News

Releases & Models

CONFIRMED: Rime launches Prism, a speech-to-speech model that kills the robotic pause in AI phone calls

Rime has launched Prism, a speech-to-speech model built for voice agents that cannot sound robotic: it listens to raw audio, replies in under a second, and avoids dead air on the call.

Ilustração editorial de uma atendente com headset conversando com uma presença de IA representada por ondas sonoras luminosas, simbolizando um modelo de fala para fala em agente de voz

The end of the robotic pause

Rime, an AI voice startup, has launched Prism, a speech-to-speech model built for voice agents that answer calls for banks, hospitals or any call center that cannot afford to sound artificial. Instead of turning speech into text, sending that text to a language model and only then generating audio, Prism listens to the caller's raw audio directly, reads the tone of voice and replies in the same pass, as if it heard, thought and spoke as a single piece.

How this changes a real call

The classic voice agent architecture stacks three separate services, each with its own latency: speech recognition, a language model and voice synthesis. Prism replaces that chain with a single model that, according to Rime's own documentation, starts speaking about 0.57 seconds after the caller finishes talking, as measured on Artificial Analysis's speech-to-speech benchmark. When it needs to check a backend system, Prism says something natural like let me check that instead of leaving dead air, which is exactly the moment that makes a caller ask if anyone is still on the line.

Transcript alongside the audio, not just the voice

Every spoken reply from Prism comes with its matching text, enabling live captions and an audit log, a sensitive point for teams operating in regulated sectors like banking and healthcare. Rime also says client companies can edit the transcript and block any phrase the agent should not say, without stopping the ongoing call. It is the kind of compliance control that speech-to-speech models used to trade away in the name of natural conversation.

The real limit: English only, for now

Prism currently works only in English, with four proprietary voices named aurora, wawona, brando and astra, which do not mix with Rime's own catalog of more than 600 text-to-speech voices. The connection runs over a dedicated WebSocket hosted in the United States, with a continuous speaking turn capped at 55 seconds and a session that closes after 5 minutes of silence. Teams already running pipelines on OpenAI's Realtime API have a direct migration path, since Prism uses that interface's own event names.

Why this matters for the voice market

The voice agent market is in a race that includes Rime, ElevenLabs, OpenAI and a line of text-to-speech and speech-to-speech startups. Rime's bet is not to win on voice count, it is to win on latency and context continuity during the call, the two factors that decide whether a customer hangs up or stays on the line until the problem is solved. That competes directly with cascade architectures that stack transcription, a language model and voice synthesis as separate services, each one adding delay.

Sources

Rime: official Prism documentation; TestingCatalog: Prism launch announcement; Rime: official voice products page.