shiro_voice (昇腾 NPU 适配版)

pip install torch==2.9.0
pip install torch_npu==2.9.0.post1
pip install librosa
pip install pyworld
pip install soundfile

# 通过 ModelScope 下载
pip install modelscope
python -c "from modelscope import snapshot_download; snapshot_download('Y39c5bb/shiro_voice', local_dir='./models/Y39c5bb/shiro_voice')"

import torch
import torch_npu
from sovits_model import Synthesizer

DEVICE = "npu:0" if torch.npu.is_available() else "cpu"

# 加载模型
checkpoint = torch.load("models/Y39c5bb/shiro_voice/sovits/G_0.pth",
                         map_location="cpu", weights_only=False)
model = Synthesizer()
model.load_state_dict(checkpoint["model"], strict=True)
model = model.to(DEVICE)
model.eval()

# 准备输入 (Content features, F0, UV)
c = torch.randn(1, 256, 100)         # 语音内容特征
f0 = torch.rand(1, 1, 100)           # 基频 (F0)
uv = torch.zeros(1, 100, dtype=torch.long)  # 有声/无声标记
g = torch.randn(1, 256, 1)           # 说话人嵌入

with torch.no_grad():
    audio, _ = model(c, f0, uv, g=g, noise_scale=0.667)

print(f"生成音频: {audio.shape}")  # [1, 1, N*512]

# 测试推理 (随机输入)
python inference_npu.py --test-only

# 语音转换
python inference_npu.py --input input.wav --output output.wav

# CPU 基线
python inference_npu.py --cpu --test-only

python evaluate_accuracy.py

python benchmark_perf.py

Synthesizer
├── pre: Conv1d(256→192)
├── emb_uv: Embedding(2→192)
├── enc_p (PitchEncoder)
│   ├── f0_emb: Embedding(256→192)
│   ├── enc_ (ConformerEncoder ×6)
│   │   ├── RelativeAttention (2 heads, window=4)
│   │   └── FFN (192→768→192)
│   └── proj: Conv1d(192→384)
├── flow (ResidualCouplingBlock ×4)
│   └── WN (WaveNet, dilation=1)
├── f0_decoder
│   ├── SimpleAttention (2 heads, 6 layers)
│   └── proj: Conv1d(192→1)
└── dec (HiFiGANGenerator)
    ├── ConvTranspose1d ×5 (upsample: 8,8,2,2,2)
    ├── ResBlock ×15 (kernels: 3,7,11)
    └── conv_post: Conv1d(512→1)

@misc{sovits-svc,
  author = {SoVITS Community},
  title = {SoVITS-SVC: Singing Voice Conversion},
  year = {2024},
}