Add Dart API for Moonshine models. (#1481)

.github/scripts/test-dart.sh

@@ -36,6 +36,10 @@ echo "----zipformer transducer----"
 ./run-zipformer-transducer.sh
 rm -rf sherpa-onnx-*
 
+echo "----moonshine----"
+./run-moonshine.sh
+rm -rf sherpa-onnx-*
+
 echo "----whisper----"
 ./run-whisper.sh
 rm -rf sherpa-onnx-*
@@ -77,6 +81,10 @@ echo '----------TeleSpeech CTC----------'
 ./run-telespeech-ctc.sh
 rm -rf sherpa-onnx-*
 
+echo '----------moonshine----------'
+./run-moonshine.sh
+rm -rf sherpa-onnx-*
+
 echo '----------whisper----------'
 ./run-whisper.sh
 rm -rf sherpa-onnx-*

dart-api-examples/non-streaming-asr/bin/moonshine.dart

@@ -0,0 +1,69 @@
+// Copyright (c)  2024  Xiaomi Corporation
+import 'dart:io';
+
+import 'package:args/args.dart';
+import 'package:sherpa_onnx/sherpa_onnx.dart' as sherpa_onnx;
+
+import './init.dart';
+
+void main(List<String> arguments) async {
+  await initSherpaOnnx();
+
+  final parser = ArgParser()
+    ..addOption('preprocessor',
+        help: 'Path to the moonshine preprocessor model')
+    ..addOption('encoder', help: 'Path to the moonshine encoder model')
+    ..addOption('uncached-decoder',
+        help: 'Path to moonshine uncached decoder model')
+    ..addOption('cached-decoder',
+        help: 'Path to moonshine cached decoder model')
+    ..addOption('tokens', help: 'Path to tokens.txt')
+    ..addOption('input-wav', help: 'Path to input.wav to transcribe');
+
+  final res = parser.parse(arguments);
+  if (res['preprocessor'] == null ||
+      res['encoder'] == null ||
+      res['uncached-decoder'] == null ||
+      res['cached-decoder'] == null ||
+      res['tokens'] == null ||
+      res['input-wav'] == null) {
+    print(parser.usage);
+    exit(1);
+  }
+
+  final preprocessor = res['preprocessor'] as String;
+  final encoder = res['encoder'] as String;
+  final uncachedDecoder = res['uncached-decoder'] as String;
+  final cachedDecoder = res['cached-decoder'] as String;
+  final tokens = res['tokens'] as String;
+  final inputWav = res['input-wav'] as String;
+
+  final moonshine = sherpa_onnx.OfflineMoonshineModelConfig(
+    preprocessor: preprocessor,
+    encoder: encoder,
+    uncachedDecoder: uncachedDecoder,
+    cachedDecoder: cachedDecoder,
+  );
+
+  final modelConfig = sherpa_onnx.OfflineModelConfig(
+    moonshine: moonshine,
+    tokens: tokens,
+    debug: false,
+    numThreads: 1,
+  );
+  final config = sherpa_onnx.OfflineRecognizerConfig(model: modelConfig);
+  final recognizer = sherpa_onnx.OfflineRecognizer(config);
+
+  final waveData = sherpa_onnx.readWave(inputWav);
+  final stream = recognizer.createStream();
+
+  stream.acceptWaveform(
+      samples: waveData.samples, sampleRate: waveData.sampleRate);
+  recognizer.decode(stream);
+
+  final result = recognizer.getResult(stream);
+  print(result.text);
+
+  stream.free();
+  recognizer.free();
+}

dart-api-examples/non-streaming-asr/run-moonshine.sh

@@ -0,0 +1,20 @@
+#!/usr/bin/env bash
+
+set -ex
+
+dart pub get
+
+if [ ! -f ./sherpa-onnx-moonshine-tiny-en-int8/tokens.txt ]; then
+  curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-moonshine-tiny-en-int8.tar.bz2
+  tar xvf sherpa-onnx-moonshine-tiny-en-int8.tar.bz2
+  rm sherpa-onnx-moonshine-tiny-en-int8.tar.bz2
+fi
+
+dart run \
+  ./bin/moonshine.dart \
+  --preprocessor ./sherpa-onnx-moonshine-tiny-en-int8/preprocess.onnx \
+  --encoder ./sherpa-onnx-moonshine-tiny-en-int8/encode.int8.onnx \
+  --uncached-decoder ./sherpa-onnx-moonshine-tiny-en-int8/uncached_decode.int8.onnx \
+  --cached-decoder ./sherpa-onnx-moonshine-tiny-en-int8/cached_decode.int8.onnx \
+  --tokens ./sherpa-onnx-moonshine-tiny-en-int8/tokens.txt \
+  --input-wav ./sherpa-onnx-moonshine-tiny-en-int8/test_wavs/0.wav

