligned_lat;
     if (model_->winfo_) {
         WordAlignLattice(rlat, *model_->trans_model_, *model_->winfo_, 0, &aligned_lat);
     } else {
-        aligned_lat = rlat;
+        CopyLatticeForMbr(rlat, &aligned_lat);
     }
 
     MinimumBayesRisk mbr(aligned_lat);
@@ -522,7 +542,7 @@ static bool CompactLatticeToWordAlignmentWeight(const CompactLattice &clat,
 }
 
 
-const char *KaldiRecognizer::NbestResult(CompactLattice &clat)
+const char *Recognizer::NbestResult(CompactLattice &clat)
 {
     Lattice lat;
     Lattice nbest_lat;
@@ -533,15 +553,15 @@ const char *KaldiRecognizer::NbestResult(CompactLattice &clat)
     fst::ConvertNbestToVector(nbest_lat, &nbest_lats);
 
     json::JSON obj;
-    std::stringstream ss;
     for (int k = 0; k < nbest_lats.size(); k++) {
 
       Lattice nlat = nbest_lats[k];
-      RmEpsilon(&nlat);
-      CompactLattice nclat;
-      CompactLattice aligned_nclat;
-      ConvertLattice(nlat, &nclat);
 
+      CompactLattice nclat;
+      fst::Invert(&nlat);
+      DeterminizeLattice(nlat, &nclat);
+
+      CompactLattice aligned_nclat;
       if (model_->winfo_) {
           WordAlignLattice(nclat, *model_->trans_model_, *model_->winfo_, 0, &aligned_nclat);
       } else {
@@ -559,7 +579,7 @@ const char *KaldiRecognizer::NbestResult(CompactLattice &clat)
       stringstream text;
       json::JSON entry;
 
-      for (int i = 0; i < words.size(); i++) {
+      for (int i = 0, first = 1; i < words.size(); i++) {
         json::JSON word;
         if (words[i] == 0)
             continue;
@@ -569,8 +589,12 @@ const char *KaldiRecognizer::NbestResult(CompactLattice &clat)
             word["end"] = samples_round_start_ / sample_frequency_ + (frame_offset_ + begin_times[i] + lengths[i]) * 0.03;
             entry["result"].append(word);
         }
-        if (i)
+
+        if (first)
+          first = 0;
+        else
           text << " ";
+
         text << model_->word_syms_->Find(words[i]);
       }
 
@@ -582,56 +606,122 @@ const char *KaldiRecognizer::NbestResult(CompactLattice &clat)
     return StoreReturn(obj.dump());
 }
 
-const char* KaldiRecognizer::GetResult()
+const char *Recognizer::NlsmlResult(CompactLattice &clat)
+{
+    Lattice lat;
+    Lattice nbest_lat;
+    std::vector<Lattice> nbest_lats;
+
+    ConvertLattice (clat, &lat);
+    fst::ShortestPath(lat, &nbest_lat, max_alternatives_);
+    fst::ConvertNbestToVector(nbest_lat, &nbest_lats);
+
+    std::stringstream ss;
+    ss << "<?xml version=\"1.0\"?>\n";
+    ss << "<result grammar=\"default\">\n";
+
+    for (int k = 0; k < nbest_lats.size(); k++) {
+
+      Lattice nlat = nbest_lats[k];
+
+      CompactLattice nclat;
+      fst::Invert(&nlat);
+      DeterminizeLattice(nlat, &nclat);
+
+      CompactLattice aligned_nclat;
+      if (model_->winfo_) {
+          WordAlignLattice(nclat, *model_->trans_model_, *model_->winfo_, 0, &aligned_nclat);
+      } else {
+          aligned_nclat = nclat;
+      }
+
+      std::vector<int32> words;
+      std::vector<int32> begin_times;
+      std::vector<int32> lengths;
+      CompactLattice::Weight weight;
+
+      CompactLatticeToWordAlignmentWeight(aligned_nclat, &words, &begin_times, &lengths, &weight);
+      float likelihood = -(weight.Weight().Value1() + weight.Weight().Value2());
+
+      stringstream text;
+      for (int i = 0, first = 1; i < words.size(); i++) {
+        if (words[i] == 0)
+            continue;
+
+        if (first)
+          first = 0;
+        else
+          text << " ";
+
+        text << model_->word_syms_->Find(words[i]);
+      }
+
+      ss << "<interpretation grammar=\"default\" confidence=\"" << likelihood << "\">\n";
+      ss << "<input mode=\"speech\">" << text.str() << "</input>\n";
+      ss << "<instance>" << text.str() << "</instance>\n";
+      ss << "</interpretation>\n";
+    }
+    ss << "</result>\n";
+
+    return StoreReturn(ss.str());
+}
+
+const char* Recognizer::GetResult()
 {
     if (decoder_->NumFramesDecoded() == 0) {
         return StoreEmptyReturn();
     }
 
-    kaldi::CompactLattice clat;
-    kaldi::CompactLattice rlat;
-    decoder_->GetLattice(true, &clat);
+    // Original from decoder, subtracted graph weight, rescored with carpa, rescored with rnnlm
+    CompactLattice clat, slat, tlat, rlat;
 
-    if (model_->rnnlm_lm_fst_) {
-        kaldi::ComposeLatticePrunedOptions compose_opts;
-        compose_opts.lattice_compose_beam = 3.0;
-        compose_opts.max_arcs = 3000;
+    clat = decoder_->GetLattice(decoder_->NumFramesDecoded(), true);
 
-        TopSortCompactLatticeIfNeeded(&clat);
-        fst::ComposeDeterministicOnDemandFst<fst::StdArc> combined_lms(lm_to_subtract_det_scale, lm_to_add);
-        CompactLattice composed_clat;
-        ComposeCompactLatticePruned(compose_opts, clat,
-                                    &combined_lms, &rlat);
-        lm_to_add_orig->Clear();
-    } else if (model_->std_lm_fst_) {
-        Lattice lat1;
+    if (lm_to_subtract_ && carpa_to_add_) {
+        Lattice lat, composed_lat;
 
-        ConvertLattice(clat, &lat1);
-        fst::ScaleLattice(fst::GraphLatticeScale(-1.0), &lat1);
-        fst::ArcSort(&lat1, fst::OLabelCompare<kaldi::LatticeArc>());
-        kaldi::Lattice composed_lat;
-        fst::Compose(lat1, *lm_fst_, &composed_lat);
+        // Delete old score
+        ConvertLattice(clat, &lat);
+        fst::ScaleLattice(fst::GraphLatticeScale(-1.0), &lat);
+        fst::Compose(lat, *lm_to_subtract_, &composed_lat);
         fst::Invert(&composed_lat);
-        kaldi::CompactLattice determinized_lat;
-        DeterminizeLattice(composed_lat, &determinized_lat);
-        fst::ScaleLattice(fst::GraphLatticeScale(-1), &determinized_lat);
-        fst::ArcSort(&determinized_lat, fst::OLabelCompare<kaldi::CompactLatticeArc>());
+        DeterminizeLattice(composed_lat, &slat);
+        fst::ScaleLattice(fst::GraphLatticeScale(-1.0), &slat);
 
-        kaldi::ConstArpaLmDeterministicFst const_arpa_fst(model_->const_arpa_);
-        kaldi::CompactLattice composed_clat;
-        kaldi::ComposeCompactLatticeDeterministic(determinized_lat, &const_arpa_fst, &composed_clat);
-        kaldi::Lattice composed_lat1;
-        ConvertLattice(composed_clat, &composed_lat1);
-        fst::Invert(&composed_lat1);
-        DeterminizeLattice(composed_lat1, &rlat);
+        // Add CARPA score
+        TopSortCompactLatticeIfNeeded(&slat);
+        ComposeCompactLatticeDeterministic(slat, carpa_to_add_, &tlat);
+
+        // Rescore with RNNLM score on top if needed
+        if (rnnlm_to_add_scale_) {
+             ComposeLatticePrunedOptions compose_opts;
+             compose_opts.lattice_compose_beam = 3.0;
+             compose_opts.max_arcs = 3000;
+             fst::ComposeDeterministicOnDemandFst<StdArc> combined_rnnlm(carpa_to_add_scale_, rnnlm_to_add_scale_);
+
+             TopSortCompactLatticeIfNeeded(&tlat);
+             ComposeCompactLatticePruned(compose_opts, tlat,
+                                         &combined_rnnlm, &rlat);
+             rnnlm_to_add_->Clear();
+        } else {
+             rlat = tlat;
+        }
     } else {
         rlat = clat;
     }
 
-    fst::ScaleLattice(fst::GraphLatticeScale(0.9), &rlat); // Apply rescoring weight
+    // Pruned composition can return empty lattice. It should be rare
+    if (rlat.Start() != 0) {
+       return StoreEmptyReturn();
+    }
+
+    // Apply rescoring weight
+    fst::ScaleLattice(fst::GraphLatticeScale(0.9), &rlat);
 
     if (max_alternatives_ == 0) {
         return MbrResult(rlat);
+    } else if (nlsml_) {
+        return NlsmlResult(rlat);
     } else {
         return NbestResult(rlat);
     }
@@ -639,7 +729,7 @@ const char* KaldiRecognizer::GetResult()
 }
 
 
-const char* KaldiRecognizer::PartialResult()
+const char* Recognizer::PartialResult()
 {
     if (state_ != RECOGNIZER_RUNNING) {
         return StoreEmptyReturn();
@@ -647,30 +737,75 @@ const char* KaldiRecognizer::PartialResult()
 
     json::JSON res;
 
-    if (decoder_->NumFramesDecoded() == 0) {
-        res["partial"] = "";
-        return StoreReturn(res.dump());
-    }
+    if (partial_words_) {
 
