also on a custom tuned model I got my response time down to 700ms on vapi, that's pretty damn close to realistic phone conversations