Also 575ms is unrealistic you’re going to get 700-900 for what I call medium level agents (4o mini cluster, cartesia, deepgram nova 3) which works perfect for most use cases and for premium voice models you’ll get 800ms-1.1 seconds