Making band parts from a recording: what the five lanes can and cannot be.

The five lanes, measured

The test file is a controlled multi-tone mix — several tones sounding together, 4.40 seconds, mono, 22,050 Hz — at /examples/midilace-mix-test.wav, with the exported MIDI beside it. It stands in for any single-file band recording. Below is the export sorted into its five pitch lanes, which is exactly how you would start assigning parts.

Two numbers tell you how much work the lanes hide. First, the seven events on A3: the measured spectrum shows one continuous 219.8 Hz tone across the whole file, so the lane is one held note chopped into seven. Second, the durations: all thirteen events add up to 7.394 seconds inside a 4.40-second recording — 1.68× — because lanes overlap in time. Both are visible before any note is checked by ear.

Four reasons a lane is not yet a part

1. The low lane is chopped. Seven A3 events start at 0.000, 0.986, 1.823, 1.986, 2.985, 3.821 and 3.902 s. The gaps — 0.986, 0.837, 0.163, 0.999, 0.836, 0.081 s — are irregular, so no beat rule rejoins them. Merging same-pitch events separated by less than about 100 ms turns the lane into one held note spanning 91% of the recording, which a bass player can read.

2. Octave doubles look like two players. A3 at 219.8 Hz and A4 at 439.7 Hz carry near-identical measured power (1.11×10¹ on both). The A4 lane might be a second instrument, or the doubling of the A3, or an artefact of the sustained tone. Mute it in the export, play the original, and keep it only if a sound disappears.

3. Two events on one lane can be one gesture split. The C5 lane's two events start 0.756 s apart (0.986 and 1.742), and the muted-guitar test shows the same pattern worse: four real notes became 23 events across 13 pitches. A lane's event count is an upper bound on notes, not the note count.

4. Nothing in the data says who plays what. The export has pitch, onset and duration. Which lane is the guitar, which is the keyboard, and whether the top two lanes are one player or two are questions the recording answers by ear — the lane list only narrows them.

From five lanes to five parts

1. Give each player their lane as a starting file, labelled by pitch. A four-piece band reading "A3 — 7 events", "A4 — 2", "C5 — 2", "D5 — 1", "E5 — 1" has something concrete to react to, and each player knows their own part.
2. Bass merges first. Sort the A3 lane, merge gaps under ~100 ms, hand over one held line. On the measured file that is 4.006 seconds of one note — most of the recording.
3. Octave-check the A4 lane before anyone learns it. Mute-and-listen, as above. Doubles that survive the check belong to whoever plays them.
4. Middle voice gets a listen-through. The two C5 events are candidates for the rhythm part; confirm against the recording, because split events are common — the muted test turned four notes into 23.
5. Top lanes go out nearly as they are, with a length caveat. D5 and E5 are single events of 0.882 s and 0.812 s. In a real song the top lane is the melody, and single events there are the most trustworthy part of the export — but the lane stops when the events stop, so anyone reading it knows it is a fragment, not a full part.
6. Collect the corrections back into one file. Each player's fixes — merged bass, deleted phantom octave, real chord tones — become the corrected sketch. That file, not the raw export, is the "parts" deliverable.

The honest summary: the recording gives you the raw material for parts in one pass, and the parts themselves come from five short per-lane sessions plus the band's own knowledge. Faster than transcribing from nothing, slower than the "instant per-instrument sheet music" the question usually hopes for.