How AI audio-to-MIDI transcription works

What makes this different from a tuner

A tuner detects one fundamental frequency at a time — good for a sustained note, not a full melody with rhythm and rests. The audio-to-MIDI transcriber uses an AI model trained on thousands of real recordings, able to recognize where each note starts and ends, even when several notes sound at once (chords).

Why it runs in your browser instead of a server

The model is under 1 MB and runs with TensorFlow.js directly on your device. That means two things: your audio is never uploaded anywhere, and it doesn't depend on a server being available or usage limits — it works just as well offline once the page has loaded once.

What it's good at, and what it isn't

It works best with one sound source at a time: a voice, a guitar, a piano. It doesn't separate instruments mixed together in a full recording (that needs a different technology, source separation). It also doesn't transcribe untuned percussion like a drum kit — the model looks for musical notes, not rhythmic hits.

What to do with the result

The MIDI file you download has no sound in it — it's instructions for which note, when, and for how long. Open it in any notation software (to see the sheet music) or in a DAW (to assign any virtual instrument and play it back).