Claude Opus dissects agent music synthesis: chroma analysis and spectrograms as stand-ins for hearing
teortaxesTex · x · 2026-10-18
Follow-up with Opus's analysis of the earlier 'procedural composer' session: audio here was mostly a text/coding task — WAVs are just numbers built from sines, decays and filtered noise, squarely in-distribution for a code-trained model, and the genre did much of the aesthetic work. The real generalization is self-built stand-ins for hearing: chroma analysis for tuning, stereo correlation for width, band energies for balance, and rendering spectrograms as images for its vision input — a workflow nobody gave it, like a deaf engineer working from meters. Its confusion marks where the stand-ins run out: it couldn't judge loudness, envelope dips, or that the unmeasured tanh 'tape saturation' stage cut 2.5 dB off peaks and kept RMS hot. A separate audit replayed the log and found the synthesis honest — bit-identical output, no cheating — with only the final report misdescribing the mastering stage.
Related event: AI Agent Writes Procedural Vaporwave Music Without Hearing(2 posts)→
More from coding & agent
- Jim Koppel poses a classic engineering problem: importing just one function from a bloated file — jimmykoppel · 2026-10-18
- The Long Walk: an interactive stroll through 300,000 years of human history, built with Runway MCP in Replit — tlakomy · 2026-10-18
- Stately.ai MCP earns early rave review from developer after live demo worked — TejasKumar_ · 2026-10-18
- Rust-based Git server on S3 handles 7,600+ concurrent pushes per second — blaizedsouza · 2026-10-18
- Chaining 2 PCs, a MacBook and 2 iPhones Runs 48-Expert MoE at 37 tok/s on 60k Context — StayLameBro · 2026-10-18
- Google, OpenAI, Meta among 13 applicants for .agent top-level domain — gaganghotra_ · 2026-10-18