Pick a letter or a short word, start the camera, and hold the sign steady. Practice uses the same recognition system as the live workspace, so results reflect real model behavior — including its mistakes.