Design notes for the ML analysis pipeline behind Sounds & Recreation — how a recorded tone is decomposed into the parameters that drive a synthesizer. A set of specialist experts, not a packaged library.
Each expert tackles one dimension of the inverse-synthesis problem — going from a recorded sound back to the settings that would produce it.
Recovers the envelope — attack, decay, sustain and release — from the note's amplitude over time.
Characterizes vibrato, tremolo and LFO movement — the periodic motion that gives a patch life.
Infers the underlying engine and oscillator content — subtractive, FM, organ or sample-based.
Combines the experts into a parameter prediction that a synth model can actually apply.
Notebook-style explorations on real songs, with the figures the design decisions are based on.
Findings here inform the algorithms exposed by audio-analysis-mcp.
This work shapes how the analysis stage turns audio into synth parameters.