cstr/mini-bs-roformer-GGUF

Mini-BS-RoFormer V2.5 — GGUF

GGUF conversions of HiDolen/Mini-BS-RoFormer-V2.5-46.2M (46.2M parameter source separation model) for use with CrispASR.

Mini-BS-RoFormer V2.5 separates 44.1 kHz stereo audio into 4 stems:

1. bass
2. drums
3. other
4. vocals

Runs 100% natively in C++ via ggml in CrispASR — zero Python, zero PyTorch at inference.

Artifacts

File	Precision	Size	Description
mini-bs-roformer-f16.gguf	F16	~88 MB	Full-precision baseline
mini-bs-roformer-q8_0.gguf	Q8_0	~50 MB	High-fidelity quantized model (recommended)
mini-bs-roformer-q4_k.gguf	Q4_K	~37 MB	Fast, compact quantized model

Usage with CrispASR

CLI

# Separate all 4 stems (bass, drums, other, vocals)
crispasr --separate -m mini-bs-roformer-q8_0.gguf -f song.wav

# Separate only vocals stem
crispasr --separate -m mini-bs-roformer-q4_k.gguf -f song.wav --stems vocals

# Auto-download from HuggingFace via CrispASR registry
crispasr --separate -m mini-bs-roformer -f song.wav

Python API

from crispasr import Session

# Load model (stereo 44.1 kHz)
with Session('mini-bs-roformer-q8_0.gguf') as s:
    stems = s.separate(audio_pcm)  # returns dict with bass, drums, other, vocals

Original Model & License

• Original model: HiDolen/Mini-BS-RoFormer-V2.5-46.2M
• License: MIT