Compose
Beyond the piano
MAiSTRO's own model composes notes, and a soundfont turns those notes into sound. That is exactly the right tool for classical piano and exactly the wrong one for music that lives in timbre — a synth pad, a drum break, a guitar tone. Nothing in a piano-note vocabulary can express “dusty Rhodes through a tape delay.” These two models can.
Text to audio — Meta MusicGen
A transformer over the tokens of an audio codec, not over notes. It predicts the next frame of compressed waveform, so it can render any genre it heard in training, at the cost of never producing a score you can edit. Runs on the Python backend.
Checking the audio stack…
In the browser — Google Magenta MelodyRNN
The same job as MAiSTRO's LSTM — continue a melody, one note at a time — but the weights download to your browser and the inference runs on your GPU through TensorFlow.js. No Python process, no cold start, no server bill. The trade is a few megabytes on first use and whatever your laptop can manage.