-    kaldi::Lattice lat;
-    decoder_->GetBestPath(false, &lat);
-    vector<kaldi::int32> alignment, words;
-    LatticeWeight weight;
-    GetLinearSymbolSequence(lat, &alignment, &words, &weight);
-
-    ostringstream text;
-    for (size_t i = 0; i < words.size(); i++) {
-        if (i) {
-            text << " ";
+        if (decoder_->NumFramesInLattice() == 0) {
+            res["partial"] = "";
+            return StoreReturn(res.dump());
         }
-        text << model_->word_syms_->Find(words[i]);
+
+        CompactLattice clat;
+        CompactLattice aligned_lat;
+
+        clat = decoder_->GetLattice(decoder_->NumFramesInLattice(), false);
+        if (model_->winfo_) {
+            WordAlignLatticePartial(clat, *model_->trans_model_, *model_->winfo_, 0, &aligned_lat);
+        } else {
+            CopyLatticeForMbr(clat, &aligned_lat);
+        }
+
+        MinimumBayesRisk mbr(aligned_lat);
+        const vector<BaseFloat> &conf = mbr.GetOneBestConfidences();
+        const vector<int32> &words = mbr.GetOneBest();
+        const vector<pair<BaseFloat, BaseFloat> > &times = mbr.GetOneBestTimes();
+
+        int size = words.size();
+
+        stringstream text;
+
+        // Create JSON object
+        for (int i = 0; i < size; i++) {
+            json::JSON word;
+
+            word["word"] = model_->word_syms_->Find(words[i]);
+            word["start"] = samples_round_start_ / sample_frequency_ + (frame_offset_ + times[i].first) * 0.03;
+            word["end"] = samples_round_start_ / sample_frequency_ + (frame_offset_ + times[i].second) * 0.03;
+            word["conf"] = conf[i];
+            res["partial_result"].append(word);
+
+            if (i) {
+                text << " ";
+            }
+            text << model_->word_syms_->Find(words[i]);
+        }
+        res["partial"] = text.str();
+
+    } else {
+
+        if (decoder_->NumFramesDecoded() == 0) {
+            res["partial"] = "";
+            return StoreReturn(res.dump());
+        }
+        Lattice lat;
+        decoder_->GetBestPath(false, &lat);
+        vector<kaldi::int32> alignment, words;
+        LatticeWeight weight;
+        GetLinearSymbolSequence(lat, &alignment, &words, &weight);
+
+        ostringstream text;
+        for (size_t i = 0; i < words.size(); i++) {
+            if (i) {
+                text << " ";
+            }
+            text << model_->word_syms_->Find(words[i]);
+        }
+        res["partial"] = text.str();
     }
-    res["partial"] = text.str();
 
     return StoreReturn(res.dump());
 }
 
-const char* KaldiRecognizer::Result()
+const char* Recognizer::Result()
 {
     if (state_ != RECOGNIZER_RUNNING) {
         return StoreEmptyReturn();
@@ -680,7 +815,7 @@ const char* KaldiRecognizer::Result()
     return GetResult();
 }
 
-const char* KaldiRecognizer::FinalResult()
+const char* Recognizer::FinalResult()
 {
     if (state_ != RECOGNIZER_RUNNING) {
         return StoreEmptyReturn();
@@ -708,7 +843,7 @@ const char* KaldiRecognizer::FinalResult()
     return last_result_.c_str();
 }
 
-void KaldiRecognizer::Reset()
+void Recognizer::Reset()
 {
     if (state_ == RECOGNIZER_RUNNING) {
         decoder_->FinalizeDecoding();
@@ -717,17 +852,25 @@ void KaldiRecognizer::Reset()
     state_ = RECOGNIZER_ENDPOINT;
 }
 
-const char *KaldiRecognizer::StoreEmptyReturn()
+const char *Recognizer::StoreEmptyReturn()
 {
     if (!max_alternatives_) {
         return StoreReturn("{\"text\": \"\"}");
+    } else if (nlsml_) {
+        return StoreReturn("<?xml version=\"1.0\"?>\n"
+                           "<result grammar=\"default\">\n"
+                           "<interpretation confidence=\"1.0\">\n"
+                           "<instance/>\n"
+                           "<input><noinput/></input>\n"
+                           "</interpretation>\n"
+                           "</result>\n");
     } else {
         return StoreReturn("{\"alternatives\" : [{\"text\": \"\", \"confidence\" : 1.0}] }");
     }
 }
 
 // Store result in recognizer and return as const string
-const char *KaldiRecognizer::StoreReturn(const string &res)
+const char *Recognizer::StoreReturn(const string &res)
 {
     last_result_ = res;
     return last_result_.c_str();

src/recognizer.h

@@ -33,22 +33,24 @@
 
 using namespace kaldi;
 
-enum KaldiRecognizerState {
+enum RecognizerState {
     RECOGNIZER_INITIALIZED,
     RECOGNIZER_RUNNING,
     RECOGNIZER_ENDPOINT,
     RECOGNIZER_FINALIZED
 };
 
-class KaldiRecognizer {
+class Recognizer {
     public:
-        KaldiRecognizer(Model *model, float sample_frequency);
-        KaldiRecognizer(Model *model, float sample_frequency, SpkModel *spk_model);
-        KaldiRecognizer(Model *model, float sample_frequency, char const *grammar);
-        ~KaldiRecognizer();
+        Recognizer(Model *model, float sample_frequency);
+        Recognizer(Model *model, float sample_frequency, SpkModel *spk_model);
+        Recognizer(Model *model, float sample_frequency, char const *grammar);
+        ~Recognizer();
         void SetMaxAlternatives(int max_alternatives);
         void SetSpkModel(SpkModel *spk_model);
         void SetWords(bool words);
+        void SetPartialWords(bool partial_words);
+        void SetNLSML(bool nlsml);
         bool AcceptWaveform(const char *data, int len);
         bool AcceptWaveform(const short *sdata, int len);
         bool AcceptWaveform(const float *fdata, int len);
@@ -69,9 +71,10 @@ class KaldiRecognizer {
         const char *StoreReturn(const string &res);
         const char *MbrResult(CompactLattice &clat);
         const char *NbestResult(CompactLattice &clat);
+        const char *NlsmlResult(CompactLattice &clat);
 
         Model *model_ = nullptr;
-        SingleUtteranceNnet3Decoder *decoder_ = nullptr;
+        SingleUtteranceNnet3IncrementalDecoder *decoder_ = nullptr;
         fst::LookaheadFst<fst::StdArc, int32> *decode_fst_ = nullptr;
         fst::StdVectorFst *g_fst_ = nullptr; // dynamically constructed grammar
         OnlineNnet2FeaturePipeline *feature_pipeline_ = nullptr;
@@ -82,17 +85,20 @@ class KaldiRecognizer {
         OnlineBaseFeature *spk_feature_ = nullptr;
 
         // Rescoring
-        fst::ArcMapFst<fst::StdArc, kaldi::LatticeArc, fst::StdToLatticeMapper<kaldi::BaseFloat> > *lm_fst_ = nullptr;
-
+        fst::ArcMapFst<fst::StdArc, LatticeArc, fst::StdToLatticeMapper<BaseFloat> > *lm_to_subtract_ = nullptr;
+        kaldi::ConstArpaLmDeterministicFst *carpa_to_add_ = nullptr;
+        fst::ScaleDeterministicOnDemandFst *carpa_to_add_scale_ = nullptr;
         // RNNLM rescoring
-        kaldi::rnnlm::RnnlmComputeStateInfo *info = nullptr;
-        fst::ScaleDeterministicOnDemandFst *lm_to_subtract_det_scale = nullptr;
-        fst::BackoffDeterministicOnDemandFst<fst::StdArc> *lm_to_subtract_det_backoff = nullptr;
-        kaldi::rnnlm::KaldiRnnlmDeterministicFst* lm_to_add_orig = nullptr;
-        fst::DeterministicOnDemandFst<fst::StdArc> *lm_to_add = nullptr;
+        kaldi::rnnlm::KaldiRnnlmDeterministicFst* rnnlm_to_add_ = nullptr;
+        fst::DeterministicOnDemandFst<fst::StdArc> *rnnlm_to_add_scale_ = nullptr;
+        kaldi::rnnlm::RnnlmComputeStateInfo *rnnlm_info_ = nullptr;
 
+
+        // Other
         int max_alternatives_ = 0; // Disable alternatives by default
         bool words_ = false;
+        bool partial_words_ = false;
+        bool nlsml_ = false;
 
         float sample_frequency_;
         int32 frame_offset_;
@@ -100,7 +106,7 @@ class KaldiRecognizer {
         int64 samples_processed_;
         int64 samples_round_start_;
 
-        KaldiRecognizerState state_;
+        RecognizerState state_;
         string last_result_;
 };
 

src/spk_model.h

@@ -22,7 +22,7 @@
 
 using namespace kaldi;
 
-class KaldiRecognizer;
+class Recognizer;
 
 class SpkModel {
 
@@ -32,7 +32,7 @@ public:
     void Unref();
 
 protected:
-    friend class KaldiRecognizer;
+    friend class Recognizer;
     ~SpkModel() {};
 
     kaldi::nnet3::Nnet speaker_nnet;

src/vosk_api.cc

@@ -13,12 +13,14 @@
 // limitations under the License.
 
