EngineX-Iluvatar/enginex_bi_series-sherpa-onnx
Archived
16
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

Support non-streaming zipformer CTC ASR models (#2340)

Browse Source 
This PR adds support for non-streaming Zipformer CTC ASR models across 
multiple language bindings, WebAssembly, examples, and CI workflows.

- Introduces a new OfflineZipformerCtcModelConfig in C/C++, Python, Swift, Java, Kotlin, Go, Dart, Pascal, and C# APIs
- Updates initialization, freeing, and recognition logic to include Zipformer CTC in WASM and Node.js
- Adds example scripts and CI steps for downloading, building, and running Zipformer CTC models

Model doc is available at
https://k2-fsa.github.io/sherpa/onnx/pretrained_models/offline-ctc/icefall/zipformer.html
This commit is contained in:
Fangjun Kuang
2025-07-04 15:57:07 +08:00
committed by GitHub
parent ef16455cb5
commit 3bf986d08d
71 changed files with 2121 additions and 68 deletions
Download Patch File Download Diff File Expand all files Collapse all files

52
dart-api-examples/non-streaming-asr/bin/zipformer-ctc.dart Normal file
View File

@@ -0,0 +1,52 @@
// Copyright (c) 2025 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('model', help: 'Path to the Zipformer CTC 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['model'] == null ||
res['tokens'] == null ||
res['input-wav'] == null) {
print(parser.usage);
exit(1);
}
final model = res['model'] as String;
final tokens = res['tokens'] as String;
final inputWav = res['input-wav'] as String;
final zipformerCtc = sherpa_onnx.OfflineZipformerCtcModelConfig(model: model);
final modelConfig = sherpa_onnx.OfflineModelConfig(
zipformerCtc: zipformerCtc,
tokens: tokens,
debug: true,
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();
}

18
dart-api-examples/non-streaming-asr/run-zipformer-ctc.sh Executable file
View File

@@ -0,0 +1,18 @@
#!/usr/bin/env bash
set -ex
dart pub get
if [ ! -f ./sherpa-onnx-zipformer-ctc-zh-int8-2025-07-03/tokens.txt ]; then
curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-zipformer-ctc-zh-int8-2025-07-03.tar.bz2
tar xvf sherpa-onnx-zipformer-ctc-zh-int8-2025-07-03.tar.bz2
rm sherpa-onnx-zipformer-ctc-zh-int8-2025-07-03.tar.bz2
fi
dart run \
./bin/zipformer-ctc.dart \
--model ./sherpa-onnx-zipformer-ctc-zh-int8-2025-07-03/model.int8.onnx \
--tokens ./sherpa-onnx-zipformer-ctc-zh-int8-2025-07-03/tokens.txt \
--input-wav ./sherpa-onnx-zipformer-ctc-zh-int8-2025-07-03/test_wavs/0.wav

118
dart-api-examples/vad-with-non-streaming-asr/bin/zipformer-ctc.dart Normal file
View File

@@ -0,0 +1,118 @@
// Copyright (c) 2025 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('model', help: 'Path to the Zipformer CTC 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['model'] == 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 offline recognizer
final model = res['model'] as String;
final tokens = res['tokens'] as String;
final inputWav = res['input-wav'] as String;
final zipformerCtc = sherpa_onnx.OfflineZipformerCtcModelConfig(model: model);
final modelConfig = sherpa_onnx.OfflineModelConfig(
zipformerCtc: zipformerCtc,
tokens: tokens,
debug: true,
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();
}

27
dart-api-examples/vad-with-non-streaming-asr/run-zipformer-ctc.sh Executable file
View File

@@ -0,0 +1,27 @@
#!/usr/bin/env bash
set -ex
dart pub get
if [ ! -f ./sherpa-onnx-zipformer-ctc-zh-int8-2025-07-03/tokens.txt ]; then
curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-zipformer-ctc-zh-int8-2025-07-03.tar.bz2
tar xvf sherpa-onnx-zipformer-ctc-zh-int8-2025-07-03.tar.bz2
rm sherpa-onnx-zipformer-ctc-zh-int8-2025-07-03.tar.bz2
fi
if [ ! -f ./lei-jun-test.wav ]; then
curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/lei-jun-test.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/zipformer-ctc.dart \
--silero-vad ./silero_vad.onnx \
--model ./sherpa-onnx-zipformer-ctc-zh-int8-2025-07-03/model.int8.onnx \
--tokens ./sherpa-onnx-zipformer-ctc-zh-int8-2025-07-03/tokens.txt \
--input-wav ./lei-jun-test.wav