The voice mode will be available in "a few weeks". So at the moment it's not using the end-to-end model but whisper->gtp-4o->tts