AI Music Repair Studio

Inspiration

Musicians, arrangers, and students can hear that a passage feels unusual without quickly knowing which notes or patterns deserve a closer look. I wanted a focused second pair of eyes that explains what it noticed without pretending to replace musical judgment.

What it does

AI Music Repair Studio is a local-first MIDI review desk. It parses MIDI in the browser, maps notes to an SVG piano roll, identifies repetition, large melodic leaps, and density spikes, and explains each review candidate in plain language. A constrained repair is shown as a visible diff before it is applied. Before/After comparison, Undo, local playback, and repaired MIDI export keep the workflow inspectable and reversible.

The app does not claim objective musical correctness. It highlights places a human may want to listen and decide.

How it was built

The MVP uses Vite, React, TypeScript, @tonejs/midi, SVG, Web Audio oscillators, Vitest, ESLint, and the TypeScript compiler. MIDI stays local in the browser and no application API key or remote AI endpoint is required.

The combined generated sample loads with 12 notes and 5 review candidates. Applying the density candidate removes only two selected notes, producing an After state with 10 notes and 2 changes. The exported result round-trips as parseable MIDI with PPQ 480, 2 tracks, and 10 notes.

How Codex and GPT-5.6 helped

Codex using GPT-5.6 served as the development orchestrator: it read the specification and checklist, guided focused implementation of parsing, analysis, repair, UI, playback, export, and error handling, ran the test matrix, exercised the Chrome demo, inspected the repaired MIDI, and prepared evidence-backed submission materials. Human judgment remained responsible for scope, wording, asset rights, listening, and publication.

Challenges and lessons

The main challenges were keeping tick-based MIDI timing canonical, making density heuristics useful without overstating them, preserving immutable Before/After state, and making the demo stable when browser audio is unavailable. Deterministic generated fixtures and visible diffs made the behavior easier to test and explain.

Limits

This is a deterministic heuristic review assistant, not an automatic music critic or unconstrained MIDI rewriting service. Browser audio and the final visual framing still need human review.

Built With

• vite