The first really big adoption of conversational voice AI was in simple, customer-facing interactions: customer support, scheduling, and semi-scripted outbound calls to do things like remind people about appointments or qualify sales leads.
There's now a lot of interesting new stuff being built that pushes into more complex use cases and architectures.
Voice experiences have to: 1) be fast, responsive, and natural, 2) perform tool calls and instruction following very accurately across long, multi-turn interactions.
Most of the voice applications we build, today, use multiple models doing different kinds of inference in parallel, to achieve both 1 and 2. Here's a demo from Jon and Dom showing multiple people talking, with a voice agent in the call weighing in only when it's spoken to directly, or there's a semantic gap in the conversation. This uses Jev alongside a couple of LLMs.