wav2vec2-indonesian-javanese-sundanese on Ascend NPU

# 方法1：使用hf-mirror镜像
export HF_ENDPOINT=https://hf-mirror.com
python3 -c "
from huggingface_hub import snapshot_download
snapshot_download(
    'indonesian-nlp/wav2vec2-indonesian-javanese-sundanese',
    local_dir='./wav2vec2-indonesian-javanese-sundanese'
)
"

# 方法2：直接使用 transformers（需要设置 HF_ENDPOINT）
export HF_ENDPOINT=https://hf-mirror.com
python3 -c "
from transformers import Wav2Vec2ForCTC, Wav2Vec2ProcessorWithLM
model = Wav2Vec2ForCTC.from_pretrained('indonesian-nlp/wav2vec2-indonesian-javanese-sundanese')
processor = Wav2Vec2ProcessorWithLM.from_pretrained('indonesian-nlp/wav2vec2-indonesian-javanese-sundanese')
"

# 基础依赖
pip install torch_npu transformers soundfile librosa

# kenlm 需从源码安装 (用于4gram语言模型解码)
pip install https://github.com/kpu/kenlm/archive/master.zip

# 单文件推理 (NPU后端)
python3 inference.py --audio audio.wav --backend npu

# CPU后端 (精度对比)
python3 inference.py --audio audio.wav --backend cpu

# 长音频分块推理
python3 inference.py --audio long_audio.wav --chunk-length-s 10.0

# 详细模式
python3 inference.py --audio audio.wav --verbose

import torch
import torch_npu
import numpy as np
import soundfile as sf
from transformers import Wav2Vec2ForCTC, Wav2Vec2ProcessorWithLM

MODEL_PATH = "./wav2vec2-indonesian-javanese-sundanese"

# 加载模型
processor = Wav2Vec2ProcessorWithLM.from_pretrained(MODEL_PATH)
model = Wav2Vec2ForCTC.from_pretrained(MODEL_PATH).npu().eval()

# 加载音频
audio, sr = sf.read("audio.wav")
if len(audio.shape) > 1:
    audio = audio.mean(axis=1)
if sr != 16000:
    import librosa
    audio = librosa.resample(audio, orig_sr=sr, target_sr=16000)

# NPU推理
inputs = processor(audio.astype(np.float32), sampling_rate=16000, return_tensors="pt", padding=True)
with torch.no_grad():
    logits = model(inputs.input_values.npu()).logits.cpu().numpy()

# CTC解码 (含4gram语言模型)
text = processor.decode(logits[0]).text
print(f"识别结果: {text}")

# 验证NPU可用
python3 -c "import torch_npu; print(torch.npu.is_available(), torch.npu.get_device_name(0))"

# 模型加载测试
python3 -c "
from transformers import Wav2Vec2ForCTC
model = Wav2Vec2ForCTC.from_pretrained('./wav2vec2-indonesian-javanese-sundanese')
print(f'Params: {sum(p.numel() for p in model.parameters()):,}')
"

# 运行完整评测
cd eval && python3 benchmark.py