Critical ivector bugfix and version 0.3.15

Thanks to Markus Elfring

.gitignore

@@ -5,6 +5,9 @@
 # Java class files
 *.class
 
+# Object files
+*.o
+
 # Gradle files
 .gradle/
 build/

README.md

@@ -1,26 +1,25 @@
-Vosk is an open source speech recognition toolkit which supports 9
-languages - English, German, French, Spanish, Portuguese, Chinese,
-Russian, Turkish, Vietnamese. Vosk works offline with small (50 Mb), but
-accurate model, zero-latency response with streaming API, reconfigurable
+# About
+
+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.
+
+Vosk models are small (50 Mb) but provide continuous large vocabulary
+transcription, zero-latency response with streaming API, reconfigurable
 vocabulary and speaker identification.
 
-### Installation and usage
+Speech recognition bindings implemented for various programming languages
+like Python, Java, Node.JS, C#, C++ and others.
 
-For Vosk installation instructions, examples and turorial and documentation visit https://alphacephei.com/vosk
+Vosk supplies speech recognition for chatbots, smart home appliances,
+virtual assistants. It can also create subtitles for movies,
+transcription for lectures and interviews.
 
-### Build
+Vosk scales from small devices like Raspberry Pi or Android smartphone to
+big clusters.
 
-[![Build Status](https://travis-ci.com/alphacep/vosk-api.svg?branch=master)](https://travis-ci.com/alphacep/vosk-api)
+# Documentation
 
-### Models for different languages
-
-For information about models see [the documentation on available models](https://alphacephei.com/vosk/models.html).
-
-### Contact Us
-
-If you have any questions, feel free to:
-
-   * Post an issue here on github
-   * Send us an e-mail at [contact@alphacephei.com](mailto:contact@alphacephei.com)
-   * Join our group dedicated to speech recognition on Telegram [@speech_recognition](https://t.me/speech_recognition)
-   * We have a Wechat group which is pretty big, so it is invitation-only. Mail us to join the group and provide some information about yourself.
+For installation instructions, examples and documentation visit [Vosk
+Website](https://alphacephei.com/vosk).

android/CMakeLists.txt

@@ -22,6 +22,8 @@ set(LIB_ROOT "${PROJECT_SOURCE_DIR}/build/kaldi_${KALDI_SUFFIX}/local")
 set(API_SOURCES
 "${PROJECT_SOURCE_DIR}/../src/kaldi_recognizer.cc"
 "${PROJECT_SOURCE_DIR}/../src/kaldi_recognizer.h"
+"${PROJECT_SOURCE_DIR}/../src/language_model.cc"
+"${PROJECT_SOURCE_DIR}/../src/language_model.h"
 "${PROJECT_SOURCE_DIR}/../src/model.cc"
 "${PROJECT_SOURCE_DIR}/../src/model.h"
 "${PROJECT_SOURCE_DIR}/../src/spk_model.cc"
@@ -32,14 +34,14 @@ set(API_SOURCES
 
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3 -DFST_NO_DYNAMIC_LINKING")
 
-add_library( kaldi_jni SHARED
+add_library( vosk_jni SHARED
              build/generated-src/cpp/vosk_wrap.cc
              ${API_SOURCES}
 )
 
 include_directories("${PROJECT_SOURCE_DIR}/../src" "build/kaldi_${KALDI_SUFFIX}/kaldi/src" "build/kaldi_${KALDI_SUFFIX}/local/include")
 
-target_link_libraries( kaldi_jni
+target_link_libraries( vosk_jni
              ${KALDI_ROOT}/src/online2/kaldi-online2.a
              ${KALDI_ROOT}/src/decoder/kaldi-decoder.a
              ${KALDI_ROOT}/src/ivector/kaldi-ivector.a

android/build.gradle

@@ -5,9 +5,17 @@ buildscript {
     }
     dependencies {
         classpath 'com.android.tools.build:gradle:3.5.3'
+        classpath 'com.jfrog.bintray.gradle:gradle-bintray-plugin:1.8.5'
     }
 }
 
+plugins {
+    id "com.jfrog.bintray" version "1.8.5"
+}
+
+def archiveName = "vosk-android"
+def libVersion = "0.3.15"
+
 allprojects {
   repositories {
       google()
@@ -16,15 +24,16 @@ allprojects {
 }
 
 apply plugin: 'com.android.library'
+apply plugin: 'maven-publish'
 
 android {
     compileSdkVersion 29
     defaultConfig {
         minSdkVersion 21
         targetSdkVersion 29
-        versionCode 5
-        versionName "5.2"
-        setProperty("archivesBaseName", "kaldi-android-$versionName")
+        versionCode 6
+        versionName = libVersion
+        archivesBaseName = archiveName
         externalNativeBuild {
             cmake {
                 arguments "-DCMAKE_VERBOSE_MAKEFILE=ON", "-DANDROID_ARM_NEON=TRUE", "-DCMAKE_CXX_FLAGS_RELEASE=-O3"
@@ -46,6 +55,39 @@ android {
     }
 }
 
+Properties properties = new Properties()
+properties.load(project.rootProject.file('local.properties').newDataInputStream())
+
+bintray {
+    user = properties.getProperty("bintray.user")
+    key = properties.getProperty("bintray.apikey")
+    pkg {
+        repo = 'vosk'
+        name = 'vosk-android'
+        userOrg = "alphacep"
+        licenses = ['Apache2.0']
+        websiteUrl = 'https://github.com/alphacep/vosk-api'
+        issueTrackerUrl = 'https://github.com/alphacep/vosk-api/issues'
+        vcsUrl = 'https://github.com/alphacep/vosk-api'
+        version {
+            name = libVersion
+            vcsTag = libVersion
+        }
+    }
+    publications = ['aar']
+}
+
+publishing {
+    publications {
+        aar(MavenPublication) {
+            groupId 'com.alphacep'
+            artifactId archiveName
+            version libVersion
+            artifact("$buildDir/outputs/aar/$archiveName-release.aar")
+        }
+    }
+}
+
 task swig {
     doLast {
         mkdir 'build/generated-src/java'

android/src/main/java/org/kaldi/SpeechService.java

@@ -29,19 +29,18 @@ import android.os.Looper;
 import android.util.Log;
 
 /**
- * Main class to access recognizer functions. After configuration this class
- * starts a listener thread which records the data and recognizes it using
- * VOSK engine. Recognition events are passed to a client using
+ * Service that records audio in a thread, passes it to a recognizer and emits 
+ * recognition results. Recognition events are passed to a client using
  * {@link RecognitionListener}
- * 
+ *
  */
-public class SpeechRecognizer {
+public class SpeechService {
 
-    protected static final String TAG = SpeechRecognizer.class.getSimpleName();
+    protected static final String TAG = SpeechService.class.getSimpleName();
 
     private final KaldiRecognizer recognizer;
 
-    private final int sampleRate;        
+    private final int sampleRate;
     private final static float BUFFER_SIZE_SECONDS = 0.4f;
     private int bufferSize;
     private final AudioRecord recorder;
@@ -53,33 +52,18 @@ public class SpeechRecognizer {
     private final Collection<RecognitionListener> listeners = new HashSet<RecognitionListener>();
     
     /**
-     * Creates speech recognizer. Recognizer holds the AudioRecord object, so you 
+     * Creates speech service. Service holds the AudioRecord object, so you 
      * need to call {@link release} in order to properly finalize it.
      * 
      * @throws IOException thrown if audio recorder can not be created for some reason.
      */
-    public SpeechRecognizer(Model model) throws IOException {
-        recognizer = new KaldiRecognizer(model, 16000.0f);
-        sampleRate = 16000;
-        bufferSize = Math.round(sampleRate * BUFFER_SIZE_SECONDS);
-        recorder = new AudioRecord(
-                AudioSource.VOICE_RECOGNITION, sampleRate,
-                AudioFormat.CHANNEL_IN_MONO,
-                AudioFormat.ENCODING_PCM_16BIT, bufferSize * 2);
+    public SpeechService(KaldiRecognizer recognizer, float sampleRate) throws IOException {
+        this.recognizer = recognizer;
+        this.sampleRate = (int)sampleRate;
 
-        if (recorder.getState() == AudioRecord.STATE_UNINITIALIZED) {
-            recorder.release();
-            throw new IOException(
-                    "Failed to initialize recorder. Microphone might be already in use.");
-        }
-    }
-
-    public SpeechRecognizer(Model model, SpkModel spkModel) throws IOException {
-        recognizer = new KaldiRecognizer(model, spkModel, 16000.0f);
-        sampleRate = 16000;
-        bufferSize = Math.round(sampleRate * BUFFER_SIZE_SECONDS);
+        bufferSize = Math.round(this.sampleRate * BUFFER_SIZE_SECONDS);
         recorder = new AudioRecord(
-                AudioSource.VOICE_RECOGNITION, sampleRate,
+                AudioSource.VOICE_RECOGNITION, this.sampleRate,
                 AudioFormat.CHANNEL_IN_MONO,
                 AudioFormat.ENCODING_PCM_16BIT, bufferSize * 2);
 
