Sheet music for every instrument: one pass gives you a sketch, not parts.

What one pass produces on a mix

The test file is a controlled multi-tone mix, not a real band: several tones sounding together, 4.40 seconds, mono, 22,050 Hz. It is the site's own /examples/midilace-mix-test.wav. The exported MIDI and the spectral measurement of the same audio are both downloadable below, so every number here can be recounted.

So the pitch list was right. Every one of the five exported pitches appears in the measured spectrum. What did not survive is the structure — how many instruments are playing, which note belongs to which, and when a tone starts and stops. That is the part you actually wanted.

The three things that break part structure

1. A sustained tone gets chopped. The low tone spans the whole file, yet it arrives as seven notes. The gaps between them are irregular — 0.986, 0.837, 0.163, 0.999, 0.836 and 0.081 seconds — so no single rule about "one note per beat" will rejoin them. This is the single most common reason a bass or organ part looks like a flurry of short notes instead of a held line.

2. An octave double is indistinguishable from a second instrument. A3 at 219.8 Hz and A4 at 439.7 Hz carry almost the same measured power, 1.11×10¹ on both. If your mix has a bass and a guitar an octave above it, the export contains both and gives you nothing to tell a doubled octave from two players. Sorting your MIDI by pitch and assuming the top line is a separate instrument is a guess, not a reading.

3. Inner voices go first. The hardest-measured case on this site is not the mix but a 12-second solo acoustic passage with every note annotated: 45 notes, 41 found within 80 ms, four missed, 27 extra events, mean timing error 15 to 17 ms. Two of those four misses happened while two or three notes rang at once; the other two were single notes that fell just below the detection threshold. A solo instrument still loses notes inside chords. A full band mix has that problem everywhere, not occasionally.

A full mix is harder than the solo case, not easier — more simultaneous sources, denser harmonics, and no single line for the model to lock onto. Treat the solo case as the ceiling on what a clean recording can do, not the floor.

The workflow that does give you parts

1. Convert one instrument at a time. If you have stems or a multitrack, run each on its own. A single instrument is the case that measured well — 41 of 45 annotated notes — and each run gives you a part you can read as a part. Six solo passes beat one mix pass, even though it is six times the clicks.
2. If you only have the mix, work in 10 to 20 second windows. Fewer simultaneous notes per window means fewer chopped tones and fewer duplicates, and you can check each window against the original before moving on.
3. Rejoin the chopped low line first. Sort by pitch, find repeated notes at the same pitch, and merge the ones separated by less than about 100 milliseconds. On this test file that turns seven A3 events into roughly one held note, and it is the difference between a bass part and noise.
4. Deal with the octave layer by listening, not by pitch. Play the original, then play the export with the suspected duplicate line muted. If nothing disappears from the recording, the line was an artefact.
5. Write the inner voices yourself. Where the recording has chords, expect to supply the middle notes by hand. On the measured case, three notes in twelve seconds was the whole edit.
6. Keep the original beside the export. Any event you cannot hear in the recording is safe to delete, and any note you can hear but cannot find in the export is one you have to add.

What you get at the end of that is a sketch you have corrected into something playable — which is a different claim from "sheet music for every instrument", and it is the honest one.