07 / Project / Jan 2026
SilentSpeaker
A real-time ASL accessibility platform that translates sign to text or speech, and text or speech back to sign, right in the browser.
- Model accuracy
- ~94%
- Labeled gestures collected
- 7,000+
- Hackathon build
- 36h
- Translation directions
- 2
Overview
Most ASL translation tools either work poorly or ask you to upload video to a server. In 36 hours at HackTheBias, I led a team of four to build SilentSpeaker: a real-time accessibility platform that translates sign language into text or speech, and text or speech back into sign, running in an ordinary web browser.
Sign to text and speech
We trained an ASL model with TensorFlow.js and MediaPipe that recognizes alphabets and numbers fully on the client, with no uploads, so it is private and fast. The team manually collected and verified more than 7,000 labeled hand-gesture samples, and three training iterations took accuracy from about 40% to 60% and finally to about 94%.
Text and speech to sign
The reverse direction runs on a FastAPI backend that turns English into a signing video. It matches multi-word phrases against the sign dataset, falls back to finger spelling letter by letter when a word has no full sign, then stitches the clips together with natural pauses.