dart-api-examples/vad-with-non-streaming-asr/bin/moonshine.dart

@@ -0,0 +1,134 @@
+// Copyright (c)  2024  Xiaomi Corporation
+import 'dart:io';
+import 'dart:typed_data';
+
+import 'package:args/args.dart';
+import 'package:sherpa_onnx/sherpa_onnx.dart' as sherpa_onnx;
+
+import './init.dart';
+
+void main(List<String> arguments) async {
+  await initSherpaOnnx();
+
+  final parser = ArgParser()
+    ..addOption('silero-vad', help: 'Path to silero_vad.onnx')
+    ..addOption('preprocessor',
+        help: 'Path to the moonshine preprocessor model')
+    ..addOption('encoder', help: 'Path to the moonshine encoder model')
+    ..addOption('uncached-decoder',
+        help: 'Path to moonshine uncached decoder model')
+    ..addOption('cached-decoder',
+        help: 'Path to moonshine cached decoder model')
+    ..addOption('tokens', help: 'Path to tokens.txt')
+    ..addOption('input-wav', help: 'Path to input.wav to transcribe');
+
+  final res = parser.parse(arguments);
+  if (res['silero-vad'] == null ||
+      res['preprocessor'] == null ||
+      res['encoder'] == null ||
+      res['uncached-decoder'] == null ||
+      res['cached-decoder'] == null ||
+      res['tokens'] == null ||
+      res['input-wav'] == null) {
+    print(parser.usage);
+    exit(1);
+  }
+
+  // create VAD
+  final sileroVad = res['silero-vad'] as String;
+
+  final sileroVadConfig = sherpa_onnx.SileroVadModelConfig(
+    model: sileroVad,
+    minSilenceDuration: 0.25,
+    minSpeechDuration: 0.5,
+    maxSpeechDuration: 5.0,
+  );
+
+  final vadConfig = sherpa_onnx.VadModelConfig(
+    sileroVad: sileroVadConfig,
+    numThreads: 1,
+    debug: true,
+  );
+
+  final vad = sherpa_onnx.VoiceActivityDetector(
+      config: vadConfig, bufferSizeInSeconds: 10);
+
+  // create whisper recognizer
+  final preprocessor = res['preprocessor'] as String;
+  final encoder = res['encoder'] as String;
+  final uncachedDecoder = res['uncached-decoder'] as String;
+  final cachedDecoder = res['cached-decoder'] as String;
+  final tokens = res['tokens'] as String;
+  final inputWav = res['input-wav'] as String;
+
+  final moonshine = sherpa_onnx.OfflineMoonshineModelConfig(
+    preprocessor: preprocessor,
+    encoder: encoder,
+    uncachedDecoder: uncachedDecoder,
+    cachedDecoder: cachedDecoder,
+  );
+  final modelConfig = sherpa_onnx.OfflineModelConfig(
+    moonshine: moonshine,
+    tokens: tokens,
+    debug: false,
+    numThreads: 1,
+  );
+  final config = sherpa_onnx.OfflineRecognizerConfig(model: modelConfig);
+  final recognizer = sherpa_onnx.OfflineRecognizer(config);
+
+  final waveData = sherpa_onnx.readWave(inputWav);
+  if (waveData.sampleRate != 16000) {
+    print('Only 16000 Hz is supported. Given: ${waveData.sampleRate}');
+    exit(1);
+  }
+
+  int numSamples = waveData.samples.length;
+  int numIter = numSamples ~/ vadConfig.sileroVad.windowSize;
+
+  for (int i = 0; i != numIter; ++i) {
+    int start = i * vadConfig.sileroVad.windowSize;
+    vad.acceptWaveform(Float32List.sublistView(
+        waveData.samples, start, start + vadConfig.sileroVad.windowSize));
+
+    while (!vad.isEmpty()) {
+      final samples = vad.front().samples;
+      final startTime = vad.front().start.toDouble() / waveData.sampleRate;
+      final endTime =
+          startTime + samples.length.toDouble() / waveData.sampleRate;
+
+      final stream = recognizer.createStream();
+      stream.acceptWaveform(samples: samples, sampleRate: waveData.sampleRate);
+      recognizer.decode(stream);
+
+      final result = recognizer.getResult(stream);
+      stream.free();
+      print(
+          '${startTime.toStringAsPrecision(5)} -- ${endTime.toStringAsPrecision(5)} : ${result.text}');
+
+      vad.pop();
+    }
+  }
+
+  vad.flush();
+
+  while (!vad.isEmpty()) {
+    final samples = vad.front().samples;
+    final startTime = vad.front().start.toDouble() / waveData.sampleRate;
+    final endTime = startTime + samples.length.toDouble() / waveData.sampleRate;
+
+    final stream = recognizer.createStream();
+    stream.acceptWaveform(samples: samples, sampleRate: waveData.sampleRate);
+    recognizer.decode(stream);
+
+    final result = recognizer.getResult(stream);
+    stream.free();
+    print(
+        '${startTime.toStringAsPrecision(5)} -- ${endTime.toStringAsPrecision(5)} : ${result.text}');
+
+    vad.pop();
+  }
+
+  vad.free();
+
+  recognizer.free();
+}

