MuScriptor

MuScriptor converts music audio into note events and a standard MIDI file.

What it does in Miso

One task, Transcribe to MIDI, on the sound page.

What comes back

Note events and a .mid file, and no audio at all. This is the only task in Miso that works this way.

That is why a transcription is not a take. It cannot be played by the dock, drawn as a waveform, exported as WAV or MP3, separated or mixed. It lives under the take it was read from, with a preview that plays from the stored note list.

The notes are stored beside the file rather than re-derived, because the preview plays from them and parsing MIDI back into notes the model already handed over would be work in a circle.

The lead-in

MuScriptor drops a note that starts at exactly zero seconds.

Miso puts a second of silence in front of the audio before sending it, then takes that offset back off the note times when it stores the result. You will not see this happen.

This makes transcription the only task whose source is modified before the model sees it for a reason other than sample rate.

Getting good results

Run it on a single instrument. A separated bass or vocal stem transcribes far more cleanly than a full mix. See Splitting into stems.

Turning a transcription into a score

The most useful thing to do with a transcription inside Miso is hand it to YuE2 as a melody.

Miso turns the note events into an ABC score: the highest note wherever notes overlap, quantized to sixteenths against a tempo, barred, and tied across bar lines so every bar adds up.

MuScriptor supplies pitch and times in seconds and nothing else, so tempo, key and meter are inferred and shown in a dialog you can correct. Check the tempo. A tempo read at half speed still writes the right pitches and produces a cover that is hard to diagnose afterwards.

See Scores with YuE2.