 #include "vosk_api.h"
-#include "kaldi_recognizer.h"
+
+#include "recognizer.h"
 #include "model.h"
 #include "spk_model.h"
 
 #if HAVE_CUDA
 #include "cudamatrix/cu-device.h"
+#include "batch_recognizer.h"
 #endif
 
 #include <string.h>
@@ -27,11 +29,18 @@ using namespace kaldi;
 
 VoskModel *vosk_model_new(const char *model_path)
 {
-    return (VoskModel *)new Model(model_path);
+    try {
+        return (VoskModel *)new Model(model_path);
+    } catch (...) {
+        return nullptr;
+    }
 }
 
 void vosk_model_free(VoskModel *model)
 {
+    if (model == nullptr) {
+       return;
+    }
     ((Model *)model)->Unref();
 }
 
@@ -42,82 +51,126 @@ int vosk_model_find_word(VoskModel *model, const char *word)
 
 VoskSpkModel *vosk_spk_model_new(const char *model_path)
 {
-    return (VoskSpkModel *)new SpkModel(model_path);
+    try {
+        return (VoskSpkModel *)new SpkModel(model_path);
+    } catch (...) {
+        return nullptr;
+    }
 }
 
 void vosk_spk_model_free(VoskSpkModel *model)
 {
+    if (model == nullptr) {
+       return;
+    }
     ((SpkModel *)model)->Unref();
 }
 
 VoskRecognizer *vosk_recognizer_new(VoskModel *model, float sample_rate)
 {
-    return (VoskRecognizer *)new KaldiRecognizer((Model *)model, sample_rate);
+    try {
+        return (VoskRecognizer *)new Recognizer((Model *)model, sample_rate);
+    } catch (...) {
+        return nullptr;
+    }
 }
 
 VoskRecognizer *vosk_recognizer_new_spk(VoskModel *model, float sample_rate, VoskSpkModel *spk_model)
 {
-    return (VoskRecognizer *)new KaldiRecognizer((Model *)model, sample_rate, (SpkModel *)spk_model);
+    try {
+        return (VoskRecognizer *)new Recognizer((Model *)model, sample_rate, (SpkModel *)spk_model);
+    } catch (...) {
+        return nullptr;
+    }
 }
 
 VoskRecognizer *vosk_recognizer_new_grm(VoskModel *model, float sample_rate, const char *grammar)
 {
-    return (VoskRecognizer *)new KaldiRecognizer((Model *)model, sample_rate, grammar);
+    try {
+        return (VoskRecognizer *)new Recognizer((Model *)model, sample_rate, grammar);
+    } catch (...) {
+        return nullptr;
+    }
 }
 
 void vosk_recognizer_set_max_alternatives(VoskRecognizer *recognizer, int max_alternatives)
 {
-    ((KaldiRecognizer *)recognizer)->SetMaxAlternatives(max_alternatives);
+    ((Recognizer *)recognizer)->SetMaxAlternatives(max_alternatives);
 }
 
 void vosk_recognizer_set_words(VoskRecognizer *recognizer, int words)
 {
-    ((KaldiRecognizer *)recognizer)->SetWords((bool)words);
+    ((Recognizer *)recognizer)->SetWords((bool)words);
+}
+
+void vosk_recognizer_set_partial_words(VoskRecognizer *recognizer, int partial_words)
+{
+    ((Recognizer *)recognizer)->SetPartialWords((bool)partial_words);
+}
+
+void vosk_recognizer_set_nlsml(VoskRecognizer *recognizer, int nlsml)
+{
+    ((Recognizer *)recognizer)->SetNLSML((bool)nlsml);
 }
 
 void vosk_recognizer_set_spk_model(VoskRecognizer *recognizer, VoskSpkModel *spk_model)
 {
-    ((KaldiRecognizer *)recognizer)->SetSpkModel((SpkModel *)spk_model);
+    if (recognizer == nullptr || spk_model == nullptr) {
+       return;
+    }
+    ((Recognizer *)recognizer)->SetSpkModel((SpkModel *)spk_model);
 }
 
 int vosk_recognizer_accept_waveform(VoskRecognizer *recognizer, const char *data, int length)
 {
-    return ((KaldiRecognizer *)(recognizer))->AcceptWaveform(data, length);
+    try {
+        return ((Recognizer *)(recognizer))->AcceptWaveform(data, length);
+    } catch (...) {
+        return -1;
+    }
 }
 
 int vosk_recognizer_accept_waveform_s(VoskRecognizer *recognizer, const short *data, int length)
 {
-    return ((KaldiRecognizer *)(recognizer))->AcceptWaveform(data, length);
+    try {
+        return ((Recognizer *)(recognizer))->AcceptWaveform(data, length);
+    } catch (...) {
+        return -1;
+    }
 }
 
 int vosk_recognizer_accept_waveform_f(VoskRecognizer *recognizer, const float *data, int length)
 {
-    return ((KaldiRecognizer *)(recognizer))->AcceptWaveform(data, length);
+    try {
+        return ((Recognizer *)(recognizer))->AcceptWaveform(data, length);
+    } catch (...) {
+        return -1;
+    }
 }
 
 const char *vosk_recognizer_result(VoskRecognizer *recognizer)
 {
-    return ((KaldiRecognizer *)recognizer)->Result();
+    return ((Recognizer *)recognizer)->Result();
 }
 
 const char *vosk_recognizer_partial_result(VoskRecognizer *recognizer)
 {
-    return ((KaldiRecognizer *)recognizer)->PartialResult();
+    return ((Recognizer *)recognizer)->PartialResult();
 }
 
 const char *vosk_recognizer_final_result(VoskRecognizer *recognizer)
 {
-    return ((KaldiRecognizer *)recognizer)->FinalResult();
+    return ((Recognizer *)recognizer)->FinalResult();
 }
 
 void vosk_recognizer_reset(VoskRecognizer *recognizer)
 {
-    ((KaldiRecognizer *)recognizer)->Reset();
+    ((Recognizer *)recognizer)->Reset();
 }
 
 void vosk_recognizer_free(VoskRecognizer *recognizer)
 {
-    delete (KaldiRecognizer *)(recognizer);
+    delete (Recognizer *)(recognizer);
 }
 
 void vosk_set_log_level(int log_level)
@@ -128,6 +181,8 @@ void vosk_set_log_level(int log_level)
 void vosk_gpu_init()
 {
 #if HAVE_CUDA
+//    kaldi::CuDevice::EnableTensorCores(true);
+//    kaldi::CuDevice::EnableTf32Compute(true);
     kaldi::CuDevice::Instantiate().SelectGpuId("yes");
     kaldi::CuDevice::Instantiate().AllowMultithreading();
 #endif
@@ -139,3 +194,89 @@ void vosk_gpu_thread_init()
     kaldi::CuDevice::Instantiate();
 #endif
 }
+
+VoskBatchModel *vosk_batch_model_new()
+{
+#if HAVE_CUDA
+    return (VoskBatchModel *)(new BatchModel());
+#else
+    return NULL;
+#endif
+}
+
+void vosk_batch_model_free(VoskBatchModel *model)
+{
+#if HAVE_CUDA
+    delete ((BatchModel *)model);
+#endif
+}
+
+void vosk_batch_model_wait(VoskBatchModel *model)
+{
+#if HAVE_CUDA
+    ((BatchModel *)model)->WaitForCompletion();
+#endif
+}
+
+VoskBatchRecognizer *vosk_batch_recognizer_new(VoskBatchModel *model, float sample_rate)
+{
+#if HAVE_CUDA
+    return (VoskBatchRecognizer *)(new BatchRecognizer((BatchModel *)model, sample_rate));
+#else
+    return NULL;
+#endif
+}
+
+void vosk_batch_recognizer_free(VoskBatchRecognizer *recognizer)
+{
+#if HAVE_CUDA
+    delete ((BatchRecognizer *)recognizer);
+#endif
+}
+
+void vosk_batch_recognizer_accept_waveform(VoskBatchRecognizer *recognizer, const char *data, int length)
+{
+#if HAVE_CUDA
+    ((BatchRecognizer *)recognizer)->AcceptWaveform(data, length);
+#endif
+}
+
+void vosk_batch_recognizer_set_nlsml(VoskBatchRecognizer *recognizer, int nlsml)
+{
+#if HAVE_CUDA
+    ((BatchRecognizer *)recognizer)->SetNLSML((bool)nlsml);
+#endif
+}
+
+void vosk_batch_recognizer_finish_stream(VoskBatchRecognizer *recognizer)
+{
+#if HAVE_CUDA
+    ((BatchRecognizer *)recognizer)->FinishStream();
+#endif
+}
+
+const char *vosk_batch_recognizer_front_result(VoskBatchRecognizer *recognizer)
+{
+#if HAVE_CUDA
+    return ((BatchRecognizer *)recognizer)->FrontResult();
+#else
+    return NULL;
+#endif
+}
+
+void vosk_batch_recognizer_pop(VoskBatchRecognizer *recognizer)
+{
+#if HAVE_CUDA
+    ((BatchRecognizer *)recognizer)->Pop();
+#endif
+}
+
+
+int vosk_batch_recognizer_get_pending_chunks(VoskBatchRecognizer *recognizer)
+{
+#if HAVE_CUDA
+    return ((BatchRecognizer *)recognizer)->GetNumPendingChunks();
+#else
+    return 0;
+#endif
+}

src/vosk_api.h

@@ -40,10 +40,21 @@ typedef struct VoskSpkModel VoskSpkModel;
 typedef struct VoskRecognizer VoskRecognizer;
 
 
+/**
+ * Batch model object
+ */
+typedef struct VoskBatchModel VoskBatchModel;
+
+/**
+ * Batch recognizer object
+ */
+typedef struct VoskBatchRecognizer VoskBatchRecognizer;
+
+
 /** Loads model data from the file and returns the model object
  *
  * @param model_path: the path of the model on the filesystem
- @ @returns model object */
+ * @returns model object or NULL if problem occured */
 VoskModel *vosk_model_new(const char *model_path);
 
