speech to speech
A model that listens to audio and answers with audio directly, without converting your words to text in between.
Older voice assistants worked in three steps: turn your speech into text, send that text to a language model, then read the answer back out loud with a separate voice. A speech to speech model does the whole thing itself. It takes the sound in and produces sound out, which removes two handoffs and the delay that comes with them.
The difference is easy to hear. Because nothing gets flattened into text along the way, the model can pick up on tone, interruptions and the fact that you started a sentence over, and it can begin answering before you have quite finished. That is why a speech to speech assistant feels like a phone call while the older kind feels like taking turns with a walkie talkie.