schism-audio/mini-bs-roformer-v2-coreml

Mini-BS-RoFormer V2 (46.8M) — Core ML

HiDolen/Mini-BS-RoFormer-V2-46.8M (2025 Band-Split RoPE Transformer with dual-resolution STFT, 4-stem music source separation) for Core ML on Apple devices. CC-BY-NC-4.0 (non-commercial) — for commercial use, see the MIT-licensed 18M or HTDemucs Core ML ports in this org.

Converted from the same verified reference used by the schism-mlx MLX ports. Two variants per model:

File	Precision	Compute units	chunk residual
BSRoformerV2_Core_fp16.mlpackage	FLOAT16	ALL (ANE)	−72.9 dB
BSRoformerV2_Core_fp32.mlpackage	FLOAT32	CPU+GPU	−132.7 dB

Verified on-device-equivalently via coremltools on an M5 Max, against the reference implementation on real audio. fp16 is ANE-eligible and recommended for iPhone / iPad; fp32 is the tight-parity fallback.

Download

.mlpackage bundles must be materialized as real files — the Core ML compiler rejects the symlinks that a default snapshot_download creates in the Hugging Face cache:

from huggingface_hub import snapshot_download
path = snapshot_download("schism-audio/mini-bs-roformer-v2-coreml", local_dir="./mini-bs-roformer-v2-coreml")

(or hf download schism-audio/mini-bs-roformer-v2-coreml --local-dir ./mini-bs-roformer-v2-coreml). Swift hosts downloading files directly are unaffected.

I/O contract

• input spec: (1, 690, 8196) float32 — merged analysis STFT of one 8 s / 352800-sample chunk: n_fft 4096, hop 512, hann, unnormalized, layout (C, 2049, 690) complex → (690, 2049*2*2) with frequency outer, channel, re/im innermost
• output mask: (1, 4, 690, 4100) float32 — per-stem complex mask at the synthesis resolution (F=1025); unmerge → complex, multiply with a second STFT of the same chunk at n_fft 2048 (same hop/window), iSTFT (n_fft 2048, hop 512, unnormalized) → stems — the host runs two STFTs per chunk
• full tracks: 8 s / 352800-sample chunks, 50% overlap, linear fade window (fade = chunk/10), fold-normalized overlap-add (see separate() in schism-mlx bs_roformer/model.py)
• stem order is bass / drums / other / vocals — not the Demucs order

DSP frontend (host-side)

The Core ML graph contains the network only. The host implements the audio frontend and must match schism_mlx.audio numerically — test_vectors_*.npz in this repo holds deterministic input/output pairs (float32; match within ~1e-4 relative to be interchangeable with what these models were verified against). test_vectors_pipeline.npz additionally pins the full-track orchestration around the model call (chunk grid, blending, normalization) via a deterministic mock model — scaled-down sizes, same algorithm, generated by the shipped schism-mlx separate() / windowing code. test_vectors_integration.npz goes one step further: a full synthetic track plus the verified reference output, for end-to-end host validation (the schism-dsp integration tests run the published fp32 core through it). A validated Swift implementation (Accelerate; modules SchismDSP and SchismPipeline) is available at schism-audio/schism-dsp, tested against these exact vectors.

License

CC-BY-NC-4.0, inherited from the source weights (weights cast bf16 → fp32/fp16 at conversion). Model by HiDolen; Core ML conversion by schism-audio.