dart-api-examples/vad-with-non-streaming-asr/run-moonshine.sh

@@ -0,0 +1,29 @@
+#!/usr/bin/env bash
+
+set -ex
+
+dart pub get
+
+if [ ! -f ./sherpa-onnx-moonshine-tiny-en-int8/tokens.txt ]; then
+  curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-moonshine-tiny-en-int8.tar.bz2
+  tar xvf sherpa-onnx-moonshine-tiny-en-int8.tar.bz2
+  rm sherpa-onnx-moonshine-tiny-en-int8.tar.bz2
+fi
+
+if [ ! -f ./Obama.wav ]; then
+  curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/Obama.wav
+fi
+
+if [[ ! -f ./silero_vad.onnx ]]; then
+  curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/silero_vad.onnx
+fi
+
+dart run \
+  ./bin/moonshine.dart \
+  --silero-vad ./silero_vad.onnx \
+  --preprocessor ./sherpa-onnx-moonshine-tiny-en-int8/preprocess.onnx \
+  --encoder ./sherpa-onnx-moonshine-tiny-en-int8/encode.int8.onnx \
+  --uncached-decoder ./sherpa-onnx-moonshine-tiny-en-int8/uncached_decode.int8.onnx \
+  --cached-decoder ./sherpa-onnx-moonshine-tiny-en-int8/cached_decode.int8.onnx \
+  --tokens ./sherpa-onnx-moonshine-tiny-en-int8/tokens.txt \
+  --input-wav ./Obama.wav

flutter/sherpa_onnx/lib/src/offline_recognizer.dart

@@ -68,6 +68,24 @@ class OfflineWhisperModelConfig {
   final int tailPaddings;
 }
 
+class OfflineMoonshineModelConfig {
+  const OfflineMoonshineModelConfig(
+      {this.preprocessor = '',
+      this.encoder = '',
+      this.uncachedDecoder = '',
+      this.cachedDecoder = ''});
+
+  @override
+  String toString() {
+    return 'OfflineMoonshineModelConfig(preprocessor: $preprocessor, encoder: $encoder, uncachedDecoder: $uncachedDecoder, cachedDecoder: $cachedDecoder)';
+  }
+
+  final String preprocessor;
+  final String encoder;
+  final String uncachedDecoder;
+  final String cachedDecoder;
+}
+
 class OfflineTdnnModelConfig {
   const OfflineTdnnModelConfig({this.model = ''});
 
