---

name: scienceclaw-benchmark-for36 description: "Use when a task asks to separate stereo music mixtures into four stems (vocals, drums, bass, other) and deliver the estimated stems, optionally with training mixtures plus reference stems, judged by BSSEval-v4 source-to-distortion ratio (SDR, dB), as in MUSDB18 music source separation (FoR36 of the companion ScienceClaw-Eval benchmark)." metadata: { "openclaw": { "emoji": "📊" } }

Music source separation: vocals / drums / bass / other

Task

• Input: stereo mixtures (k, n, 2) (linear PCM, sample rate given by the task); optionally training mixtures with stems (m, 4, n, 2).
• Deliverable: float estimates (k, 4, n, 2), axis 1 in the order (vocals, drums, bass, other), same rate and length as the input, on the amplitude scale of the mixture. Finite, and no estimated target exactly zero over a whole 1-s window (the metric skips such windows).
• Quality: mean over the four targets of the median over items of the median over 1-s windows of SDR = 10 log10(sum ref^2 / sum (est - ref)^2), dB, higher is better (scilib.audiosep.sdr_scores).

Routes

• Frozen neural separators: scilib.audiosep_pretrained.separate_pretrained(mixtures, model, sample_rate) with htdemucs, htdemucs_ft or mdx_extra (asset demucs); scilib.scnet_pretrained.separate_pretrained(mixtures, model="mimo_scnet_small") (asset scnet_mimo_small). Operators music_source_separation_htdemucs, music_source_separation_htdemucs_ft, music_source_separation_scnet. Check audiosep_pretrained.available(model), scnet_pretrained.available() and scienceclaw_tools(operation=weights) first.
• Trained CPU baseline: audiosep.separate(train_mixtures, train_stems, mixtures) (STFT ratio masks from boosted trees; operator music_source_separation_softmask; at least 2 training excerpts). audiosep.cross_validate(..., groups=track_ids) gives grouped k-fold SDR. Score with audiosep.sdr_scores (operator music_separation_sdr).

Pitfalls

• SDR is not scale invariant: an almost-zero estimate scores about 0 dB and copying the mixture into every stem scores strongly negative, so a level change alone moves the score. audiosep.gain_only and cross_validate report mixture_sdr, null_sdr and gain_only_sdr beside the separator; separate level effect from separation and never submit near-silent stems.
• Pretrained overlap: htdemucs was trained on MUSDB18-HQ plus further songs; assume the other Demucs bags overlap likewise. MUSDB18 training tracks (including visible training or validation excerpts) were seen in training and score higher than unseen songs; disclose this, and check the training data of any other checkpoint (e.g. SCNet) before calling it clean.
• Never use reference stems or oracle masks of the items being separated to build the deliverable; keep all excerpts of a track in one fold.
• The wrappers already return the input length and the order (vocals, drums, bass, other); do not permute again. A silent reference stem gives NaN windows that the median ignores.