 
@@ -66,7 +77,7 @@ int vosk_model_find_word(VoskModel *model, const char *word);
 /** Loads speaker model data from the file and returns the model object
  *
  * @param model_path: the path of the model on the filesystem
- * @returns model object */
+ * @returns model object or NULL if problem occured */
 VoskSpkModel *vosk_spk_model_new(const char *model_path);
 
 
@@ -79,9 +90,13 @@ void vosk_spk_model_free(VoskSpkModel *model);
 
 /** Creates the recognizer object
  *
- *  The recognizers process the speech and return text using shared model data 
- *  @param sample_rate The sample rate of the audio you going to feed into the recognizer
- *  @returns recognizer object */
+ *  The recognizers process the speech and return text using shared model data
+ *  @param model       VoskModel containing static data for recognizer. Model can be
+ *                     shared across recognizers, even running in different threads.
+ *  @param sample_rate The sample rate of the audio you going to feed into the recognizer.
+ *                     Make sure this rate matches the audio content, it is a common
+ *                     issue causing accuracy problems.
+ *  @returns recognizer object or NULL if problem occured */
 VoskRecognizer *vosk_recognizer_new(VoskModel *model, float sample_rate);
 
 
@@ -90,9 +105,13 @@ VoskRecognizer *vosk_recognizer_new(VoskModel *model, float sample_rate);
  *  With the speaker recognition mode the recognizer not just recognize
  *  text but also return speaker vectors one can use for speaker identification
  *
- *  @param sample_rate The sample rate of the audio you going to feed into the recognizer
+ *  @param model       VoskModel containing static data for recognizer. Model can be
+ *                     shared across recognizers, even running in different threads.
+ *  @param sample_rate The sample rate of the audio you going to feed into the recognizer.
+ *                     Make sure this rate matches the audio content, it is a common
+ *                     issue causing accuracy problems.
  *  @param spk_model speaker model for speaker identification
- *  @returns recognizer object */
+ *  @returns recognizer object or NULL if problem occured */
 VoskRecognizer *vosk_recognizer_new_spk(VoskModel *model, float sample_rate, VoskSpkModel *spk_model);
 
 
@@ -106,11 +125,15 @@ VoskRecognizer *vosk_recognizer_new_spk(VoskModel *model, float sample_rate, Vos
  *  Only recognizers with lookahead models support this type of quick configuration.
  *  Precompiled HCLG graph models are not supported.
  *
- *  @param sample_rate The sample rate of the audio you going to feed into the recognizer
+ *  @param model       VoskModel containing static data for recognizer. Model can be
+ *                     shared across recognizers, even running in different threads.
+ *  @param sample_rate The sample rate of the audio you going to feed into the recognizer.
+ *                     Make sure this rate matches the audio content, it is a common
+ *                     issue causing accuracy problems.
  *  @param grammar The string with the list of phrases to recognize as JSON array of strings,
  *                 for example "["one two three four five", "[unk]"]".
  *
- *  @returns recognizer object */
+ *  @returns recognizer object or NULL if problem occured */
 VoskRecognizer *vosk_recognizer_new_grm(VoskModel *model, float sample_rate, const char *grammar);
 
 
@@ -174,6 +197,17 @@ void vosk_recognizer_set_max_alternatives(VoskRecognizer *recognizer, int max_al
  */
 void vosk_recognizer_set_words(VoskRecognizer *recognizer, int words);
 
+/** Like above return words and confidences in partial results
+ *
+ * @param partial_words - boolean value
+ */
+void vosk_recognizer_set_partial_words(VoskRecognizer *recognizer, int partial_words);
+
+/** Set NLSML output
+ * @param nlsml - boolean value
+ */
+void vosk_recognizer_set_nlsml(VoskRecognizer *recognizer, int nlsml);
+
 
 /** Accept voice data
  *
@@ -181,7 +215,9 @@ void vosk_recognizer_set_words(VoskRecognizer *recognizer, int words);
  *
  *  @param data - audio data in PCM 16-bit mono format
  *  @param length - length of the audio data
- *  @returns true if silence is occured and you can retrieve a new utterance with result method */
+ *  @returns 1 if silence is occured and you can retrieve a new utterance with result method 
+ *           0 if decoding continues
+ *           -1 if exception occured */
 int vosk_recognizer_accept_waveform(VoskRecognizer *recognizer, const char *data, int length);
 
 
@@ -271,6 +307,44 @@ void vosk_gpu_init();
  */
 void vosk_gpu_thread_init();
 
+/** Creates the batch recognizer object
+ *
+ *  @returns model object or NULL if problem occured */
+VoskBatchModel *vosk_batch_model_new();
+
+/** Releases batch model object */
+void vosk_batch_model_free(VoskBatchModel *model);
+
+/** Wait for the processing */
+void vosk_batch_model_wait(VoskBatchModel *model);
+
+/** Creates batch recognizer object
+ *  @returns recognizer object or NULL if problem occured */
+VoskBatchRecognizer *vosk_batch_recognizer_new(VoskBatchModel *model, float sample_rate);
+ 
+/** Releases batch recognizer object */
+void vosk_batch_recognizer_free(VoskBatchRecognizer *recognizer);
+
+/** Accept batch voice data */
+void vosk_batch_recognizer_accept_waveform(VoskBatchRecognizer *recognizer, const char *data, int length);
+
+/** Set NLSML output
+ * @param nlsml - boolean value
+ */
+void vosk_batch_recognizer_set_nlsml(VoskBatchRecognizer *recognizer, int nlsml);
+
+/** Closes the stream */
+void vosk_batch_recognizer_finish_stream(VoskBatchRecognizer *recognizer);
+
+/** Return results */
+const char *vosk_batch_recognizer_front_result(VoskBatchRecognizer *recognizer);
+
+/** Release and free first retrieved result */
+void vosk_batch_recognizer_pop(VoskBatchRecognizer *recognizer);
+
+/** Get amount of pending chunks for more intelligent waiting */
+int vosk_batch_recognizer_get_pending_chunks(VoskBatchRecognizer *recognizer);
+
 #ifdef __cplusplus
 }
 #endif

training/README.md

@@ -0,0 +1,3 @@
+A proper simple setup to train a Vosk model
+
+More documentation later

training/RESULTS

@@ -0,0 +1,2 @@
+#!/bin/bash
+for x in exp/*/decode* exp/chain*/*/decode*; do [ -d $x ] && [[ $x =~ "$1" ]] && grep WER $x/wer_* | utils/best_wer.sh; done

training/RESULTS.txt

@@ -0,0 +1,2 @@
+%WER 14.10 [ 2839 / 20138, 214 ins, 487 del, 2138 sub ] exp/chain/tdnn/decode_test/wer_11_0.0
+%WER 12.67 [ 2552 / 20138, 215 ins, 406 del, 1931 sub ] exp/chain/tdnn/decode_test_rescore/wer_11_0.0

training/cmd.sh

@@ -0,0 +1,5 @@
+export train_cmd="run.pl"
+export decode_cmd="run.pl"
+export mkgraph_cmd="run.pl"
+export cuda_cmd="run.pl"
+export get_egs_cmd="run.pl"

training/conf/mfcc.conf

@@ -0,0 +1,7 @@
+--use-energy=false
+--num-mel-bins=40
+--num-ceps=40
+--low-freq=20
+--high-freq=-400
+--allow-upsample=true
+--allow-downsample=true

training/local/chain/run_ivector_common.sh

@@ -0,0 +1,86 @@
+#!/bin/bash
+
+set -euo pipefail
+
+# This script is called from local/chain_${suffix}/run_tdnn.sh and
+# local/chain_${suffix}/run_tdnn.sh (and may eventually be called by more
+# scripts).  It contains the common feature preparation and
+# iVector-related parts of the script.  See those scripts for examples
+# of usage.
+
+stage=0
+train_set=train
+gmm=tri3
+suffix=""
+
+. ./cmd.sh
+. ./path.sh
+. utils/parse_options.sh
+
+gmm_dir=exp/${gmm}
+ali_dir=exp/${gmm}_ali
+
+if [ $stage -le 4 ]; then
+  echo "$0: computing a subset of data to train the diagonal UBM."
+  # We'll use about a quarter of the data.
+  mkdir -p exp/chain${suffix}/diag_ubm
+  temp_data_root=exp/chain${suffix}/diag_ubm
+
+  num_utts_total=$(wc -l <data/${train_set}/utt2spk)
+  num_utts=$[$num_utts_total/4]
+  utils/data/subset_data_dir.sh data/${train_set} \
+     $num_utts ${temp_data_root}/${train_set}_subset
+
+  echo "$0: computing a PCA transform from the data."
+  steps/online/nnet2/get_pca_transform.sh --cmd "$train_cmd" \
+      --splice-opts "--left-context=3 --right-context=3" \
+      --max-utts 10000 --subsample 2 \
+       ${temp_data_root}/${train_set}_subset \
+       exp/chain${suffix}/pca_transform
+
+  echo "$0: training the diagonal UBM."
+  # Use 512 Gaussians in the UBM.
+  steps/online/nnet2/train_diag_ubm.sh --cmd "$train_cmd" --nj 10 \
+    --num-frames 700000 \
+    --num-threads 8 \
+    ${temp_data_root}/${train_set}_subset 512 \
+    exp/chain${suffix}/pca_transform exp/chain${suffix}/diag_ubm
+fi
+
+if [ $stage -le 5 ]; then
+  # Train the iVector extractor.  Use all of the speed-perturbed data since iVector extractors
+  # can be sensitive to the amount of data.  The script defaults to an iVector dimension of
+  # 100.
+  echo "$0: training the iVector extractor"
+  steps/online/nnet2/train_ivector_extractor.sh --cmd "$train_cmd" --nj 2 \
+     --ivector-dim 40 \
+     data/${train_set} exp/chain${suffix}/diag_ubm \
+     exp/chain${suffix}/extractor || exit 1;
+fi
+
+
+if [ $stage -le 6 ]; then
+  # We extract iVectors on the speed-perturbed training data after combining
+  # short segments, which will be what we train the system on.  With
+  # --utts-per-spk-max 2, the script pairs the utterances into twos, and treats
+  # each of these pairs as one speaker; this gives more diversity in iVectors..
+  # Note that these are extracted 'online'.
+
+  # note, we don't encode the 'max2' in the name of the ivectordir even though
+  # that's the data we extract the ivectors from, as it's still going to be
+  # valid for the non-'max2' data, the utterance list is the same.
+
+  ivectordir=exp/chain${suffix}/ivectors_${train_set}
+
+  # having a larger number of speakers is helpful for generalization, and to
+  # handle per-utterance decoding well (iVector starts at zero).
+  temp_data_root=${ivectordir}
+  utils/data/modify_speaker_info.sh --utts-per-spk-max 2 \
+    data/${train_set} ${temp_data_root}/${train_set}_max2
+
+  steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj 10 \
+    ${temp_data_root}/${train_set}_max2 \
+    exp/chain${suffix}/extractor $ivectordir
+fi
+
+exit 0

