home/ modules/ training-and-tuning-source-separation-models

Training and Tuning Source-Separation Models

L4Performance 6h K · AI & real-time generative AV 9 atoms

Learning objectives

learner can assemble a custom bag-of-models with per-source weights for a target stem set
learner can apply remix, BPM/pitch-aligned, and pitch/tempo augmentation to build realistic training data
learner can evaluate separation quality against a standard benchmark using SDR and the shift trick

Capstone — one whole task that evidences the objectives

Design and document a training recipe for an improved separation model targeting your genre: specify a custom model bag with per-source weights, an augmentation pipeline (remix + BPM/pitch alignment + on-the-fly pitch/tempo shift), the MUSDB-HQ benchmark protocol, and how the shift trick and SDR will be used to validate gains.

Prerequisite modules

Separating Stems from Any Track with Demucs

Off-the-shelf Demucs is tuned for MUSDB’s rock/pop centre of gravity. If you live-code UK garage, footwork, or dub techno, the stems you pull for resampling and live remixing inherit that bias: mushy sub-bass splits, hi-hat bleed into “other”. This module builds the whole task of fixing that — designing a complete, documented training recipe that would measurably improve separation for your genre, the kind of recipe you could hand to a GPU box (or a collaborator) and defend every choice in it.

The arc starts supported: reproduce a known configuration by writing a custom bag YAML with per-source weights (the atom on combining checkpoints into custom model bags is your JIT how-to), then run it against MUSDB-HQ tracks and score it, leaning on the shift-trick atom for cheap SDR gains at test time. From there you work up the data side: first plain remix augmentation, then the more musical BPM- and pitch-aligned cross-mixing, then consistent on-the-fly pitch/tempo shifting — each atom a concrete technique you slot into the pipeline. The unsupported capstone asks you to compose all of these into one coherent recipe targeted at your genre, with a validation protocol you specify in advance.

Every required atom gates that recipe: you cannot specify per-source weights without the bag mechanics, cannot justify the augmentation pipeline without the three augmentation concepts, and cannot define validation without MUSDB-HQ, SDR, and the shift trick. The supporting atoms enrich your judgement rather than gate it — knowing why official models are ensembles, and why hybrid waveform/spectrogram branches are complementary, helps you pick which checkpoints belong in your bag and anticipate where your genre will stress each domain.