The Sign Language AI Dataset Landscape in 2026

There have never been more sign language datasets than in 2026. Academic corpora keep growing, big tech is mining signing video at scale, and new companies are being founded on the promise of sign language data. This is good news. For years, the entire field could be summarized as "there is no data."

And yet, if you are an AI lab trying to add a sign language modality to your model today, you will run into the same wall we describe in Why Data, Not the Model: almost none of this data is something you can actually train a commercial system on.

I run a sign language data company, so read this with that in mind. But I want to map the landscape honestly, including the projects I admire and the ones that compete with us directly, because the gaps only become visible when you put everything side by side.

The academic corpora

WLASL (2020) is still the most cited: around 21,000 clips covering 2,000 isolated signs, compiled from web sources. It kick-started modern sign language recognition research. It is also licensed for non-commercial computational use only, suffers from dead links as source videos disappear, and contains isolated dictionary-style signs rather than natural signing.

How2Sign (2021) brought continuous signing: more than 80 hours of instructional content in a studio setting, from 11 signers. Excellent for research on continuous recognition and translation; non-commercial license.

ASL Citizen (2023, Microsoft Research) crowdsourced 83,000 webcam clips of 2,731 isolated signs from 52 signers. A landmark in participatory collection, and licensed for non-revenue-generating research only.

YouTube-SL-25 (2024, Google) is the scale play: more than 3,000 hours across 25+ sign languages, mined from YouTube with caption alignment. But it ships as video IDs (the content belongs to the uploaders), and the alignment between captions and signing is weak by construction.

ASL STEM Wiki (2024, Microsoft Research) is over 300 hours of 254 Wikipedia STEM articles interpreted into ASL by 37 certified interpreters. Consent-based, IRB-approved, genuinely useful for STEM education research.

The TUB Sign Language Corpus Collection (2025) aggregated over 1,300 hours across 12 sign languages from broadcast material, aligned with 1.3 million subtitles, including the first parallel corpora for eight Latin American sign languages.

The industry efforts

GoSign.AI is the newest entrant in our own lane: a Deaf-led company paying contributors to record short sign clips at scale, with the stated goal of building the largest commercially available sign language dataset. I am glad they exist. Deaf-led companies proving there is a market for sign language data grows the category for everyone, and their factory model answers a real question: how do you collect volume ethically, with paid Deaf contributors instead of scraped videos?

There are more adjacent efforts (learning platforms, avatar companies, translation apps), but they are building products, not training corpora. This map covers the data layer.

What the map actually shows

Put all of this together and three structural gaps appear.

1. Almost nothing is commercially usable. WLASL, How2Sign, and ASL Citizen are all research-only. YouTube-SL-25 does not own its own videos. Broadcast corpora inherit broadcast rights. If you are a foundation model lab with a legal team, your list of trainable sign language datasets is close to empty, which is why we wrote Why Foundation Models Need ASL Training Data.

2. Most of the signing is interpretation, not language. Broadcast corpora and STEM Wiki capture interpreters translating spoken content in real time. Valuable, but structurally different from how Deaf people actually sign. Learning platforms capture students. Very little of the landscape captures fluent, native signing as it is naturally produced. That distinction is not a detail; it is the difference between training a model on a language and training it on a translation artifact.

3. Hours are not structure. Nearly every large corpus is video plus loosely aligned text. No sign-level segmentation. No gloss labels. No intent structure. No systematic annotation of facial expressions, which carry grammar in ASL. A model trained on subtitle-aligned video learns English-ordered signing with the grammar stripped out. Scale without structure reproduces the same failure at a larger size. That is the annotation problem we broke down in From Raw Video to Labeled Signs.

Where CLERC sits

CLERC is a bet on the opposite corner of the map: depth first.

Our corpus is built exclusively with native Deaf signers (no interpreters, no learners) and annotated at the sign level by expert annotators: sign boundaries, gloss labels, dominant hand, facial expressions, and a structured intent taxonomy, with quality tiers and live validation infrastructure on top. It grows every week.

We publish an open, citable slice of it: EPEE, a benchmark subset of 600 clips from four native Deaf signers with full keypoint data and 150 parallel phrases, with a DOI (10.5281/zenodo.20268565) and distribution on Hugging Face. The full corpus is licensed commercially, with tiers designed for AI labs, because the point of a dataset you can actually license is that you can actually ship with it.

To be equally honest about our own position on the map: in raw hours, CLERC is smaller than the scraped and broadcast corpora. That is deliberate. The field does not lack video; it lacks structured, rights-clean, natively signed data. The realistic recipe for a serious sign language model in 2026 is both: scale corpora for pretraining, and structured native-signer data for alignment, fine-tuning, and evaluation.

The honest conclusion

The landscape is the healthiest it has ever been. Academic corpora keep the research community moving, big tech is waking up to the modality, and new Deaf-led companies are proving the category. All of that is good news.

But the gap between "sign language datasets exist" and "a lab can train and ship a sign language product" is still wide open in 2026. Closing it takes native signers, expert annotation, and licensing built for commercial AI. Those are the three things almost no one on this map is doing at once.

That corner of the map is where we build. Start with the data.