@@ -187,9 +171,9 @@ public class SpeechRecognizer {
     public void shutdown() {
         recorder.release();
     }
-    
+
     private final class RecognizerThread extends Thread {
-        
+
         private int remainingSamples;
         private int timeoutSamples;
         private final static int NO_TIMEOUT = -1;

c/Makefile

@@ -0,0 +1,50 @@
+KALDI_ROOT=$(HOME)/travis/kaldi
+
+VOSK_SOURCES= \
+    ../src/kaldi_recognizer.cc \
+    ../src/language_model.cc \
+    ../src/model.cc \
+    ../src/spk_model.cc \
+    ../src/vosk_api.cc
+
+CFLAGS=-g -O2 -DFST_NO_DYNAMIC_LINKING -I../src -I$(KALDI_ROOT)/src -I$(KALDI_ROOT)/tools/openfst/include
+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/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 \
+	$(KALDI_ROOT)/tools/OpenBLAS/libopenblas.a \
+	$(KALDI_ROOT)/tools/openfst/lib/libfst.a \
+	$(KALDI_ROOT)/tools/openfst/lib/libfstngram.a
+
+all: test_vosk test_vosk_speaker
+
+test_vosk: test_vosk.o libvosk.a
+	g++ $^ -o $@ $(LIBS) -lgfortran -lpthread
+
+test_vosk_speaker: test_vosk_speaker.o libvosk.a
+	g++ $^ -o $@ $(LIBS) -lgfortran -lpthread
+
+libvosk.a: $(VOSK_SOURCES:.cc=.o)
+	ar rcs $@ $^
+
+%.o: %.c
+	g++ $(CFLAGS) -c -o $@ $<
+
+%.o: %.cc
+	g++ -std=c++11 $(CFLAGS) -c -o $@ $<
+
+clean:
+	rm -f *.o *.a test_vosk

c/test_vosk.c

@@ -0,0 +1,28 @@
+#include <vosk_api.h>
+#include <stdio.h>
+
+int main() {
+    FILE *wavin;
+    char buf[3200];
+    int nread, final;
+
+    VoskModel *model = vosk_model_new("model");
+    VoskRecognizer *recognizer = vosk_recognizer_new(model, 16000.0);
+
+    wavin = fopen("test.wav", "rb");
+    fseek(wavin, 44, SEEK_SET);
+    while (!feof(wavin)) {
+         nread = fread(buf, 1, sizeof(buf), wavin);
+         final = vosk_recognizer_accept_waveform(recognizer, buf, nread);
+         if (final) {
+             printf("%s\n", vosk_recognizer_result(recognizer));
+         } else {
+             printf("%s\n", vosk_recognizer_partial_result(recognizer));
+         }
+    }
+    printf("%s\n", vosk_recognizer_final_result(recognizer));
+
+    vosk_recognizer_free(recognizer);
+    vosk_model_free(model);
+    return 0;
+}

c/test_vosk_speaker.c

@@ -0,0 +1,30 @@
+#include <vosk_api.h>
+#include <stdio.h>
+
+int main() {
+    FILE *wavin;
+    char buf[3200];
+    int nread, final;
+
+    VoskModel *model = vosk_model_new("model");
+    VoskSpkModel *spk_model = vosk_spk_model_new("spk-model");
+    VoskRecognizer *recognizer = vosk_recognizer_new_spk(model, spk_model, 16000.0);
+
+    wavin = fopen("test.wav", "rb");
+    fseek(wavin, 44, SEEK_SET);
+    while (!feof(wavin)) {
+         nread = fread(buf, 1, sizeof(buf), wavin);
+         final = vosk_recognizer_accept_waveform(recognizer, buf, nread);
+         if (final) {
+             printf("%s\n", vosk_recognizer_result(recognizer));
+         } else {
+             printf("%s\n", vosk_recognizer_partial_result(recognizer));
+         }
+    }
+    printf("%s\n", vosk_recognizer_final_result(recognizer));
+
+    vosk_recognizer_free(recognizer);
+    vosk_spk_model_free(spk_model);
+    vosk_model_free(model);
+    return 0;
+}

csharp/Makefile

@@ -20,9 +20,12 @@ KALDI_LIBS = \
              ${KALDI_ROOT}/src/util/kaldi-util.a \
              ${KALDI_ROOT}/src/base/kaldi-base.a \
              ${KALDI_ROOT}/tools/openfst/lib/libfst.a \
-             ${KALDI_ROOT}/tools/openfst/lib/libfstngram.a \
-             ${KALDI_ROOT}/tools/OpenBLAS/libopenblas.a \
-             -lgfortran -lstdc++
+             ${KALDI_ROOT}/tools/openfst/lib/libfstngram.a
+
+# On Linux
+MATH_LIBS = ${KALDI_ROOT}/tools/OpenBLAS/libopenblas.a -lgfortran
+# On OSX
+# MATH_LIBS = -framework Accelerate
 
 all: test.exe
 
@@ -32,16 +35,20 @@ test.exe: libkaldiwrap.so test.cs
 VOSK_SOURCES = \
 	vosk_wrap.c \
 	../src/kaldi_recognizer.cc \
-	../src/kaldi_recognizer.h \
+	../src/language_model.cc \
 	../src/model.cc \
-	../src/model.h \
 	../src/spk_model.cc \
-	../src/spk_model.h \
-	../src/vosk_api.cc \
-	../src/vosk_api.h
+	../src/vosk_api.cc
 
-libkaldiwrap.so: $(VOSK_SOURCES)
-	$(CXX) -fpermissive $(CFLAGS) $(CPPFLAGS) -shared -o $@ $(VOSK_SOURCES) $(KALDI_LIBS)
+VOSK_HEADERS = \
+	../src/kaldi_recognizer.h \
+	../src/language_model.h \
+	../src/model.h \
+	../src/vosk_api.h \
+	../src/spk_model.h
+
+libkaldiwrap.so: $(VOSK_SOURCES) $(VOSK_HEADERS)
+	$(CXX) -fpermissive $(CFLAGS) $(CPPFLAGS) -shared -o $@ $(VOSK_SOURCES) $(KALDI_LIBS) $(MATH_LIBS)
 
 vosk_wrap.c: ../src/vosk.i
 	mkdir -p gen

doc/accuracy.md

@@ -1 +1 @@
-See https://alphacephei.com/vosk/accuracy.html
+See https://alphacephei.com/vosk/accuracy

doc/adaptation.md

@@ -1 +1 @@
-See https://alphacephei.com/vosk/adaptation.html
+See https://alphacephei.com/vosk/adaptation

doc/models.md

@@ -1 +1 @@
-See https://alphacephei.com/vosk/models.html
+See https://alphacephei.com/vosk/models

ios/VoskApiTest/Vosk/vosk_api.h

@@ -12,37 +12,199 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+/* This header contains the C API for Vosk speech recognition system */
 
-#ifndef _VOSK_API_H_
-#define _VOSK_API_H_
+#ifndef VOSK_API_H
+#define VOSK_API_H
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
+/** Model stores all the data required for recognition
+ *  it contains static data and can be shared across processing
+ *  threads. */
 typedef struct VoskModel VoskModel;
+
+
+/** Speaker model is the same as model but contains the data
+ *  for speaker identification. */
 typedef struct VoskSpkModel VoskSpkModel;
+
+
+/** Recognizer object is the main object which processes data.
+ *  Each recognizer usually runs in own thread and takes audio as input.
+ *  Once audio is processed recognizer returns JSON object as a string
+ *  which represent decoded information - words, confidences, times, n-best lists,
+ *  speaker information and so on */
 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 */
 VoskModel *vosk_model_new(const char *model_path);
+
+
+/** Releases the model memory
+ *
+ *  The model object is reference-counted so if some recognizer
+ *  depends on this model, model might still stay alive. When
+ *  last recognizer is released, model will be released too. */
 void vosk_model_free(VoskModel *model);
 
