Rime launches voice AI that lets companies check replies before they are spoken

Rime launched Prism on October 28, 2026, promising businesses voice AI that hears conversational cues while letting them check replies before callers hear them. The model can be customized and privately deployed. That control is the core proposition; the evidence for better call handling rests on Rime’s own benchmark and customer-trial results.

A checkpoint between listening and speaking

Prism is a speech-to-speech model: it takes spoken input and produces a spoken response. Rime says it trains the listening, response-writing and speaking functions together, rather than requiring businesses to assemble separate speech-recognition, language-model and speech-generation systems. It also continues listening while it speaks.

The distinction is where text enters the process. Rime says Prism makes its response decision from audio, retaining information such as hesitation, frustration, laughter and background sounds. It then exposes the proposed words at a text checkpoint. Company rules can inspect or rewrite that text before it becomes speech.

CEO Lily Clifford frames this as an alternative to two compromises: reducing a conversation to a transcript before deciding what to say, or processing audio directly without an editable response checkpoint. That is Rime’s argument for the architecture, not proof that every competing system makes the same tradeoff.

• Response checks: apply content and safety rules before speech, with a text record for auditing.
• Pronunciation controls: guide how names and account-specific terms sound.
• Connected actions: manage software-tool requests, which Rime says can run alongside speech.

Customization extends beyond a prompt

Rime says companies can fine-tune Prism on their own conversations and terminology, then run it on-premises or in a private cloud. According to AI Insider’s launch coverage, businesses can also adjust vocabulary, pronunciation and response latency. The deployment options give customers a way to keep the system within their own infrastructure.

Some of the customization tooling remains in development. Rime says it is building self-service training interfaces with beta customers for supervised fine-tuning and reinforcement learning—two methods for adapting model behavior. It invites users to try Prism for free, but that invitation should not be confused with a claim that the self-service training tools are finished.

Dialpad worked with Rime as a design partner and plans to deploy Prism across its agentic AI products, according to AI Insider. That is a launch-linked adoption plan, rather than a completed rollout with disclosed operating results.

Faster speech is not the highest task score

Rime tested Prism using the evaluation setup behind Artificial Analysis’s τ-Voice benchmark, which measures voice-agent tool use and instruction following on simulated airline, retail and telecom customer-service calls. Rime ran Prism without a separate reasoning step before speaking and reported an overall score of 56.6.

The company reports 0.57 seconds to first audio in that evaluation. Two models scored higher, but both used reasoning before speaking and took 1.24 to 1.35 seconds to start talking. Those results expose a tradeoff inside Rime’s own comparison: the fastest response and the strongest task performance were not the same outcome.

The launch page separately advertises responses in about 300 milliseconds. That is not the same figure as its benchmark’s time to first audio. For the task comparison, 0.57 seconds is the reported measurement tied to the evaluation.

Rime also describes adapting Prism for ordering-agent company Revmo using 10 hours of production calls, software-tool traces and transcripts. It built a reinforcement-learning environment around that material. On Revmo calls, Rime says 94% of orders were placed only after the caller explicitly agreed, compared with 57% for OpenAI.

That measures a specific behavior—getting permission before placing an order—not overall call quality. AI Insider reports that Rime did not disclose the comparison’s call count, methodology or which OpenAI model it used. Those omissions limit how far businesses can generalize the result beyond the customer example.