training/local/chain/run_tdnn.sh

@@ -0,0 +1,165 @@
+#!/bin/bash
+
+# Set -e here so that we catch if any executable fails immediately
+set -euo pipefail
+
+# (some of which are also used in this script directly).
+stage=-1
+decode_nj=10
+train_set=train
+gmm=tri3
+nnet3_affix=
+suffix=
+
+# The rest are configs specific to this script.  Most of the parameters
+# are just hardcoded at this level, in the commands below.
+affix=   # affix for the TDNN directory name
+tree_affix=
+train_stage=-10
+get_egs_stage=-10
+decode_iter=
+
+# training options
+# training chunk-options
+chunk_width=140,100,160
+common_egs_dir=
+xent_regularize=0.1
+dropout_schedule='0,0@0.20,0.5@0.50,0'
+
+# training options
+srand=0
+remove_egs=true
+
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+# Problem: We have removed the "train_" prefix of our training set in
+# the alignment directory names! Bad!
+gmm_dir=exp/$gmm
+ali_dir=exp/${gmm}_ali
+tree_dir=exp/chain${suffix}/tree${tree_affix:+_$tree_affix}
+lang=data/lang_chain${suffix}
+lat_dir=exp/chain${suffix}/${gmm}_${train_set}_lats
+dir=exp/chain${suffix}/tdnn${affix}
+train_data_dir=data/${train_set}
+
+for f in $gmm_dir/final.mdl $train_data_dir/feats.scp $ali_dir/ali.1.gz; do
+  [ ! -f $f ] && echo "$0: expected file $f to exist" && exit 1
+done
+
+if [ $stage -le 9 ]; then
+    local/chain/run_ivector_common.sh \
+                          --train-set ${train_set} \
+                          --gmm ${gmm} \
+                          --suffix "${suffix}"
+fi
+
+if [ $stage -le 10 ]; then
+  echo "$0: creating lang directory $lang with chain-type topology"
+  rm -rf $lang
+  cp -r data/lang $lang
+  silphonelist=$(cat $lang/phones/silence.csl) 
+  nonsilphonelist=$(cat $lang/phones/nonsilence.csl) 
+  steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
+fi
+
+if [ $stage -le 11 ]; then
+    steps/align_fmllr_lats.sh --nj 20 --cmd "$train_cmd" ${train_data_dir} \
+         data/lang $gmm_dir $lat_dir
+fi
+
+if [ $stage -le 12 ]; then
+  steps/nnet3/chain/build_tree.sh \
+    --frame-subsampling-factor 3 \
+    --context-opts "--context-width=2 --central-position=1" \
+    --cmd "$train_cmd" 2500 ${train_data_dir} \
+    $lang $ali_dir $tree_dir
+fi
+
+if [ $stage -le 13 ]; then
+  echo "$0: creating neural net configs using the xconfig parser";
+
+  num_targets=$(tree-info $tree_dir/tree | grep num-pdfs | awk '{print $2}')
+  learning_rate_factor=$(echo "print (0.5/$xent_regularize)" | python)
+
+  affine_opts="l2-regularize=0.008 dropout-proportion=0.0 dropout-per-dim=true dropout-per-dim-continuous=true"
+  tdnnf_opts="l2-regularize=0.008 dropout-proportion=0.0 bypass-scale=0.75"
+  linear_opts="l2-regularize=0.008 orthonormal-constraint=-1.0"
+  prefinal_opts="l2-regularize=0.008"
+  output_opts="l2-regularize=0.002"
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+
+  input dim=40 name=ivector
+  input dim=40 name=input
+
+  idct-layer name=idct input=input dim=40 cepstral-lifter=22 affine-transform-file=$dir/configs/idct.mat
+  batchnorm-component name=batchnorm0 input=idct
+  spec-augment-layer name=spec-augment freq-max-proportion=0.5 time-zeroed-proportion=0.2 time-mask-max-frames=20
+  delta-layer name=delta input=spec-augment
+
+  no-op-component name=input2 input=Append(delta, ReplaceIndex(ivector, t, 0))
+
+  # the first splicing is moved before the lda layer, so no splicing here
+  relu-batchnorm-dropout-layer name=tdnn1 $affine_opts dim=512 input=input2
+  tdnnf-layer name=tdnnf2 $tdnnf_opts dim=512 bottleneck-dim=96 time-stride=1
+  tdnnf-layer name=tdnnf3 $tdnnf_opts dim=512 bottleneck-dim=96 time-stride=1
+  tdnnf-layer name=tdnnf4 $tdnnf_opts dim=512 bottleneck-dim=96 time-stride=1
+  tdnnf-layer name=tdnnf5 $tdnnf_opts dim=512 bottleneck-dim=96 time-stride=0
+  tdnnf-layer name=tdnnf6 $tdnnf_opts dim=512 bottleneck-dim=96 time-stride=3
+  tdnnf-layer name=tdnnf7 $tdnnf_opts dim=512 bottleneck-dim=96 time-stride=3
+  tdnnf-layer name=tdnnf8 $tdnnf_opts dim=512 bottleneck-dim=96 time-stride=3
+  tdnnf-layer name=tdnnf9 $tdnnf_opts dim=512 bottleneck-dim=96 time-stride=3
+  tdnnf-layer name=tdnnf10 $tdnnf_opts dim=512 bottleneck-dim=96 time-stride=3
+  tdnnf-layer name=tdnnf11 $tdnnf_opts dim=512 bottleneck-dim=96 time-stride=3
+  tdnnf-layer name=tdnnf12 $tdnnf_opts dim=512 bottleneck-dim=96 time-stride=3
+  linear-component name=prefinal-l dim=192 $linear_opts
+
+  ## adding the layers for chain branch
+  prefinal-layer name=prefinal-chain input=prefinal-l $prefinal_opts small-dim=192 big-dim=512
+  output-layer name=output include-log-softmax=false dim=$num_targets $output_opts
+
+  # adding the layers for xent branch
+  prefinal-layer name=prefinal-xent input=prefinal-l $prefinal_opts small-dim=192 big-dim=512
+  output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor $output_opts
+
+EOF
+  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
+fi
+
+if [ $stage -le 14 ]; then
+  steps/nnet3/chain/train.py --stage $train_stage \
+    --cmd "$cuda_cmd" \
+    --feat.online-ivector-dir exp/chain${suffix}/ivectors_${train_set} \
+    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
+    --chain.xent-regularize $xent_regularize \
+    --chain.leaky-hmm-coefficient 0.1 \
+    --chain.l2-regularize 0.0 \
+    --chain.apply-deriv-weights false \
+    --chain.lm-opts="--num-extra-lm-states=2000" \
+    --egs.cmd "$get_egs_cmd" \
+    --egs.dir "$common_egs_dir" \
+    --egs.stage $get_egs_stage \
+    --egs.opts "--frames-overlap-per-eg 0 --constrained false" \
+    --egs.chunk-width $chunk_width \
+    --trainer.dropout-schedule $dropout_schedule \
+    --trainer.add-option="--optimization.memory-compression-level=2" \
+    --trainer.num-chunk-per-minibatch 64 \
+    --trainer.frames-per-iter 2500000 \
+    --trainer.num-epochs 20 \
+    --trainer.optimization.num-jobs-initial 1 \
+    --trainer.optimization.num-jobs-final 1 \
+    --trainer.optimization.initial-effective-lrate 0.001 \
+    --trainer.optimization.final-effective-lrate 0.0001 \
+    --trainer.max-param-change 2.0 \
+    --cleanup.remove-egs $remove_egs \
+    --feat-dir $train_data_dir \
+    --tree-dir $tree_dir \
+    --lat-dir $lat_dir \
+    --dir $dir  || exit 1;
+fi