+
+/** 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 */
 VoskSpkModel *vosk_spk_model_new(const char *model_path);
+
+
+/** Releases the model memory
+ *
+ *  The model object is reference-counted so if some recognizer
+ *  depends on this model, model might still stay alive. When
+ *  last recognizer is released, model will be released too. */
 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 */
 VoskRecognizer *vosk_recognizer_new(VoskModel *model, float sample_rate);
+
+
+/** Creates the recognizer object with speaker recognition
+ *
+ *  With the speaker recognition mode the recognizer not just recognize
+ *  text but also return speaker vectors one can use for speaker identification
+ *
+ *  @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);
+
+
+/** Creates the recognizer object with the grammar
+ *
+ *  Sometimes when you want to improve recognition accuracy and when you don't need
+ *  to recognize large vocabulary you can specify a list of words to recognize. This
+ *  will improve recognizer speed and accuracy but might return [unk] if user said
+ *  something different.
+ *
+ *  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 grammar The string with the list of words to recognize, for example "one two three four five [unk]"
+ *
+ *  @returns recognizer object */
 VoskRecognizer *vosk_recognizer_new_grm(VoskModel *model, float sample_rate, const char *grammar);
+
+
+/** 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 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);
+
+
+/** 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>
+ * {
+ *   "result" : [{
+ *       "conf" : 1.000000,
+ *       "end" : 1.110000,
+ *       "start" : 0.870000,
+ *       "word" : "what"
+ *     }, {
+ *       "conf" : 1.000000,
+ *       "end" : 1.530000,
+ *       "start" : 1.110000,
+ *       "word" : "zero"
+ *     }, {
+ *       "conf" : 1.000000,
+ *       "end" : 1.950000,
+ *       "start" : 1.530000,
+ *       "word" : "zero"
+ *     }, {
+ *       "conf" : 1.000000,
+ *       "end" : 2.340000,
+ *       "start" : 1.950000,
+ *       "word" : "zero"
+ *     }, {
+ *       "conf" : 1.000000,
+ *      "end" : 2.610000,
+ *       "start" : 2.340000,
+ *       "word" : "one"
+ *     }],
+ *   "text" : "what zero zero zero one"
+ *  }
+ * </pre>
+ */
 const char *vosk_recognizer_result(VoskRecognizer *recognizer);
+
+
+/** Returns partial speech recognition
+ *
+ * @returns partial speech recognition text which is not yet finalized.
+ *          result may change as recognizer process more data.
+ *
+ * <pre>
+ * {
+ *  "partial" : "cyril one eight zero"
+ * }
+ * </pre>
+ */
 const char *vosk_recognizer_partial_result(VoskRecognizer *recognizer);
+
+
+/** Returns speech recognition result. Same as result, but doesn't wait for silence
+ *  You usually call it in the end of the stream to get final bits of audio. It
+ *  flushes the feature pipeline, so all remaining audio chunks got processed.
+ *
+ *  @returns speech result in JSON format.
+ */
 const char *vosk_recognizer_final_result(VoskRecognizer *recognizer);
+
+
+/** Releases recognizer object
+ *
+ *  Underlying model is also unreferenced and if needed released */
 void vosk_recognizer_free(VoskRecognizer *recognizer);
 
+/** Set log level for Kaldi messages
+ *
+ *  @param log_level the level
+ *     0 - default value to print info and error messages but no debug
+ *     less than 0 - don't print info messages
+ *     greather than 0 - more verbose mode
+ */
+void vosk_set_log_level(int log_level);
+
 #ifdef __cplusplus
 }
 #endif
 
-#endif /* _VOSK_API_H_ */
+#endif /* VOSK_API_H */

java/Makefile

@@ -30,6 +30,8 @@ VOSK_SOURCES = \
 	vosk_wrap.cc \
 	../src/kaldi_recognizer.cc \
 	../src/kaldi_recognizer.h \
+	../src/language_model.cc \
+	../src/language_model.h \
 	../src/model.cc \
 	../src/model.h \
 	../src/spk_model.cc \

java/test/DecoderTest.java

@@ -14,9 +14,6 @@ import org.kaldi.SpkModel;
 import org.kaldi.Vosk;
 
 public class DecoderTest {
-    static {
-        System.loadLibrary("vosk_jni");
-    }
 
     public static void main(String args[]) throws IOException {
         Vosk.SetLogLevel(-10);

nodejs/README.md

@@ -9,7 +9,7 @@ Still, you need swig of newest version 4.0.1
 Build like this
 
 ```
-npm install --kaldi_root=/home/suser/kaldi
+npm install --kaldi_root=/home/user/kaldi
 ```
 
 Then test with

nodejs/binding.gyp

@@ -5,22 +5,32 @@
       'sources': [
          '../src/kaldi_recognizer.cc',
          '../src/model.cc',
+         '../src/language_model.cc',
          '../src/spk_model.cc',
          '../src/vosk_api.cc',
          'vosk_wrap.cc',
       ],
       'cflags': [
-           '-std=c++11',
-           '-DFST_NO_DYNAMIC_LINKING',
-           '-Wno-deprecated-declarations',
-           '-Wno-sign-compare',
-           '-Wno-unused-local-typedefs',
-           '-Wno-ignored-quaifiers',
-           '-Wno-extra',
+          '-std=c++11',
+          '-DFST_NO_DYNAMIC_LINKING',
+          '-Wno-deprecated-declarations',
+          '-Wno-sign-compare',
+          '-Wno-unused-local-typedefs',
+          '-Wno-ignored-quaifiers',
+          '-Wno-extra',
       ],
       'cflags_cc!' : [
-           '-fno-rtti',
-           '-fno-exceptions',
+          '-fno-rtti',
+          '-fno-exceptions',
+      ],
+      'conditions': [
+          ['OS == "mac"', {
+              'xcode_settings': {
+                  'GCC_ENABLE_CPP_EXCEPTIONS': 'YES',
+                  'GCC_ENABLE_CPP_RTTI': 'YES',
+                  'CLANG_CXX_LANGUAGE_STANDARD': 'c++11'
+              }
+          }]
       ],
       'actions': [
         {

nodejs/example/test-speaker.js

@@ -1,47 +1,90 @@
-#!/usr/bin/env node
+const wav = require('wav')
+const fs = require('fs')
 
-const fs = require("fs");
-const { Readable } = require("stream");
-const wav = require("wav");
-
-const { Model, KaldiRecognizer, SpkModel } = require("..");
+const {Readable} = require('stream')
+const {Model, KaldiRecognizer, SpkModel} = require('..')
 
 try {
-    fs.accessSync("model", fs.constants.R_OK);
-} catch(err) {
-    console.error("Please download the model from https://github.com/alphacep/kaldi-android-demo/releases and unpack as 'model' in the current folder.");
-    process.exit(1);
+    fs.accessSync('model', fs.constants.R_OK)
+} catch (err) {
+    console.error("Please download the model from https://github.com/alphacep/kaldi-android-demo/releases and unpack as 'model' in the current folder.")
+    process.exit(1)
 }
 
 try {
-    fs.accessSync("model-spk", fs.constants.R_OK);
-} catch(err) {
-    console.error("Please download the speaker model from https://github.com/alphacep/kaldi-android-demo/releases and unpack as 'model-spk' in the current folder.");
-    process.exit(1);
+    fs.accessSync('model-spk', fs.constants.R_OK)
+} catch (err) {
+    console.error("Please download the speaker model from https://github.com/alphacep/kaldi-android-demo/releases and unpack as 'model-spk' in the current folder.")
+    process.exit(1)
 }
 
+const wfStream = fs.createReadStream('test.wav', { highWaterMark: 4096 })
+const wfReader = new wav.Reader()
 
-const wfStream = fs.createReadStream("test.wav", {'highWaterMark': 4096});
-const wfReader = new wav.Reader();
+const model = new Model('model')
+const spkModel = new SpkModel('model-spk')
 
-const model = new Model("model");
-const spkModel = new SpkModel("model-spk");
+const spk_sig = [-0.56648, 0.030579, 1.730239, 0.239899, -1.194183,
+0.251954, 0.540388, -0.971872, -0.020963, 0.085036, 0.563973, -0.019682,
+-0.597381, 1.094719, -0.566738, 0.29819, 0.171165, 0.370341, -0.033539,
+-0.09757, 1.228286, 0.485949, 0.427826, 0.147762, -0.015112, 0.599513,
+-2.040655, -0.490882, 0.440161, -0.072991, 0.835955, -0.496124, 0.952978,
+0.85356, -1.096116, 0.107764, -0.385486, 1.410305, 0.609147, -0.457014,
+-1.542864, 0.343669, 0.171913, -0.627281, -1.281781, -1.134276,
+-0.639895, 1.190183, -0.700537, 1.063457, 0.206946, 0.342198, -1.165625,
+1.475955, -0.089007, -2.555155, 0.551438, -0.212736, 1.025625, -1.631965,
+-0.716256, -1.295995, 1.554956, -1.866009, -1.010782, -1.43231, 0.109027,
+2.123925, 1.703283, -0.784997, 2.730568, 0.755113, 0.0617, 0.128955,
+-0.054047, 1.359119, -0.611666, -1.105754, -0.631353, 0.052109, 0.729386,
+-0.769876, 1.250235, -1.463298, 0.648176, -0.73239, -0.385239,
+-1.661856, 0.602106, -0.45567, -1.438431, -0.836673, 0.033557, 0.373597,
+-1.343341, -0.181095, 0.237287, -0.522005, -1.722836, 0.932333,
+-0.092861, -0.219254, 0.476182, 1.033803, -1.633563, -0.874341, 1.039064,
+-1.758573, -0.838422, -0.324336, -0.924634, 1.962594, 2.152814, 1.2521,
+-0.46172, -1.50271, 1.685691, 0.403097, -0.819042, 0.866403, -0.591716,
+-0.578645, -0.553839, 0.381861, -1.051647, -1.477578, 0.524005, 0.925245]
+
+function dotp(x, y) {
+    function dotp_sum(a, b) {
+        return a + b
+    }
+    function dotp_times(a, i) {
+        return x[i] * y[i]
+    }
+    return x.map(dotp_times).reduce(dotp_sum, 0)
+}
+
+function cosineSimilarity(A, B) {
+    var similarity =
+        dotp(A, B) / (Math.sqrt(dotp(A, A)) * Math.sqrt(dotp(B, B)))
+        return similarity
+}
+
+function cosine_dist(x, y) {
+    return 1 - cosineSimilarity(x, y)
+}
 
