Bump JNA requirements to 5.18.1 and update version

Co-authored-by: ajaso <github@ajaso.net>

CMakeLists.txt

@@ -10,6 +10,7 @@ add_library(vosk
   src/recognizer.cc
   src/spk_model.cc
   src/vosk_api.cc
+  src/postprocessor.cc
 )
 
 find_package(kaldi REQUIRED)

android/build.gradle

@@ -4,13 +4,13 @@ buildscript {
         mavenCentral()
     }
     dependencies {
-        classpath 'com.android.tools.build:gradle:7.4.0'
-        classpath 'com.vanniktech:gradle-maven-publish-plugin:0.24.0'
+        classpath 'com.android.tools.build:gradle:8.13.0'
+        classpath 'com.vanniktech:gradle-maven-publish-plugin:0.34.0'
     }
 }
 
 allprojects {
-    version = '0.3.47'
+    version = '0.3.75'
 }
 
 subprojects {
@@ -24,7 +24,7 @@ subprojects {
     }
 
     mavenPublishing {
-        publishToMavenCentral(com.vanniktech.maven.publish.SonatypeHost.S01, false)
+        publishToMavenCentral()
         signAllPublications()
     }
 

android/lib/build-vosk.sh

@@ -29,7 +29,7 @@ set -x
 OS_NAME=`echo $(uname -s) | tr '[:upper:]' '[:lower:]'`
 ANDROID_TOOLCHAIN_PATH=$ANDROID_NDK_HOME/toolchains/llvm/prebuilt/${OS_NAME}-x86_64
 WORKDIR_BASE=`pwd`/build
-PATH=$ANDROID_NDK_HOME/toolchains/llvm/prebuilt/${OS_NAME}-x86_64/bin:$PATH
+PATH=$ANDROID_TOOLCHAIN_PATH/bin:$PATH
 OPENFST_VERSION=1.8.0
 
 for arch in armeabi-v7a arm64-v8a x86_64 x86; do
@@ -45,6 +45,7 @@ case $arch in
           CC=armv7a-linux-androideabi21-clang
           CXX=armv7a-linux-androideabi21-clang++
           ARCHFLAGS="-mfloat-abi=softfp -mfpu=neon"
+          PAGESIZE_LDFLAGS=""
           ;;
     arm64-v8a)
           BLAS_ARCH=ARMV8
@@ -54,6 +55,8 @@ case $arch in
           CC=aarch64-linux-android21-clang
           CXX=aarch64-linux-android21-clang++
           ARCHFLAGS=""
+          # Ensure compatibility with 16KiB page size devices
+          PAGESIZE_LDFLAGS="-Wl,-z,common-page-size=4096 -Wl,-z,max-page-size=16384"
           ;;
     x86_64)
           BLAS_ARCH=ATOM
@@ -63,6 +66,7 @@ case $arch in
           CC=x86_64-linux-android21-clang
           CXX=x86_64-linux-android21-clang++
           ARCHFLAGS=""
+          PAGESIZE_LDFLAGS=""
           ;;
     x86)
           BLAS_ARCH=ATOM
@@ -72,6 +76,7 @@ case $arch in
           CC=i686-linux-android21-clang
           CXX=i686-linux-android21-clang++
           ARCHFLAGS=""
+          PAGESIZE_LDFLAGS=""
           ;;
 esac
 
@@ -79,16 +84,16 @@ mkdir -p $WORKDIR/local/lib
 
 # openblas first
 cd $WORKDIR
-git clone -b v0.3.13 --single-branch https://github.com/xianyi/OpenBLAS
-make -C OpenBLAS TARGET=$BLAS_ARCH ONLY_CBLAS=1 AR=$AR CC=$CC HOSTCC=gcc ARM_SOFTFP_ABI=1 USE_THREAD=0 NUM_THREADS=1 -j4
+git clone -b v0.3.20 --single-branch https://github.com/xianyi/OpenBLAS
+make -C OpenBLAS TARGET=$BLAS_ARCH ONLY_CBLAS=1 AR=$AR CC=$CC HOSTCC=gcc ARM_SOFTFP_ABI=1 USE_THREAD=0 NUM_THREADS=1 -j 8
 make -C OpenBLAS install PREFIX=$WORKDIR/local
 
 # CLAPACK
 cd $WORKDIR
 git clone -b v3.2.1  --single-branch https://github.com/alphacep/clapack
 mkdir -p clapack/BUILD && cd clapack/BUILD
-cmake -DCMAKE_C_FLAGS=$ARCHFLAGS -DCMAKE_C_COMPILER_TARGET=$HOST \
-    -DCMAKE_C_COMPILER=$CC -DCMAKE_SYSTEM_NAME=Generic -DCMAKE_AR=$ANDROID_NDK_HOME/toolchains/llvm/prebuilt/${OS_NAME}-x86_64/bin/$AR \
+cmake -DCMAKE_C_FLAGS="$ARCHFLAGS" -DCMAKE_C_COMPILER_TARGET=$HOST \
+    -DCMAKE_C_COMPILER=$CC -DCMAKE_SYSTEM_NAME=Generic -DCMAKE_AR=$ANDROID_TOOLCHAIN_PATH/bin/$AR \
     -DCMAKE_TRY_COMPILE_TARGET_TYPE=STATIC_LIBRARY \
     -DCMAKE_CROSSCOMPILING=True ..
 make -j 8 -C F2CLIBS/libf2c
@@ -118,7 +123,7 @@ CXX=$CXX AR=$AR RANLIB=$RANLIB CXXFLAGS="$ARCHFLAGS -O3 -DFST_NO_DYNAMIC_LINKING
     --fst-root=${WORKDIR}/local --fst-version=${OPENFST_VERSION}
 make -j 8 depend
 cd $WORKDIR/kaldi/src
-make -j 8 online2 lm rnnlm
+make -j 8 online2 rnnlm
 
 # Vosk-api
 cd $WORKDIR
@@ -129,7 +134,7 @@ make -j 8 -C ${WORKDIR_BASE}/../../../src \
     OPENFST_ROOT=${WORKDIR}/local \
     OPENBLAS_ROOT=${WORKDIR}/local \
     CXX=$CXX \
-    EXTRA_LDFLAGS="-llog -static-libstdc++ -Wl,-soname,libvosk.so"
+    EXTRA_LDFLAGS="-llog -static-libstdc++ -Wl,-soname,libvosk.so ${PAGESIZE_LDFLAGS}"
 cp $WORKDIR/vosk/libvosk.so $WORKDIR/../../src/main/jniLibs/$arch/libvosk.so
 
 done

android/lib/build.gradle

@@ -4,14 +4,14 @@ def pomDescription = "Vosk speech recognition library for Android"
 
 android {
     namespace 'org.vosk'
-    compileSdkVersion 33
+    compileSdkVersion 36
     defaultConfig {
         minSdkVersion 21
-        targetSdkVersion 33
+        targetSdkVersion 36
         versionCode 10
         versionName = version
         archivesBaseName = archiveName
-        ndkVersion = "25.2.9519653"
+        ndkVersion = "28.2.13676358"
     }
     compileOptions {
         sourceCompatibility JavaVersion.VERSION_1_8
@@ -25,7 +25,7 @@ task buildVosk(type: Exec) {
 }
 
 dependencies {
-    api 'net.java.dev.jna:jna:5.13.0@aar'
+    api 'net.java.dev.jna:jna:5.18.1@aar'
 }
 
 //preBuild.dependsOn buildVosk

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

@@ -56,8 +56,18 @@ public class LibVosk {
 
     public static native void vosk_recognizer_reset(Pointer recognizer);
 
+    public static native void vosk_recognizer_set_endpointer_mode(Pointer recognizer, int mode);
+
+    public static native void vosk_recognizer_set_endpointer_delays(Pointer recognizer, float t_start_max, float t_end, float t_max);
+
     public static native void vosk_recognizer_free(Pointer recognizer);
 
+    public static native Pointer vosk_text_processor_new(String verbalizer, String tagger);
+
+    public static native void vosk_text_processor_free(Pointer processor);
+
+    public static native String vosk_text_processor_itn(Pointer processor, String input);
+
     /**
      * Set log level for Kaldi messages.
      *

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

@@ -236,6 +236,34 @@ public class Recognizer extends PointerType implements AutoCloseable {
         LibVosk.vosk_recognizer_reset(this.getPointer());
     }
 
+    /**
+     * Endpointer delay mode
+     */
+    public class EndpointerMode {
+        public static final int DEFAULT = 0;
+        public static final int SHORT = 1;
+        public static final int LONG = 2;
+        public static final int VERY_LONG = 3;
+    }
+
+    /**
+     * Configures endpointer mode for recognizer
+     */
+    public void setEndpointerMode(int mode) {
+        LibVosk.vosk_recognizer_set_endpointer_mode(this.getPointer(), mode);
+    }
+
+    /**
+     * Set endpointer delays
+     *
+     * @param t_start_max     timeout for stopping recognition in case of initial silence (usually around 5.0)
+     * @param t_end           timeout for stopping recognition in milliseconds after we recognized something (usually around 0.5 - 1.0)
+     * @param t_max           timeout for forcing utterance end in milliseconds (usually around 20-30)
+     **/
+    public void setEndpointerDelays(float t_start_max, float t_end, float t_max) {
+        LibVosk.vosk_recognizer_set_endpointer_delays(this.getPointer(), t_start_max, t_end, t_max);
+    }
+
     /**
      * Releases recognizer object.
      * Underlying model is also unreferenced and if needed, released.

