Attempt to build manylinux wheel

Allow to use vosk-server for transcription

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

@@ -4,14 +4,15 @@ 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. More to come.
+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

@@ -10,7 +10,7 @@ buildscript {
 }
 
 allprojects {
-    version = '0.3.41'
+    version = '0.3.43'
 }
 
 subprojects {

csharp/demo/VoskDemo.csproj

@@ -11,7 +11,7 @@
   </PropertyGroup>
 
   <ItemGroup>
-    <PackageReference Include="Vosk" Version="0.3.41" />
+    <PackageReference Include="Vosk" Version="0.3.43" />
   </ItemGroup>
 
 </Project>

csharp/nuget/Vosk.nuspec

@@ -2,7 +2,7 @@
 <package>
   <metadata>
     <id>Vosk</id>
-    <version>0.3.41</version>
+    <version>0.3.43</version>
     <authors>Alpha Cephei Inc</authors>
     <owners>Alpha Cephei Inc</owners>
     <license type="expression">Apache-2.0</license>

go/README.md

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

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

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

java/demo/build.gradle

@@ -12,5 +12,5 @@ repositories {
 
 dependencies {
     implementation group: 'net.java.dev.jna', name: 'jna', version: '5.7.0'
-    implementation group: 'com.alphacephei', name: 'vosk', version: '0.3.41+'
+    implementation group: 'com.alphacephei', name: 'vosk', version: '0.3.43+'
 }

java/lib/build.gradle

@@ -16,7 +16,7 @@ repositories {
 
 archivesBaseName = 'vosk'
 group = 'com.alphacephei'
-version = '0.3.41'
+version = '0.3.43'
 
 mavenPublish {
     group = 'com.alphacephei'

nodejs/README.md

@@ -13,7 +13,8 @@ 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. More to come.
+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/package.json

@@ -1,6 +1,6 @@
 {
   "name": "vosk",
-  "version": "0.3.41",
+  "version": "0.3.43",
   "description": "Node binding for continuous offline voice recoginition with Vosk library.",
   "repository": {
     "type": "git",

python/README.md

@@ -4,7 +4,8 @@ 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. More to come.
+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_webvtt.py

@@ -62,6 +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/setup.py

@@ -32,6 +32,8 @@ else:
                 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="0.3.41",
+    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",
@@ -68,7 +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', 'requests', 'tqdm'],
-    install_requires=['cffi>=1.0', 'requests', 'tqdm'],
+    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

@@ -86,7 +86,7 @@ class Model(object):
             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(f"vosk-model(-small)?-{lang}", model)]
+            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)
@@ -98,8 +98,8 @@ class Model(object):
             return Path(directory, result_model[0])
 
     def download_model(self, model_name):
-        if not MODEL_DIRS[3].exists():
-            MODEL_DIRS[3].mkdir()
+        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)

python/vosk/transcriber/cli.py

@@ -2,6 +2,7 @@
 
 import logging
 import argparse
+import os
 
 from pathlib import Path
 from vosk import list_models, list_languages
@@ -12,6 +13,9 @@ parser = argparse.ArgumentParser(
 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')
@@ -33,6 +37,9 @@ parser.add_argument(
 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')
@@ -52,27 +59,27 @@ def main():
         return
 
     if not args.input:
-        logging.info('Please specify input file or directory')
+        logging.info("Please specify input file or directory")
         exit(1)
 
     if not Path(args.input).exists():
-        logging.info('File %s does not exist, please specify an existing file/directory' % (args.input))
-        exit(1)
-
-    if args.output !='' and not Path(args.output).exists():
-        logging.info('Output %s does not exist, please specify an existing file' % (args.output))
+        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() and Path(args.output).is_dir():
-        transcriber.process_dir(args)
-        return
-    elif Path(args.input).is_file() and (args.output=='' or Path(args.output).is_file()):
-        transcriber.process_file(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, input and output must be same type')
+        logging.info("Wrong arguments")
         exit(1)
 
