MIDI quantization: snapping onsets to a grid, measured.

Seven onsets, one grid, two collisions

The export of the site's 4.40-second multi-tone test file (/examples/midilace-mix-test.wav, MIDI below) contains seven A3 events. Their onsets, snapped to the nearest quarter note at 120 BPM — a 0.5-second grid — produce the table below. The file itself has no tempo event, so 120 BPM is the standard default a player would apply.

Read the fourth column as the cost: quantization at this grid resolves seven onsets into five positions — 0.000, 1.000, 2.000, 3.000, 4.000 s. The two colliding pairs were 0.163 s and 0.081 s apart. Those gaps are the only thing distinguishing them from single events, and a grid that coarse wipes them out.

What a 0.500-second grid does to four measured takes

The mix file is synthetic, so the grid test was repeated on the site's real recordings — including the annotated ground truth of the 12-second guitar case. Every row is the same operation: snap every onset to the nearest 0.500 s point and count what survives.

The last row is the one to sit with: the human annotation is not on the grid either. The annotated notes of the 12-second guitar case were placed by ear against the recording, and snapping them to 0.500 s quarters would move some of them by up to 0.240 s. If a hand-made answer key does not live on a 0.5-second grid, a default quantize pass is not "correcting" the export toward the truth — it is imposing a meter the performance may not have.

What the grid fixes, and what it hides

It fixes real jitter. The corrections in the first table are up to 179 ms, which is far beyond musical timing — the annotated guitar case on this site found mean timing error of 15 to 17 ms on notes it did find, and treats 80 ms as its match window. Onsets sitting 98 to 179 ms off the grid are more likely estimation noise than played rubato, and quantization is the right tool for them.

It merges events that were separate. Two pairs above now share a grid point. If the export had split one gesture into two — the 0.081-second gap is a strong candidate — quantization makes the split permanent and invisible. Nothing in the file records that a merge happened.

It does not rejoin a chopped tone by itself. The seven A3 events are one continuous 219.8 Hz tone in the measured spectrum. Quantizing their onsets changes their start times; it does not shorten their count. Rejoining needs a second step — merge same-pitch events on the same grid point — and that step is a musical judgement, not a grid setting.

It cannot invent a meter. The grid assumes the music lives on your chosen grid. A swung feel, a rubato intro, or the irregular gaps measured here (0.986, 0.837, 0.163, 0.999, 0.836, 0.081 s) do not; quantizing them flattens exactly the detail you were trying to read.

How to quantize without lying to yourself

1. Choose the grid from the recording, not the export. Tap the beat while listening. If the take swings, quantize at half the grid and check by ear, or do not quantize at all.
2. Quantize a copy, keep the raw events. The measured export is the evidence; the quantized version is an interpretation. Comparing them is how you find what the grid erased.
3. Merge same-pitch events that share a grid point — deliberately. On this file that is the two colliding pairs. Decide each one by ear: two chord tones of one gesture, or two real notes?
4. Apply the 80 ms test to anything you keep. This site's annotated case calls a note found when it lands within 80 ms of the annotation. Events that only make sense inside 80 ms of each other are artefacts far more often than notes.
5. Re-check the low lane last. The sustained-tone problem — a held note chopped into seven — is a duration problem, not an onset problem. Quantization never fixes it; merging does.

Used that way, quantization is a cleanup of estimation noise, measured here at up to 179 ms per onset. Used without the merge step and the listening checks, it quietly manufactures notes that were never played separately.