Audio to MIDI Conversion in Your Browser

Choose an MP3, WAV, M4A or OGG file up to ten minutes long. The page decodes it, mixes it down to mono at 22,050 Hz — the rate the transcription model expects — and then walks through the audio in thirty-second blocks, keeping two seconds of overlap on each side of every block.

That overlap matters more than it sounds. Deciding where a note starts and ends needs a few frames of context on both sides, so a note that happens to cross a block boundary would otherwise be split into two short fragments. The overlapping regions are transcribed and then discarded, leaving one clean set of notes.

Progress is reported per block, so a long file gives you a moving percentage rather than a frozen spinner. The first run of the session also downloads the model weights, which is the slowest part; after that, further files start converting immediately.

The model was trained to follow pitched, sustained sound, so it is strongest on a single clear instrument: solo piano, a sung or hummed melody, an acoustic guitar line, a bass part, a woodwind or string melody. On that material the note starts, pitches and lengths are usually close enough to edit rather than redo.

Three things degrade the result. Dense full-band mixes give the model several pitched instruments at once, and it will interleave them into one track rather than separating them — run the stem splitter first if you want a single part. Percussion has no stable pitch to track, so drums come back as scattered low notes rather than a usable groove. Heavy distortion and thick reverb blur the onsets, which shows up as extra short notes around the real ones.

Pitch bends are captured where the model detects them, so slides and vibrato survive into the MIDI as bend events rather than being flattened. That is useful for guitar and vocal lines, and occasionally worth deleting in a notation program where you only want discrete pitches.

The download is a standard MIDI file, so anything that reads MIDI will open it: a DAW, a notation program, a hardware sequencer, a mobile sketch app. The tempo written into the file is estimated from the spacing between note starts, folded into the usual 60–190 range; it only controls where bar lines land in your editor and does not move the notes themselves, so correcting it is harmless.

Expect to tidy the result rather than use it untouched. The two edits that matter most are deleting the stray short notes that cluster around loud onsets, and extending notes the model cut early in reverberant passages. Both are a few minutes of work on a typical melody.

Your recent conversions stay listed on this page so you can re-download a MIDI file without converting the same audio twice.