Back to the converter

How to get better audio to MIDI results

Which recordings convert to usable MIDI, which never will, and the settings that fix the four most common failures.

When you are done: A clear sense of whether your material can convert at all, and the two or three settings to reach for when it half-works.

The material matters more than the settings

No setting rescues a recording the model cannot hear into. Roughly, in descending order of how well it converts:

  1. Clean, single instrument, one note at a time. A dry piano take, a DI guitar line, a hummed melody. This is the case the tool is built for.
  2. Single instrument playing chords. Usable, but expect missed inner voices and extra notes from overtones.
  3. Two instruments together. Both land on one MIDI track, interleaved. Sometimes useful as a sketch.
  4. A finished, mastered mix. Expect a rough outline of the loudest line and a lot of noise. This is not a limitation to configure around; it is what the technology does.

Four things that hurt, in order

1. Reverb and delay

The worst offender by a distance. A tail blurs where one note ends and the next begins, so the model reports notes that overlap, repeat, or never stop. If you have access to the dry source, use it. If you do not, raise the minimum note length and turn on “trim overlaps”, and accept that note ends will need work.

2. Distortion and heavy compression

Both generate harmonics that read as real notes, usually an octave or a fifth above what was played. Narrow the pitch range to the instrument’s actual range and a lot of this disappears.

3. Several instruments at once

There is no separation step here: everything lands on one track. Convert stems individually if you have them. If you only have the mix, pick the section where the part you want is most exposed.

4. Very fast passages

Below roughly 60–80 ms per note, onsets start merging. Rolls, trills and fast runs come out as fewer, longer notes. Lowering the minimum note length helps a little, at the cost of more false notes everywhere else.

What to change, by symptom

Far too many short notes

Raise sensitivity towards 0.6–0.7, and set remove short notes to 100–150 ms. If they cluster at the very bottom or top of the range, narrow the pitch range instead — that is usually rumble or harmonics, not performance.

Notes are missing

Lower sensitivity towards 0.25–0.35. If quiet notes are the ones vanishing, check that remove quiet notes is off. If whole phrases are missing, the source is probably too dense at that point.

Notes cut off too early, or run on forever

That is note sustain. Lower it to hold notes longer, raise it to end them sooner. On material with a lot of ring-out, pair a higher value with trim overlaps.

One pitch repeats rapidly where you played one long note

A false onset on a sustained note. Set merge repeated notes to 80–150 ms. This is very common on piano with pedal and on bowed strings.

Timing feels wrong in the DAW

Check the tempo value before exporting. It is estimated from note spacing and can land on half or double the real tempo. If you know the tempo, type it in. Only then reach for quantize, and keep it low — at 100% you throw away the feel that made you record the take.

A workflow that saves time

  1. Convert fifteen seconds first, not the whole thing. If fifteen seconds are unusable, sixty will be too.
  2. Judge it by ear with the Original / MIDI switch, not by looking at the piano roll. A roll that looks busy can sound fine, and vice versa.
  3. Fix detection first (sensitivity, sustain, pitch range), cleanup second. Cleanup cannot recover a note that was never detected.
  4. Export, open it in your DAW, and play it against the original audio there. Problems that were inaudible in a short loop show up quickly.

When to stop

If after a couple of minutes of adjustment the draft still needs more editing than playing the part in by hand would, stop and play it in. This tool is worth using when it saves you time, and it will not always.