training/local/data_prep.sh

@@ -0,0 +1 @@
+../../../librispeech/s5/local/data_prep.sh

training/local/download_and_untar.sh

@@ -0,0 +1,100 @@
+#!/usr/bin/env bash
+
+# Copyright   2014  Johns Hopkins University (author: Daniel Povey)
+#             2017  Luminar Technologies, Inc. (author: Daniel Galvez)
+# Apache 2.0
+
+remove_archive=false
+
+if [ "$1" == --remove-archive ]; then
+  remove_archive=true
+  shift
+fi
+
+if [ $# -ne 3 ]; then
+  echo "Usage: $0 [--remove-archive] <data-base> <url-base> <corpus-part>"
+  echo "e.g.: $0 /export/a05/dgalvez/ www.openslr.org/resources/31 dev-clean-2"
+  echo "With --remove-archive it will remove the archive after successfully un-tarring it."
+  echo "<corpus-part> can be one of: dev-clean-2, test-clean-5, dev-other, test-other,"
+  echo "          train-clean-100, train-clean-360, train-other-500."
+fi
+
+data=$1
+url=$2
+part=$3
+
+if [ ! -d "$data" ]; then
+  echo "$0: no such directory $data"
+  exit 1;
+fi
+
+data=$(readlink -f $data)
+
+part_ok=false
+list="dev-clean-2 train-clean-5"
+for x in $list; do
+  if [ "$part" == $x ]; then part_ok=true; fi
+done
+if ! $part_ok; then
+  echo "$0: expected <corpus-part> to be one of $list, but got '$part'"
+  exit 1;
+fi
+
+if [ -z "$url" ]; then
+  echo "$0: empty URL base."
+  exit 1;
+fi
+
+if [ -f $data/LibriSpeech/$part/.complete ]; then
+  echo "$0: data part $part was already successfully extracted, nothing to do."
+  exit 0;
+fi
+
+
+#sizes="126046265 332747356"
+sizes="126046265 332954390"
+
+if [ -f $data/$part.tar.gz ]; then
+  size=$(/bin/ls -l $data/$part.tar.gz | awk '{print $5}')
+  size_ok=false
+  for s in $sizes; do if [ $s == $size ]; then size_ok=true; fi; done
+  if ! $size_ok; then
+    echo "$0: removing existing file $data/$part.tar.gz because its size in bytes $size"
+    echo "does not equal the size of one of the archives."
+    rm $data/$part.tar.gz
+  else
+    echo "$data/$part.tar.gz exists and appears to be complete."
+  fi
+fi
+
+if [ ! -f $data/$part.tar.gz ]; then
+  if ! which wget >/dev/null; then
+    echo "$0: wget is not installed."
+    exit 1;
+  fi
+  full_url=$url/$part.tar.gz
+  echo "$0: downloading data from $full_url.  This may take some time, please be patient."
+
+  cd $data
+  if ! wget --no-check-certificate $full_url; then
+    echo "$0: error executing wget $full_url"
+    exit 1;
+  fi
+  cd -
+fi
+
+cd $data
+
+if ! tar -xvzf $part.tar.gz; then
+  echo "$0: error un-tarring archive $data/$part.tar.gz"
+  exit 1;
+fi
+
+touch $data/LibriSpeech/$part/.complete
+
+echo "$0: Successfully downloaded and un-tarred $data/$part.tar.gz"
+
+if $remove_archive; then
+  echo "$0: removing $data/$part.tar.gz file since --remove-archive option was supplied."
+  rm $data/$part.tar.gz
+fi

training/local/download_lm.sh

@@ -0,0 +1,73 @@
+#!/usr/bin/env bash
+
+# Copyright 2014  Vassil Panayotov
+#           2017  Daniel Povey
+# Apache 2.0
+
+if [ $# -ne "3" ]; then
+  echo "Usage: $0 <base-url> <download_dir> <local?"
+  echo "e.g.: $0 http://www.openslr.org/resources/11 ./corpus/ data/local/lm"
+  exit 1
+fi
+
+base_url=$1
+dst_dir=$2
+local_dir=$3
+
+# given a filename returns the corresponding file size in bytes
+# The switch cases below can be autogenerated by entering the data directory and running:
+# for f in *; do echo "\"$f\") echo \"$(du -b $f | awk '{print $1}')\";;"; done
+function filesize() {
+  case $1 in
+    "3-gram.arpa.gz") echo "759636181";;
+    "3-gram.pruned.1e-7.arpa.gz") echo "34094057";;
+    "3-gram.pruned.3e-7.arpa.gz") echo "13654242";;
+    "librispeech-lexicon.txt") echo "5627653";;
+    "librispeech-vocab.txt") echo "1737588";;
+    *) echo "";;
+  esac
+}
+
+function check_and_download () {
+  [[ $# -eq 1 ]] || { echo "check_and_download() expects exactly one argument!"; return 1; }
+  fname=$1
+  echo "Downloading file '$fname' into '$dst_dir'..."
+  expect_size="$(filesize $fname)"
+  [[ ! -z "$expect_size" ]] || { echo "Unknown file size for '$fname'"; return 1; }
+  if [[ -s $dst_dir/$fname ]]; then
+    # In the following statement, the first version works on linux, and the part
+    # after '||' works on Linux.
+    f=$dst_dir/$fname
+    fsize=$(set -o pipefail; du -b $f 2>/dev/null | awk '{print $1}' || stat '-f %z' $f)
+    if [[ "$fsize" -eq "$expect_size" ]]; then
+      echo "'$fname' already exists and appears to be complete"
+      return 0
+    else
+      echo "WARNING: '$fname' exists, but the size is wrong - re-downloading ..."
+    fi
+  fi
+  wget --no-check-certificate -O $dst_dir/$fname $base_url/$fname || {
+    echo "Error while trying to download $fname!"
+    return 1
+  }
+  f=$dst_dir/$fname
+  # In the following statement, the first version works on linux, and the part after '||'
+  # works on Linux.
+  fsize=$(set -o pipefail; du -b $f 2>/dev/null | awk '{print $1}' || stat '-f %z' $f)
+  [[ "$fsize" -eq "$expect_size" ]] || { echo "$fname: file size mismatch!"; return 1; }
+  return 0
+}
+
+mkdir -p $dst_dir $local_dir
+
+for f in 3-gram.pruned.1e-7.arpa.gz 3-gram.pruned.3e-7.arpa.gz  \
+         librispeech-vocab.txt librispeech-lexicon.txt; do
+  check_and_download $f || exit 1
+done
+
+dst_dir=$(readlink -f $dst_dir)
+ln -sf $dst_dir/3-gram.pruned.1e-7.arpa.gz $local_dir/lm_tgmed.arpa.gz
+ln -sf $dst_dir/3-gram.pruned.3e-7.arpa.gz $local_dir/lm_tgsmall.arpa.gz
+ln -sf $dst_dir/librispeech-lexicon.txt $local_dir/librispeech-lexicon.txt
+ln -sf $dst_dir/librispeech-vocab.txt $local_dir/librispeech-vocab.txt
+exit 0

training/local/prepare_dict.sh

@@ -0,0 +1,18 @@
+#!/usr/bin/env bash
+
+echo "Preparing phone lists and lexicon"
+
+src_dir=$1
+dst_dir=$2
+
+mkdir -p $dst_dir
+cat $src_dir/librispeech-lexicon.txt | sed 's:[012]::g' > $dst_dir/lexicon_raw_nosil.txt
+
+(echo SIL; echo SPN;) > $dst_dir/silence_phones.txt
+echo SIL > $dst_dir/optional_silence.txt
+echo "" > $dst_dir/extra_questions
+
+cat $dst_dir/lexicon_raw_nosil.txt |  awk '{ for(n=2;n<=NF;n++){ phones[$n] = 1; }} END{for (p in phones) print p;}' | \
+  grep -v SIL | sort > $dst_dir/nonsilence_phones.txt
+
+(echo '!SIL SIL'; echo '<UNK> SPN'; ) | cat - $dst_dir/lexicon_raw_nosil.txt | sort | uniq >$dst_dir/lexicon.txt