+    transcriber.process_task_list(args, task_list)    
+
 if __name__ == "__main__":
     main()

python/vosk/transcriber/transcriber.py

@@ -4,10 +4,17 @@ 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:
 
@@ -18,24 +25,62 @@ class Transcriber:
     def recognize_stream(self, rec, stream):
         tot_samples = 0
         result = []
+
         while True:
-            data = stream.stdout.read(4000)
+            data = stream.stdout.read(CHUNK_SIZE)
+
             if len(data) == 0:
                 break
+
+            tot_samples += len(data)
             if rec.AcceptWaveform(data):
-                tot_samples += len(data)
-                result.append(json.loads(rec.Result())) 
-        result.append(json.loads(rec.FinalResult()))
+                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),
@@ -44,46 +89,84 @@ class Transcriber:
                             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):
-        stream = subprocess.Popen(
-            ['ffmpeg', '-nostdin', '-loglevel', 'quiet', '-i', 
-            infile, 
-            '-ar', '16000','-ac', '1', '-f', 's16le', '-'], 
-            stdout=subprocess.PIPE)
+        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)
 
-    def process_entry(self, inputdata):
-        logging.info(f'Recognizing {inputdata[0]}')
+    async def server_worker(self):
+        while True:
+            try:
+                input_file, output_file = self.queue.get_nowait()
+            except:
+                break
 
-        rec = KaldiRecognizer(self.model, 16000)
+            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)
-
-        stream = self.resample_ffmpeg(inputdata[0])
         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)
-        return final_result, tot_samples
 
-
-    def process_directory(self,args):
-        task_list = [(Path(args.input, fn), Path(args.output, Path(fn).stem).with_suffix('.' + args.output_type)) for fn in os.listdir(args.input)]
-        with Pool() as pool:
-            pool.map(self.process_entry, file_list)
-
-    def process_file(self, args):
-        start_time = timer()
-        final_result, tot_samples = self.process_entry([args.input, args.output])
         elapsed = timer() - start_time
-        logging.info(f'''Execution time: {elapsed:.3f} sec; xRT: {format(tot_samples / 16000.0 / float(elapsed), '.3f')}''')
+        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/recognizer.cc

@@ -418,14 +418,23 @@ bool Recognizer::GetSpkVector(Vector<BaseFloat> &out_xvector, int *num_spk_frame
     return true;
 }
 
+// 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);
@@ -739,7 +748,11 @@ const char* Recognizer::PartialResult()
         CompactLattice aligned_lat;
 
         clat = decoder_->GetLattice(decoder_->NumFramesInLattice(), false);
-        WordAlignLatticePartial(clat, *model_->trans_model_, *model_->winfo_, 0, &aligned_lat);
+        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();

travis/Dockerfile.dockcross

@@ -29,7 +29,7 @@ RUN cd /opt \
     && git clone -b v3.2.1  --single-branch https://github.com/alphacep/clapack \
     && echo ${OPENBLAS_ARGS} \
     && make -C OpenBLAS ONLY_CBLAS=1 ${OPENBLAS_ARGS} HOSTCC=gcc USE_LOCKING=1 USE_THREAD=0 all \
-    && make -C OpenBLAS PREFIX=$(pwd)/OpenBLAS/install install \
+    && 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 \

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

@@ -9,6 +9,7 @@ RUN yum -y update && yum -y install \
     autoconf \
     libtool \
     cmake \
+    libffi-devel \
     && yum clean all
 
