Push-to-talk audio is already recorded, so it doesn't need a WebSocket. Transcribe short clips in one HTTP request, and see where the latency hides.