The Rubicon of Real-Time Voice-to-Voice

In a year we'll have audio books that I wouldn't be surprised if humans actually preferred the audio books that are generated over the ones that are required by humans. We will cross a Rubicon beyond which certain elements of society will forever be changed. There's like layers of abstraction that make it really difficult to shave the last few milliseconds. The realism goes up with larger models. Larger models are slow to serve. And so 300 milliseconds is not bad for certain experiences and certain bits and pieces. It is terrible for real-time communication because you probably want that to be under 100 milliseconds ideally.