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

@@ -0,0 +1,21 @@
+package org.vosk;
+
+import com.sun.jna.PointerType;
+
+public class TextProcessor extends PointerType implements AutoCloseable {
+    public TextProcessor() {
+    }
+
+    public TextProcessor(String verbalizer, String tagger) {
+        super(LibVosk.vosk_text_processor_new(verbalizer, tagger));
+    }
+
+    @Override
+    public void close() {
+        LibVosk.vosk_text_processor_free(this.getPointer());
+    }
+
+    public String itn(String input) {
+        return LibVosk.vosk_text_processor_itn(this.getPointer(), input);
+    }
+}

android/model-en/build.gradle

@@ -4,10 +4,10 @@ def pomDescription = "Small English model for Android"
 
 android {
     namespace "org.vosk"
-    compileSdkVersion 33
+    compileSdkVersion 36
     defaultConfig {
         minSdkVersion 21
-        targetSdkVersion 33
+        targetSdkVersion 36
         versionCode 10
         versionName = version
         archivesBaseName = archiveName

csharp/demo/VoskDemo.cs

@@ -28,6 +28,9 @@ public class VoskDemo
    {
         // Demo float array
         VoskRecognizer rec = new VoskRecognizer(model, 16000.0f);
+
+        rec.SetEndpointerMode(EndpointerMode.LONG);
+
         using(Stream source = File.OpenRead("test.wav")) {
             byte[] buffer = new byte[4096];
             int bytesRead;

csharp/demo/VoskDemo.csproj

@@ -2,7 +2,7 @@
 
   <PropertyGroup>
     <OutputType>Exe</OutputType>
-    <TargetFramework>net5.0</TargetFramework>
+    <TargetFramework>net8.0</TargetFramework>
     <RootNamespace>VoskDemo</RootNamespace>
   </PropertyGroup>
 
@@ -11,7 +11,7 @@
   </PropertyGroup>
 
   <ItemGroup>
-    <PackageReference Include="Vosk" Version="0.3.45" />
+    <PackageReference Include="Vosk" Version="0.3.75" />
   </ItemGroup>
 
 </Project>

csharp/nuget/Vosk.csproj

@@ -0,0 +1,17 @@
+<Project Sdk="Microsoft.NET.Sdk">
+
+  <PropertyGroup>
+    <TargetFramework>net8.0</TargetFramework>
+    <ImplicitUsings>enable</ImplicitUsings>
+    <Nullable>enable</Nullable>
+    <PackageId>Vosk</PackageId>
+    <Version>0.3.75</Version>
+    <authors>Alpha Cephei Inc</authors>
+    <owners>Alpha Cephei Inc</owners>
+  </PropertyGroup>
+
+  <Target Name="CopyFiles" AfterTargets="Build">
+    <Copy SourceFiles="bin/Release/net8.0/Vosk.dll" DestinationFolder="lib/net8.0" />
+  </Target>
+
+</Project>

csharp/nuget/Vosk.nuspec

@@ -2,7 +2,7 @@
 <package>
   <metadata>
     <id>Vosk</id>
-    <version>0.3.45</version>
+    <version>0.3.75</version>
     <authors>Alpha Cephei Inc</authors>
     <owners>Alpha Cephei Inc</owners>
     <license type="expression">Apache-2.0</license>
@@ -23,10 +23,10 @@ Vosk scales from small devices like Raspberry Pi or Android smartphone to big cl
     <copyright>Copyright 2020-2050 Alpha Cephei Inc</copyright>
     <tags>speech recognition voice stt asr speech-to-text ai offline privacy</tags>
     <dependencies>
-      <group targetFramework=".NETStandard2.0"/>
+      <group targetFramework="net8.0"/>
     </dependencies>
   </metadata>
   <files>
-    <file src="**" exclude="src/*.cs;build.sh;**/.keep-me;*.nupkg" />
+    <file src="**" exclude="bin/**;obj/**;build.sh;src/*.cs;*.nupkg;**/.keep-me" />
   </files>
 </package>

csharp/nuget/build.sh

@@ -1,2 +1,2 @@
-mcs -out:lib/netstandard2.0/Vosk.dll -target:library src/*.cs
-nuget pack
+rm -rf bin lib obj
+/home/shmyrev/local/dotnet/dotnet pack Vosk.csproj -p:NuspecFile=Vosk.nuspec -o .

csharp/nuget/src/VoskPINVOKE.cs

@@ -65,6 +65,12 @@ class VoskPINVOKE {
   [global::System.Runtime.InteropServices.DllImport("libvosk", EntryPoint="vosk_recognizer_reset")]
   public static extern void VoskRecognizer_Reset(global::System.Runtime.InteropServices.HandleRef jarg1);
 
+  [global::System.Runtime.InteropServices.DllImport("libvosk", EntryPoint="vosk_recognizer_set_endpointer_mode")]
+  public static extern void VoskRecognizer_SetEndpointerMode(global::System.Runtime.InteropServices.HandleRef jarg1, int jarg2);
+
+  [global::System.Runtime.InteropServices.DllImport("libvosk", EntryPoint="vosk_recognizer_set_endpointer_delays")]
+  public static extern void VoskRecognizer_SetEndpointerDelays(global::System.Runtime.InteropServices.HandleRef jarg1, float jarg2, float jarg3, float jarg4);
+
   [global::System.Runtime.InteropServices.DllImport("libvosk", EntryPoint="vosk_set_log_level")]
   public static extern void SetLogLevel(int jarg1);
 
@@ -107,7 +113,6 @@ class VoskPINVOKE {
   [global::System.Runtime.InteropServices.DllImport("libvosk", EntryPoint = "vosk_batch_recognizer_get_pending_chunks")]
   public static extern int VoskBatchRecognizer_GetPendingChunks(global::System.Runtime.InteropServices.HandleRef jarg1);
 
-
-    }
+}
 
 }

csharp/nuget/src/VoskRecognizer.cs

@@ -1,5 +1,12 @@
 namespace Vosk {
 
+public enum EndpointerMode {
+    DEFAULT = 0,
+    SHORT = 1,
+    LONG = 2,
+    VERY_LONG = 3
+}
+
 public class VoskRecognizer : System.IDisposable {
   private System.Runtime.InteropServices.HandleRef handle;
 
@@ -91,6 +98,14 @@ public class VoskRecognizer : System.IDisposable {
     VoskPINVOKE.VoskRecognizer_Reset(handle);
   }
 
+  public void SetEndpointerMode(EndpointerMode mode) {
+    VoskPINVOKE.VoskRecognizer_SetEndpointerMode(handle, (int) mode);
+  }
+
+  public void SetEndpointerDelays(float t_start_max, float t_end, float t_max) {
+    VoskPINVOKE.VoskRecognizer_SetEndpointerDelays(handle, t_start_max, t_end, t_max);
+  }
+
 }
 
