I built NanoMaestro Realtime, a small AI music model designed for real-time piano music generation on CPU directly in the browser. It is around 50MB with about 13M parameters, exported as an ONNX model and run locally using Transformers.js with WASM, so it does not require a GPU, server backend, or heavy setup. Webdemo(HF spaces)
The model itself is a compact 2-layer LSTM with a 1024-dimensional hidden state. It continuously generates piano events in real time, It is so lightweight that it can run real-time on older Raspberry Pi models. Model
Impressive work! Running real-time piano generation entirely in the browser on CPU is a great achievement. The small model size and Raspberry Pi compatibility make it especially interesting for creative and educational applications. I’d love to hear more about the training process and how well it maintains musical coherence over longer performances. Thanks for sharing!