 wfReader.on('format', async ({ audioFormat, sampleRate, channels }) => {
     if (audioFormat != 1 || channels != 1) {
-        console.error("Audio file must be WAV format mono PCM.");
-        process.exit(1);
+        console.error('Audio file must be WAV format mono PCM.')
+        process.exit(1)
     }
-    const rec = new KaldiRecognizer(model, spkModel, sampleRate);
+    const rec = new KaldiRecognizer(model, spkModel, sampleRate)
     for await (const data of new Readable().wrap(wfReader)) {
-        const result = await rec.AcceptWaveform(data);
-        if (result != 0) {
-           console.log(await rec.Result());
+        const endOfSpeech = await rec.AcceptWaveform(data)
+        if (endOfSpeech) {
+            res = await JSON.parse(rec.Result());
+            console.log(res)
+            console.log('X-vector:', JSON.stringify(res['spk']))
+            console.log('Speaker distance:', cosine_dist(spk_sig, res['spk']))
         } else {
-           console.log(await rec.PartialResult());
+            console.log(await rec.PartialResult())
         }
     }
-    console.log(await rec.FinalResult());
-});
+    res = await JSON.parse(rec.FinalResult());
+    console.log(res)
+    console.log('X-vector:', JSON.stringify(res['spk']))
+    console.log('Speaker distance:', cosine_dist(spk_sig, res['spk']))
+})
 
-wfStream.pipe(wfReader);
+wfStream.pipe(wfReader)

nodejs/package.json

@@ -1,7 +1,7 @@
 {
   "name": "vosk",
-  "version": "0.3.8",
-  "description": "Node binding for continuous voice recoginition through pocketsphinx.",
+  "version": "0.3.15",
+  "description": "Node binding for continuous offline voice recoginition with Vosk library.",
   "repository": {
     "type": "git",
     "url": "git://github.com/alphacep/vosk-api.git"

python/README.md

@@ -1,3 +1,23 @@
-Python module for vosk-api
+This is a Python module for Vosk.
 
-See for details https://github.com/alphacep/vosk-api
+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.
+
+Vosk models are small (50 Mb) but provide continuous large vocabulary
+transcription, zero-latency response with streaming API, reconfigurable
+vocabulary and speaker identification.
+
+Vosk supplies speech recognition for chatbots, smart home appliances,
+virtual assistants. It can also create subtitles for movies,
+transcription for lectures and interviews.
+
+Vosk scales from small devices like Raspberry Pi or Android smartphone to
+big clusters.
+
+# Documentation
+
+For installation instructions, examples and documentation visit [Vosk
+Website](https://alphacephei.com/vosk). See also our project on
+[Github](https://github.com/alphacep/vosk-api).

python/example/test_ffmpeg.py

@@ -9,7 +9,7 @@ import subprocess
 SetLogLevel(0)
 
 if not os.path.exists("model"):
-    print ("Please download the model from https://github.com/alphacep/vosk-api/blob/master/doc/models.md and unpack as 'model' in the current folder.")
+    print ("Please download the model from https://alphacephei.com/vosk/models and unpack as 'model' in the current folder.")
     exit (1)
 
 sample_rate=16000

python/example/test_microphone.py

@@ -4,7 +4,7 @@ from vosk import Model, KaldiRecognizer
 import os
 
 if not os.path.exists("model"):
-    print ("Please download the model from https://github.com/alphacep/vosk-api/blob/master/doc/models.md and unpack as 'model' in the current folder.")
+    print ("Please download the model from https://alphacephei.com/vosk/models and unpack as 'model' in the current folder.")
     exit (1)
 
 import pyaudio

python/example/test_simple.py

@@ -8,7 +8,7 @@ import wave
 SetLogLevel(0)
 
 if not os.path.exists("model"):
-    print ("Please download the model from https://github.com/alphacep/vosk-api/blob/master/doc/models.md and unpack as 'model' in the current folder.")
+    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")

python/example/test_speaker.py

@@ -11,11 +11,11 @@ model_path = "model"
 spk_model_path = "model-spk"
 
 if not os.path.exists(model_path):
-    print ("Please download the model from https://github.com/alphacep/vosk-api/blob/master/doc/models.md and unpack as {} in the current folder.".format(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://github.com/alphacep/vosk-api/blob/master/doc/models.md and unpack as {} in the current folder.".format(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)
 
 wf = wave.open(sys.argv[1], "rb")
@@ -30,7 +30,7 @@ rec = KaldiRecognizer(model, spk_model, wf.getframerate())
 
 # We compare speakers with cosine distance. We can keep one or several fingerprints for the speaker in a database
 # to distingusih among users.
-spk_sig = [4.658117, 1.277387, 3.346158, -1.473036, -2.15727, 2.461757, 3.76756, -1.241252, 2.333765, 0.642588, -2.848165, 1.229534, 3.907015, 1.726496, -1.188692, 1.16322, -0.668811, -0.623309, 4.628018, 0.407197, 0.089955, 0.920438, 1.47237, -0.311365, -0.437051, -0.531738, -1.591781, 3.095415, 0.439524, -0.274787, 4.03165, 2.665864, 4.815553, 1.581063, 1.078242, 5.017717, -0.089395, -3.123428, 5.34038, 0.456982, 2.465727, 2.131833, 4.056272, 1.178392, -2.075712, -1.568503, 0.847139, 0.409214, 1.84727, 0.986758, 4.222116, 2.235512, 1.369377, 4.283126, 2.278125, -1.467577, -0.999971, 3.070041, 1.462214, 0.423204, 2.143578, 0.567174, -2.294655, 1.864723, 4.307356, 2.610872, -1.238721, 0.551861, 2.861954, 0.59613, -0.715396, -1.395357, 2.706177, -2.004444, 2.055255, 0.458283, 1.231968, 3.48234, 2.993858, 0.402819, 0.940885, 0.360162, -2.173674, -2.504609, 0.329541, 3.653913, 3.638025, -1.406409, 2.14059, 1.662765, -0.991323, 0.770921, 0.010094, 3.775469, 1.847511, 2.074432, -1.928593, 0.807414, 2.964505, 0.128597, 1.297962, 2.645227, 0.136405, -2.543087, 0.932246, 2.405783, -2.122267, 3.044013, 0.486728, 4.395338, 0.474267, 0.781297, 1.694144, -0.831078, -0.462362, -0.964715, 3.187863, 6.008708, 1.725954, 3.667886, -1.467623, 3.370667, 2.72555, -0.796541, 2.416543, 0.675401, -0.737634, -1.709676]
+spk_sig = [-1.110417,0.09703002,1.35658,0.7798632,-0.305457,-0.339204,0.6186931,-0.4521213,0.3982236,-0.004530723,0.7651616,0.6500852,-0.6664245,0.1361499,0.1358056,-0.2887807,-0.1280468,-0.8208137,-1.620276,-0.4628615,0.7870904,-0.105754,0.9739769,-0.3258137,-0.7322628,-0.6212429,-0.5531687,-0.7796484,0.7035915,1.056094,-0.4941756,-0.6521456,-0.2238328,-0.003737517,0.2165709,1.200186,-0.7737719,0.492015,1.16058,0.6135428,-0.7183084,0.3153541,0.3458071,-1.418189,-0.9624157,0.4168292,-1.627305,0.2742135,-0.6166027,0.1962581,-0.6406527,0.4372789,-0.4296024,0.4898657,-0.9531326,-0.2945702,0.7879696,-1.517101,-0.9344181,-0.5049928,-0.005040941,-0.4637912,0.8223695,-1.079849,0.8871287,-0.9732434,-0.5548235,1.879138,-1.452064,-0.1975368,1.55047,0.5941782,-0.52897,1.368219,0.6782904,1.202505,-0.9256122,-0.9718158,-0.9570228,-0.5563112,-1.19049,-1.167985,2.606804,-2.261825,0.01340385,0.2526799,-1.125458,-1.575991,-0.363153,0.3270262,1.485984,-1.769565,1.541829,0.7293826,0.1743717,-0.4759418,1.523451,-2.487134,-1.824067,-0.626367,0.7448186,-1.425648,0.3524166,-0.9903384,3.339342,0.4563958,-0.2876643,1.521635,0.9508078,-0.1398541,0.3867955,-0.7550205,0.6568405,0.09419366,-1.583935,1.306094,-0.3501927,0.1794427,-0.3768163,0.9683866,-0.2442541,-1.696921,-1.8056,-0.6803037,-1.842043,0.3069353,0.9070363,-0.486526]
 
 def cosine_dist(x, y):
     nx = np.array(x)
@@ -45,9 +45,12 @@ while True:
         res = json.loads(rec.Result())
         print ("Text:", res['text'])
         print ("X-vector:", res['spk'])
-        print ("Speaker distance:", cosine_dist(spk_sig, res['spk']))
+        print ("Speaker distance:", cosine_dist(spk_sig, res['spk']), "based on", res['spk_frames'], "frames")
+
+print ("Note that second distance is not very reliable because utterance is too short. Utterances longer than 4 seconds give better xvector")
 