 }

go/batch.go

@@ -0,0 +1,99 @@
+package vosk
+
+// #cgo CPPFLAGS: -I ${SRCDIR}/../src
+// #cgo !windows LDFLAGS: -L ${SRCDIR}/../src -lvosk -ldl -lpthread
+// #cgo windows LDFLAGS: -L ${SRCDIR}/../src -lvosk -lpthread
+// #include <stdlib.h>
+// #include <vosk_api.h>
+import "C"
+import "unsafe"
+
+// VoskBatchModel contains a reference to the C VoskBatchModel
+type VoskBatchModel struct {
+	model *C.struct_VoskBatchModel
+}
+
+// NewBatchModel creates a new VoskBatchModel instance
+func NewBatchModel(modelPath string) (*VoskBatchModel, error) {
+	cmodelPath := C.CString(modelPath)
+	defer C.free(unsafe.Pointer(cmodelPath))
+	internal := C.vosk_batch_model_new(cmodelPath)
+	model := &VoskBatchModel{model: internal}
+	return model, nil
+}
+
+func (m *VoskBatchModel) Free() {
+	C.vosk_batch_model_free(m.model)
+}
+
+func (m *VoskBatchModel) Wait() {
+	C.vosk_batch_model_wait(m.model);
+}
+
+func freeBatchModel(model *VoskBatchModel) {
+	C.vosk_batch_model_free(model.model)
+}
+
+// VoskBatchRecognizer contains a reference to the C VoskBatchRecognizer
+type VoskBatchRecognizer struct {
+	rec *C.struct_VoskBatchRecognizer
+}
+
+func freeBatchRecognizer(recognizer *VoskBatchRecognizer) {
+	C.vosk_batch_recognizer_free(recognizer.rec)
+}
+
+func (r *VoskBatchRecognizer) Free() {
+	C.vosk_batch_recognizer_free(r.rec)
+}
+
+// NewBatchRecognizer creates a new VoskBatchRecognizer instance
+func NewBatchRecognizer(model *VoskBatchModel, sampleRate float64) (*VoskBatchRecognizer, error) {
+	internal := C.vosk_batch_recognizer_new(model.model, C.float(sampleRate))
+	rec := &VoskBatchRecognizer{rec: internal}
+	return rec, nil
+}
+
+// AcceptWaveform accepts and processes a new chunk of the voice data.
+func (r *VoskBatchRecognizer) AcceptWaveform(buffer []byte) {
+	cbuf := C.CBytes(buffer)
+	defer C.free(cbuf)
+	C.vosk_batch_recognizer_accept_waveform(r.rec, (*C.char)(cbuf), C.int(len(buffer)))
+}
+
+/** Set NLSML output
+ * @param nlsml - boolean value
+ */
+//void vosk_batch_recognizer_set_nlsml(VoskBatchRecognizer *recognizer, int nlsml);
+
+func (r *VoskBatchRecognizer) SetNlsml(nlsml int) {
+	C.vosk_batch_recognizer_set_nlsml(r.rec, C.int(nlsml))
+}
+
+/** Closes the stream */
+//void vosk_batch_recognizer_finish_stream(VoskBatchRecognizer *recognizer);
+
+func (r *VoskBatchRecognizer) FinishStream() {
+	C.vosk_batch_recognizer_finish_stream(r.rec)
+}
+
+/** Return results */
+//const char *vosk_batch_recognizer_front_result(VoskBatchRecognizer *recognizer);
+
+func (r *VoskBatchRecognizer) FrontResult() string {
+	return C.GoString(C.vosk_batch_recognizer_front_result(r.rec))
+}
+
+/** Release and free first retrieved result */
+//void vosk_batch_recognizer_pop(VoskBatchRecognizer *recognizer);
+
+func (r *VoskBatchRecognizer) Pop() {
+	C.vosk_batch_recognizer_pop(r.rec)
+}
+
+/** Get amount of pending chunks for more intelligent waiting */
+//int vosk_batch_recognizer_get_pending_chunks(VoskBatchRecognizer *recognizer);
+func (r *VoskBatchRecognizer) GetPendingChunks() int {
+	i := C.vosk_batch_recognizer_get_pending_chunks(r.rec)
+	return int(i)
+}

go/batch_example/README.md

@@ -0,0 +1,5 @@
+This example expects a `s16le` converted audio file and converts it to text in a
+manner that imitates the Python example of [test_gpu_batch.py](../python/example/test_gpu_batch.py).
+
+Note that the `libvosk.so` must be in the library path. This was successfully tested on
+Ubuntu 24.04 with Go 1.18, gcc-11, NVIDIA driver 570.172.08.

go/batch_example/test_batch.go

@@ -0,0 +1,54 @@
+package main
+
+import (
+	"flag"
+	"fmt"
+	"io"
+	"log"
+	"os"
+
+	vosk "github.com/alphacep/vosk-api/go"
+)
+
+func main() {
+	var filename string
+	flag.StringVar(&filename, "f", "", "file to transcribe")
+	flag.Parse()
+
+	vosk.GPUInit()
+
+	model, err := vosk.NewBatchModel("model")
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	rec, err := vosk.NewBatchRecognizer(model, 16000.0)
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	file, err := os.Open(filename)
+	if err != nil {
+		panic(err)
+	}
+	defer file.Close()
+
+	buf := make([]byte, 8000)
+
+	for {
+		if _, err := file.Read(buf); err != nil {
+			if err != io.EOF {
+				log.Fatal(err)
+			}
+
+			break
+		}
+		rec.AcceptWaveform(buf)
+		model.Wait()
+		if rec.FrontResult() != "" {
+			fmt.Println(rec.FrontResult())
+			rec.Pop()
+		}
+	}
+	// Is this needed? rec.FinishStream()
+}

go/example/test_simple.go

@@ -16,6 +16,8 @@ func main() {
 	flag.StringVar(&filename, "f", "", "file to transcribe")
 	flag.Parse()
 
+	vosk.GPUInit()
+
 	model, err := vosk.NewModel("model")
 	if err != nil {
 		log.Fatal(err)

go/vosk.go

@@ -125,6 +125,16 @@ func (r *VoskRecognizer) SetPartialWords(words int) {
 	C.vosk_recognizer_set_partial_words(r.rec, C.int(words))
 }
 
+// SetEndpointerDelays sets the recognition timeouts, where startMax
+// is the timeout for stopping recognition in case of initial silence
+// (usually around 5), end is the timeout for stopping recognition
+// in milliseconds after we recognized something (usually around 0.5-1.0),
+// and max is the timeout for forcing utterance end in milliseconds
+// (usually around 20-30).
+func (r *VoskRecognizer) SetEndpointerDelays(startMax, end, max float64) {
+	C.vosk_recognizer_set_endpointer_delays(r.rec, C.float(startMax), C.float(end), C.float(max))
+}
+
 // AcceptWaveform accepts and processes a new chunk of the voice data.
 func (r *VoskRecognizer) AcceptWaveform(buffer []byte) int {
 	cbuf := C.CBytes(buffer)

java/demo/build.gradle

@@ -11,5 +11,5 @@ repositories {
 }
 
 dependencies {
-    implementation group: 'com.alphacephei', name: 'vosk', version: '0.3.45'
+    implementation group: 'com.alphacephei', name: 'vosk', version: '0.3.75'
 }

java/lib/build.gradle

@@ -16,7 +16,7 @@ repositories {
 
 archivesBaseName = 'vosk'
 group = 'com.alphacephei'
-version = '0.3.45'
+version = '0.3.75'
 
 mavenPublish {
     group = 'com.alphacephei'
@@ -25,7 +25,7 @@ mavenPublish {
 }
 
 dependencies {
-    api group: 'net.java.dev.jna', name: 'jna', version: '5.13.0'
+    api group: 'net.java.dev.jna', name: 'jna', version: '5.18.1'
     testImplementation 'junit:junit:4.13'
 }
 

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

@@ -82,8 +82,18 @@ public class LibVosk {
 
     public static native void vosk_recognizer_reset(Pointer recognizer);
 
+    public static native void vosk_recognizer_set_endpointer_mode(Pointer recognizer, int mode);
+
+    public static native void vosk_recognizer_set_endpointer_delays(Pointer recognizer, float t_start_max, float t_end, float t_max);
+
     public static native void vosk_recognizer_free(Pointer recognizer);
 
+    public static native Pointer vosk_text_processor_new(String verbalizer, String tagger);
+
+    public static native void vosk_text_processor_free(Pointer processor);
+
+    public static native String vosk_text_processor_itn(Pointer processor, String input);
+
     /**
      * Set log level for Kaldi messages.
      *

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

@@ -236,6 +236,34 @@ public class Recognizer extends PointerType implements AutoCloseable {
         LibVosk.vosk_recognizer_reset(this.getPointer());
     }
 
+    /**
+     * Endpointer delay mode
+     */
+    public class EndpointerMode {
+        public static final int DEFAULT = 0;
+        public static final int SHORT = 1;
+        public static final int LONG = 2;
+        public static final int VERY_LONG = 3;
+    }
+
+    /**
+     * Configures endpointer mode for recognizer
+     */
+    public void setEndpointerMode(int mode) {
+        LibVosk.vosk_recognizer_set_endpointer_mode(this.getPointer(), mode);
+    }
+
+    /**
+     * Set endpointer delays
+     *
+     * @param t_start_max     timeout for stopping recognition in case of initial silence (usually around 5.0)
+     * @param t_end           timeout for stopping recognition in milliseconds after we recognized something (usually around 0.5 - 1.0)
+     * @param t_max           timeout for forcing utterance end in milliseconds (usually around 20-30)
+     **/
+    public void setEndpointerDelays(float t_start_max, float t_end, float t_max) {
+        LibVosk.vosk_recognizer_set_endpointer_delays(this.getPointer(), t_start_max, t_end, t_max);
+    }
+
     /**
      * Releases recognizer object.
      * Underlying model is also unreferenced and if needed, released.

