Voice agents sound simple when you explain the pipeline: speech comes in, text gets processed, and speech comes back out. But how smooth is the  actual conversation?

In this interview, Luke Farrell talks with Kræn Hansen (ElevenLabs) about the problems users notice first when talking to voice agents. Latency matters, especially perceived latency: if users hear silence, they assume the connection dropped. If they hear cues that something is happening, the interaction feels intact.

The conversation gets into turn-taking, interruption handling, and audio feedback. It covers how voice agents can use spoken cues, background sounds, and better intent detection to avoid awkward cutoffs. It also gets into the tradeoff between speed, expressiveness, and safety when the agent is used in more sensitive settings.

The interview also covers implementation. Kræn walks through a layered SDK approach: a universal JavaScript core, React abstractions on top, and a React Native package for platform-specific audio input.