Live demo · v3.1.1
Sign, then listen
Allow the camera, tap Start signing, and keep your hands in the frame. For the dedicated model, sign one isolated sign from the 200-word vocab, or several with a short pause after each (hello, lunch, Coca-Cola…). Songs and fluent conversation use Gemini video instead. Unmute so the English can be spoken automatically.
v3.1.1What's new in v3.1.1Patch notes
Longer webcam clips now stay small enough to reach Gemini video, and a few signs in a row are less likely to be thrown away for one soft piece. One-word clips are unchanged.
- The camera records at 640×480 / 15 fps / 480 kbps so a 30-second clip stays under Vercel’s upload limit. If a recording is still too large, the app says so instead of failing later.
- Gemini video retries once with the other container type (webm ↔ mp4) when the first mime guess is wrong. Unknown magic bytes no longer block the fallback.
- Several-signs reading pads more generously, drops wrist twitches, and uses slightly looser gates (or soft top-1 guesses) so a real webcam phrase can still become English. Isolated-sign gates stay 55% / 0.15 / 0.75.
- If translation still fails, the error names one tip — pause between signs, or sign one short word — and does not blame Google.