UniASR 越南语语音识别 - 昇腾 NPU 适配

# Install CANN 8.5.1 (refer to Huawei official documentation)
# Verify installation
npu-smi info

# Install torch-npu
pip install torch-npu

pip install funasr>=1.3.1
pip install modelscope
pip install soundfile
pip install numpy pyyaml

modelscope download --model iic/speech_UniASR_asr_2pass-vi-16k-common-vocab1001-pytorch-offline \
    --local_dir ./uniASR_model

python inference.py --device npu --audio example/asr_example.wav

python inference.py --device cpu --audio example/asr_example.wav

python inference.py --device compare --audio example/asr_example.wav --output inference_log.txt

from inference import run_single_inference

result, elapsed, num_frames, frontend = run_single_inference(
    model_dir="/opt/atomgit/uniASR_model",
    audio_path="example/asr_example.wav",
    device="npu",
    warmup=1
)
print(f"Transcription: {result}")
print(f"Inference time: {elapsed:.3f}s")

Audio (16kHz WAV)
    │
    ▼
WavFrontend (80-dim fbank + LFR)
    │ 560-dim features
    ▼
SANM Encoder (SpecAugLFR + stride_conv)
    │
    ▼
CIF PredictorV2 ──► pre_acoustic_embeds, scama_mask
    │
    ▼
SANM Encoder2 (Big-Chunk)
    │
    ▼
FSMN-SCAMA Decoder + Beam Search (beam_size=5)
    │
    ▼
Token Decoding (CharTokenizer, 1002 vocab)
    │
    ▼
Vietnamese Text

uniASR_model/
├── inference.py          # NPU-adapted inference script
├── README.md             # This documentation
├── config.yaml           # Model architecture config
├── decoding.yaml         # Beam search config
├── model.pb              # PyTorch state_dict (~367 MB)
├── configuration.json    # ModelScope metadata
├── example/
│   └── asr_example.wav   # Test audio sample
└── inference_log.txt     # Accuracy comparison output

@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}
}