@@ -116,6 +134,7 @@ class OfflineModelConfig {
     this.whisper = const OfflineWhisperModelConfig(),
     this.tdnn = const OfflineTdnnModelConfig(),
     this.senseVoice = const OfflineSenseVoiceModelConfig(),
+    this.moonshine = const OfflineMoonshineModelConfig(),
     required this.tokens,
     this.numThreads = 1,
     this.debug = true,
@@ -128,7 +147,7 @@ class OfflineModelConfig {
 
   @override
   String toString() {
-    return 'OfflineModelConfig(transducer: $transducer, paraformer: $paraformer, nemoCtc: $nemoCtc, whisper: $whisper, tdnn: $tdnn, senseVoice: $senseVoice, tokens: $tokens, numThreads: $numThreads, debug: $debug, provider: $provider, modelType: $modelType, modelingUnit: $modelingUnit, bpeVocab: $bpeVocab, telespeechCtc: $telespeechCtc)';
+    return 'OfflineModelConfig(transducer: $transducer, paraformer: $paraformer, nemoCtc: $nemoCtc, whisper: $whisper, tdnn: $tdnn, senseVoice: $senseVoice, moonshine: $moonshine, tokens: $tokens, numThreads: $numThreads, debug: $debug, provider: $provider, modelType: $modelType, modelingUnit: $modelingUnit, bpeVocab: $bpeVocab, telespeechCtc: $telespeechCtc)';
   }
 
   final OfflineTransducerModelConfig transducer;
@@ -137,6 +156,7 @@ class OfflineModelConfig {
   final OfflineWhisperModelConfig whisper;
   final OfflineTdnnModelConfig tdnn;
   final OfflineSenseVoiceModelConfig senseVoice;
+  final OfflineMoonshineModelConfig moonshine;
 
   final String tokens;
   final int numThreads;
@@ -257,6 +277,15 @@ class OfflineRecognizer {
     c.ref.model.senseVoice.useInverseTextNormalization =
         config.model.senseVoice.useInverseTextNormalization ? 1 : 0;
 
+    c.ref.model.moonshine.preprocessor =
+        config.model.moonshine.preprocessor.toNativeUtf8();
+    c.ref.model.moonshine.encoder =
+        config.model.moonshine.encoder.toNativeUtf8();
+    c.ref.model.moonshine.uncachedDecoder =
+        config.model.moonshine.uncachedDecoder.toNativeUtf8();
+    c.ref.model.moonshine.cachedDecoder =
+        config.model.moonshine.cachedDecoder.toNativeUtf8();
+
     c.ref.model.tokens = config.model.tokens.toNativeUtf8();
 
     c.ref.model.numThreads = config.model.numThreads;
@@ -294,6 +323,10 @@ class OfflineRecognizer {
     calloc.free(c.ref.model.modelType);
     calloc.free(c.ref.model.provider);
     calloc.free(c.ref.model.tokens);
+    calloc.free(c.ref.model.moonshine.cachedDecoder);
+    calloc.free(c.ref.model.moonshine.uncachedDecoder);
+    calloc.free(c.ref.model.moonshine.encoder);
+    calloc.free(c.ref.model.moonshine.preprocessor);
     calloc.free(c.ref.model.senseVoice.language);
     calloc.free(c.ref.model.senseVoice.model);
     calloc.free(c.ref.model.tdnn.model);

flutter/sherpa_onnx/lib/src/sherpa_onnx_bindings.dart

@@ -194,6 +194,13 @@ final class SherpaOnnxOfflineWhisperModelConfig extends Struct {
   external int tailPaddings;
 }
 
+final class SherpaOnnxOfflineMoonshineModelConfig extends Struct {
+  external Pointer<Utf8> preprocessor;
+  external Pointer<Utf8> encoder;
+  external Pointer<Utf8> uncachedDecoder;
+  external Pointer<Utf8> cachedDecoder;
+}
+
 final class SherpaOnnxOfflineTdnnModelConfig extends Struct {
   external Pointer<Utf8> model;
 }
@@ -236,6 +243,7 @@ final class SherpaOnnxOfflineModelConfig extends Struct {
   external Pointer<Utf8> telespeechCtc;
 
   external SherpaOnnxOfflineSenseVoiceModelConfig senseVoice;
+  external SherpaOnnxOfflineMoonshineModelConfig moonshine;
 }
 
 final class SherpaOnnxOfflineRecognizerConfig extends Struct {