Let's Dialog: Building the Next Voice Agents: Inside the Model (Dialog RSN-1)

Inside Dialog-RSN-1: The audio-native model for real conversation

AI models can write production code, draft contracts, and solve PhD-level math problems, yet on a live phone call they struggle with something anyone does without thinking: knowing when to speak. In a real-time call, a late answer leaves the caller sitting in silence, wondering if anyone’s there.

PolyAI’s dialog research team built Dialog-RSN-1 for exactly that problem. It hears the caller’s audio and responds in a single pass, with no separate speech-recognition step, in under 300 milliseconds. In our tests of real enterprise calls, it beats every other model, fast enough to keep a live conversation going.

Join the researchers who built it for an honest look at how it works on real call audio. You’ll hear how it distinguishes a pause from the end of a turn, handles noise and interruptions, and decides when a turn needs real thought.

This is Part 1 of Let’s Dialog: Building the Next Voice Agents. In Part 2, we build a live agent on Dialog-RSN-1.

What you’ll take away

• Hear what audio-native actually changes. What you gain by removing the separate speech-recognition stage, and what a transcript throws away. Demonstrated on real calls.
• See how it handles messy, real calls. Turn-taking, barge-in, crosstalk, and noisy phone lines, and why we train for them rather than patch around them.
• Learn how selective reasoning keeps calls fast. How the model decides per turn whether to think, so most turns stay instant and hard turns get real deliberation.
• Leave with a checklist. Questions to ask any voice AI vendor about turn-taking, latency, and audio handling, and where the model ends and the harness begins.