Informed / Model-Based Source Separation

Definition

Source separation guided by external side information — a musical score, MIDI file, user annotations, or prior knowledge about the sources — rather than operating purely blind from the mixture.

Key Ideas

• Score-informed: use sheet music or MIDI as a guide — know what notes each instrument plays and when. Simplifies separation dramatically.
• Model-based: prior parametric models of instrument sound (e.g., harmonic series, formant structure) constrain the separation.
• Contrasts with blind source separation (no prior information) and query-based separation (open-vocabulary text query).
• Practical when side information is available — e.g., studio recordings where MIDI tracks exist, or transcription tasks where a rough transcription can bootstrap separation.

Relationships

• Contrasts with ../concepts/query-based-source-separation — informed uses structured data, query-based uses open-vocabulary text
• Related to ../concepts/permutation-invariant-training — informed separation doesn't have the permutation problem
• Practical pipeline: informed separation -> transcription -> refined separation

Sources

None ingested yet — seed batch setup.