sathvik77/htdemucs-core-onnx

htdemucs core — ONNX (external STFT)

Demucs v4 htdemucs (4 stems) exported to ONNX without the STFT/iSTFT (complex FFT ops are not ONNX-exportable). The caller computes the STFT and iSTFT outside the graph — see export_onnx.py / verify_onnx.py in the karoke-maker repo for the reference implementation and the parity gate this model passed (0.00% rel err vs PyTorch, voice-leak 0.055 == PyTorch reference).

I/O contract

• mix float32 (1, 2, 343980) — raw 44.1kHz stereo waveform segment (7.8s)
• mag float32 (1, 4, 2048, 336) — CaC spectrogram of the same segment (nfft 4096, hop 1024, periodic Hann, torch.stft normalized=True, channels ordered [L.re, L.im, R.re, R.im])
• spec_out float32 (1, 4, 4, 2048, 336) — per-source CaC spectrograms
• wave_out float32 (1, 4, 2, 343980) — per-source time-branch waveforms

Sources: [drums, bass, other, vocals]. Reconstruction: stems = istft(spec_out) + wave_out.

Weights: Meta AI htdemucs release (CC-BY-NC 4.0).