Analysis updated 2026-05-18
Watch an AI model compose and perform original piano music live in your browser.
Explore how a small transformer model can run entirely on a graphics card inside a web page.
See generated music rendered as live sheet music with a clickable playhead in the classic view.
Study the WebGPU and WGSL techniques used to run an AI model without any inference library.
| n8python/claude-recital | abhishek-kumar09/mern-live-js-june-2021 | abhishek-kumar09/z | |
|---|---|---|---|
| Stars | 1 | 1 | 1 |
| Language | HTML | HTML | HTML |
| Last pushed | — | 2021-09-13 | 2019-12-23 |
| Maintenance | — | Dormant | Dormant |
| Setup difficulty | easy | moderate | easy |
| Complexity | 3/5 | 2/5 | 1/5 |
| Audience | general | developer | general |
Figures from each repo's GitHub metadata at analysis time.
Requires a WebGPU-capable browser, Chrome on Apple Silicon Macs is recommended for best performance.
claude-recital is a browser-based demo that generates original piano music using a small AI model, then shows a 3D avatar performing it live as if at a recital. Everything, including running the AI model itself, happens directly in your web browser using WebGPU, a technology that lets browsers use your graphics card for heavy computation, rather than sending anything to a server. The model behind the music is a 50 million parameter neural network that predicts music one raw byte at a time, which is a fairly small model by AI standards, built and run using hand-written low-level graphics code instead of any existing AI inference library. It was trained to remember long musical passages and uses several tricks, such as compressing its weights and processing tokens in the fewest steps possible, to run quickly on a graphics card inside a browser tab. There is also a classic view of the demo that shows live sheet music with a moving playhead you can click to jump to different points in the piece. To try it yourself, you serve the folder over a simple local web server and open it in a WebGPU-capable browser, with Chrome on Apple Silicon Macs specifically recommended for the best experience. The README is careful about licensing. The musical compositions used for training are in the public domain, but some of the recorded performance data used to teach the model its style comes from datasets with their own restrictive licenses, including a non-commercial one from Google Magenta. Because of this, the authors describe the whole project, meaning the trained model and the demo built around it, as a non-commercial research demo rather than something meant for commercial use. The piano sound samples and the three.js graphics library used for the 3D avatar carry their own separate open licenses.
A browser demo where a small AI model composes original piano music in real time, performed live by a 3D avatar using WebGPU.
Mainly HTML. The stack also includes WebGPU, WGSL, three.js.
Non-commercial research demo: the model and demo can be used and explored, but not for commercial purposes, due to restrictive licenses on some of the training data.
Setup difficulty is rated easy, with roughly 5min to a first successful run.
Mainly general.
This repo across BitVibe Labs
Verify against the repo before relying on details.