Topic: Classical & Pre-Deep-Learning Techniques

Overview

Foundational approaches to source separation and audio analysis that predate or complement deep learning. Covers auditory scene analysis (psychological and computational), sinusoidal modeling, matrix factorization methods, and model-based informed separation.

Sub-topics / Concepts

• ../concepts/auditory-scene-analysis — Bregman Auditory Scene Analysis (ASA): psychological framework for how humans parse complex auditory scenes.
• ../concepts/computational-auditory-scene-analysis — CASA: computational implementation of ASA principles.
• ../concepts/common-fate-principle — Gestalt principle: components changing in synchrony perceived as one source.
• ../concepts/sinusoidal-modeling — McAulay-Quatieri (1986). Audio as sum of time-varying sinusoids.
• ../concepts/non-negative-matrix-factorization-audio — Smaragdis and Brown (2003). NMF on magnitude spectrograms.
• ../concepts/shift-invariant-nmf — Convolutive NMF, sparse NMF variants.
• ../concepts/informed-model-based-separation — Separation guided by side information (score, MIDI, annotations).

Key Entities

• ../entities/albert-bregman — Author of "Auditory Scene Analysis" (1990). Foundational figure.
• ../entities/paris-smaragdis — NMF for audio (2003), shift-invariant NMF, deep separation.
• ../entities/john-hershey — Deep clustering, MixIT, permutation invariant training. Spans classical and deep eras.

Sources

None ingested yet — seed batch setup.

Open Questions

• How much of CASA theory is encoded in modern deep separation architectures?
• Can NMF-based approaches serve as lightweight baselines for neural transcription?
• Does sinusoidal modeling still have a role in hybrid systems?