---
title: usvc-pretrained
canonical_url: "https://www.modelscope.cn/models/baicai1145/usvc-pretrained"
md_url: "https://www.modelscope.cn/models/baicai1145/usvc-pretrained.md"
repository: baicai1145/usvc-pretrained
last_updated: 2026-08-13
pipeline_tag: text-to-speech
tasks:
  - text-to-speech
downloads: 1
stars: 0
---

# usvc-pretrained

> usvc-pretrained - baicai1145 在 ModelScope 开源的模型。Pretrained weights for usvc (Unified SVC framework).

- **Repository**: baicai1145/usvc-pretrained
- **Tasks**: text-to-speech
- **Downloads**: 1
- **Stars**: 0
- **Last updated**: 2026-08-13

Source: https://www.modelscope.cn/models/baicai1145/usvc-pretrained

---

# usvc-pretrained

Pretrained weights for [usvc](https://github.com/baicai1145/usvc) (Unified SVC framework).

## Base models

| Directory | Description |
|-----------|-------------|
| `contentvec/` | ContentVec encoder (HF format, 768-dim, layer 12) |
| `nsf_hifigan/` | PC-NSF-HiFiGAN vocoder (openvpi 2025.02+, pitch-conditioned) |
| `hubert/` | HubertSoft encoder (so-vits-svc 4.0, 256-dim) |
| `rmvpe/` | RMVPE F0 extractor |
| `fcpe/` | FCPE F0 extractor |

## SVC base models

### ddsp6

Organized by version:

```
ddsp6/
├── v6.1/          # ContentVec (hop 320)
│   ├── 10x1024/          # 10 layers × 1024 chans
│   ├── 10x1024_attn/     # + attention
│   ├── 10x1024_noise/    # + noise injection
│   ├── 10x1024_attn_noise/
│   ├── 6x2048/           # 6 layers × 2048 chans
│   ├── 6x2048_attn/
│   ├── 6x2048_noise/
│   └── 6x2048_attn_noise/
├── v6.3/          # ContentVec TTA-2x (hop 160), LYNXNet2
│   ├── 6x512_6x512/      # aux 6×512, velocity 6×512
│   ├── 6x512_6x768/
│   ├── 6x512_6x1024/
│   ├── 6x512_10x2048/
│   ├── 6x1024_6x1024/
│   ├── 10x1024_6x512/
│   ├── 10x1024_6x1024/
│   └── ...
└── v6.0/          # DDSP-SVC 6.0 (ContentVec, pre-6.1)
    ├── 12x1024/
    └── 6x512/
```

### rift_dit

```
rift_dit/
├── dit-512-8.ckpt    # 512-dim, 8 layers
├── dit-768-12.ckpt   # 768-dim, 12 layers
└── dit-1024-16.ckpt  # 1024-dim, 16 layers
```

### sovits

```
sovits/
├── cvec768l12_2023/      # ContentVec 768-dim, layer 12 (2023)
└── cvec768l12_20241221/ # BigDataset, 131 speakers (2024.12.21)
```

### diffusion_shallow

```
diffusion_shallow/
└── contentvec768l12/   # Diffusion shallow-diffusion enhancer base
                         # (for sovits/rvc, not standalone inference)
```

### rvc

```
rvc_v1/             # RVC v1 (HubertSoft 256-dim)
├── 32k_f0/
├── 40k_f0/
└── 48k_f0/

rvc_v2/             # RVC v2 (ContentVec 768-dim)
├── 32k_f0/
├── 40k_f0/
└── 48k_f0/
```

### reflow_vae

```
reflow_vae/
└── contentvec768l12/
```

## Download

```bash
# HuggingFace
huggingface-cli download usvc/usvc-pretrained --local-dir pretrained

# ModelScope
modelscope download --model usvc/usvc-pretrained --local_dir pretrained
```