Synthetic Mixing / Data Augmentation Pipelines

Definition

Creating training data for source separation by artificially mixing isolated stems with randomized gains, effects, and spatialization. Multiplies available training data and improves generalization.

Key Ideas

• Start with isolated stems (from datasets like MUSDB18, Slakh2100, or solo recordings).
• Apply random gains, panning, EQ, reverb (via ../entities/pedalboard, ../entities/echothief, ../entities/openair) to each stem.
• Sum into mixtures — ground truth stems are known by construction.
• Slakh2100 takes this further — synthesizes stems from MIDI with virtual instruments, then mixes.
• Critical for domains with limited multi-track data (e.g., guitar transcription from GuitarSet).

Relationships

• Uses ../entities/pedalboard, ../entities/echothief, ../entities/openair
• Related to ../concepts/mixture-invariant-training — MixIT can train without isolated stems
• Datasets: ../entities/musdb18, ../entities/slakh2100, ../entities/guitarset

Sources

None ingested yet — seed batch setup.