F5-TTS Emilia-ZH-EN Ascend NPU 适配

# 验证 NPU 可用性
python -c "import torch; import torch_npu; print(torch.npu.device_count())"
# 输出: 2

# 验证 NPU 型号
python -c "import torch; print(torch.npu.get_device_name(0))"
# 输出: Ascend910_9362

pip install f5-tts safetensors omegaconf soundfile torch_npu

# 从 ModelScope 下载
pip install modelscope
modelscope download --model SWivid/F5-TTS_Emilia-ZH-EN --local_dir ./ckpts/F5TTS_v1_Base

# 基本推理命令
python inference.py \
  --ref_audio /path/to/reference.wav \
  --ref_text "参考音频对应的文本内容。" \
  --gen_text "这是要在昇腾NPU上生成的语音合成文本。" \
  --output output.wav

# 完整参数示例
python inference.py \
  --ref_audio /opt/atomgit/pretrained_models/CosyVoice-300M/asset/zero_shot_prompt.wav \
  --ref_text "希望你以后能够做的比我还好呦。" \
  --gen_text "在当今快速发展的科技时代，人工智能正在深刻改变着我们的生活方式。" \
  --output output.wav \
  --nfe_step 32 \
  --cfg_strength 2.0 \
  --speed 1.0

from inference import patch_mel_spec_for_npu, patch_vocoder_for_npu, load_model_npu
from f5_tts.infer.utils_infer import load_vocoder
import soundfile as sf

# 加载模型
model = load_model_npu(
    ckpt_path="ckpts/F5TTS_v1_Base/model_1250000.safetensors",
    vocab_file="ckpts/F5TTS_v1_Base/vocab.txt",
    device="npu:0",
)
vocoder = load_vocoder(vocoder_name="vocos", is_local=False, local_path="", device="npu:0")
patch_vocoder_for_npu(vocoder)

# 推理... (参考 inference.py 中的 infer_on_npu 函数)

def patch_mel_spec_for_npu(model):
    """将 MelSpec 计算移至 CPU 执行，规避 NPU stft 系列算子问题。"""
    original_forward = model.mel_spec.forward
    dummy = torch.randn(1, 1200).float()
    try:
        original_forward(dummy)  # 预热 CPU 缓存
    except Exception:
        pass

    def patched_forward(wav):
        orig_device, orig_dtype = wav.device, wav.dtype
        wav_cpu = wav.detach().cpu().float()
        mel_cpu = original_forward(wav_cpu)
        return mel_cpu.to(device=orig_device, dtype=orig_dtype)
    model.mel_spec.forward = patched_forward

def patch_vocoder_for_npu(vocoder):
    """将 Vocos ISTFT 移至 CPU 执行，规避 aclnnUnfoldGrad bug。"""
    try:
        istft_module = vocoder.head.istft
    except AttributeError:
        return
    istft_module.to("cpu")
    original_forward = istft_module.forward
    def patched_forward(spec):
        spec_cpu = spec.detach().cpu()  # 保持 complex dtype
        result = original_forward(spec_cpu)
        return result.to(device=spec.device)
    istft_module.forward = patched_forward

def load_audio(path, target_sr=None):
    """使用 soundfile 加载音频（替代 torchaudio.load）。"""
    data, sr = sf.read(path)
    if data.ndim == 1:
        data = data.reshape(1, -1)
    audio = torch.from_numpy(data.astype(np.float32))
    if target_sr is not None and sr != target_sr:
        from scipy import signal
        new_len = int(audio.shape[-1] * target_sr / sr)
        audio = torch.from_numpy(signal.resample(audio.numpy(), new_len, axis=-1).astype(np.float32))
    return audio, (target_sr or sr)

# 完整评测（精度 + 性能）
python evaluate_npu.py \
  --ckpt_path ckpts/F5TTS_v1_Base/model_1250000.safetensors \
  --ref_audio /path/to/real_reference_speech.wav \
  --output_dir ./eval_results \
  --warmup 2 \
  --benchmark 5

# 仅性能评测
python evaluate_npu.py --skip_accuracy --warmup 2 --benchmark 5

# 仅精度评测
python evaluate_npu.py --skip_performance

# 查看评测报告
cat eval_results/evaluation_report.json

@misc{f5tts,
  title={F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching},
  author={Yushen Chen and Zhikang Niu and et al.},
  year={2024},
}