dsh-ai-voice-cover
vamyane/dsh-ai-voice-cover
AI 翻唱合成:给一首歌加约 10 秒参考人声,产出用该音色翻唱的干声与成品;只换音色不写旋律,所以旋律严格对齐原唱。支持零样本即时出歌、音色库建档复用、自动半音移调并同步伴奏,自带隔离 venv,GPU 加速、纯 CPU 也能跑。
安装
dsh plugin --profile web add github:vamyane/dsh-ai-voice-coverREADME
dsh-ai-voice-cover
AI voice cover for DeepSeek Harness. One install gives the agent the ai-voice-cover skill:
feed it a song plus about 10 seconds of reference audio and it produces a cover sung in that voice.
What it does
Input:
- A target song — A the original track with vocals, or B accompaniment + guide vocal
- Reference audio (≈10s, any content; or a whole clean recording library for fine-tuning)
Output:
- Dry vocal
output/dry_vocal.wav— the vocal converted to the reference timbre, no accompaniment - Final song
output/final_song.wav— dry vocal mixed with the accompaniment - Manifest
output/manifest.json— auto-resolved parameters and engine path
The guarantee: the melody stays aligned with the original — only timbre is swapped; pitch and rhythm come from the source performance.
Capabilities
- Zero-shot covers from 10 seconds, or archive a clean recording set as a reusable voice
- Automatic semi-tone shift with accompaniment sync, so a female voice is not sung back as a male one
- Dual engine with automatic fallback: Seed-VC first, GPT-SoVITS on failure
- Isolated Python venv (never pollutes system Python), GPU accelerated and CPU capable
- Pre-delivery self-check: duration, stereo, chroma melody similarity, pitch-range fit, no clipping
Install
dsh plugin --profile desktop add dsh-ai-voice-cover
Refresh the page (or restart DSH) and ai-voice-cover appears in the skill catalog. The plugin market
installs it the same way.
Environment
| Component | Version | Notes |
|---|---|---|
| DeepSeek Harness | >= 0.2.2-rc.1 < 0.3.0-0 | Uses the ctx.skills registry |
| Python | 3.10+ | The skill creates its own isolated venv |
| ffmpeg | recent | Bootstrapped on first setup, or use the system one |
| GPU | optional | CUDA is faster; CPU works with reduced diffusion steps |
First setup downloads several GB of model weights (Seed-VC / GPT-SoVITS).
How it is built
- Skill-bundle plugin: no UI, no
/apiroutes; one bundled skill provider. - Never shadows your skills: bundled rank 600, so a same-named skill in a project or
~/.dsh/skillswins. - The plugin itself never downloads anything: all downloads happen when the user runs the skill's setup script.
- Zero runtime dependencies: node built-ins only.
License
MIT for the plugin and the skill's code/docs. Third-party engines and models downloaded at runtime
(Seed-VC, GPT-SoVITS, demucs, PyTorch, librosa, …) remain under their own licenses — see NOTICE.md
and the skill's DISCLAIMER.md.