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

@@ -15,8 +15,10 @@ import javax.sound.sampled.UnsupportedAudioFileException;
 
 import org.vosk.LogLevel;
 import org.vosk.Recognizer;
+import org.vosk.Recognizer.EndpointerMode;
 import org.vosk.LibVosk;
 import org.vosk.Model;
+import org.vosk.TextProcessor;
 
 public class DecoderTest {
 
@@ -95,9 +97,24 @@ public class DecoderTest {
         Assert.assertTrue(true);
     }
 
+    @Test
+    public void decoderEndpointerDelays() throws IOException, UnsupportedAudioFileException {
+        try (Model model = new Model("model");
+            Recognizer recognizer = new Recognizer(model, 16000)) {
+            recognizer.setEndpointerMode(EndpointerMode.VERY_LONG);
+            recognizer.setEndpointerDelays(5.0f, 3.0f, 50.0f);
+	}
+        Assert.assertTrue(true);
+    }
 
     @Test(expected = IOException.class)
     public void decoderTestException() throws IOException {
         Model model = new Model("model_missing");
     }
+
+    @Test
+    public void testItn() throws IOException {
+        TextProcessor p = new TextProcessor("model/itn/en_itn_tagger.fst", "model/itn/en_itn_verbalizer.fst");
+        System.out.println(p.itn("as easy as one two three"));
+    }
 }

kotlin/build.gradle.kts

@@ -1,5 +1,6 @@
 import org.jetbrains.dokka.gradle.DokkaTask
-import org.jetbrains.kotlin.config.JvmTarget
+import org.jetbrains.kotlin.gradle.ExperimentalKotlinGradlePluginApi
+import org.jetbrains.kotlin.gradle.dsl.JvmTarget
 
 /*
  * Copyright 2020 Alpha Cephei Inc. & Doomsdayrs
@@ -18,15 +19,15 @@ import org.jetbrains.kotlin.config.JvmTarget
  */
 
 plugins {
-	kotlin("multiplatform") version "1.8.10"
+	kotlin("multiplatform") version "2.0.0"
 	id("com.android.library")
 	`maven-publish`
-	id("org.jetbrains.dokka") version "1.7.20"
-	kotlin("plugin.serialization") version "1.8.10"
+	id("org.jetbrains.dokka") version "1.9.20"
+	kotlin("plugin.serialization") version "2.0.0"
 }
 
 group = "com.alphacephei"
-version = "0.4.0-alpha0"
+version = "0.3.75"
 
 repositories {
 	google()
@@ -67,9 +68,11 @@ fun org.jetbrains.kotlin.gradle.dsl.KotlinMultiplatformExtension.native(
 
 kotlin {
 	jvm {
-		compilations.all {
-			kotlinOptions.jvmTarget = JvmTarget.JVM_11.description
+		@OptIn(ExperimentalKotlinGradlePluginApi::class)
+		compilerOptions {
+			jvmTarget.set(JvmTarget.JVM_17)
 		}
+
 		testRuns["test"].executionTask.configure {
 			useJUnitPlatform()
 			environment("MODEL", "VOSK_MODEL")
@@ -80,7 +83,7 @@ kotlin {
 		}
 	}
 
-	android {
+	androidTarget {
 		publishAllLibraryVariants()
 	}
 
@@ -102,6 +105,16 @@ kotlin {
 			}
 		}
 
+
+	@OptIn(ExperimentalKotlinGradlePluginApi::class)
+	applyDefaultHierarchyTemplate {
+		withJvm()
+		withAndroidTarget()
+
+		if (enableNative)
+			withNative()
+	}
+
 	publishing {
 		publications {
 			withType<MavenPublication> {
@@ -130,13 +143,13 @@ kotlin {
 		}
 	}
 
-	val jna_version = "5.13.0"
-	val coroutines_version = "1.6.4"
+	val jna_version = "5.14.0"
+	val coroutines_version = "1.7.3"
 
 	sourceSets {
 		val commonMain by getting {
 			dependencies {
-				api("org.jetbrains.kotlinx:kotlinx-serialization-json:1.4.1")
+				api("org.jetbrains.kotlinx:kotlinx-serialization-json:1.7.0")
 				api("org.jetbrains.kotlinx:kotlinx-coroutines-core:$coroutines_version")
 			}
 		}
@@ -161,7 +174,7 @@ kotlin {
 				api("net.java.dev.jna:jna:$jna_version@aar")
 			}
 		}
-		val androidTest by getting {
+		val androidUnitTest by getting {
 			dependencies {
 				implementation("junit:junit:4.13.2")
 			}
@@ -170,15 +183,16 @@ kotlin {
 }
 
 android {
-	compileSdk = 33
+	namespace = "com.alphacephei.library"
+	compileSdk = 34
 	sourceSets["main"].manifest.srcFile("src/androidMain/AndroidManifest.xml")
 	defaultConfig {
 		minSdk = 24
-		targetSdk = 33
+		targetSdk = 34
 	}
 	compileOptions {
-		sourceCompatibility = JavaVersion.VERSION_11
-		targetCompatibility = JavaVersion.VERSION_11
+		sourceCompatibility = JavaVersion.VERSION_17
+		targetCompatibility = JavaVersion.VERSION_17
 	}
 	publishing {
 		multipleVariants {

kotlin/settings.gradle.kts

@@ -23,7 +23,7 @@ pluginManagement {
 	resolutionStrategy {
 		eachPlugin {
 			if (requested.id.namespace == "com.android") {
-				useModule("com.android.tools.build:gradle:7.3.0")
+				useModule("com.android.tools.build:gradle:8.3.0")
 			}
 		}
 	}

kotlin/src/commonMain/kotlin/org/vosk/EndPointerMode.kt

@@ -0,0 +1,27 @@
+/*
+ * Copyright 2024 Alpha Cephei Inc.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.vosk
+
+/**
+ * VoskEpMode
+ */
+enum class EndPointerMode {
+	ANSWER_DEFAULT,
+	ANSWER_SHORT,
+	ANSWER_LONG,
+	ANSWER_VERY_LONG
+}

kotlin/src/commonMain/kotlin/org/vosk/Recognizer.kt

@@ -151,6 +151,7 @@ expect class Recognizer : Freeable {
 	 *     }],
 	 * </pre>
 	 *
+	 * C equivalent = vosk_recognizer_set_words
 	 * @param words - boolean value
 	 */
 	fun setOutputWordTimes(words: Boolean)
@@ -168,6 +169,23 @@ expect class Recognizer : Freeable {
 	 */
 	fun setNLSML(nlsml: Boolean)
 
+
+	/**
+	 * Set endpointer scaling factor
+	 *
+	 * @param mode Endpointer mode
+	 **/
+	fun setEndPointerMode(mode: EndPointerMode)
+
+	/**
+	 * Set endpointer delays
+	 *
+	 * @param tStartMax timeout for stopping recognition in case of initial silence (usually around 5.0)
+	 * @param tEnd      timeout for stopping recognition in milliseconds after we recognized something (usually around 0.5 - 1.0)
+	 * @param tMax      timeout for forcing utterance end in milliseconds (usually around 20-30)
+	 **/
+	fun setEndPointerDelays(tStartMax: Float, tEnd: Float, tMax: Float)
+
 	/**
 	 * Accept voice data
 	 *

kotlin/src/commonMain/kotlin/org/vosk/TextProcessor.kt

@@ -0,0 +1,32 @@
+/*
+ * Copyright 2024 Alpha Cephei Inc.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.vosk
+
+/**
+ * Inverse text normalization
+ *
+ * @since 2024/06/19
+ * @constructor Create text processor
+ */
+expect class TextProcessor constructor(tagger: Char, verbalizer: Char) : Freeable {
+
+	/** Release text processor */
+	override fun free()
+
+	/** Convert string */
+	fun itn(input: Char): Char
+}

kotlin/src/commonMain/kotlin/org/vosk/exception/IOException.kt

@@ -19,4 +19,4 @@ package org.vosk.exception
 /**
  * Internal common IO exception. On JVM this is just a type alias.
  */
-expect open class IOException(message: String? = null) : Exception
+expect open class IOException(message: String?) : Exception

kotlin/src/jvmMain/kotlin/org/vosk/LibVosk.kt

@@ -42,14 +42,13 @@ internal object LibVosk {
 
 	@Throws(IOException::class)
 	private fun unpackDll(targetDir: File, lib: String) {
-		val source: InputStream =
-			Vosk::class.java.getResourceAsStream("/win32-x86-64/$lib.dll")!!
-
-		Files.copy(
-			source,
-			File(targetDir, "$lib.dll").toPath(),
-			StandardCopyOption.REPLACE_EXISTING
-		)
+		Vosk::class.java.getResourceAsStream("/win32-x86-64/$lib.dll")!!.use {
+			Files.copy(
+				it,
+				File(targetDir, "$lib.dll").toPath(),
+				StandardCopyOption.REPLACE_EXISTING
+			)
+		}
 	}
 
