BS-RoFormer — Vocal Extraction

Unofficial HuggingFace packaging of BS-RoFormer (Band-Split RoFormer), using the ViperX checkpoint. Adds a from_pretrained interface and native Apple Silicon (MPS) support.

Extracts vocals from a music track. Outputs exactly 2 stems:

Stem	Description
vocals	Isolated vocal track
other	Everything else (mix − vocals)

Quick start

pip install transformers torch librosa soundfile \
            einops beartype rotary-embedding-torch

from transformers import AutoModel

model = AutoModel.from_pretrained("puar-playground/bs-roformer", trust_remote_code=True)
stems = model.separate("song.wav", output_dir="output/")
# output/vocals.wav  — isolated vocals
# output/other.wav   — instrumental

Device selection

# Auto-detect: CUDA → MPS (Apple Silicon) → CPU
model = AutoModel.from_pretrained("puar-playground/bs-roformer", trust_remote_code=True)

# Explicit
model = AutoModel.from_pretrained("puar-playground/bs-roformer", trust_remote_code=True,
                                  device="mps")   # or "cuda", "cpu"

Repository layout

bs-roformer/
├── modeling_bs_roformer.py   # BSRoFormerConfig + BSRoFormerForSourceSeparation
├── config.json
├── requirements.txt
├── bs_roformer.yaml          # model architecture config
├── bs_roformer.ckpt          # model weights (Git LFS)
└── bs_roformer_src/          # vendored BS-RoFormer source

Credits

• BS-RoFormer architecture: lucidrains/BS-RoFormer
• Checkpoint: BS-RoFormer (ViperX edition) from ZFTurbo/Music-Source-Separation-Training
• Paper: Music Source Separation with Band-Split RoPE Transformer (Chen et al., 2024)