Use a guided turn-based voice path instead of a raw websocket experiment.

This screen keeps the realtime test loop simple: choose languages, pick devices, capture one turn, then inspect transcript, translation, and playback in one place.