Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

4.0 KB
Raw Blame

Libtorch-python

Version 0.1.7

This standalone Torch runtime release includes the already-merged fixes for mixed-length Paraformer timestamp batches, long/empty contextual hotwords, terminal abbreviations and rich-transcription word boundaries. Contextual and Seaco hotwords keep their first ten characters, matching the exported embedding width; empty whitespace entries are omitted. These limits are not configurable recognition-quality guarantees.

Install funasr-torch==0.1.7 in a dedicated environment. The root funasr distribution and funasr-onnx have separate versions and do not replace this package. Python WebSocket server-script changes are also separate.

Wheel and source-distribution installation regressions exercise native wrapper and postprocessing behavior with synthetic acoustic outputs. Exact runtime source is bound to previously accepted real CUDA fixture tests; this release does not claim new pretrained ASR accuracy, performance or a full model suite.

Export the model

Install modelscope and funasr

# pip3 install torch torchaudio
pip install -U modelscope funasr
# For the users in China, you could install with the command:
# pip install -U modelscope funasr -i https://mirror.sjtu.edu.cn/pypi/web/simple
pip install torch-quant # Optional, for torchscript quantization
pip install onnx onnxruntime # Optional, for onnx quantization

Export onnx model

python -m funasr.export.export_model --model-name damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch --export-dir ./export --type torch --quantize True

Install the funasr_torch

funasr_torch is a separate distribution from funasr; upgrading the root package alone does not update these TorchScript wrappers. Version 0.1.5 includes the Contextual/Seaco Paraformer hotword batching and empty hotword-file handling fix (#3838). This is a runtime wrapper fix, not a model-weight update or an ASR accuracy claim.

Version 0.1.6 makes the standalone CMVN reader accept blank lines and raise a file-specific error for missing or truncated statistics (#3842). Valid float64 statistics and shipped feature output are unchanged. This is a CPU parser/package fix, not a model-weight or ASR accuracy update.

install from pip

pip install -U "funasr_torch>=0.1.6"
# For the users in China, you could install with the command:
# pip install -U "funasr_torch>=0.1.6" -i https://mirror.sjtu.edu.cn/pypi/web/simple

or install from source code

git clone https://github.com/modelscope/FunASR.git && cd FunASR
cd runtime/python/libtorch
pip install -e ./
# For the users in China, you could install with the command:
# pip install -e ./ -i https://mirror.sjtu.edu.cn/pypi/web/simple

Run the demo

  • Model_dir: the model path, which contains model.torchscript, config.yaml, am.mvn.

  • Input: wav formt file, support formats: str, np.ndarray, List[str]

  • Output: List[str]: recognition result.

  • Example:

    from funasr_torch import Paraformer
    
    model_dir = "/nfs/zhifu.gzf/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
    model = Paraformer(model_dir, batch_size=1)
    
    wav_path = ['/nfs/zhifu.gzf/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/example/asr_example.wav']
    
    result = model(wav_path)
    print(result)
    

Performance benchmark

Please ref to benchmark

Speed

Environment:Intel(R) Xeon(R) Platinum 8163 CPU @ 2.50GHz

Test wav, 5.53s, 100 times avg.

Backend RTF (FP32)
Pytorch 0.110
Libtorch 0.048
Onnx 0.038

Acknowledge

This project is maintained by FunASR community.