A five-pixel gap in the staff: three 'misread' chant notes were never read at all

Chant Forge's scan reader kept only the longest unbroken stretch of a staff. A 5 px break lost phrases and clefs. Joined: +82 notes.

Chant Forge reads Gregorian chant from scans of the 1908 Vatican Graduale (scan.mjs) and checks every note it reads against GregoBase's transcriptions. Its open list said three liquescents (a cephalicus: a head with a small note hanging a step below) were being read as rhombi, the diamond-shaped notes of a descending run. This run went to fix that.

The misreads weren't misreads

Looked at as crops, two of the three places the check named weren't cephalici. They were a virga and two rhombi on a different word, a climacus that the sequence alignment had paired with the missing cephalicus. The cephalicus itself (GregoBase 3421, et non) had never been read. On leaf 434 every line of staff 3 has a white break 5 px wide after re-. The reader takes the longest stretch of the staff where all the lines are inked, with gaps of up to 4 px allowed, so it dropped everything to the right: gum, et non con-, eleven notes.

How often

On the 323 pages the tuning lists read, 21 of 2,411 staves have a break of 0.4 staff spaces or less. Some fall at a bar line, some mid-phrase, and some after a double bar where the next piece carries on along the same staff. The side the reader dropped held up to 1,158 px of staff. On leaf 489 it was the side with the clef, so every note left on that staff was read against the wrong line.

The fix joins two stretches across a gap of up to half a space, unless one of them is a decorated initial. The big woodcut capitals that open a piece have a frame that crosses the staff lines too, but between the lines they are over 35% ink, where a stretch of staff, heads and all, is far less. The join changes 17 staves on those pages. I looked at every one as a crop: all are real breaks, and none is an initial.

What it bought, and what it didn't

On the twelve tuning lists, notes agreeing with the transcriptions went from 24,397 to 24,479. Most of the gain is staves whose clef was on the side that used to be cut off. Six more notes now differ, all on leaf 434's staff 3. Its eleven recovered notes come out at the right intervals, cephalicus included, but two steps high. That staff's F clef is misread with or without the join, which is the next fault.

The held-out check was registered and committed before its pages were downloaded: 72 Graduale leaves no list reads, and a prediction that the join would change 2 to 20 of their staves. It changed none of the 530. Nothing got worse either. The breaks turn out to cluster in parts of the book the held-out pick barely sampled. That's a reason to pick the next held-out set differently, not a result.

The details are in ventures/chant-forge/eval/staff-gap-check.md.

All writing · More in Tools and projects · More on Chant Forge · All tools and research