 res = json.loads(rec.FinalResult())
 print ("Text:", res['text'])
-print ("Speaker distance:", cosine_dist(spk_sig, res['spk']))
+print ("X-vector:", res['spk'])
+print ("Speaker distance:", cosine_dist(spk_sig, res['spk']), "based on", res['spk_frames'], "frames")
 

python/example/test_srt.py

@@ -0,0 +1,49 @@
+#!/usr/bin/env python3
+
+from vosk import Model, KaldiRecognizer, SetLogLevel
+import sys
+import os
+import wave
+import subprocess
+import srt
+import json
+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")
+rec = KaldiRecognizer(model, sample_rate)
+
+process = subprocess.Popen(['ffmpeg', '-loglevel', 'quiet', '-i',
+                            sys.argv[1],
+                            '-ar', str(sample_rate) , '-ac', '1', '-f', 's16le', '-'],
+                            stdout=subprocess.PIPE)
+
+
+def transcribe():
+    results = []
+    subs = []
+    while True:
+       data = process.stdout.read(4000)
+       if len(data) == 0:
+           break
+       if rec.AcceptWaveform(data):
+           results.append(rec.Result())
+    results.append(rec.FinalResult())
+
+    for i, res in enumerate(results):
+       jres = json.loads(res)
+       s = srt.Subtitle(index=i, 
+                   content=jres['text'], 
+                   start=datetime.timedelta(seconds=jres['result'][0]['start']), 
+                   end=datetime.timedelta(seconds=jres['result'][-1]['end']))
+       subs.append(s)
+    return subs
+
+print (srt.compose(transcribe()))
+

python/example/test_text.py

@@ -6,7 +6,7 @@ import json
 import os
 
 if not os.path.exists("model"):
-    print ("Please download the model from https://github.com/alphacep/vosk-api/blob/master/doc/models.md and unpack as 'model' in the current folder.")
+    print ("Please download the model from https://alphacephei.com/vosk/models and unpack as 'model' in the current folder.")
     exit (1)
 
 

python/example/test_words.py

@@ -6,7 +6,7 @@ import os
 import wave
 
 if not os.path.exists("model"):
-    print ("Please download the model from https://github.com/alphacep/vosk-api/blob/master/doc/models.md and unpack as 'model' in the current folder.")
+    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")
@@ -15,8 +15,9 @@ if wf.getnchannels() != 1 or wf.getsampwidth() != 2 or wf.getcomptype() != "NONE
     exit (1)
 
 model = Model("model")
-# You can also specify the possible word list
-rec = KaldiRecognizer(model, wf.getframerate(), "zero oh one two three four five six seven eight nine")
+
+# 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]"]')
 
 while True:
     data = wf.readframes(4000)

python/setup.py

@@ -52,14 +52,14 @@ kaldi_libraries = []
 
 if sys.platform.startswith('darwin'):
     kaldi_link_args.extend(['-Wl,-undefined,dynamic_lookup', '-framework', 'Accelerate'])
-elif kaldi_mkl != None:
+elif kaldi_mkl == "1":
     kaldi_link_args.extend(['-L/opt/intel/mkl/lib/intel64', '-Wl,-rpath=/opt/intel/mkl/lib/intel64'])
     kaldi_libraries.extend(['mkl_rt', 'mkl_intel_lp64', 'mkl_core', 'mkl_sequential'])
 else:
     kaldi_static_libs.append('tools/OpenBLAS/libopenblas.a')
     kaldi_libraries.append('gfortran')
 
-sources = ['kaldi_recognizer.cc', 'model.cc', 'spk_model.cc', 'vosk_api.cc', 'vosk.i']
+sources = ['kaldi_recognizer.cc', 'model.cc', 'spk_model.cc', 'vosk_api.cc', 'language_model.cc', 'vosk.i']
 
 vosk_ext = Extension('vosk._vosk',
                     define_macros = [('FST_NO_DYNAMIC_LINKING', '1')],
@@ -72,11 +72,11 @@ vosk_ext = Extension('vosk._vosk',
                     extra_compile_args = ['-std=c++11', '-Wno-sign-compare', '-Wno-unused-variable', '-Wno-unused-local-typedefs'])
 
 setuptools.setup(
-    name="vosk", # Replace with your own username
-    version="0.3.9",
+    name="vosk",
+    version="0.3.15",
     author="Alpha Cephei Inc",
     author_email="contact@alphacephei.com",
-    description="API for Kaldi and Vosk",
+    description="Offline open source speech recognition API based on Kaldi and Vosk",
     long_description=long_description,
     long_description_content_type="text/markdown",
     url="https://github.com/alphacep/vosk-api",

src/kaldi_recognizer.cc

@@ -16,6 +16,7 @@
 #include "json.h"
 #include "fstext/fstext-utils.h"
 #include "lat/sausages.h"
+#include "language_model.h"
 
 using namespace fst;
 using namespace kaldi::nnet3;
@@ -57,28 +58,49 @@ KaldiRecognizer::KaldiRecognizer(Model *model, float sample_frequency, char cons
     feature_pipeline_ = new kaldi::OnlineNnet2FeaturePipeline (model_->feature_info_);
     silence_weighting_ = new kaldi::OnlineSilenceWeighting(*model_->trans_model_, model_->feature_info_.silence_weighting_config, 3);
 
-    g_fst_ = new StdVectorFst();
     if (model_->hcl_fst_) {
-        g_fst_->AddState();
-        g_fst_->SetStart(0);
-        g_fst_->AddState();
-        g_fst_->SetFinal(1, fst::TropicalWeight::One());
-        g_fst_->AddArc(1, StdArc(0, 0, fst::TropicalWeight::One(), 0));
+        json::JSON obj;
+        obj = json::JSON::Load(grammar);
 
-        // Create simple word loop FST
-        stringstream ss(grammar);
-        string token;
+        if (obj.length() <= 0) {
+            KALDI_WARN << "Expecting array of strings, got: '" << grammar << "'";
+        } else {
+            KALDI_LOG << obj;
 
-        while (getline(ss, token, ' ')) {
-            int32 id = model_->word_syms_->Find(token);
-            g_fst_->AddArc(0, StdArc(id, id, fst::TropicalWeight::One(), 1));
+            LanguageModelOptions opts;
+
+            opts.ngram_order = 2;
+            opts.discount = 0.5;
+
+            LanguageModelEstimator estimator(opts);
+            for (int i = 0; i < obj.length(); i++) {
+                bool ok;
+                string line = obj[i].ToString(ok);
+                if (!ok) {
+                    KALDI_ERR << "Expecting array of strings, got: '" << obj << "'";
+                }
+
+                std::vector<int32> sentence;
+                stringstream ss(line);
+                string token;
+                while (getline(ss, token, ' ')) {
+                    int32 id = model_->word_syms_->Find(token);
+                    if (id == kNoSymbol) {
+                        KALDI_WARN << "Ignoring word missing in vocabulary: '" << token << "'";
+                    } else {
+                        sentence.push_back(id);
+                    }
+                }
+                estimator.AddCounts(sentence);
+            }
+            g_fst_ = new StdVectorFst();
+            estimator.Estimate(g_fst_);
+
+            decode_fst_ = LookaheadComposeFst(*model_->hcl_fst_, *g_fst_, model_->disambig_);
         }
-        ArcSort(g_fst_, ILabelCompare<StdArc>());
-
-        decode_fst_ = LookaheadComposeFst(*model_->hcl_fst_, *g_fst_, model_->disambig_);
     } else {
         decode_fst_ = NULL;
-        KALDI_ERR << "Can't create decoding graph";
+        KALDI_WARN << "Runtime graphs are not supported by this model";
     }
 
     decoder_ = new kaldi::SingleUtteranceNnet3Decoder(model_->nnet3_decoding_config_,
@@ -125,9 +147,9 @@ KaldiRecognizer::KaldiRecognizer(Model *model, SpkModel *spk_model, float sample
 }
 
 KaldiRecognizer::~KaldiRecognizer() {
+    delete decoder_;
     delete feature_pipeline_;
     delete silence_weighting_;
-    delete decoder_;
     delete g_fst_;
     delete decode_fst_;
     delete spk_feature_;
@@ -164,12 +186,8 @@ void KaldiRecognizer::CleanUp()
     delete silence_weighting_;
     silence_weighting_ = new kaldi::OnlineSilenceWeighting(*model_->trans_model_, model_->feature_info_.silence_weighting_config, 3);
 
-    if (spk_feature_) {
-        delete spk_feature_;
-        spk_feature_ = new OnlineMfcc(spk_model_->spkvector_mfcc_opts);
-    }
-
-    frame_offset_ += decoder_->NumFramesDecoded();
+    if (decoder_)
+       frame_offset_ += decoder_->NumFramesDecoded();
 
     // Each 10 minutes we drop the pipeline to save frontend memory in continuous processing
     // here we drop few frames remaining in the feature pipeline but hope it will not
@@ -177,8 +195,9 @@ void KaldiRecognizer::CleanUp()
 
