Rime's Prism writes text before it speaks, so rules can check it

Rime, a San Francisco voice AI company founded in 2022, launched Prism on October 28. Prism is a speech-to-speech model: it takes a caller's audio in and produces spoken replies, instead of chaining separate transcription, language-model, and text-to-speech services. Rime says one WebSocket connection does the work of all three, and that Prism hears the caller's audio itself rather than reading only a transcript.

The part Rime leads with is the checkpoint in the middle. Prism produces text before it produces audio, Rime says, so a company's own rules can check the reply before a caller hears it, and every spoken reply arrives with its text for captions and logs. Rime calls Prism the first fine-tunable frontier speech-to-speech model; that “first” is Rime's description, not an independent finding. As CEO Lily Clifford put it in the announcement: “A generic model with a custom prompt is not your voice. It’s a rental with your logo on it.”

Rime reports that Prism produces its first audio about 0.57 seconds after a caller finishes a turn, and that it scores 56.6 on the tau-Voice benchmark, which simulates customer-service calls. Rime's launch post says it ran Prism itself on the harness Artificial Analysis uses for its speech-to-speech leaderboard, with no reasoning step before it speaks. Prism did not appear in the Artificial Analysis table checked for this story, and neither result has been independently verified. On that leaderboard, time to first audio is an average over its Big Bench Audio questions.

Rime says the 56.6 is ahead of real-time models from OpenAI and Google that do not reason first, by 17 to 30 points, and that Prism reaches first audio in about half the time of GPT-Live-1, GPT-Realtime-2.1, and Gemini 3.8 Live. Those comparisons are Rime's. Artificial Analysis's table lists several models with higher tau-Voice scores, among them GPT-Live-1 (Astra, medium) at 74.5 and Grok Voice 3.0 at 71.6; Rime says the models that score above Prism reason before they speak. Rime's launch post also says Prism responds in “about 300ms”; the sources checked do not define that figure or say how it relates to the 0.57-second benchmark number.

In a trial with Revmo, Rime says it fine-tuned Prism on about ten hours of Revmo's production ordering calls and compared it with the OpenAI model behind Revmo's agents on the same calls. On Rime's measure, whether the agent asked “shall I place it?” and got a yes before placing the order, Rime says 94% of Prism's orders met it, against 57% for OpenAI. Rime also says order confirmations and pickup-method completions reached 100% against 68%, calls free of hallucinations 85% against 35%, and calls without dead air 88% against 66%. Rime did not disclose how many calls were in the trial or which OpenAI model was used, as The AI Insider notes, so read those figures as one vendor's trial result, not proof.

Prism is English only at launch, with four voices - aurora, wawona, brando, and astra - and Rime's docs list the practical limits: a caller's turn ends after 55 seconds of continuous speech, a conversation closes after 5 minutes idle, and context is 65,536 tokens. Rime says companies can run Prism in a private cloud or on their own infrastructure, and that it is building self-serve fine-tuning tools with beta customers. Dialpad, a design partner, says it plans to roll Prism out across its agentic products. The launch materials checked list no pricing.

Open questions: whether Dialpad and other customers report similar completion rates on live calls, and whether independent benchmarks reproduce the 0.57-second figure outside Rime's own runs. Rime's stated pitch is a voice model a company can inspect, tune, and run where its data lives.