BS-RoFormer SW (6 stems), Core ML mask network
Core ML conversion (fp16, macOS 15+) of the BS-RoFormer SW six-stem checkpoint (bass, drums, other, vocals, guitar, piano), used by Gridshift's stem split.
Only the mask network is converted. Input spec [1, 2050, 1151, 2] is the
stereo STFT (n_fft 2048, hop 512, periodic Hann, reflect padding, 1025 bins,
channels interleaved per bin, real/imag last) of a 588 800-sample window at
44.1 kHz. Output mask [1, 6, 2050, 1151, 2] is one complex mask per stem in
the same layout; multiply with the spectrum and inverse-STFT to get the stems.
Mask error against the PyTorch checkpoint: -66 dB.
Licence
The original checkpoint's licence is unknown. It was published on Hugging Face by jarredou (BS-ROFO-SW-Fixed); that account is gone and its mirrors list "license: unknown". The BS-RoFormer architecture code is MIT; that licence does not cover the weights. No rights beyond those of the original are granted here.