     // Also restart if we retrieved final result already
 
-    if (frame_offset_ > 20000 || state_ == RECOGNIZER_FINALIZED) {
+    if (decoder_ == NULL || state_ == RECOGNIZER_FINALIZED || frame_offset_ > 20000) {
         samples_round_start_ += samples_processed_;
+        samples_processed_ = 0;
         frame_offset_ = 0;
 
         delete decoder_;
@@ -190,6 +209,11 @@ void KaldiRecognizer::CleanUp()
             *model_->decodable_info_,
             model_->hclg_fst_ ? *model_->hclg_fst_ : *decode_fst_,
             feature_pipeline_);
+
+        if (spk_model_) {
+            delete spk_feature_;
+            spk_feature_ = new OnlineMfcc(spk_model_->spkvector_mfcc_opts);
+        }
     } else {
         decoder_->InitDecoding(frame_offset_);
     }
@@ -292,7 +316,9 @@ static void RunNnetComputation(const MatrixBase<BaseFloat> &features,
     xvector->CopyFromVec(cu_output.Row(0));
 }
 
-void KaldiRecognizer::GetSpkVector(Vector<BaseFloat> &xvector)
+#define MIN_SPK_FEATS 50
+
+bool KaldiRecognizer::GetSpkVector(Vector<BaseFloat> &out_xvector, int *num_spk_frames)
 {
     vector<int32> nonsilence_frames;
     if (silence_weighting_->Active() && feature_pipeline_->NumFramesReady() > 0) {
@@ -302,17 +328,36 @@ void KaldiRecognizer::GetSpkVector(Vector<BaseFloat> &xvector)
                                           &nonsilence_frames);
     }
 
-    int num_frames = spk_feature_->NumFramesReady();
+    int num_frames = spk_feature_->NumFramesReady() - frame_offset_ * 3;
     Matrix<BaseFloat> mfcc(num_frames, spk_feature_->Dim());
+
+    // Not very efficient, would be nice to have faster search
+    int num_nonsilence_frames = 0;
+    Vector<BaseFloat> feat(spk_feature_->Dim());
+
     for (int i = 0; i < num_frames; ++i) {
        if (std::find(nonsilence_frames.begin(),
-                     nonsilence_frames.end(), i % 3) == nonsilence_frames.end())
+                     nonsilence_frames.end(), i / 3) == nonsilence_frames.end()) {
            continue;
-       Vector<BaseFloat> feat(spk_feature_->Dim());
-       spk_feature_->GetFrame(i, &feat);
-       mfcc.CopyRowFromVec(feat, i);
+       }
+
+       spk_feature_->GetFrame(i + frame_offset_ * 3, &feat);
+       mfcc.CopyRowFromVec(feat, num_nonsilence_frames);
+       num_nonsilence_frames++;
     }
+
+    *num_spk_frames = num_nonsilence_frames;
+
+    // Don't extract vector if not enough data
+    if (num_nonsilence_frames < MIN_SPK_FEATS) {
+        return false;
+    }
+
+    mfcc.Resize(num_nonsilence_frames, spk_feature_->Dim(), kCopyData);
+
     SlidingWindowCmnOptions cmvn_opts;
+    cmvn_opts.center = true;
+    cmvn_opts.cmn_window = 300;
     Matrix<BaseFloat> features(mfcc.NumRows(), mfcc.NumCols(), kUndefined);
     SlidingWindowCmn(cmvn_opts, mfcc, &features);
 
@@ -320,7 +365,22 @@ void KaldiRecognizer::GetSpkVector(Vector<BaseFloat> &xvector)
     nnet3::CachingOptimizingCompilerOptions compiler_config;
     nnet3::CachingOptimizingCompiler compiler(spk_model_->speaker_nnet, opts.optimize_config, compiler_config);
 
+    Vector<BaseFloat> xvector;
     RunNnetComputation(features, spk_model_->speaker_nnet, &compiler, &xvector);
+
+    // Whiten the vector with global mean and transform and normalize mean
+    xvector.AddVec(-1.0, spk_model_->mean);
+
+    out_xvector.Resize(spk_model_->transform.NumRows(), kSetZero);
+    out_xvector.AddMatVec(1.0, spk_model_->transform, kNoTrans, xvector, 0.0);
+
+    BaseFloat norm = out_xvector.Norm(2.0);
+    BaseFloat ratio = norm / sqrt(out_xvector.Dim()); // how much larger it is
+                                                  // than it would be, in
+                                                  // expectation, if normally
+    out_xvector.Scale(1.0 / ratio);
+
+    return true;
 }
 
 const char* KaldiRecognizer::GetResult()
@@ -355,7 +415,7 @@ const char* KaldiRecognizer::GetResult()
         DeterminizeLattice(composed_lat1, &clat);
     }
 
-    fst::ScaleLattice(fst::LatticeScale(9.0, 10.0), &clat);
+    fst::ScaleLattice(fst::GraphLatticeScale(0.9), &clat); // Apply rescoring weight
     CompactLattice aligned_lat;
     if (model_->winfo_) {
         WordAlignLattice(clat, *model_->trans_model_, *model_->winfo_, 0, &aligned_lat);
@@ -392,9 +452,12 @@ const char* KaldiRecognizer::GetResult()
 
     if (spk_model_) {
         Vector<BaseFloat> xvector;
-        GetSpkVector(xvector);
-        for (int i = 0; i < xvector.Dim(); i++) {
-            obj["spk"].append(xvector(i));
+        int num_spk_frames;
+        if (GetSpkVector(xvector, &num_spk_frames)) {
+            for (int i = 0; i < xvector.Dim(); i++) {
+                obj["spk"].append(xvector(i));
+            }
+            obj["spk_frames"] = num_spk_frames;
         }
     }
 
@@ -454,7 +517,21 @@ const char* KaldiRecognizer::FinalResult()
     decoder_->AdvanceDecoding();
     decoder_->FinalizeDecoding();
     state_ = RECOGNIZER_FINALIZED;
-    return GetResult();
+    GetResult();
+
+    // Free some memory while we are finalized, next
+    // iteration will reinitialize them anyway
+    delete decoder_;
+    delete feature_pipeline_;
+    delete silence_weighting_;
+    delete spk_feature_;
+
+    feature_pipeline_ = NULL;
+    silence_weighting_ = NULL;
+    decoder_ = NULL;
+    spk_feature_ = NULL;
+
+    return last_result_.c_str();
 }
 
 // Store result in recognizer and return as const string

src/kaldi_recognizer.h

@@ -12,6 +12,9 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#ifndef VOSK_KALDI_RECOGNIZER_H
+#define VOSK_KALDI_RECOGNIZER_H
+
 #include "base/kaldi-common.h"
 #include "util/common-utils.h"
 #include "fstext/fstext-lib.h"
@@ -55,7 +58,7 @@ class KaldiRecognizer {
         void CleanUp();
         void UpdateSilenceWeights();
         bool AcceptWaveform(Vector<BaseFloat> &wdata);
-        void GetSpkVector(Vector<BaseFloat> &xvector);
+        bool GetSpkVector(Vector<BaseFloat> &out_xvector, int *frames);
         const char *GetResult();
         const char *StoreReturn(const string &res);
 
@@ -80,3 +83,5 @@ class KaldiRecognizer {
         KaldiRecognizerState state_;
         string last_result_;
 };
+
+#endif /* VOSK_KALDI_RECOGNIZER_H */

