README.md

---
license: mit
tags:
- music-source-separation
- bs-roformer
- gguf
- stem-separation
- source-separation
- pytorch
- transformers
---
# BS-Roformer Multi-Stem Model (GGUF Q8)

A quantized GGUF version of the BS-Roformer model, optimized for C++ inference via [Fork-BSRoformer.cpp](https://github.com/victorlfdev/Fork-BSRoformer.cpp). Supports multi-stem separation (drums, bass, vocals, others).

## Model description

BS-Roformer (Band Split RoFormer) is a neural network architecture for music source separation. It splits the audio frequency spectrum into multiple bands and processes each band independently using a transformer encoder with self-attention mechanisms. The model was originally derived from [Suno AI's Bark](https://github.com/suno/bark) project (text-to-music generation), where it was used internally for music source separation.

This checkpoint has been quantized to Q8 format using the GGUF library for efficient inference in C++ environments, reducing memory usage while maintaining high separation quality.

### Training Data

This model was trained using the framework from [Music Source Separation Training](https://github.com/ZFTurbo/Music-Source-Separation-Training), which is a PyTorch-based training framework for music source separation models. The training data consists of publicly available music datasets used by the community training efforts documented in that repository.

## Original Research

- **BS-Roformer (Band Split RoFormer)**: Architecture derived from Suno AI's Bark project for music source separation
- **Music Source Separation Training**: [ZFTurbo/Music-Source-Separation-Training](https://github.com/ZFTurbo/Music-Source-Separation-Training)

## Credits

| Credit | Link |
|--------|------|
| **Suno AI** | Original creators of the BS-Roformer / Mel-Band-Roformer architecture via the Bark project | [GitHub](https://github.com/suno/bark) |
| **ZFTurbo (Vladislav Sukachov)** | Music Source Separation Training framework and community model training | [GitHub](https://github.com/ZFTurbo/Music-Source-Separation-Training) |
| **anvuew** | Trained BS-RoFormer checkpoint (SDR 12.45) | [HuggingFace](https://huggingface.co/anvuew/BS-RoFormer) |
| **GaboxR67** | Mel-Band-Roformer checkpoints | [HuggingFace](https://huggingface.co/GaboxR67/MelBandRoformers) |
| **沉默の金 (chenmozhijin)** | Fork-BSRoformer.cpp — C++ GGUF inference engine | [GitHub](https://github.com/chenmozhijin/Fork-BSRoformer.cpp) |
| **ggerganov** | GGML library for efficient tensor computation | [GitHub](https://github.com/ggerganov/ggml) |
| **dr_libs** | Lightweight audio decoding library | [GitHub](https://github.com/mackron/dr_libs) |

## Usage

### Via Fork-BSRoformer.cpp

Download the compiled binary and run:

```bash
./bs_roformer-cli -m bs-roformer-multi-q8.gguf -a input.wav -o output.wav
See Fork-BSRoformer.cpp (https://github.com/victorlfdev/Fork-BSRoformer.cpp) for full CLI options and usage.
Via Python
from bs_roformer_cpp_cli import BsRoformerCppCLI

cli = BsRoformerCppCLI(model_path="./bs-roformer-multi-q8.gguf", device="cuda")
cli.process("input.wav", "output.wav")
Model Architecture
- Type: Band Split RoFormer (transformer-based music source separator)
- Quantization: Q8 (8-bit uniform quantization via GGUF)
- Stems: 4 (drums, bass, vocals, other)
- Input: Mono/stereo WAV audio (any sample rate, resampled internally)
- Output: 4-channel separated stems (WAV format)
License
This model is shared for research and educational purposes. The underlying BS-Roformer architecture and training methodology are derived from community efforts referenced above. Redistribution of trained weights should comply with the original training data licenses.
Acknowledgements
- ggerganov/ggml (https://github.com/ggerganov/ggml) — Efficient tensor library
- ZFTurbo/Music-Source-Separation-Training (https://github.com/ZFTurbo/Music-Source-Separation-Training) — PyTorch reference implementation
- dr_libs (https://github.com/mackron/dr_libs) — Lightweight audio library