BSRoformer-Wind-CoreML

The wind stem (brass and woodwinds) of the MVSep Mega 53-stem BS-RoFormer, converted to Core ML for slurper, where it produces the "horns" stem.

• bsr_wind_fp16.mlpackage: the graph frames[1,2,690,2048] -> recon[1,2,690,2048] for one 8 s chunk at 44.1 kHz, framed with a 2048-sample periodic Hann window at hop 512: windowed DFT, band split, the axial rotary transformer, mask estimator, complex mask multiply and inverse DFT, all inside the model in float16. The host reflect-pads and frames the chunk and overlap-adds the output divided by the summed squared window.
• golden_raw.f32, golden_horns.f32: an 8 s stereo chunk (channel-major float32) and PyTorch's output for it, used by slurper's tests. Core ML matches PyTorch on it at cosine 0.99998 (44.7 dB SDR).

Produced by scripts/convert_bs_roformer.py in the slurper repository from the single-stem repack noblebarkrr/BS-Roformer-MVSep-Mega-53-stems (v1/bs_mega_53stem_wind_mvsep.ckpt), with the model code from Music-Source-Separation-Training (MIT). See the upstream release for the weights' terms.

The golden chunk is 8 s from 0:32 of "Blues for Mundy" by The Airmen of Note, The United States Air Force Band (composed by Rick Whitehead), from 60 Years of the Airmen of Note (2011), via Wikimedia Commons. As a work of the US federal government the recording is in the public domain.