Under the hood
How the sign-validation engine works
Every practice attempt runs through a computer-vision pipeline that turns a webcam recording into an accuracy score \u2014 all on your device.
On-device hand tracking
MediaPipe’s HandLandmarker detects 21 landmarks per hand from your webcam, running entirely in the browser — no video leaves your device.
Landmark normalization
Each frame is re-centered on the wrist and scaled by the distance to a stable knuckle, so hand size and position in frame never affect the score.
Rotation alignment
Frames are rotated to a common orientation, so the same hand shape scores consistently whether your hand is tilted, high, or low.
Motion sequence matching
Your recording is compared to the reference sign frame-by-frame using Dynamic Time Warping — an algorithm that elastically aligns two sequences even when performed at different speeds.
Speed-invariant scoring
The match cost is averaged over the actual aligned path length, so signing faster or slower than the reference no longer changes your accuracy score.
Facial expression scoring
A face landmarker extracts blendshape coefficients — eyebrows, mouth, eyes — and compares them to the reference, since expression carries meaning in BSL.
Two-handed signs are scored against primary and secondary reference hand sequences, with a graceful fallback when only one hand is shown.