Every enterprise voice AI vendor has been selling the same choice. Turn the call into text first, and the AI loses the caller's tone of voice. Keep the audio the whole time instead, and you lose control of the voice, plus a chunk of your compute budget. PolyAI released a new model this week, Dialog-RSN-1, that it says avoids both problems.

PolyAI's own framing lines up with what it told investors eighteen months ago. Nvidia's venture arm backed PolyAI's Series D on the idea that real-time voice AI needs a lot of graphics processing power (PolyAI, 2026). Dialog-RSN-1 is PolyAI's attempt to need less of it, at least for the always-on kind.

Two ways to build a voice AI, one shared cost problem

A cascaded voice system works in steps. It turns the caller's speech into text, then hands that text to a large language model to figure out a reply. The problem is the text is only ever one guess at what the caller said. Tone, hesitation, a misheard word: all of it disappears once the words hit the page. Making a cascade work well means hand-tuning it for every use case, and those tweaks rarely carry over to a different one.

Speech-to-speech models fix the tone problem by keeping the audio the whole way through. OpenAI's GPT Realtime and Google's Gemini Live both hear the raw sound and speak back in raw audio. But the voice they speak in is baked into the model, so a brand cannot fully control it. And the always-on versions keep a graphics processing unit running for the length of every call, which is the most expensive way to run this kind of AI. That expensive setup sits underneath a lot of enterprise voice AI pricing today.

PolyAI claims some of those models may not be doing what they advertise. In its testing, GPT Realtime 2.1 scored about the same as cascaded systems on audio tests, which PolyAI reads as a sign it may be a well-engineered cascade rather than a model that reasons over raw audio (PolyAI, 2026). That claim comes from PolyAI's own testing, so a reader cannot check it independently.

Audio in, speech out, power used only when needed

Dialog-RSN-1 combines listening and deciding what to say into one model. It hears the raw sound of the call, but hands the talking to a separate text-to-speech system. PolyAI's bet is that listening is where an audio-aware model matters most, and that a good text-to-speech system, one you can prompt to sound a certain way, gets you most of the benefit without giving up control of the voice.

Instead of running all the time, Dialog-RSN-1 turns on when it is needed. A separate, simple system watches for sounds and the clock, and pings the model whenever something might be worth responding to. That system is set to check often, since a false check costs almost nothing. The model itself makes the real decision about whether to speak. This setup lets PolyAI free up the graphics processing unit between checks instead of holding it for the whole call.

PolyAI wants Dialog-RSN-1 to respond in under 300 milliseconds on Nvidia A100 chips, which keeps the model's size in a specific range: big enough to reason well, small enough to stay fast (PolyAI, 2026). A bigger, slower model would defeat the purpose of a real-time voice agent, no matter how well it scores otherwise.

The numbers behind the announcement

PolyAI built its own test, called Dialog-Eval, to check whether an agent can get what it needs from a caller and knows when to hand the call off to a human. Among models fast enough to run in real time, Dialog-RSN-1 scored the highest overall. The biggest gap showed up on tasks that need audio, like catching an accent or an unusual term from context (PolyAI, 2026).

Two real customer results back up the lab numbers. A national restaurant group resolved 11 percent more calls without handing off to a human after switching to Dialog-RSN-1. A large insurance provider cut its response time by 37 percent after the same switch (PolyAI, 2026).

An 11 percent gain in calls resolved without a human is a staffing number before it is a model number.

Speech recognition accuracy tells a similar story. PolyAI gave the same conversation context to competing transcription tools and its own model. Dialog-RSN-1 still made the fewest errors, and even OpenAI's gpt-4o-transcribe improved once it got that same context, dropping from a 7.8 percent error rate to 6.9 percent (PolyAI, 2026).

None of these numbers come from an outside source.

PolyAI graded its own model on its own test, using its own customers' calls.

What this means if you're comparing voice AI vendors

Companies buying voice AI usually judge it on two separate questions: how good does it sound, and how much does it cost to run. Dialog-RSN-1 is PolyAI's argument that those two questions have gotten tied together by how the models were built, an architecture choice rather than a fixed law of physics. If a check-in model can match the quality of an always-on model at scale, the cost math that has shaped speech-to-speech pricing stops being a fixed rule of the category.

This release only covers English. PolyAI says the underlying model can handle other languages without extra training, and has already picked up non-English speech in production. But right now, a company running calls in multiple languages has to choose between Dialog-RSN-1 for English and PolyAI's separate Raven 3.5 model for everything else (PolyAI, 2026). That's an extra integration to plan for.