Can AI edit sheet music? Mechanically yes — and the verification stays human.

The edit list, measured

Every number below comes from this site's own test files, each downloadable with its export and — where one exists — its annotated ground truth. This is the work any transcription produces before it is a chart.

Read as a workload, it is bounded: roughly one edit decision per exported event, plus one merge per chopped lane. That is the honest size of "AI edits your sheet music" — the mechanical half can be automated; the by-ear half is yours.

Two synthetic takes, and the stacking the numbers reveal

The muted and ringing guitar takes exist to make one variable visible: the same four annotated notes (A3 to E4), played once into a synthetic guitar-like tone, muted and ringing. Both exports found all four annotated notes within 80 ms — and both added far more than they found.

The simultaneity counts show where the growth sits: in the muted export, 21 of the 23 events sound while at least three others sound (the export's busiest moments hold six events at once); in the ringing export, 25 of 29 do. The extra events are not random noise spread over time — they cluster into chords around the four real notes, which is exactly why deleting them is a by-ear job: a cluster of six simultaneous events can be one real note plus its decay, or a real chord, and only the recording tells you which.

The three judgements that stay with you

1. Which extra events are real. The guitar export carries 27 events the annotation does not have. Some are octave doublings of real notes, some are split tails, some are harmonics that sound fine in a mix. Each one is deleted or kept by playing the recording — no rule derived from the file settles it, because the file is the thing in question.

2. Which missing notes to add. Four annotated notes are absent from the guitar export, and they are missing because they sounded inside chords or below the detection threshold — the exact conditions a model cannot see. Adding them means reading the ground truth or knowing the song. An editor that invents them produces a plausible wrong chart.

3. Whether a merge is musical. The seven A3 events merge to one on a duration argument (4.006 s against one continuous 219.8 Hz tone). But the two pairs of onsets 0.163 s and 0.081 s apart could be played flourishes rather than split artefacts. On this synthetic file they are split artefacts; on your recording the answer comes from your ears.

An editing workflow that verifies as it goes

1. Fix your expectation first. Before opening any editor, write down what you can hear: how many players, the bass line, the top note. That list is your ground truth, and it is what prevents a tooled-up wrong answer from surviving.
2. Sort by pitch and deal lane by lane. The measured failures cluster by lane: the low lane splits (7 events, merge to 1), the high lanes gain octave artefacts (voice export up to E7), the middle lanes split inside chords.
3. Merge with a stated threshold and a listen. Same pitch, gap under ~100 ms → candidate merge; play the window before you commit it. On the measured mix this is the whole bass fix.
4. Delete by ear, not by rule. For each extra event: play the window with the event, play it without. Unhearable events go. That is the 27-event check on the guitar case.
5. Re-add missing notes where the recording has them. If you can hear a note that is not in the export, it goes in — the four missed guitar notes are exactly this case, and no detector improvement measured here would have found them all.
6. Compare against the answer key where you have one. The annotated files use an 80 ms match window with mean timing error of 15 to 19 ms on found notes — a fair criterion for "did the edit restore the note", not for "is the timing perfect".

So: yes, AI can edit sheet music — the merge, the sorting, the pitch-based re-adds are all mechanical and all automatable. The measurement above is the reminder that the automatable part is only half the job, and the half that stays human is the half that makes the chart true.