 	init {
@@ -57,6 +56,7 @@ internal object LibVosk {
 			Platform.isAndroid() -> {
 				Native.register(LibVosk::class.java, "vosk")
 			}
+
 			Platform.isWindows() -> {
 				// We have to unpack dependencies
 				try {
@@ -79,6 +79,7 @@ internal object LibVosk {
 					Native.register(LibVosk::class.java, "libvosk");
 				}
 			}
+
 			else -> {
 				Native.register(LibVosk::class.java, "vosk");
 			}
@@ -194,4 +195,19 @@ internal object LibVosk {
 	external fun vosk_batch_recognizer_pop(recognizer: BatchRecognizer)
 
 	external fun vosk_batch_recognizer_get_pending_chunks(recognizer: BatchRecognizer): Int
+
+	external fun vosk_text_processor_new(tagger: Char, verbalizer: Char): Pointer
+
+	external fun vosk_text_processor_free(processor: TextProcessor)
+
+	external fun vosk_text_processor_itn(processor: TextProcessor, input: Char): Char
+
+	external fun vosk_recognizer_set_endpointer_mode(recognizer: Recognizer, ordinal: Int)
+
+	external fun vosk_recognizer_set_endpointer_delays(
+		recognizer: Recognizer,
+		tStartMax: Float,
+		tEnd: Float,
+		tMax: Float
+	)
 }

kotlin/src/jvmMain/kotlin/org/vosk/Recognizer.kt

@@ -327,4 +327,27 @@ actual class Recognizer : Freeable, PointerType, AutoCloseable {
 		free()
 	}
 
+	/**
+	 * Set endpointer scaling factor
+	 *
+	 * @param mode Endpointer mode
+	 **/
+	actual fun setEndPointerMode(mode: EndPointerMode) {
+		LibVosk.vosk_recognizer_set_endpointer_mode(this, mode.ordinal)
+	}
+
+	/**
+	 * Set endpointer delays
+	 *
+	 * @param tStartMax timeout for stopping recognition in case of initial silence (usually around 5.0)
+	 * @param tEnd      timeout for stopping recognition in milliseconds after we recognized something (usually around 0.5 - 1.0)
+	 * @param tMax      timeout for forcing utterance end in milliseconds (usually around 20-30)
+	 **/
+	actual fun setEndPointerDelays(
+		tStartMax: Float,
+		tEnd: Float,
+		tMax: Float
+	) {
+		LibVosk.vosk_recognizer_set_endpointer_delays(this, tStartMax, tEnd, tMax)
+	}
 }

kotlin/src/jvmMain/kotlin/org/vosk/TextProcessor.kt

@@ -0,0 +1,51 @@
+/*
+ * Copyright 2024 Alpha Cephei Inc.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.vosk
+
+import com.sun.jna.PointerType
+
+/**
+ * Inverse text normalization
+ *
+ * @since 2024/06/19
+ */
+actual class TextProcessor :
+	Freeable, PointerType, AutoCloseable {
+
+	/**
+	 * Create text processor
+	 */
+	actual constructor(tagger: Char, verbalizer: Char) :
+			super(LibVosk.vosk_text_processor_new(tagger, verbalizer))
+
+
+	/** Release text processor */
+	actual override fun free() {
+		LibVosk.vosk_text_processor_free(this)
+	}
+
+	/** Convert string */
+	actual fun itn(input: Char): Char =
+		LibVosk.vosk_text_processor_itn(this, input)
+
+	/**
+	 * @see free
+	 */
+	override fun close() {
+		free()
+	}
+}

kotlin/src/nativeInterop/cinterop/libvosk.def

@@ -22,4 +22,4 @@ linkerOpts.linux = \
 
 linkerOpts.linux_x64 = \
 	-L/usr/lib64/ \
-	-L/usr/local/lib64/
+	-L/usr/local/lib64/

nodejs/index.js

@@ -69,16 +69,18 @@ const vosk_recognizer_ptr = ref.refType(vosk_recognizer);
 
 let soname;
 if (os.platform() == 'win32') {
-    // Update path to load dependent dlls
-    let currentPath = process.env.Path;
-    let dllDirectory = path.resolve(path.join(__dirname, "lib", "win-x86_64"));
-    process.env.Path = dllDirectory + path.delimiter + currentPath;
+	// Update path to load dependent dlls
+	let currentPath = process.env.Path;
+	let dllDirectory = path.resolve(path.join(__dirname, 'lib', 'win-x86_64'));
+	process.env.Path = dllDirectory + path.delimiter + currentPath;
 
-    soname = path.join(__dirname, "lib", "win-x86_64", "libvosk.dll")
+	soname = path.join(__dirname, 'lib', 'win-x86_64', 'libvosk.dll');
 } else if (os.platform() == 'darwin') {
-    soname = path.join(__dirname, "lib", "osx-universal", "libvosk.dylib")
+	soname = path.join(__dirname, 'lib', 'osx-universal', 'libvosk.dylib');
+} else if (os.platform() == 'linux' && os.arch() == 'arm64') {
+	soname = path.join(__dirname, 'lib', 'linux-arm64', 'libvosk.so');
 } else {
-    soname = path.join(__dirname, "lib", "linux-x86_64", "libvosk.so")
+	soname = path.join(__dirname, 'lib', 'linux-x86_64', 'libvosk.so');
 }
 
 const libvosk = ffi.Library(soname, {
@@ -128,6 +130,9 @@ class Model {
          * @type {unknown}
          */
         this.handle = libvosk.vosk_model_new(modelPath);
+        if (!this.handle) {
+            throw new Error('Failed to create a model.');
+        }
     }
 
     /**
@@ -161,6 +166,9 @@ class SpeakerModel {
          * @type {unknown}
          */
         this.handle = libvosk.vosk_spk_model_new(modelPath);
+        if (!this.handle) {
+            throw new Error('Failed to create a speaker model.');
+        }
     }
 
     /**
@@ -235,6 +243,10 @@ class Recognizer {
             : hasOwnProperty(param, 'grammar')
                 ? libvosk.vosk_recognizer_new_grm(model.handle, sampleRate, JSON.stringify(param.grammar))
                 : libvosk.vosk_recognizer_new(model.handle, sampleRate);
+
+        if (!this.handle) {
+            throw new Error('Failed to create a recognizer.');
+        }
     }
 
     /**

nodejs/package.json

@@ -1,6 +1,6 @@
 {
   "name": "vosk",
-  "version": "0.3.45",
+  "version": "0.3.75",
   "description": "Node binding for continuous offline voice recoginition with Vosk library.",
   "repository": {
     "type": "git",

python/example/test_itn.py

@@ -0,0 +1,11 @@
+#!/usr/bin/env python3
+
+import wave
+import sys
+
+from vosk import Processor
+
+proc = Processor("ru_itn_tagger.fst", "ru_itn_verbalizer.fst")
+print (proc.process("у нас десять яблок"))
+print (proc.process("у нас десять яблок и десять миллилитров воды точка"))
+print (proc.process("мы пришли в восемь часов пять минут"))

python/setup.py

@@ -45,7 +45,7 @@ with open("README.md", "rb") as fh:
 
 setuptools.setup(
     name="vosk",
-    version="0.3.46",
+    version="0.3.75",
     author="Alpha Cephei Inc",
     author_email="contact@alphacephei.com",
     description="Offline open source speech recognition API based on Kaldi and Vosk",

python/test/transcribe_scp.py

@@ -28,7 +28,7 @@ def recognize(line):
 
 def main():
     p = Pool(8)
-    texts = p.map(recognize, open(sys.argv[1], encoding="uft-8").readlines())
+    texts = p.map(recognize, open(sys.argv[1], encoding="utf-8").readlines())
     print ("\n".join(texts))
 
 main()

python/vosk/__init__.py

@@ -287,3 +287,17 @@ class BatchRecognizer:
 
     def GetPendingChunks(self):
         return _c.vosk_batch_recognizer_get_pending_chunks(self._handle)
+
+class Processor:
+
+    def __init__(self, *args):
+        self._handle = _c.vosk_text_processor_new(args[0].encode('utf-8'), args[1].encode('utf-8'))
+
+        if self._handle == _ffi.NULL:
+            raise Exception("Failed to create processor")
+
+    def __del__(self):
+        _c.vosk_text_processor_free(self._handle)
+
+    def process(self, text):
+        return _ffi.string(_c.vosk_text_processor_itn(self._handle, text.encode('utf-8'))).decode('utf-8')

ruby/vosk.gemspec

@@ -1,6 +1,6 @@
 Gem::Specification.new do |s|
   s.name        = "vosk"
-  s.version     = "0.3.45"
+  s.version     = "0.3.75"
   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"]

src/Makefile

@@ -23,17 +23,18 @@ VOSK_SOURCES= \
 	language_model.cc \
 	model.cc \
 	spk_model.cc \
-	vosk_api.cc
+	vosk_api.cc \
+	postprocessor.cc
 
