WeNet U2++ Conformer AISHELL ASR (离线版) - 昇腾 NPU 适配

# 安装依赖
pip install torch torchaudio torch_npu
pip install onnx onnx2torch onnxruntime
pip install soundfile librosa modelscope numpy Pillow

python3 inference.py --device cpu

python3 inference.py --device npu

python3 compare_cpu_npu.py

import torch
from onnx2torch import convert as onnx2torch_convert
import torchaudio.compliance.kaldi as kaldi
import soundfile as sf

# 1. 加载 ONNX 模型并转换为 PyTorch
encoder = onnx2torch_convert("encoder.onnx").eval()
ctc = onnx2torch_convert("ctc.onnx").eval()

# 2. 将模型迁移到 NPU（如果可用）
if torch.npu.is_available():
    encoder = encoder.npu()
    ctc = ctc.npu()

# 3. 提取音频 FBank 特征
waveform, sr = sf.read("audio.wav")
waveform = torch.from_numpy(waveform).float().unsqueeze(0)
waveform = waveform * (1 << 15)
feats = kaldi.fbank(waveform, num_mel_bins=80, frame_length=25,
                    frame_shift=10, dither=0.0, sample_frequency=16000)
# feats: [T, 80]

# 4. 编码器推理
chunk = feats.unsqueeze(0)  # [1, T, 80]
offset = torch.tensor([0], dtype=torch.long)
att_cache = torch.zeros(12, 4, 0, 128)
cnn_cache = torch.zeros(12, 1, 256, 7)

if torch.npu.is_available():
    chunk = chunk.npu()
    offset = offset.npu()
    att_cache = att_cache.npu()
    cnn_cache = cnn_cache.npu()

with torch.no_grad():
    encoder_out = encoder(chunk, offset, att_cache, cnn_cache)[0]

# 5. CTC 解码获取字符概率
with torch.no_grad():
    ctc_probs = ctc(encoder_out)

# 6. CTC 贪心解码
max_idx = torch.argmax(ctc_probs, dim=-1).squeeze(0).tolist()
# 去除 blank 和重复
result = []
prev = -1
for idx in max_idx:
    if idx != prev and idx != 0:  # 0 = blank
        result.append(idx)
    prev = idx
# 映射为文字
char_dict = {0: '', 1: '<unk>', 2: '一', ...}
text = ''.join([char_dict.get(i, '') for i in result])
print(text)

# NPU 推理
python3 inference.py --device npu

# CPU 推理
python3 inference.py --device cpu