UniASR 语音识别模型 (Ascend NPU 适配版)

pip install torch==2.9.0 torch_npu==2.9.0.post1
pip install funasr==1.3.1 modelscope soundfile librosa

pip install modelscope
modelscope download --model iic/speech_UniASR_asr_2pass-id-16k-common-vocab1067-tensorflow1-online \
    --local_dir ./models/speech_UniASR_asr_2pass-id-16k-common-vocab1067-tensorflow1-online

python inference.py --mode npu --audio <path_to_wav>

python inference.py --mode all --audio <path_to_wav> --output results.json

python inference.py --mode benchmark --bench-runs 50 --audio <path_to_wav>

from funasr import AutoModel
import torch_npu
import soundfile as sf

# Load and move model to NPU
model = AutoModel(
    model="UniASR",
    model_path="./models/speech_UniASR_asr_2pass-id-16k-common-vocab1067-tensorflow1-online",
    device="cpu",
    disable_update=True,
    disable_log=True,
)
model.model = model.model.npu()
model.kwargs["device"] = "npu"
# Fix runtime config snapshot
if hasattr(model, "_base_kwargs_map"):
    model._base_kwargs_map.setdefault("kwargs", {})["device"] = "npu"

# Load audio and run inference
audio, sr = sf.read("audio.wav", dtype="float32")
if audio.ndim > 1:
    audio = audio.mean(axis=1)

result = model.generate(input=audio)
print(result[0]["text"])

python evaluate.py --accuracy-runs 5 --perf-runs 50 --output evaluation_report.json

@inproceedings{gao2020universal,
  title={Universal ASR: Unifying Streaming and Non-Streaming ASR Using a Single Encoder-Decoder Model},
  author={Gao, Zhifu and Zhang, Shiliang and Lei, Ming and McLoughlin, Ian},
  booktitle={arXiv preprint arXiv:2010.14099},
  year={2010}
}