Skip to content

Live recognition

Perform a sign in front of your camera and see what the model reads. Everything runs locally in your browser — no video is uploaded.

Why the vocabulary is small

The model can only recognize signs it has seen enough examples of. It currently covers a handful of signs, because that is what the training set holds. This is exactly the bottleneck contributions solve: more varied, approved clips provide training material. New signs also need model training and evaluation.

Add a sign

Getting the best results

  • Sit so your head and both hands are in frame.
  • Start and finish each sign with your hands at rest — that is how signs get separated.
  • Avoid a busy background directly behind your hands.
  • Lower the confidence threshold to see what the model is considering.