belentani7 Shao - High-Fidelity Music Generation
Open-source system for high-fidelity music generation using unified acoustic-token pipeline. Generates complete musical works from text descriptions and lyrics.
When to Use
- User needs the highest quality open-source music generation
- User wants complete musical works (not just loops)
- User wants text + lyric controlled music
- User has NVIDIA GPU with 24GB+ VRAM
Key Features
- Unified acoustic-token representation
- 64-layer RVQ acoustic token hierarchy
- Two-stage generation pipeline
- Complete musical works (not short clips)
- Text and lyric control
- Coarse-to-fine acoustic detail generation
Requirements
- NVIDIA GPU with 24GB+ VRAM (RTX 4090 or higher)
- Docker and NVIDIA Container Toolkit
- CUDA-compatible NVIDIA driver
- Python and Node.js
Installation
git clone https://github.com/Shao-Music-AI/Shao.git
cd Shao
# Follow ENVIRONMENT_SETUP.md for Docker setup
Quick Start
# Docker-based inference
docker build -t shao .
docker run --gpus all -it shao
Architecture
| Stage | Description |
|---|---|
| Backbone | Generates coarse acoustic tokens |
| Super-Resolution | Completes higher RVQ token layers |
| Decoder | Reconstructs audio waveform |
Online Demo
Model Weights
- HuggingFace: liujiafeng/Shao-MusicGeneration-v1.0
- arXiv: https://arxiv.org/abs/2605.01790
Related
song-generation: Tencent SongGeneration/LeVoheartmula: HeartMuLa music modelsrvc: Voice conversion and singing