Paraformer-zh-streaming (NPU Adapted)

# 安装依赖
pip install funasr torch_npu soundfile librosa editdistance

# 安装 ModelScope (用于模型下载)
pip install modelscope

# ModelScope SDK 下载
modelscope download --model shuai1618/paraformer-zh-streaming --local_dir ./paraformer-zh-streaming

# 或通过 Git
git clone https://www.modelscope.cn/shuai1618/paraformer-zh-streaming.git

from funasr.auto.auto_model import AutoModel
import soundfile as sf

# 加载模型到 NPU
model = AutoModel(
    model="./paraformer-zh-streaming",
    device="npu:0",    # 使用昇腾 NPU
    disable_pbar=True,
)

# 加载音频 (16kHz, mono)
audio, sr = sf.read("your_audio.wav")

# 推理
cache = {}
result = model.generate(input=audio, cache=cache, is_final=True)
print(f"识别结果: {result[0]['text']}")

# 运行推理
python inference.py --audio your_audio.wav --device npu:0 --output result.json

# 运行完整评测
python benchmark.py --audio your_audio.wav --output benchmark_result.json --runs 10

CPU Reference : 嗯 (与 NPU 输出完全一致)
NPU Output    : 嗯
CER           : 0.000000 (0.0000%)

paraformer-zh-streaming/
├── inference.py          # NPU 推理脚本
├── benchmark.py          # 精度/性能评测脚本
├── benchmark_result.json # 评测结果
├── inference_result.json # 推理结果
├── config.yaml           # 模型配置
├── configuration.json    # 模型元信息
├── model.pt              # 模型权重
├── tokens.json           # 词表
├── seg_dict              # 分词词典
├── am.mvn                # 均值方差归一化
└── README.md             # 本文档

@inproceedings{gao2023paraformer,
  title={Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition},
  author={Gao, Zhifu and Zhang, Shiliang and Lei, Ming and Chen, Xie},
  booktitle={INTERSPEECH},
  year={2023}
}