Expand description
Learned (demucs / open-unmix / band-split RNN class) source separation — fail-closed stub.
A learned separator estimates per-source masks (or waveforms) from a neural network whose
weights are not present in this build. Per the project’s honesty rule, this module must never
fabricate stems: with no weights loaded there is no faithful separation to return, so the only
correct behaviour is to abstain. separate_learned therefore always returns
AudioError::BackendUnavailable (the NeedsWeights signal).
When real weights are wired (a P64 head, gated on the principal per ADR 007 / delivery plan
Waves 7–8), this stub is replaced by the actual inference path. Until then, callers get an
explicit, honest failure rather than plausible-looking noise. For weight-free separation use
the classical masking in crate::models::separation::apply_soft_mask.
Functions§
- separate_
learned - Attempt learned, demucs-class source separation of a mixture magnitude spectrum into
n_stemsstems written contiguously intoout_stems(n_stems × mixture_mag.len()).