audio2ay

The AY chip is the language.

Convert instrumental audio into General Instrument AY-3-8910 register streams via analysis-by-synthesis: search the legal AY register space with a cycle-accurate emulator in the loop, scoring candidates by perceptual similarity to the source rather than by parameter distance.

Resolution

This research effort has been discontinued and has reached a dead end. No further development, updates, or maintenance are planned. For practical use and continued development, please refer to audio2ay3.

See design/PLAN.md, design/ARCHITECTURE.md, design/AY_REFERENCE.md, design/LOSS.md and design/ROADMAP.md.

Install (Windows / PowerShell)

py -3.12 -m venv .venv
.\.venv\Scripts\python.exe -m pip install -e .

Usage

# activate the venv first (once per terminal session)
.\.venv\Scripts\Activate.ps1

# audio -> 50 Hz register stream (+ optional mp3 preview)
audio2ay2 convert samples\short\01_arpeggio_mono.wav -o out.ay --preview out.mp3

# render a stream back to audio
audio2ay2 preview out.ay -o out.mp3

# convert + compare + validation report (audio + loss plot + score)
audio2ay2 validate samples\short\01_arpeggio_mono.wav -o reports\arp

# inspect a stream
audio2ay2 info out.ay

Key options for convert: --format raw|psg|ym, --profile balanced|melodic|percussive, --optimizer local|beam|annealing|genetic, --proposals N, --pyramid 160,80,40,20, --clock, --fps, --chip ay|ym2149, --seed.

Pipeline

audio -> features -> diverse proposals -> [coarse-to-fine multi-start optimizer
         (emulator + perceptual loss in the loop)] -> 50 Hz register stream
         -> export (raw / PSG / YM) and/or preview (wav / mp3)

The optimizer is built from three decoupled interfaces — MoveGenerator, Evaluator, SearchStrategy — and operates only on the semantic AYState; raw registers R0..R13 exist solely at the export boundary.