training/local/score.sh

@@ -0,0 +1,155 @@
+#!/bin/bash
+# Copyright 2012-2014  Johns Hopkins University (Author: Daniel Povey, Yenda Trmal)
+# Apache 2.0
+
+[ -f ./path.sh ] && . ./path.sh
+
+# begin configuration section.
+cmd=run.pl
+stage=0
+decode_mbr=false
+reverse=false
+stats=true
+beam=6
+word_ins_penalty=0.0
+min_lmwt=7
+max_lmwt=13
+iter=final
+#end configuration section.
+
+echo "$0 $@"  # Print the command line for logging
+[ -f ./path.sh ] && . ./path.sh
+. parse_options.sh || exit 1;
+
+if [ $# -ne 3 ]; then
+  echo "Usage: local/score.sh [--cmd (run.pl|queue.pl...)] <data-dir> <lang-dir|graph-dir> <decode-dir>"
+  echo " Options:"
+  echo "    --cmd (run.pl|queue.pl...)      # specify how to run the sub-processes."
+  echo "    --stage (0|1|2)                 # start scoring script from part-way through."
+  echo "    --decode_mbr (true/false)       # maximum bayes risk decoding (confusion network)."
+  echo "    --min_lmwt <int>                # minumum LM-weight for lattice rescoring "
+  echo "    --max_lmwt <int>                # maximum LM-weight for lattice rescoring "
+  echo "    --reverse (true/false)          # score with time reversed features "
+  exit 1;
+fi
+
+data=$1
+lang_or_graph=$2
+dir=$3
+
+symtab=$lang_or_graph/words.txt
+
+for f in $symtab $dir/lat.1.gz $data/text; do
+  [ ! -f $f ] && echo "score.sh: no such file $f" && exit 1;
+done
+
+
+ref_filtering_cmd="cat"
+[ -x local/wer_output_filter ] && ref_filtering_cmd="local/wer_output_filter"
+[ -x local/wer_ref_filter ] && ref_filtering_cmd="local/wer_ref_filter"
+hyp_filtering_cmd="cat"
+[ -x local/wer_output_filter ] && hyp_filtering_cmd="local/wer_output_filter"
+[ -x local/wer_hyp_filter ] && hyp_filtering_cmd="local/wer_hyp_filter"
+
+
+if $decode_mbr ; then
+  echo "$0: scoring with MBR, word insertion penalty=$word_ins_penalty"
+else
+  echo "$0: scoring with word insertion penalty=$word_ins_penalty"
+fi
+
+
+mkdir -p $dir/scoring_kaldi
+cat $data/text | $ref_filtering_cmd > $dir/scoring_kaldi/test_filt.txt || exit 1;
+
+if [ $stage -le 0 ]; then
+
+  for wip in $(echo $word_ins_penalty | sed 's/,/ /g'); do
+    mkdir -p $dir/scoring_kaldi/penalty_$wip/log
+
+    if $decode_mbr ; then
+      $cmd LMWT=$min_lmwt:$max_lmwt $dir/scoring_kaldi/penalty_$wip/log/best_path.LMWT.log \
+        acwt=\`perl -e \"print 1.0/LMWT\"\`\; \
+        lattice-scale --inv-acoustic-scale=LMWT "ark:gunzip -c $dir/lat.*.gz|" ark:- \| \
+        lattice-add-penalty --word-ins-penalty=$wip ark:- ark:- \| \
+        lattice-prune --beam=$beam ark:- ark:- \| \
+        lattice-mbr-decode  --word-symbol-table=$symtab \
+        ark:- ark,t:- \| \
+        utils/int2sym.pl -f 2- $symtab \| \
+        $hyp_filtering_cmd '>' $dir/scoring_kaldi/penalty_$wip/LMWT.txt || exit 1;
+
+    else
+      $cmd LMWT=$min_lmwt:$max_lmwt $dir/scoring_kaldi/penalty_$wip/log/best_path.LMWT.log \
+        lattice-scale --inv-acoustic-scale=LMWT "ark:gunzip -c $dir/lat.*.gz|" ark:- \| \
+        lattice-add-penalty --word-ins-penalty=$wip ark:- ark:- \| \
+        lattice-best-path --word-symbol-table=$symtab ark:- ark,t:- \| \
+        utils/int2sym.pl -f 2- $symtab \| \
+        $hyp_filtering_cmd '>' $dir/scoring_kaldi/penalty_$wip/LMWT.txt || exit 1;
+    fi
+
+    if $reverse; then # rarely-used option, ignore this.
+      for lmwt in `seq $min_lmwt $max_lmwt`; do
+        mv $dir/scoring_kaldi/penalty_$wip/$lmwt.txt $dir/scoring_kaldi/penalty_$wip/$lmwt.txt.orig
+        awk '{ printf("%s ",$1); for(i=NF; i>1; i--){ printf("%s ",$i); } printf("\n"); }' \
+          <$dir/scoring_kaldi/penalty_$wip/$lmwt.txt.orig >$dir/scoring_kaldi/penalty_$wip/$lmwt.txt
+      done
+    fi
+
+    $cmd LMWT=$min_lmwt:$max_lmwt $dir/scoring_kaldi/penalty_$wip/log/score.LMWT.log \
+      cat $dir/scoring_kaldi/penalty_$wip/LMWT.txt \| \
+      compute-wer --text --mode=present \
+      ark:$dir/scoring_kaldi/test_filt.txt  ark,p:- ">&" $dir/wer_LMWT_$wip || exit 1;
+
+  done
+fi
+
+
+
+if [ $stage -le 1 ]; then
+
+  for wip in $(echo $word_ins_penalty | sed 's/,/ /g'); do
+    for lmwt in $(seq $min_lmwt $max_lmwt); do
+      # adding /dev/null to the command list below forces grep to output the filename
+      grep WER $dir/wer_${lmwt}_${wip} /dev/null
+    done
+  done | utils/best_wer.sh  >& $dir/scoring_kaldi/best_wer || exit 1
+
+  best_wer_file=$(awk '{print $NF}' $dir/scoring_kaldi/best_wer)
+  best_wip=$(echo $best_wer_file | awk -F_ '{print $NF}')
+  best_lmwt=$(echo $best_wer_file | awk -F_ '{N=NF-1; print $N}')
+
+  if [ -z "$best_lmwt" ]; then
+    echo "$0: we could not get the details of the best WER from the file $dir/wer_*.  Probably something went wrong."
+    exit 1;
+  fi
+
+  if $stats; then
+    mkdir -p $dir/scoring_kaldi/wer_details
+    echo $best_lmwt > $dir/scoring_kaldi/wer_details/lmwt # record best language model weight
+    echo $best_wip > $dir/scoring_kaldi/wer_details/wip # record best word insertion penalty
+
+    $cmd $dir/scoring_kaldi/log/stats1.log \
+      cat $dir/scoring_kaldi/penalty_$best_wip/$best_lmwt.txt \| \
+      align-text --special-symbol="'***'" ark:$dir/scoring_kaldi/test_filt.txt ark:- ark,t:- \|  \
+      utils/scoring/wer_per_utt_details.pl --special-symbol "'***'" \| tee $dir/scoring_kaldi/wer_details/per_utt \|\
+       utils/scoring/wer_per_spk_details.pl $data/utt2spk \> $dir/scoring_kaldi/wer_details/per_spk || exit 1;
+
+    $cmd $dir/scoring_kaldi/log/stats2.log \
+      cat $dir/scoring_kaldi/wer_details/per_utt \| \
+      utils/scoring/wer_ops_details.pl --special-symbol "'***'" \| \
+      sort -b -i -k 1,1 -k 4,4rn -k 2,2 -k 3,3 \> $dir/scoring_kaldi/wer_details/ops || exit 1;
+
+    $cmd $dir/scoring_kaldi/log/wer_bootci.log \
+      compute-wer-bootci --mode=present \
+        ark:$dir/scoring_kaldi/test_filt.txt ark:$dir/scoring_kaldi/penalty_$best_wip/$best_lmwt.txt \
+        '>' $dir/scoring_kaldi/wer_details/wer_bootci || exit 1;
+
+  fi
+fi
+
+# If we got here, the scoring was successful.
+# As a  small aid to prevent confusion, we remove all wer_{?,??} files;
+# these originate from the previous version of the scoring files
+rm $dir/wer_{?,??} 2>/dev/null
+
+exit 0;

training/path.sh

@@ -0,0 +1,6 @@
+export KALDI_ROOT=`pwd`/../../..
+[ -f $KALDI_ROOT/tools/env.sh ] && . $KALDI_ROOT/tools/env.sh
+export PATH=$PWD/utils/:$KALDI_ROOT/tools/openfst/bin:$PWD:$PATH
+[ ! -f $KALDI_ROOT/tools/config/common_path.sh ] && echo >&2 "The standard file $KALDI_ROOT/tools/config/common_path.sh is not present -> Exit!" && exit 1
+. $KALDI_ROOT/tools/config/common_path.sh
+export LC_ALL=C

training/run.sh

@@ -0,0 +1,100 @@
+#!/usr/bin/env bash
+
+. ./cmd.sh
+. ./path.sh
+
+stage=0
+. utils/parse_options.sh
+
+# Data preparation
+if [ $stage -le 0 ]; then
+  data_url=www.openslr.org/resources/31
+  lm_url=www.openslr.org/resources/11
+  database=corpus
+
+  mkdir -p $database
+  for part in dev-clean-2 train-clean-5; do
+    local/download_and_untar.sh $database $data_url $part
+  done
+
+  local/download_lm.sh $lm_url $database data/local/lm
+
+  local/data_prep.sh $database/LibriSpeech/train-clean-5 data/train
+  local/data_prep.sh $database/LibriSpeech/dev-clean-2 data/test
+fi
+
+# Dictionary formatting
+if [ $stage -le 1 ]; then
+  local/prepare_dict.sh data/local/lm data/local/dict
+  utils/prepare_lang.sh data/local/dict "<UNK>" data/local/lang data/lang
+fi
+
+# Extract MFCC features
+if [ $stage -le 2 ]; then
+  for task in train; do
+    steps/make_mfcc.sh --cmd "$train_cmd" --nj 10 data/$task exp/make_mfcc/$task $mfcc
+    steps/compute_cmvn_stats.sh data/$task exp/make_mfcc/$task $mfcc
+  done
+fi
+
+# Train GMM models
+if [ $stage -le 3 ]; then
+  steps/train_mono.sh --nj 10 --cmd "$train_cmd" \
+    data/train data/lang exp/mono
+
+  steps/align_si.sh  --nj 10 --cmd "$train_cmd" \
+    data/train data/lang exp/mono exp/mono_ali
+
+  steps/train_lda_mllt.sh  --cmd "$train_cmd" \
+    2000 10000 data/train data/lang exp/mono_ali exp/tri1
+
+  steps/align_si.sh --nj 10 --cmd "$train_cmd" \
+    data/train data/lang exp/tri1 exp/tri1_ali
+
+  steps/train_lda_mllt.sh --cmd "$train_cmd" \
+    2500 15000 data/train data/lang exp/tri1_ali exp/tri2
+
+  steps/align_si.sh  --nj 10 --cmd "$train_cmd" \
+    data/train data/lang exp/tri2 exp/tri2_ali
+
+  steps/train_lda_mllt.sh --cmd "$train_cmd" \
+    2500 20000 data/train data/lang exp/tri2_ali exp/tri3
+
+  steps/align_si.sh  --nj 10 --cmd "$train_cmd" \
+    data/train data/lang exp/tri3 exp/tri3_ali
+fi
+
+# Train TDNN model
+if [ $stage -le 4 ]; then
+  local/chain/run_tdnn.sh
+fi
+
+# Decode
+if [ $stage -le 5 ]; then
+
+  utils/format_lm.sh data/lang data/local/lm/lm_tgsmall.arpa.gz data/local/dict/lexicon.txt data/lang_test
+  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_test exp/chain/tdnn exp/chain/tdnn/graph
+  utils/build_const_arpa_lm.sh data/local/lm/lm_tgmed.arpa.gz \
+    data/lang data/lang_test_rescore
+
+  for task in test; do
+
+    steps/make_mfcc.sh --cmd "$train_cmd" --nj 10 data/$task exp/make_mfcc/$task $mfcc
+    steps/compute_cmvn_stats.sh data/$task exp/make_mfcc/$task $mfcc
+
+    steps/online/nnet2/extract_ivectors_online.sh --nj 10 \
+        data/${task} exp/chain/extractor \
+        exp/chain/ivectors_${task}
+
+    steps/nnet3/decode.sh --cmd $decode_cmd --num-threads 10 --nj 1 \
+         --beam 13.0 --max-active 7000 --lattice-beam 4.0 \
+         --online-ivector-dir exp/chain/ivectors_${task} \
+         --acwt 1.0 --post-decode-acwt 10.0 \
+         exp/chain/tdnn/graph data/${task} exp/chain/tdnn/decode_${task}
+
+    steps/lmrescore_const_arpa.sh data/lang_test data/lang_test_rescore \
+        data/${task} exp/chain/tdnn/decode_${task} exp/chain/tdnn/decode_${task}_rescore
+  done
+
+  bash RESULTS
+fi