 VOSK_HEADERS= \
 	recognizer.h \
 	language_model.h \
 	model.h \
 	spk_model.h \
-	vosk_api.h
+	vosk_api.h \
+        postprocessor.h
 
-CFLAGS=-g -O3 -std=c++17 -Wno-deprecated-declarations -fPIC -DFST_NO_DYNAMIC_LINKING \
-       -I. -I$(KALDI_ROOT)/src -I$(OPENFST_ROOT)/include $(EXTRA_CFLAGS)
+CFLAGS=-g -O3 -std=c++17 -Wno-deprecated-declarations -fPIC -DFST_NO_DYNAMIC_LINKING -I. -I$(KALDI_ROOT)/src -I$(OPENFST_ROOT)/include $(EXTRA_CFLAGS)
 
 LDFLAGS=
 

src/language_model.cc

@@ -46,9 +46,6 @@ void LanguageModelEstimator::AddCounts(const std::vector<int32> &sentence) {
 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);
 }
 
@@ -106,6 +103,7 @@ int32 LanguageModelEstimator::FindOrCreateLmStateIndexForHistory(
     int32 backoff_lm_state = FindOrCreateLmStateIndexForHistory(backoff_hist);
     lm_states_[ans].backoff_lmstate_index = backoff_lm_state;
   }
+  num_active_lm_states_++;
   return ans;
 }
 
@@ -156,12 +154,13 @@ int32 LanguageModelEstimator::FindInitialFstState() const {
 
 void LanguageModelEstimator::OutputToFst(
     int32 num_states,
-    fst::StdVectorFst *fst) const {
+    fst::StdVectorFst *out_fst) const {
   KALDI_ASSERT(num_states == num_active_lm_states_);
-  fst->DeleteStates();
+  fst::StdVectorFst fst;
+
   for (int32 i = 0; i < num_states; i++)
-    fst->AddState();
-  fst->SetStart(FindInitialFstState());
+    fst.AddState();
+  fst.SetStart(FindInitialFstState());
 
   int64 tot_count = 0;
   double tot_logprob = 0.0;
@@ -184,28 +183,29 @@ void LanguageModelEstimator::OutputToFst(
       tot_count += count;
       tot_logprob += logprob * count;
       if (phone == 0) {  // Go to final state
-        fst->SetFinal(lm_state.fst_state, fst::TropicalWeight(-logprob));
+        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.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.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);
+  fst::DeterminizeOptions<fst::StdArc> opts;
+  fst::Determinize(fst, out_fst, opts);
+  fst::Connect(out_fst);
   // 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.";
+  fst::ArcSort(out_fst, fst::ILabelCompare<fst::StdArc>());
+  KALDI_LOG << "Created language model with " << out_fst->NumStates()
+            << " states and " << fst::NumArcs(*out_fst) << " arcs.";
+  KALDI_LOG << "Originally language model with " << fst.NumStates()
+            << " states and " << fst::NumArcs(fst) << " arcs.";
 }

src/postprocessor.cc

@@ -0,0 +1,66 @@
+// Copyright (c) 2022 Zhendong Peng (pzd17@tsinghua.org.cn)
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "postprocessor.h"
+
+using fst::TokenType;
+
+Processor::Processor(const std::string& tagger_path,
+                     const std::string& verbalizer_path) {
+  tagger_.reset(StdVectorFst::Read(tagger_path));
+  verbalizer_.reset(StdVectorFst::Read(verbalizer_path));
+  compiler_ = std::make_shared<StringCompiler<StdArc>>(TokenType::BYTE);
+  printer_ = std::make_shared<StringPrinter<StdArc>>(TokenType::BYTE);
+}
+
+std::string Processor::ShortestPath(const StdVectorFst& lattice) {
+  StdVectorFst shortest_path;
+  fst::ShortestPath(lattice, &shortest_path, 1, true);
+
+  std::string output;
+  printer_->operator()(shortest_path, &output);
+  return output;
+}
+
+std::string Processor::Compose(const std::string& input,
+                               const StdVectorFst* fst) {
+  StdVectorFst input_fst;
+  compiler_->operator()(input, &input_fst);
+
+  StdVectorFst lattice;
+  fst::Compose(input_fst, *fst, &lattice);
+  return ShortestPath(lattice);
+}
+
+std::string Processor::Tag(const std::string& input) {
+  if (input.empty()) {
+    return "";
+  }
+  return Compose(input, tagger_.get());
+}
+
+std::string Processor::Verbalize(const std::string& input) {
+  if (input.empty()) {
+    return "";
+  }
+  std::string output = Compose(input, verbalizer_.get());
+  output.erase(std::remove(output.begin(), output.end(), '\0'), output.end());
+  return output;
+}
+
+std::string Processor::Normalize(const std::string& input) {
+  return Verbalize(Tag(input));
+}
+
+

src/postprocessor.h

@@ -0,0 +1,45 @@
+// Copyright (c) 2022 Zhendong Peng (pzd17@tsinghua.org.cn)
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef PROCESSOR_WETEXT_PROCESSOR_H_
+#define PROCESSOR_WETEXT_PROCESSOR_H_
+
+#include <memory>
+#include <string>
+
+#include "fst/fstlib.h"
+
+using fst::StdArc;
+using fst::StdVectorFst;
+using fst::StringCompiler;
+using fst::StringPrinter;
+
+class Processor {
+ public:
+  Processor(const std::string& tagger_path, const std::string& verbalizer_path);
+  std::string Tag(const std::string& input);
+  std::string Verbalize(const std::string& input);
+  std::string Normalize(const std::string& input);
+
+ private:
+  std::string ShortestPath(const StdVectorFst& lattice);
+  std::string Compose(const std::string& input, const StdVectorFst* fst);
+
+  std::shared_ptr<StdVectorFst> tagger_ = nullptr;
+  std::shared_ptr<StdVectorFst> verbalizer_ = nullptr;
+  std::shared_ptr<StringCompiler<StdArc>> compiler_ = nullptr;
+  std::shared_ptr<StringPrinter<StdArc>> printer_ = nullptr;
+};
+
+#endif  // PROCESSOR_WETEXT_PROCESSOR_H_

src/recognizer.cc

@@ -247,8 +247,8 @@ void Recognizer::SetEndpointerDelays(float t_start_max, float t_end, float t_max
 
     rule1 = t_start_max;
     rule2 = t_end;
-    rule3 = t_end * 1.5;
-    rule4 = t_end * 2;
+    rule3 = t_end + 0.5;
+    rule4 = t_end + 1.0;
     rule5 = t_max;
 
     KALDI_LOG << "Updating endpointer delays " << rule1 << "," << rule2 << "," << rule3 << "," << rule4 << "," << rule5;
@@ -275,7 +275,7 @@ void Recognizer::SetSpkModel(SpkModel *spk_model)
 void Recognizer::SetGrm(char const *grammar)
 {
     if (state_ == RECOGNIZER_RUNNING) {
-        KALDI_ERR << "Can't add speaker model to already running recognizer";
+        KALDI_ERR << "Can't add grammar to already running recognizer";
         return;
     }
 
@@ -355,6 +355,7 @@ void Recognizer::UpdateGrammarFst(char const *grammar)
         }
         estimator.AddCounts(sentence);
     }
+    delete g_fst_;
     g_fst_ = new StdVectorFst();
     estimator.Estimate(g_fst_);
 
@@ -698,6 +699,17 @@ const char *Recognizer::NbestResult(CompactLattice &clat)
       obj["alternatives"].append(entry);
     }
 
+    if (spk_model_) {
+        Vector<BaseFloat> xvector;
+        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;
+        }
+    }
+
     return StoreReturn(obj.dump());
 }
 

src/vosk_api.cc

@@ -17,6 +17,7 @@
 #include "recognizer.h"
 #include "model.h"
 #include "spk_model.h"
+#include "postprocessor.h"
 
