🎚️ moozz · stem-separator dev/ops bench moozz-lab registry · link pending v…

🎚️ Moozz · Stem Test Bench

Split a track into vocals and instrumental with Demucs on GPU. Pick the model and quality settings, optionally trim a segment, and review each run's parameters, per-step timings and output waveforms.

New test

Parameter guide
  • Modelhtdemucs_ft averages 4 fine-tuned sub-models: best quality, ~4× slower. htdemucs is a single model: still great and ~4× faster.
  • shifts — number of random-shift passes that get averaged. 0–1 = fastest; 5 (default) = cleaner separation but runtime scales roughly linearly.
  • overlap — overlap between processing windows (0–0.95). 0.8 (default) reduces edge artifacts (a faint "echo") at the cost of speed.
  • Segment — drag on the waveform to process only a slice; great for fast iteration on a chorus or a problem section.
  • Backing vocal split — splits the lead vocal from backing/harmony vocals; the backing vocal is kept as its own additive stem_vocal_backing, never discarded. Off by default — listen and decide per run.
  • Denoise — removes the noise floor from the vocal; the removed noise is kept as additive stem_noise. Off by default — listen and decide per run.
  • Dereverb strength — WPE-based reverb/echo removal from the vocal (0 disables); the removed wet tail is kept as additive stem_reverb. 1 (default, full strength) — listen and decide per run.
Examples:

Tests