STU — studio
Same conditions as training. The friendliest number a model will ever produce, and the one to be most suspicious of.
Dataset
What the models learned from.
Everything here is trained on MM-WLAuslan. Its licence is non-commercial, and that single fact shapes what this initiative is allowed to become.
MM-WLAuslan
A word-level Auslan corpus, recorded from three views.
Published at NeurIPS 2024 as a datasets and benchmarks track paper by the University of Queensland and collaborators. It is, as far as we know, the largest word-level Auslan recognition dataset that exists.
282KSign videosRecorded in a studio environment
3,215Auslan glossesCommonly used vocabulary — the model’s entire world
73SignersThe ceiling on how much signer variation we can learn
4Cameras, RGBDThree Kinect-V2 plus one RealSense, triple-view
Source: Shen et al., MM-WLAuslan: Multi-View Multi-Modal Word-Level Australian Sign Language Recognition Dataset, NeurIPS 2024 Datasets and Benchmarks Track (arXiv:2410.19488).
Test design
Four test conditions, not one.
The dataset ships four separate test subsets, each covering the full vocabulary. This is the most useful thing about it: it makes robustness measurable instead of assumed.
TED — temporal disturbance
Timing is altered. Tests whether the model learned the movement or just the tempo it was trained on.
SYN — synthetic background
Background replaced. Tests whether the model is reading the signer or reading the room.
ITW — in the wild
Real-world capture conditions. The closest thing to an honest estimate, and the number we lead with.
Licence
CC BY-NC-SA 4.0 — and we mean the NC.
MM-WLAuslan is released under Creative Commons Attribution-NonCommercial-ShareAlike 4.0. Anything trained on it inherits that constraint. Our models are therefore non-commercial research artefacts, and this initiative sits inside the Octopus Foundation rather than inside a company.
This is a real limit, not a formality. A commercial Auslan product could not be built on these weights. It would require either purpose-collected data with consent obtained for commercial use, or a separate licensing agreement with the dataset’s authors. We have deliberately not blurred that line.
Consent is not a licence checkbox
The recordings are of real people signing. A permissive licence settles what we may legally do; it does not settle what those signers understood they were agreeing to. Any future use that would generate synthetic signing from those recordings — an avatar, for instance — is a consent question first and an engineering question second.
What the data cannot tell us
A studio corpus has a studio’s blind spots.
- Isolated signs only. One sign per clip, with clean boundaries. Real signing has no gaps, and signs reshape each other where they meet.
- 73 signers is not a population. Regional variation, age variation, and the natural range of signing styles are all under-sampled relative to the real community.
- Vocabulary is a closed world. The model can only ever name one of 3,215 glosses. Fingerspelling, name signs, and depicting constructions fall outside it entirely.
- Studio framing. Even the in-the-wild subset was built by a research team, not scraped from how people actually use video.