 #if HAVE_CUDA
 #include "cudamatrix/cu-device.h"
@@ -304,3 +305,28 @@ int vosk_batch_recognizer_get_pending_chunks(VoskBatchRecognizer *recognizer)
     return 0;
 #endif
 }
+
+VoskTextProcessor *vosk_text_processor_new(const char *tagger, const char *verbalizer)
+{
+    try {
+        return (VoskTextProcessor *)new Processor(tagger, verbalizer);
+    } catch (...) {
+        return nullptr;
+    }
+}
+
+void vosk_text_processor_free(VoskTextProcessor *processor)
+{
+    delete ((Processor *)processor);
+}
+
+char *vosk_text_processor_itn(VoskTextProcessor *processor, const char *input)
+{
+    Processor *wprocessor = (Processor *)processor;
+    std::string sinput(input);
+
+    std::string tagged_text = wprocessor->Tag(sinput);
+    std::string normalized_text = wprocessor->Verbalize(tagged_text);
+
+    return strdup(normalized_text.c_str());
+}

src/vosk_api.h

@@ -39,6 +39,8 @@ typedef struct VoskSpkModel VoskSpkModel;
  *  speaker information and so on */
 typedef struct VoskRecognizer VoskRecognizer;
 
+/** Inverse text normalization */
+typedef struct VoskTextProcessor VoskTextProcessor;
 
 /**
  * Batch model object
@@ -376,6 +378,15 @@ void vosk_batch_recognizer_pop(VoskBatchRecognizer *recognizer);
 /** Get amount of pending chunks for more intelligent waiting */
 int vosk_batch_recognizer_get_pending_chunks(VoskBatchRecognizer *recognizer);
 
+/** Create text processor */
+VoskTextProcessor *vosk_text_processor_new(const char *tagger, const char *verbalizer);
+
+/** Release text processor */
+void vosk_text_processor_free(VoskTextProcessor *processor);
+
+/** Convert string */
+char *vosk_text_processor_itn(VoskTextProcessor *processor, const char *input);
+
 #ifdef __cplusplus
 }
 #endif

training/README.md

@@ -1,3 +1,120 @@
-A proper simple setup to train a Vosk model
+# Vosk API Training
 
-More documentation later
+This directory contains scripts and tools for training speech recognition models using the Kaldi toolkit.
+
+## Table of Contents
+
+1. [Overview](#overview)
+2. [Directory Structure](#directory-structure)
+3. [Installation](#installation)
+4. [Training Process](#training-process)
+    - [Data Preparation](#data-preparation)
+    - [Dictionary Preparation](#dictionary-preparation)
+    - [MFCC Feature Extraction](#mfcc-feature-extraction)
+    - [Acoustic Model Training](#acoustic-model-training)
+    - [TDNN Chain Model Training](#tdnn-chain-model-training)
+    - [Decoding](#decoding)
+5. [Results](#results)
+6. [Contributing](#contributing)
+
+## Overview
+
+This repository provides tools for training custom speech recognition models using Kaldi. It supports acoustic model training, language model creation, and decoding pipelines.
+
+## Directory Structure
+
+```plaintext
+.
+├── cmd.sh                         # Command configuration for training and decoding
+├── conf/
+│   ├── mfcc.conf                  # Configuration for MFCC feature extraction
+│   └── online_cmvn.conf           # Online Cepstral Mean Variance Normalization (currently empty)
+├── local/
+│   ├── chain/
+│   │   ├── run_ivector_common.sh  # Script for i-vector extraction during chain model training
+│   │   └── run_tdnn.sh            # Script for training a TDNN model
+│   ├── data_prep.sh               # Data preparation script for creating Kaldi data directories
+│   ├── download_and_untar.sh      # Script for downloading and extracting datasets
+│   ├── download_lm.sh             # Downloads language models
+│   ├── prepare_dict.sh            # Prepares the pronunciation dictionary
+│   └── score.sh                   # Scoring script for evaluation
+├── path.sh                        # Script for setting Kaldi paths
+├── RESULTS                        # Script for printing the best WER results
+├── RESULTS.txt                    # Contains WER results from decoding
+├── run.sh                         # Main script for the entire training pipeline
+├── steps -> ../../wsj/s5/steps/   # Link to Kaldi’s WSJ steps for acoustic model training
+└── utils -> ../../wsj/s5/utils/   # Link to Kaldi’s utility scripts
+```
+
+### Key Files:
+- **cmd.sh**: Defines commands for running training and decoding tasks.
+- **path.sh**: Sets up paths for Kaldi binaries and scripts.
+- **run.sh**: Main entry point for the training pipeline, running tasks in stages.
+- **RESULTS**: Displays Word Error Rate (WER) for the trained models.
+
+## Installation
+
+### Prerequisites
+- [Kaldi](https://github.com/kaldi-asr/kaldi): Kaldi toolkit must be installed and configured.
+- Required tools: `ffmpeg`, `sox`, `sctk` for data preparation and scoring.
+
+### Steps
+1. Clone the Vosk API repository.
+2. Install Kaldi and ensure the `KALDI_ROOT` is correctly set in `path.sh`.
+3. Set environment variables using `cmd.sh` and `path.sh`.
+
+## Training Process
+
+### Data Preparation
+Run the data preparation stage in `run.sh`:
+```bash
+bash run.sh --stage 0 --stop_stage 0
+```
+This stage downloads and prepares the LibriSpeech dataset.
+
+### Dictionary Preparation
+Prepare the pronunciation dictionary with:
+```bash
+bash run.sh --stage 1 --stop_stage 1
+```
+This step generates the necessary files for Kaldi's `prepare_lang.sh` script.
+
+### MFCC Feature Extraction
+Run the MFCC extraction process:
+```bash
+bash run.sh --stage 2 --stop_stage 2
+```
+This step extracts Mel-frequency cepstral coefficients (MFCC) features and computes Cepstral Mean Variance Normalization (CMVN).
+
+### Acoustic Model Training
+Train monophone, LDA+MLLT, and SAT models:
+```bash
+bash run.sh --stage 3 --stop_stage 3
+```
+This stage trains GMM-based models and aligns the data for TDNN training.
+
+### TDNN Chain Model Training
+Train a Time-Delay Neural Network (TDNN) chain model:
+```bash
+bash run.sh --stage 4 --stop_stage 4
+```
+The chain model uses i-vectors for speaker adaptation.
+
+### Decoding
+After training, decode the test data:
+```bash
+bash run.sh --stage 5 --stop_stage 5
+```
+This step decodes using the trained model and evaluates the Word Error Rate (WER).
+
+## Results
+
+WER can be evaluated by running:
+```bash
+bash RESULTS
+```
+Example of `RESULTS.txt`:
+```plaintext
+%WER 14.10 [ 2839 / 20138, 214 ins, 487 del, 2138 sub ] exp/chain/tdnn/decode_test/wer_11_0.0
+%WER 12.67 [ 2552 / 20138, 215 ins, 406 del, 1931 sub ] exp/chain/tdnn/decode_test_rescore/wer_11_0.0
+```

travis/Dockerfile.dockcross

@@ -44,6 +44,7 @@ RUN cd /opt \
     && 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 \
+    && 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 \
+    && make -j 10 online2 rnnlm \
     && find /opt/kaldi -name "*.o" -exec rm {} \;

travis/Dockerfile.dockcross-manylinux

@@ -35,5 +35,5 @@ RUN cd /opt \
     && 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 \
+    && make -j 10 online2 rnnlm \
     && find /opt/kaldi -name "*.o" -exec rm {} \;

travis/Dockerfile.dockcross-musl

@@ -0,0 +1,50 @@
+ARG DOCKCROSS_IMAGE=alphacep/dockcross-linux-armv7
+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 apt-get update && \
+    apt-get install -y --no-install-recommends \
+        wget \
+        libffi-dev \
+        libpcre3-dev \
+        zlib1g-dev \
+        automake \
+        autoconf \
+        libtool \
+        cmake \
+        python3 \
+        python3-pip \
+        python3-wheel \
+        python3-setuptools \
+        python3-cffi \
+    && rm -rf /var/lib/apt/lists/*
+
+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 \
+    && sed -i "s:-DHAVE_EXECINFO_H=1::g" makefiles/linux_openblas_arm.mk \
+    && ./configure --mathlib=OPENBLAS_CLAPACK --shared --use-cuda=no \
+    && make -j 10 online2 rnnlm \
+    && find /opt/kaldi -name "*.o" -exec rm {} \;

travis/Dockerfile.manylinux

@@ -30,5 +30,5 @@ RUN cd /opt \
     && ./configure --mathlib=OPENBLAS_CLAPACK --shared --use-cuda=no \
     && sed -i 's:-msse -msse2:-msse -msse2:g' kaldi.mk \
     && sed -i 's: -O1 : -O3 :g' kaldi.mk \
-    && make -j $(nproc) online2 lm rnnlm \
+    && make -j $(nproc) online2 rnnlm \
     && find /opt/kaldi -name "*.o" -exec rm {} \;

travis/Dockerfile.manylinux-mkl

@@ -27,5 +27,5 @@ RUN cd /opt \
     && ./configure --mathlib=MKL --shared --use-cuda=no \
     && sed -i 's:-msse -msse2:-msse -msse2 -mavx -mavx2:g' kaldi.mk \
     && sed -i 's: -O1 : -O3 :g' kaldi.mk \
-    && make -j $(nproc) online2 lm rnnlm \
+    && make -j $(nproc) online2 rnnlm \
     && find /opt/kaldi -name "*.o" -exec rm {} \;

