Research-grade technology built for real understanding.
Signet combines advanced AI and linguistic research to translate sign language with measured accuracy, held-out evaluation, and reproducible scoring.
Explore the science →
FROM SIGN TO MEANING
Sign video
A clip of continuous,
front-facing signing.
front-facing signing.
→
Pose tracking
543 MediaPipe Holistic landmarks per frame: body, face and both hands.
→
Translation modeling
AI models interpret movement, context, and grammar.
→
Grounding & verification
Every output is scored against the true references and against a shuffled set; the ratio is what reading the video was worth.
→
Meaning in context
That is the basic explanation for the NAP.
Real SHuBERT output on a How2Sign test clip. Scored, never asserted.
Our Technology
1. Pose Tracking
& Landmark Extraction
& Landmark Extraction
MediaPipe Holistic detects body, face and both hands to produce the 543-landmark sequence the subnet's pose contract fixes, at the clip's native frame rate.
33 body, 468 face, 21 landmarks per hand
A confidence channel marks absent parts explicitly
2. Translation Modeling
& Context Understanding
& Context Understanding
Pose-first, video-LLM and self-supervised encoder branches compete to map signing to fluent English sentences.
Public training corpora: YouTube-ASL, OpenASL, How2Sign
Uni-Sign, InternVL and SHuBERT branches served in the demo
3. Grounding
& Verification
& Verification
Every submission is scored against the true references and against a derangement of them; the ratio is what reading the video was worth.
chrF, BLEU and BLEURT against certified references
Below a 1.25× canary grounding ratio the score is zero (R14)
4. Private Benchmark
Evaluation
Evaluation
Rounds score on a private, rotating corpus that is never published. Until it exists, the demonstration round scores on the public How2Sign test split and says so.
300 clips per round across four slices
Corpus spec: certified, second-interpreter-verified references
5. Continuous Model
Improvement
Improvement
An open competition: miners publish frozen models, validators re-execute and score them every round, and every record is published with its rejection causes.
Leader changes only on a paired bootstrap at p < 0.05
Every round record published, rejections included
6. Secure & Reproducible
Execution
Execution
Models run in the validator's own sandbox with no network interface, so results are reproducible and proxying to an outside API is impossible.
Weights public, hash-pinned, open to inspection
No network interface during scoring
Built for real understanding.
Measured, not asserted.
Help build a more inclusive world with Signet.