src/language_model.cc

@@ -0,0 +1,211 @@
+// Copyright      2015   Johns Hopkins University (author: Daniel Povey)
+
+// See ../../COPYING for clarification regarding multiple authors
+//
+// 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
+//
+// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
+// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
+// MERCHANTABLITY OR NON-INFRINGEMENT.
+// See the Apache 2 License for the specific language governing permissions and
+// limitations under the License.
+
+// A modified version from chain/language-model.cc for static backoff
+
+#include <algorithm>
+#include <numeric>
+
+#include "language_model.h"
+
+using namespace kaldi;
+
+void LanguageModelEstimator::AddCounts(const std::vector<int32> &sentence) {
+  KALDI_ASSERT(opts_.ngram_order >= 2 && "--ngram-order must be >= 2");
+  int32 order = opts_.ngram_order;
+  // 0 is used for left-context at the beginning of the file.. treat it as BOS.
+  std::vector<int32> history(0);
+  std::vector<int32>::const_iterator iter = sentence.begin(),
+      end = sentence.end();
+  for (; iter != end; ++iter) {
+    KALDI_ASSERT(*iter != 0);
+    IncrementCount(history, *iter);
+    history.push_back(*iter);
+    if (history.size() >= order)
+      history.erase(history.begin());
+  }
+  // Probability of end of sentence.  This will end up getting ignored later, but
+  // it still makes a difference for probability-normalization reasons.
+  IncrementCount(history, 0);
+}
+
+void LanguageModelEstimator::IncrementCount(const std::vector<int32> &history,
+                                            int32 next_phone) {
+  int32 lm_state_index = FindOrCreateLmStateIndexForHistory(history);
+  if (lm_states_[lm_state_index].tot_count == 0) {
+    num_active_lm_states_++;
+  }
+  lm_states_[lm_state_index].AddCount(next_phone, 1);
+}
+
+void LanguageModelEstimator::SetParentCounts() {
+  int32 num_lm_states = lm_states_.size();
+  for (int32 l = 0; l < num_lm_states; l++) {
+    int32 l_iter = lm_states_[l].backoff_lmstate_index;
+    while (l_iter != -1) {
+      lm_states_[l_iter].Add(lm_states_[l]);
+      l_iter = lm_states_[l_iter].backoff_lmstate_index;
+    }
+  }
+}
+
+int32 LanguageModelEstimator::FindLmStateIndexForHistory(
+    const std::vector<int32> &hist) const {
+  MapType::const_iterator iter = hist_to_lmstate_index_.find(hist);
+  if (iter == hist_to_lmstate_index_.end())
+    return -1;
+  else
+    return iter->second;
+}
+
+int32 LanguageModelEstimator::FindNonzeroLmStateIndexForHistory(
+    std::vector<int32> hist) const {
+  while (1) {
+    int32 l = FindLmStateIndexForHistory(hist);
+    if (l == -1 || lm_states_[l].tot_count == 0) {
+      // no such state or state has zero count.
+      if (hist.empty())
+        KALDI_ERR << "Error looking up LM state index for history "
+                  << "(likely code bug)";
+      hist.erase(hist.begin());  // back off.
+    } else {
+      return l;
+    }
+  }
+}
+
+int32 LanguageModelEstimator::FindOrCreateLmStateIndexForHistory(
+    const std::vector<int32> &hist) {
+  MapType::const_iterator iter = hist_to_lmstate_index_.find(hist);
+  if (iter != hist_to_lmstate_index_.end())
+    return iter->second;
+  int32 ans = lm_states_.size();  // index of next element
+  // next statement relies on default construct of LmState.
+  lm_states_.resize(lm_states_.size() + 1);
+  lm_states_.back().history = hist;
+  hist_to_lmstate_index_[hist] = ans;
+
+  // make sure backoff_lmstate_index is set
+  if (hist.size() > 0) {
+    std::vector<int32> backoff_hist(hist.begin() + 1,
+                                    hist.end());
+    int32 backoff_lm_state = FindOrCreateLmStateIndexForHistory(backoff_hist);
+    lm_states_[ans].backoff_lmstate_index = backoff_lm_state;
+  }
+  return ans;
+}
+
+void LanguageModelEstimator::LmState::AddCount(int32 phone, int32 count) {
+  std::map<int32, int32>::iterator iter = phone_to_count.find(phone);
+  if (iter == phone_to_count.end())
+    phone_to_count[phone] = count;
+  else
+    iter->second += count;
+  tot_count += count;
+}
+
+void LanguageModelEstimator::LmState::Add(const LmState &other) {
+  KALDI_ASSERT(&other != this);
+  std::map<int32, int32>::const_iterator iter = other.phone_to_count.begin(),
+      end = other.phone_to_count.end();
+  for (; iter != end; ++iter)
+    AddCount(iter->first, iter->second);
+}
+
+int32 LanguageModelEstimator::AssignFstStates() {
+  int32 num_lm_states = lm_states_.size();
+  int32 current_fst_state = 0;
+  for (int32 l = 0; l < num_lm_states; l++) {
+    if (lm_states_[l].tot_count != 0) {
+      lm_states_[l].fst_state = current_fst_state++;
+    }
+  }
+  KALDI_ASSERT(current_fst_state == num_active_lm_states_);
+  return current_fst_state;
+}
+
+void LanguageModelEstimator::Estimate(fst::StdVectorFst *fst) {
+  KALDI_LOG << "Estimating language model with ngram-order="
+            << opts_.ngram_order << ", discount="
+            << opts_.discount;
+  SetParentCounts();
+  int32 num_fst_states = AssignFstStates();
+  OutputToFst(num_fst_states, fst);
+}
+
+int32 LanguageModelEstimator::FindInitialFstState() const {
+  std::vector<int32> history(0);
+  int32 l = FindNonzeroLmStateIndexForHistory(history);
+  KALDI_ASSERT(l != -1 && lm_states_[l].fst_state != -1);
+  return lm_states_[l].fst_state;
+}
+
+void LanguageModelEstimator::OutputToFst(
+    int32 num_states,
+    fst::StdVectorFst *fst) const {
+  KALDI_ASSERT(num_states == num_active_lm_states_);
+  fst->DeleteStates();
+  for (int32 i = 0; i < num_states; i++)
+    fst->AddState();
+  fst->SetStart(FindInitialFstState());
+
+  int64 tot_count = 0;
+  double tot_logprob = 0.0;
+
+  int32 num_lm_states = lm_states_.size();
+  // note: not all lm-states end up being 'active'.
+  for (int32 l = 0; l < num_lm_states; l++) {
+    const LmState &lm_state = lm_states_[l];
+    if (lm_state.fst_state == -1) {
+      continue;
+    }
+    int32 state_count = lm_state.tot_count;
+    KALDI_ASSERT(state_count != 0);
+    std::map<int32, int32>::const_iterator
+        iter = lm_state.phone_to_count.begin(),
+        end = lm_state.phone_to_count.end();
+    for (; iter != end; ++iter) {
+      int32 phone = iter->first, count = iter->second;
+      BaseFloat logprob = log(count * opts_.discount / state_count);
+      tot_count += count;
+      tot_logprob += logprob * count;
+      if (phone == 0) {  // Go to final state
+        fst->SetFinal(lm_state.fst_state, fst::TropicalWeight(-logprob));
+      } else {  // It becomes a transition.
+        std::vector<int32> next_history(lm_state.history);
+        next_history.push_back(phone);
+        int32 dest_lm_state = FindNonzeroLmStateIndexForHistory(next_history),
+            dest_fst_state = lm_states_[dest_lm_state].fst_state;
+        KALDI_ASSERT(dest_fst_state != -1);
+        fst->AddArc(lm_state.fst_state,
+                    fst::StdArc(phone, phone, fst::TropicalWeight(-logprob),
+                                dest_fst_state));
+      }
+    }
+    if (lm_state.backoff_lmstate_index >= 0) {
+      fst->AddArc(lm_state.fst_state, fst::StdArc(0, 0, fst::TropicalWeight(-log(1 - opts_.discount)), lm_states_[lm_state.backoff_lmstate_index].fst_state));
+    }
+  }
+  fst::Connect(fst);
+  // Make sure that Connect does not delete any states.
+  int32 num_states_connected = fst->NumStates();
+  KALDI_ASSERT(num_states_connected == num_states);
+  // arc-sort.  ilabel or olabel doesn't matter, it's an acceptor.
+  fst::ArcSort(fst, fst::ILabelCompare<fst::StdArc>());
+  KALDI_LOG << "Created language model with " << num_states
+            << " states and " << fst::NumArcs(*fst) << " arcs.";
+}