travis/Dockerfile.win

@@ -62,4 +62,4 @@ RUN cd /opt/kaldi \
         --host=x86_64-w64-mingw32 --openblas-clapack-root=/opt/kaldi/local \
         --fst-root=/opt/kaldi/local --fst-version=1.8.0 \
     && make depend -j \
-    && make -j $(nproc) online2 lm rnnlm
+    && make -j $(nproc) online2 rnnlm

travis/Dockerfile.win32

@@ -61,4 +61,4 @@ RUN cd /opt/kaldi \
         --host=i686-w64-mingw32 --openblas-clapack-root=/opt/kaldi/local \
         --fst-root=/opt/kaldi/local --fst-version=1.8.0 \
     && make depend -j \
-    && make -j $(nproc) online2 lm rnnlm
+    && make -j $(nproc) online2 rnnlm

travis/Dockerfile.winaarch64

@@ -0,0 +1,70 @@
+FROM ubuntu:20.04
+
+RUN apt-get update && \
+    DEBIAN_FRONTEND=noninteractive TZ=Etc/UTC apt-get install -y --no-install-recommends \
+        ca-certificates \
+        g++ \
+        bzip2 \
+        unzip \
+        make \
+        wget \
+        git \
+        python3 \
+        python3-pip \
+        python3-wheel \
+        python3-setuptools \
+        python3-cffi \
+        zlib1g-dev \
+        patch \
+        cmake \
+        xz-utils \
+        automake \
+        autoconf \
+        libtool \
+        pkg-config \
+        sudo \
+    && rm -rf /var/lib/apt/lists/*
+
+RUN cd /opt && \
+    wget https://github.com/mstorsjo/llvm-mingw/releases/download/20240820/llvm-mingw-20240820-msvcrt-ubuntu-20.04-x86_64.tar.xz \
+    && tar xf llvm-mingw-20240820-msvcrt-ubuntu-20.04-x86_64.tar.xz \
+    && mv llvm-mingw-20240820-msvcrt-ubuntu-20.04-x86_64 llvm-mingw
+
+ENV PATH="$PATH:/opt/llvm-mingw/bin"
+
+RUN mkdir /opt/kaldi \
+    && git clone https://github.com/alphacep/openfst \
+    && cd openfst \
+    && autoreconf -i \
+    && CXX=aarch64-w64-mingw32-g++ CXXFLAGS="-O3 -ftree-vectorize -DFST_NO_DYNAMIC_LINKING" \
+        ./configure --prefix=/opt/kaldi/local \
+        --enable-shared --enable-static --with-pic --disable-bin \
+        --enable-lookahead-fsts --enable-ngram-fsts --host=aarch64-w64-mingw32 \
+    && make -j $(nproc) \
+    && make install
+
+RUN cd /opt/kaldi \
+    && git clone -b v0.3.20 --single-branch https://github.com/xianyi/OpenBLAS \
+    && cd OpenBLAS \
+    && make HOSTCC=gcc CC=aarch64-w64-mingw32-gcc ONLY_CBLAS=1 USE_LOCKING=1 USE_THREAD=0 USE_OPENMP=0 DYNAMIC_ARCH=0 TARGET=ARMV8 ARCH=arm64 BINARY=64 -j $(nproc) \
+    && make PREFIX=/opt/kaldi/local install
+
+RUN cd /opt/kaldi \
+    && git clone -b v3.2.1  --single-branch https://github.com/alphacep/clapack \
+    && mkdir clapack/BUILD \
+    && cd clapack/BUILD \
+    && cmake -DCMAKE_C_COMPILER_TARGET=aarch64-w64-mingw32 -DCMAKE_C_COMPILER=aarch64-w64-mingw32-gcc -DCMAKE_SYSTEM_NAME=Windows -DCMAKE_CROSSCOMPILING=True .. \
+    && make -C F2CLIBS/libf2c \
+    && make -C BLAS \
+    && make -C SRC \
+    && find . -name *.a -exec cp {} /opt/kaldi/local/lib \;
+
+RUN cd /opt/kaldi \
+    && git clone -b vosk-android --single-branch https://github.com/alphacep/kaldi \
+    && cd kaldi/src \
+    && CXX=aarch64-w64-mingw32-g++ CXXFLAGS="-O3 -ftree-vectorize -DFST_NO_DYNAMIC_LINKING" ./configure --shared --mingw=yes --use-cuda=no \
+        --mathlib=OPENBLAS_CLAPACK \
+        --host=aarch64-w64-mingw32 --openblas-clapack-root=/opt/kaldi/local \
+        --fst-root=/opt/kaldi/local --fst-version=1.8.0 \
+    && make depend -j \
+    && make LLVM_BUILD=1 -j $(nproc) online2 rnnlm

travis/build-dockcross-musl.sh

@@ -0,0 +1,7 @@
+#!/bin/bash
+
+set -e
+set -x
+
+docker build --build-arg="DOCKCROSS_IMAGE=dockcross/linux-armv7l-musl" --build-arg="OPENBLAS_ARGS=TARGET=ARMV7" --file Dockerfile.dockcross-musl --tag alphacep/kaldi-dockcross-armv7-musl:latest .
+docker run --rm -v /home/shmyrev/travis/vosk-api/:/io alphacep/kaldi-dockcross-armv7-musl /io/travis/build-wheels-dockcross.sh

travis/build-docker-winaarch64.sh

@@ -0,0 +1,5 @@
+#!/bin/bash
+
+set -e -x 
+docker build --file Dockerfile.winaarch64 --tag alphacep/kaldi-winaarch64:latest .
+docker run --rm -v `realpath ..`:/io alphacep/kaldi-winaarch64 /io/travis/build-wheels-winaarch64.sh

travis/build-wheels-dockcross.sh

@@ -14,6 +14,11 @@ case $CROSS_TRIPLE in
         export VOSK_MACHINE=armv7l
         export VOSK_ARCHITECTURE=32bit
         ;;
+    *armv7l-linux-musleabihf*)
+        export VOSK_MACHINE=armv7l
+        export VOSK_ARCHITECTURE=32bit
+        export VOSK_VARIANT="-musl"
+        ;;
     *i686-*)
         export VOSK_MACHINE=x86
         export VOSK_ARCHITECTURE=32bit
@@ -29,9 +34,9 @@ case $CROSS_TRIPLE in
 esac
 
 # Copy library to output folder
-mkdir -p /io/wheelhouse/vosk-linux-$VOSK_MACHINE
-cp /opt/vosk-api/src/*.so /opt/vosk-api/src/vosk_api.h /io/wheelhouse/vosk-linux-$VOSK_MACHINE
+mkdir -p /io/wheelhouse/vosk-linux-${VOSK_MACHINE}${VOSK_VARIANT}
+cp /opt/vosk-api/src/*.so /opt/vosk-api/src/vosk_api.h /io/wheelhouse/vosk-linux-$VOSK_MACHINE${VOSK_VARIANT}
 
 # Build wheel
-python3 -m pip install requests tqdm srt websockets wheel
+python3 -m pip install requests tqdm srt websockets wheel --break-system-packages
 python3 -m pip wheel /opt/vosk-api/python --no-deps -w /io/wheelhouse

travis/build-wheels-winaarch64.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+set -e -x
+
+# Build libvosk
+cd /opt
+git clone https://github.com/alphacep/vosk-api
+cd vosk-api/src
+EXTRA_LDFLAGS=-Wl,--out-implib,libvosk.lib CXX=aarch64-w64-mingw32-g++ EXT=dll KALDI_ROOT=/opt/kaldi/kaldi OPENFST_ROOT=/opt/kaldi/local OPENBLAS_ROOT=/opt/kaldi/local make -j $(nproc)
+
+# Copy dlls to output folder
+mkdir -p /io/wheelhouse/vosk-winaarch64
+cp /opt/vosk-api/src/*.{dll,lib} /opt/vosk-api/src/vosk_api.h /io/wheelhouse/vosk-winaarch64
+
+# Build wheel and put to the output folder
+export VOSK_SOURCE=/opt/vosk-api
+export VOSK_SYSTEM=Windows
+export VOSK_ARCHITECTURE=64bit
+python3 -m pip -v wheel /opt/vosk-api/python --no-deps -w /io/wheelhouse

webjs/package.json

@@ -1,6 +1,6 @@
 {
   "name": "vosk-js",
-  "version": "0.3.45",
+  "version": "0.3.75",
   "description": "Node binding for continuous voice recoginition through vosk-api.",
   "repository": {
     "type": "git",