01 — Pipeline
Holistic landmarks
21 hand × 2, 33 body, and 468 face points per frame from MediaPipe Holistic, fused into a single feature record.
Sign language, made shareable
A bridge between signed and spoken.
BridgeTalk fuses hand tracking, body pose, and facial landmarks to recognize location-anchored signs — touching the chin for thank you, the chest for me, the ear for hear. Real signing is more than gestures in the air; we read where the hand meets the body.
~90
Curated signs
543
Landmarks / frame
26
Alphabet letters (ML)
0
Servers, accounts, uploads
02 — Anchors
14 body regions
Chin, forehead, mouth, nose, ears, cheeks, temples, chest, shoulders, neutral space, lap. Sized to each person's face width.
03 — Contact
Fingertip ↔ region
Proximity with dwell tracking. Distinguishes a sustained touch from an accidental pass-through.
04 — Vocabulary
~90 signs
Location, motion, and shape signs combined: thank you, eat, hear, think, sorry, me, you, plus numbers and fingerspelling fallback.
05 — Sentence
Real-time text
Auto-spaced output with capitalization, punctuation, undo and clear. Read aloud with the device's best voice. History saved locally.
06 — Privacy
Stays in your browser
No server uploads, no account. Holistic runs locally; only the optional alphabet ML uses your own machine's Python backend.
07 — Tunable
Tweakable in-app
Settings panel lets you adjust confidence, hold-time, cooldown, overlay layers, TTS speed, and audio feedback live.
08 — Open
Retrainable model
The alphabet model is a small RandomForest. Train it on real keypoint data with one script and drop the pickle into models/.
Honest scope
This is a real recognition engine, not a research demo of full sign-language translation. There is no pretrained, browser-ready model in 2026 that recognizes general ASL or PSL conversation — the state of the art needs server-side GPUs and still tops out around 60–70% on a few thousand isolated signs.
What you get here: ~90 signs that work reliably because their multimodal signature (handshape × location × motion) is unambiguous. The recognizer is built so a trained model can be plugged in later via a one-line hook. See the README for the roadmap.
Try these first
01
thank you
Open palm, fingertips touch chin, then move outward.
02
me
Index finger pointing at your own chest.
03
hello
Open palm at temple/forehead, small wave outward.
04
yes
Closed fist, knuckles facing camera, nod down.