 RUN cd /opt \

travis/build-dockcross-manylinux.sh

@@ -0,0 +1,7 @@
+#!/bin/bash
+
+set -e
+set -x
+
+docker build --build-arg="DOCKCROSS_IMAGE=dockcross/manylinux2014-aarch64" --build-arg="OPENBLAS_ARGS=TARGET=ARMV8" --file Dockerfile.dockcross-manylinux --tag alphacep/kaldi-dockcross-aarch64:latest .
+docker run --rm -v /home/shmyrev/travis/vosk-api/:/io alphacep/kaldi-dockcross-aarch64 /io/travis/build-wheels-dockcross.sh

travis/build-dockcross.sh

@@ -4,9 +4,13 @@ set -e
 set -x
 
 docker build --build-arg="DOCKCROSS_IMAGE=alphacep/dockcross-linux-armv7" --build-arg="OPENBLAS_ARGS=TARGET=ARMV7" --file Dockerfile.dockcross --tag alphacep/kaldi-dockcross-armv7:latest .
-docker build --build-arg="DOCKCROSS_IMAGE=dockcross/linux-arm64" --build-arg="OPENBLAS_ARGS=TARGET=ARMV8" --file Dockerfile.dockcross --tag alphacep/kaldi-dockcross-arm64:latest .
 docker build --build-arg="DOCKCROSS_IMAGE=dockcross/linux-x86" --build-arg="OPENBLAS_ARGS=TARGET=CORE2\ DYNAMIC_ARCH=1" --file Dockerfile.dockcross --tag alphacep/kaldi-dockcross-x86:latest .
+docker build --build-arg="DOCKCROSS_IMAGE=dockcross/linux-riscv64" --build-arg="OPENBLAS_ARGS=TARGET=RISCV64_GENERIC\ ARCH=riscv64" --file Dockerfile.dockcross --tag alphacep/kaldi-dockcross-riscv:latest .
 
 docker run --rm -v /home/shmyrev/travis/vosk-api/:/io alphacep/kaldi-dockcross-armv7 /io/travis/build-wheels-dockcross.sh
-docker run --rm -v /home/shmyrev/travis/vosk-api/:/io alphacep/kaldi-dockcross-arm64 /io/travis/build-wheels-dockcross.sh
 docker run --rm -v /home/shmyrev/travis/vosk-api/:/io alphacep/kaldi-dockcross-x86 /io/travis/build-wheels-dockcross.sh
+docker run --rm -v /home/shmyrev/travis/vosk-api/:/io alphacep/kaldi-dockcross-riscv /io/travis/build-wheels-dockcross.sh
+
+# We use manylinux (Centos-based image) for aarch64 instead
+# docker build --build-arg="DOCKCROSS_IMAGE=dockcross/linux-arm64" --build-arg="OPENBLAS_ARGS=TARGET=ARMV8" --file Dockerfile.dockcross --tag alphacep/kaldi-dockcross-arm64:latest .
+# docker run --rm -v /home/shmyrev/travis/vosk-api/:/io alphacep/kaldi-dockcross-arm64 /io/travis/build-wheels-dockcross.sh

travis/build-wheels-dockcross.sh

@@ -19,6 +19,9 @@ case $CROSS_TRIPLE in
     *i686-*)
         export VOSK_ARCHITECTURE=x86
         ;;
+    *riscv64-*)
+        export VOSK_ARCHITECTURE=riscv64
+        ;;
 esac
 
 # Copy library to output folder
@@ -26,5 +29,5 @@ mkdir -p /io/wheelhouse/vosk-linux-$VOSK_ARCHITECTURE
 cp /opt/vosk-api/src/*.so /opt/vosk-api/src/vosk_api.h /io/wheelhouse/vosk-linux-$VOSK_ARCHITECTURE
 
 # Build wheel
-python3 -m pip install requests tqdm
+python3 -m pip install requests tqdm srt wheel
 python3 -m pip wheel /opt/vosk-api/python --no-deps -w /io/wheelhouse

webjs/package.json

@@ -1,6 +1,6 @@
 {
   "name": "vosk-js",
-  "version": "0.3.41",
+  "version": "0.3.43",
   "description": "Node binding for continuous voice recoginition through vosk-api.",
   "repository": {
     "type": "git",