Benchmark: cross-signer generalization goes 29% to 69% as training signers go 1 to 6
v0.3 is live: 1,200 ASL clips, 6 native Deaf signers, a 201-phrase parallel grid, 128 keypoints per frame. Double the corpus and the signer pool of v0.2.
The number worth your attention is not the corpus size. It is what the corpus size does.
The curve
We trained a small BiLSTM on 1 signer, then 2, then up to 6, and tested it each time on signers held entirely outside the training set. Same 24-gloss subset, same protocol as the v0.2 measurement, 8 seeds:
| Training signers | 1 | 2 | 3 | 4 | 5 | 6 |
|---|---|---|---|---|---|---|
| Accuracy on an unseen signer | 29% | 47% | 57% | 63% | 65% | 69% |
macro-F1 over the same range: 0.17 to 0.57.
The architecture never changed. The hyperparameters never changed. The only thing that moved was how many different people the model had watched.
One signer gets you 29% on a stranger, which is to say nothing you can ship. Six gets you 69%. The curve has not flattened.
We also grew training examples independently of signer count: 22% to 71% as the pool went from 858 to 1,393 segments.
Full method, exact numbers and caveats: BENCHMARK.md. The v0.2 curve is overlaid on the v0.3 one, because the two are directly comparable by construction and you should be able to check that yourself.
Why we keep publishing this particular measurement
Sign language AI is routinely framed as an architecture problem. It is not, or at least not yet. A model that has seen one person signing has learned one person, not a language. That is not a modeling failure you fix with a better encoder. It is a data failure, it is measurable, so we measure it and publish the curve every release.
What else is in v0.3
- Signer pool 4 to 6. Corpus 600 to 1,200 clips, 200 per signer.
- Parallel grid widened 150 to 201 phrases.
phrase_idpairs renderings of the same phrase across signers directly, no clip-id arithmetic. - A second register: FOXTROT contributes 200 declarative sentences on a disjoint prompt set, so the corpus is not only short conversational phrases. Filter on
parallel. - Head silhouette re-extracted on a pose-guided crop. The old path lost up to 21% of frames on small-face clips, the new one loses none.
Two things we will not soften
This is a pilot. 1,200 clips and 6 signers is a baseline, not a production corpus. The limitations section is not decoration.
Skeletons are body data. Removing video removes the face, not the body. On this release a plain logistic regression separates the six signers with 94% accuracy from nine limb-ratio features alone, no hands, no face. We measured it rather than assume it. "Anonymized skeletons" claims more than it delivers: pseudonymization here protects names, not bodies.
CC BY-NC-SA 4.0 · DOI 10.5281/zenodo.22081248
If you are training a multimodal model and sign language is on your roadmap, we would rather hear what breaks than what works.