src/language_model.h

@@ -0,0 +1,150 @@
+// Copyright      2015  Johns Hopkins University (Author: Daniel Povey)
+
+// See ../../COPYING for clarification regarding multiple authors
+//
+// 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
+//
+// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
+// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
+// MERCHANTABLITY OR NON-INFRINGEMENT.
+// See the Apache 2 License for the specific language governing permissions and
+// limitations under the License.
+
+
+#ifndef VOSK_LANGUAGE_MODEL_H
+#define VOSK_LANGUAGE_MODEL_H
+
+#include <vector>
+#include <map>
+
+#include "base/kaldi-common.h"
+#include "util/common-utils.h"
+#include "fstext/fstext-lib.h"
+#include "lat/kaldi-lattice.h"
+
+using namespace kaldi;
+
+// Very simply lm construction with absolute discounting
+
+struct LanguageModelOptions {
+  int32 ngram_order;  // you might want to tune this
+  BaseFloat discount; // discount for backoff
+
+  LanguageModelOptions():
+      ngram_order(3),
+      discount(0.5)
+      { }
+
+  void Register(OptionsItf *opts) {
+    opts->Register("ngram-order", &ngram_order, "n-gram order for the phone "
+                   "language model used for the 'denominator model'");
+    opts->Register("discount", &discount, "Discount for backoff");
+  }
+};
+
+class LanguageModelEstimator {
+ public:
+  LanguageModelEstimator(LanguageModelOptions &opts): opts_(opts),
+                                                      num_active_lm_states_(0) {
+    KALDI_ASSERT(opts.ngram_order >= 1);
+  }
+
+  // Adds counts for this sentence.  Basically does: for each n-gram in the
+  // sentence, count[n-gram] += 1.  The only constraint on 'sentence' is that it
+  // should contain no zeros.
+  void AddCounts(const std::vector<int32> &sentence);
+
+  // Estimates the LM and outputs it as an FST.  Note: there is
+  // no concept here of backoff arcs.
+  void Estimate(fst::StdVectorFst *fst);
+
+ protected:
+  struct LmState {
+    // the phone history associated with this state (length can vary).
+    std::vector<int32> history;
+
+    // maps from
+    std::map<int32, int32> phone_to_count;
+
+    // total count of this state.  As we back off states to lower-order states
+    // (and note that this is a hard backoff where we completely remove un-needed
+    // states) this tot_count may become zero.
+    int32 tot_count;
+
+    // LM-state index of the backoff LM state (if it exists, else -1)...
+    // provided for convenience.
+    int32 backoff_lmstate_index;
+
+    // this is only set after we decide on the FST state numbering (at the end).
+    // If not set, it's -1.
+    int32 fst_state;
+
+    void AddCount(int32 phone, int32 count);
+
+    // Add the contents of another LmState.
+    void Add(const LmState &other);
+
+    LmState(): tot_count(0), backoff_lmstate_index(-1),
+               fst_state(-1) { }
+    LmState(const LmState &other):
+        history(other.history), phone_to_count(other.phone_to_count),
+        tot_count(other.tot_count),
+        backoff_lmstate_index(other.backoff_lmstate_index),
+        fst_state(other.fst_state) { }
+  };
+
+  // maps from history to int32
+  typedef unordered_map<std::vector<int32>, int32, VectorHasher<int32> > MapType;
+
+  LanguageModelOptions opts_;
+
+  MapType hist_to_lmstate_index_;
+  std::vector<LmState> lm_states_;  // indexed by lmstate_index, the LmStates.
+
+  // Keeps track of the number of lm states that have nonzero counts.
+  int32 num_active_lm_states_;
+
+
+  // adds the counts for this ngram (called from AddCounts()).
+  inline void IncrementCount(const std::vector<int32> &history,
+                             int32 next_phone);
+
+  // sets up tot_count_with_parents in all the lm-states
+  void SetParentCounts();
+
+  // Finds and returns an LM-state index for a history -- or -1 if it doesn't
+  // exist.  No backoff is done.
+  int32 FindLmStateIndexForHistory(const std::vector<int32> &hist) const;
+
+  // Finds and returns an LM-state index for a history -- and creates one if
+  // it doesn't exist -- and also creates any backoff states needed, down
+  // to history-length no_prune_ngram_order - 1.
+  int32 FindOrCreateLmStateIndexForHistory(const std::vector<int32> &hist);
+
+  // Finds and returns the most specific LM-state index for a history or
+  // backed-off versions of it, that exists and has nonzero count.  Will die if
+  // there is no such history.  [e.g. if there is no unigram backoff state,
+  // which generally speaking there won't be.]
+  int32 FindNonzeroLmStateIndexForHistory(std::vector<int32> hist) const;
+
+  // after all backoff has been done, assigns FST state indexes to all states
+  // that exist and have nonzero count.  Returns the number of states.
+  int32 AssignFstStates();
+
+  // find the FST index of the initial-state, and returns it.
+  int32 FindInitialFstState() const;
+
+  // Write to an FST
+  void OutputToFst(
+      int32 num_fst_states,
+      fst::StdVectorFst *fst) const;
+
+};
+
+#endif
+

src/model.cc

@@ -225,6 +225,8 @@ void Model::ReadDataFiles()
         ivector_extraction_opts.global_cmvn_stats_rxfilename = model_path_str_ + "/ivector/global_cmvn.stats";
         ivector_extraction_opts.diag_ubm_rxfilename = model_path_str_ + "/ivector/final.dubm";
         ivector_extraction_opts.ivector_extractor_rxfilename = model_path_str_ + "/ivector/final.ie";
+        ivector_extraction_opts.max_count = 100;
+
         feature_info_.use_ivectors = true;
         feature_info_.ivector_extractor_info.Init(ivector_extraction_opts);
     } else {

src/model.h

@@ -12,8 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#ifndef MODEL_H_
-#define MODEL_H_
+#ifndef VOSK_MODEL_H
+#define VOSK_MODEL_H
 
 #include "base/kaldi-common.h"
 #include "fstext/fstext-lib.h"
@@ -86,4 +86,4 @@ protected:
     int ref_cnt_;
 };
 
-#endif /* MODEL_H_ */
+#endif /* VOSK_MODEL_H */

src/spk_model.cc

@@ -25,6 +25,9 @@ SpkModel::SpkModel(const char *speaker_path) {
     SetDropoutTestMode(true, &speaker_nnet);
     CollapseModel(nnet3::CollapseModelConfig(), &speaker_nnet);
 
+    ReadKaldiObject(speaker_path_str + "/mean.vec", &mean);
+    ReadKaldiObject(speaker_path_str + "/transform.mat", &transform);
+
     ref_cnt_ = 1;
 }
 

src/spk_model.h

@@ -12,8 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#ifndef SPK_MODEL_H_
-#define SPK_MODEL_H_
+#ifndef VOSK_SPK_MODEL_H
+#define VOSK_SPK_MODEL_H
 
 #include "base/kaldi-common.h"
 #include "online2/online-feature-pipeline.h"
@@ -35,9 +35,12 @@ protected:
     ~SpkModel() {};
 
     kaldi::nnet3::Nnet speaker_nnet;
+    kaldi::Vector<BaseFloat> mean;
+    kaldi::Matrix<BaseFloat> transform;
+
     MfccOptions spkvector_mfcc_opts;
 
     int ref_cnt_;
 };
 
-#endif /* SPK_MODEL_H_ */
+#endif /* VOSK_SPK_MODEL_H */

src/vosk.i

@@ -34,6 +34,11 @@ import java.nio.ByteOrder;
     return AcceptWaveform(bdata, bdata.length);
   }
 %}
+%pragma(java) jniclasscode=%{
+    static {
+        System.loadLibrary("vosk_jni");
+    }
+%}
 #endif
 
 #if SWIGCSHARP

src/vosk_api.h

@@ -14,8 +14,8 @@
 
 /* This header contains the C API for Vosk speech recognition system */
 
-#ifndef _VOSK_API_H_
-#define _VOSK_API_H_
+#ifndef VOSK_API_H
+#define VOSK_API_H
 
 #ifdef __cplusplus
 extern "C" {
@@ -88,10 +88,10 @@ VoskRecognizer *vosk_recognizer_new(VoskModel *model, float sample_rate);
 VoskRecognizer *vosk_recognizer_new_spk(VoskModel *model, VoskSpkModel *spk_model, float sample_rate);
 
 
-/** Creates the recognizer object with the grammar
+/** Creates the recognizer object with the phrase list
  *
  *  Sometimes when you want to improve recognition accuracy and when you don't need
- *  to recognize large vocabulary you can specify a list of words to recognize. This
+ *  to recognize large vocabulary you can specify a list of phrases to recognize. This
  *  will improve recognizer speed and accuracy but might return [unk] if user said
  *  something different.
  *
@@ -99,7 +99,8 @@ VoskRecognizer *vosk_recognizer_new_spk(VoskModel *model, VoskSpkModel *spk_mode
  *  Precompiled HCLG graph models are not supported.
  *
  *  @param sample_rate The sample rate of the audio you going to feed into the recognizer
- *  @param grammar The string with the list of words to recognize, for example "one two three four five [unk]"
+ *  @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 */
 VoskRecognizer *vosk_recognizer_new_grm(VoskModel *model, float sample_rate, const char *grammar);
@@ -207,4 +208,4 @@ void vosk_set_log_level(int log_level);
 }
 #endif
 
-#endif /* _VOSK_API_H_ */
+#endif /* VOSK_API_H */

travis/Dockerfile.dockcross

@@ -90,7 +90,7 @@ RUN cd /opt \
     && cd kaldi/tools \
     && git clone -b v0.3.7 --single-branch https://github.com/xianyi/OpenBLAS \
     && make PREFIX=$(pwd)/OpenBLAS/install TARGET="${OPENBLAS_ARCH}" HOSTCC=gcc USE_LOCKING=1 USE_THREAD=0 -C OpenBLAS all install \
-    && git clone https://github.com/alphacep/openfst openfst \
+    && git clone -b old-gcc --single-branch https://github.com/alphacep/openfst openfst \
     && cd openfst \
     && autoreconf -i \
     && ./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 \

travis/Dockerfile.manylinux

@@ -24,7 +24,7 @@ RUN cd /opt \
     && git clone -b lookahead --single-branch https://github.com/alphacep/kaldi \
     && cd kaldi/tools \
     && git clone -b v0.3.7 --single-branch https://github.com/xianyi/OpenBLAS \
-    && make PREFIX=$(pwd)/OpenBLAS/install TARGET=NEHALEM USE_LOCKING=1 USE_THREAD=0 -C OpenBLAS all install \
+    && make PREFIX=$(pwd)/OpenBLAS/install DYNAMIC_ARCH=1 USE_LOCKING=1 USE_THREAD=0 -C OpenBLAS all install \
     && git clone https://github.com/alphacep/openfst openfst \
     && cd openfst \
     && autoreconf -i \
@@ -32,5 +32,5 @@ RUN cd /opt \
     && make -j 10 && make install \
     && cd /opt/kaldi/src \
     && ./configure --mathlib=OPENBLAS --shared --use-cuda=no \
-    && make -j 10 online2 lm \
+    && make -j 10 online2 lm rnnlm \
     && find /opt/kaldi -name "*.o" -exec rm {} \;