Prism wants to make voice agents as natural as a human conversation, without losing control of their responses.

Rime launched Prism, an S2S AI voice model with a 570 ms latency and text checkpoint, outperforming OpenAI GPT-Live-1 and Gemini 3.8 Live in τ-Voice tests.

An AI that hears frustration in someone's voice, recognizes hesitation, and adapts its response before speaking. With Prism, San Francisco-based Rime has introduced a speech-to-speech model designed to make conversational agents sound and behave more naturally without sacrificing control over what they say.

Its central proposition is to combine capabilities often treated as competing priorities: understanding audio directly, producing natural spoken responses, and allowing businesses to inspect the exact words an agent is about to deliver.

Announced on October 28, 2026, Prism targets large-scale customer-service operations, healthcare organizations, financial services, and other applications where voice agents must operate reliably.

Rime claims its model responds faster than several competing voice systems from OpenAI and Google while delivering stronger results on selected customer-service tasks. Most of the published performance comparisons, however, were conducted by Rime itself. Listening to the voice, not just its transcription

Many conventional voice agents use a three-stage architecture: speech recognition converts audio into text, a language model generates a response, and text-to-speech converts that response into audio.

This cascaded design gives developers control over each component, but it can lose information present in the original voice.

Someone saying “yes” might sound enthusiastic, uncertain, irritated, or sarcastic. The transcript is identical, but the meaning of the interaction may be different.

Prism belongs to the emerging category of speech-to-speech (S2S) systems, which process the audio signal directly rather than relying entirely on a transcription as the starting point for reasoning.

Rime says its model can respond to rhythm, intonation, laughter, hesitation, and vocal cues associated with different emotional states.

It can also coordinate spoken output with tool calls and continue processing conversational signals while generating its response.

That does not mean the model can determine someone's emotional state with certainty. Interpreting vocal cues remains probabilistic, and accuracy can vary across languages, speakers, accents, and conversational contexts. An inspectable text checkpoint before speech

One of Prism's most distinctive design choices involves the way it exposes its response to developers.

Some integrated speech-to-speech systems combine audio understanding and generation so tightly that applications have limited opportunities to inspect a complete spoken reply before it reaches the user.

Prism introduces an intermediate checkpoint.

The model decides what to say based on the audio, then emits the response as text before producing the corresponding speech.

That text is accessible to the application.

Businesses can inspect it, apply internal policies, block problematic wording, or modify parts of a reply before allowing the agent to speak.

This is particularly relevant to regulated workflows that require specific disclosures, accurate account information, or restrictions on what an automated agent may promise.

Rime also describes pronunciation controls and phonetic conditioning for names, account references, and specialized terminology.

The architecture aims to preserve the benefits of direct audio understanding while introducing a controllable text layer.

However, the existence of a checkpoint does not guarantee that every inappropriate or incorrect response will be prevented. Its effectiveness depends on the policies and validation mechanisms implemented by the application. A 300 ms headline claim, but 570 ms in the published evaluation

Latency is a central part of Rime's positioning.

In her launch announcement on X, cofounder and CEO Lily Clifford described Prism as having a 300-millisecond response time, claiming it was four times faster than OpenAI's GPT-Live.

The company's technical launch article presents different figures.

It describes response times of approximately 500 milliseconds and reports 0.57 seconds to first audio in its comparative evaluation.

Under Rime's testing conditions, that was approximately half the delay measured for GPT-Live-1, GPT-Realtime-2.1, and Gemini 3.8 Live.

The social media claim of 300 milliseconds therefore should not be treated as equivalent to the benchmark's measured 570 milliseconds.

Without a published methodology reconciling those figures, 570 milliseconds is the better-documented number for the company's comparative test.

Time to first audio is also different from total task completion time. Tool calls, telephony infrastructure, and other application components can add delays. Strong results on an Artificial Analysis-style voice benchmark

Rime evaluated Prism using a harness based on the τ-Voice benchmark employed by Artificial Analysis.

The benchmark examines voice agents on simulated customer-service tasks involving airline bookings, retail interactions, and telecommunications.

According to Rime, Prism scored 56.6 without a separate reasoning step before speaking.

The company reports that this exceeds the scores of the tested OpenAI and Google realtime models operating without pre-response reasoning by 17 to 30 points.

On telecommunications tasks, which often involve account information and phone numbers, Prism scored 82.5, compared with 45.6 for GPT-Live-1.

Rime also acknowledges that two systems using additional reasoning before speaking achieved higher overall scores than Prism. Those models required approximately 1.24 to 1.35 seconds to begin speaking, compared with Prism's 0.57 seconds.

Prism therefore does not lead every evaluated model on every metric.

More importantly, Rime ran these comparisons itself using the Artificial Analysis benchmark framework. The results should not be confused with an independent evaluation directly published on Artificial Analysis's public leaderboard. Reinforcement learning for specialized voice agents

Rime wants companies to adapt Prism beyond what can be achieved through instructions alone.

The