
Virtual Handpan
2026A 3D handpan that learns how you play: record takes, train a model in the browser, and let it improvise phrases in your own style.
Generated playing comes from three layers. A melodic two-hand generator with live human feel controls. A Learned mode driven by an n-gram model over scale steps and durations, trained in the browser on your recorded takes, so it phrases the way you do. And thumbs up and down feedback that reweights the generator toward your taste. Under it all sits a physically based WebGL pan with 93 builder scales, custom layouts, and a velocity-layered sample engine.
Composition with a language model: ask Claude for a phrase as structured JSON, given the layout, feel, and a style prompt like "slow, sparse, resolves often", with the humanize layer kept deterministic so it still sounds like a person.
- In-browser ML
- n-gram model
- React
- TypeScript
- three.js
- Web Audio