training/steps

@@ -0,0 +1 @@
+../../wsj/s5/steps/

training/utils

@@ -0,0 +1 @@
+../../wsj/s5/utils/

travis/Dockerfile.dockcross

@@ -21,16 +21,20 @@ RUN apt-get update && \
         python3-cffi \
     && rm -rf /var/lib/apt/lists/*
 
-ARG OPENBLAS_ARCH=ARMV7
-ARG ARM_HARDWARE_OPTS="-mfloat-abi=hard -mfpu=neon"
+ARG OPENBLAS_ARGS=
 RUN cd /opt \
-    && git clone -b lookahead-1.8.0 --single-branch https://github.com/alphacep/kaldi \
+    && git clone -b vosk --single-branch https://github.com/alphacep/kaldi \
     && cd kaldi/tools \
-    && git clone -b v0.3.13 --single-branch https://github.com/xianyi/OpenBLAS \
+    && git clone -b v0.3.20 --single-branch https://github.com/xianyi/OpenBLAS \
     && git clone -b v3.2.1  --single-branch https://github.com/alphacep/clapack \
-    && make -C OpenBLAS ONLY_CBLAS=1 TARGET="${OPENBLAS_ARCH}" HOSTCC=gcc USE_LOCKING=1 USE_THREAD=0 all \
-    && make -C OpenBLAS PREFIX=$(pwd)/OpenBLAS/install install \
-    && mkdir -p clapack/BUILD && cd clapack/BUILD && cmake .. && make -j 10 && find . -name "*.a" | xargs cp -t ../../OpenBLAS/install/lib \
+    && echo ${OPENBLAS_ARGS} \
+    && make -C OpenBLAS ONLY_CBLAS=1 ${OPENBLAS_ARGS} HOSTCC=gcc USE_LOCKING=1 USE_THREAD=0 all \
+    && make -C OpenBLAS ${OPENBLAS_ARGS} HOSTCC=gcc USE_LOCKING=1 USE_THREAD=0 PREFIX=$(pwd)/OpenBLAS/install install \
+    && mkdir -p clapack/BUILD && cd clapack/BUILD && cmake .. \
+    && make -j 10 -C F2CLIBS \
+    && make -j 10 -C BLAS \
+    && make -j 10 -C SRC \
+    && find . -name "*.a" | xargs cp -t ../../OpenBLAS/install/lib \
     && cd /opt/kaldi/tools \
     && git clone --single-branch https://github.com/alphacep/openfst openfst \
     && cd openfst \
@@ -39,7 +43,6 @@ RUN cd /opt \
     && make -j 10 && make install \
     && cd /opt/kaldi/src \
     && sed -i "s:TARGET_ARCH=\"\`uname -m\`\":TARGET_ARCH=$(echo $CROSS_TRIPLE|cut -d - -f 1):g" configure \
-    && sed -i "s:-mfloat-abi=hard -mfpu=neon:${ARM_HARDWARE_OPTS}:g" makefiles/linux_openblas_arm.mk \
     && sed -i "s: -O1 : -O3 :g" makefiles/linux_openblas_arm.mk \
     && ./configure --mathlib=OPENBLAS_CLAPACK --shared --use-cuda=no \
     && make -j 10 online2 lm rnnlm \

travis/Dockerfile.dockcross-manylinux

@@ -0,0 +1,39 @@
+ARG DOCKCROSS_IMAGE=dockcross/manylinux2014-aarch64
+FROM ${DOCKCROSS_IMAGE}
+
+LABEL description="A docker image for building portable Python linux binary wheels and Kaldi on other architectures"
+LABEL maintainer="contact@alphacephei.com"
+
+RUN yum -y install \
+    automake \
+    autoconf \
+    libtool \
+    libffi-devel \
+    && yum clean all
+
+ARG OPENBLAS_ARGS=
+RUN cd /opt \
+    && git clone -b vosk --single-branch https://github.com/alphacep/kaldi \
+    && cd kaldi/tools \
+    && git clone -b v0.3.20 --single-branch https://github.com/xianyi/OpenBLAS \
+    && git clone -b v3.2.1  --single-branch https://github.com/alphacep/clapack \
+    && echo ${OPENBLAS_ARGS} \
+    && make -C OpenBLAS ONLY_CBLAS=1 ${OPENBLAS_ARGS} HOSTCC=gcc USE_LOCKING=1 USE_THREAD=0 all \
+    && make -C OpenBLAS ${OPENBLAS_ARGS} HOSTCC=gcc USE_LOCKING=1 USE_THREAD=0 PREFIX=$(pwd)/OpenBLAS/install install \
+    && mkdir -p clapack/BUILD && cd clapack/BUILD && cmake .. \
+    && make -j 10 -C F2CLIBS \
+    && make -j 10 -C BLAS \
+    && make -j 10 -C SRC \
+    && find . -name "*.a" | xargs cp -t ../../OpenBLAS/install/lib \
+    && cd /opt/kaldi/tools \
+    && git clone --single-branch https://github.com/alphacep/openfst openfst \
+    && cd openfst \
+    && autoreconf -i \
+    && CFLAGS="-g -O3" ./configure --prefix=/opt/kaldi/tools/openfst --enable-static --enable-shared --enable-far --enable-ngram-fsts --enable-lookahead-fsts --with-pic --disable-bin --host=${CROSS_TRIPLE} --build=x86-linux-gnu \
+    && make -j 10 && make install \
+    && cd /opt/kaldi/src \
+    && sed -i "s:TARGET_ARCH=\"\`uname -m\`\":TARGET_ARCH=$(echo $CROSS_TRIPLE|cut -d - -f 1):g" configure \
+    && sed -i "s: -O1 : -O3 :g" makefiles/linux_openblas_arm.mk \
+    && ./configure --mathlib=OPENBLAS_CLAPACK --shared --use-cuda=no \
+    && make -j 10 online2 lm rnnlm \
+    && find /opt/kaldi -name "*.o" -exec rm {} \;

travis/Dockerfile.manylinux

@@ -9,12 +9,13 @@ RUN yum -y update && yum -y install \
     autoconf \
     libtool \
     cmake \
+    libffi-devel \
     && yum clean all
 
 RUN cd /opt \
-    && git clone -b lookahead-1.8.0 --single-branch https://github.com/alphacep/kaldi \
+    && git clone -b vosk --single-branch https://github.com/alphacep/kaldi \
     && cd /opt/kaldi/tools \
-    && git clone -b v0.3.13 --single-branch https://github.com/xianyi/OpenBLAS \
+    && git clone -b v0.3.20 --single-branch https://github.com/xianyi/OpenBLAS \
     && git clone -b v3.2.1  --single-branch https://github.com/alphacep/clapack \
     && make -C OpenBLAS ONLY_CBLAS=1 DYNAMIC_ARCH=1 TARGET=NEHALEM USE_LOCKING=1 USE_THREAD=0 all \
     && make -C OpenBLAS PREFIX=$(pwd)/OpenBLAS/install install \

travis/Dockerfile.win

@@ -39,7 +39,7 @@ RUN mkdir /opt/kaldi \
     && make install
 
 RUN cd /opt/kaldi \
-    && git clone -b v0.3.13 --single-branch https://github.com/xianyi/OpenBLAS \
+    && git clone -b v0.3.20 --single-branch https://github.com/xianyi/OpenBLAS \
     && cd OpenBLAS \
     && make HOSTCC=gcc BINARY=64 CC=x86_64-w64-mingw32-gcc ONLY_CBLAS=1 DYNAMIC_ARCH=1 TARGET=NEHALEM USE_LOCKING=1 USE_THREAD=0 -j $(nproc) \
     && make PREFIX=/opt/kaldi/local install
@@ -55,7 +55,7 @@ RUN cd /opt/kaldi \
     && find . -name *.a -exec cp {} /opt/kaldi/local/lib \;
 
 RUN cd /opt/kaldi \
-    && git clone -b android-mix --single-branch https://github.com/alphacep/kaldi \
+    && git clone -b vosk-android --single-branch https://github.com/alphacep/kaldi \
     && cd kaldi/src \
     && CXX=x86_64-w64-mingw32-g++-posix CXXFLAGS="-O3 -ftree-vectorize -DFST_NO_DYNAMIC_LINKING" ./configure --shared --mingw=yes --use-cuda=no \
         --mathlib=OPENBLAS_CLAPACK \