Why we digitise real hands instead of drawing or generating one
We tried three ways to get a believable handwriting onto a pen plotter. Two failed. This is what the third does differently, and what a plotter needs that a font does not give you.
Draft for review. Not yet approved for publication.
A pen plotter does not print letters. It moves a pen along paths. If those paths come from an ordinary font, the result looks like a font with ink. We wanted a letter that reads like a person wrote it, and it took three attempts.
A plotter needs a line, not an outline
Most handwriting fonts describe the shape of each letter as a filled outline. A plotter traces that contour with its pen and draws every stroke twice, once on each side. The letters come out hollow. Two tools we read up on that turn a photo of handwriting into a font file also work with outlines, so we expect the same result from them. So we look for fonts built from a single centre line per stroke.
Attempt 1: draw one
In September we drew a hand by hand with AI help. Each letter was a finished curve. Seen as a sample it looked tidy. Over a whole letter it looked wrong: all letters had the same height and the baseline was too straight. We judged it “too perfect”. It reads like a typeface.
Attempt 2: generate one
In October we built a movement model. For each character we set a few target points, and a simulated pen moved between them with a speed profile like a real stroke. Small changes to the targets produced variants, so one letter had up to four forms. The result was reproducible. It still did not look human, and we dropped it.
Attempt 3: use real strokes
A research group at Leipzig University published a dataset of German handwriting with pen strokes from many individual writers. We took pages from it, reduced each stroke to a centre line and filled gaps by composing the missing characters from the same writer's strokes. Each letter is kept in several genuine forms. Depending on the hand, lowercase letters average between 4.6 and 6.4 forms per letter.
That variety matters more than it sounds. A series of 500 letters where every “e” is the same shape looks printed. When the machine can choose between several real “e” shapes, repeated words stop looking identical.
Testing: reading, then paper
For each hand an independent reader had to read a hard sample: two letters, addresses, the alphabet, digits and a repeated line. Seven hands scored 8 out of 10 and seven scored 7 out of 10. Two others scored 6 and one is still a draft because its “n” reads like a “u”. We do not recommend those three.
The fourteen that passed were written on our machine and compared on paper. One of them looked like real handwriting to us when we held the sheet. A score does not decide that. We still look at paper.
Smaller things we found on the way
- Pen width. At a letter height of 4.5 mm, a 0.5 mm gel pen fills the small inner spaces of an “e” or an “a”. A ballpoint or a fine gel pen of 0.35 to 0.4 mm reads more clearly.
- Judge the whole letter, not the sample. A sample line can look excellent and a full page can still look wrong. Our drawn hand taught us that.
- Odd glyphs break silently. In some free single-line fonts, a few characters such as “ß” and the apostrophe were stored in a form our loader did not handle. A name like “Max' Backstube” came out as “Max Backstube” on an envelope. We now convert those characters when we load a font.
Licence and limits
The dataset hands are licensed under Creative Commons Attribution 4.0. Anyone may use them if they give credit, so they cannot be exclusive, and we do not present them as the handwriting of a named person. The full register, with every hand, its source and its status, is on the handwriting page.