Rime Prism: Speech-to-Speech That Writes the Reply Before It Speaks

Rime launched Prism on 28 October 2026, an end-to-end speech-to-speech model that answers caller audio in about 300ms, runs tool calls in parallel with speech, and generates an inspectable text layer so guardrails can read every word before it is spoken. It deploys in the cloud, on-prem or at a co-located endpoint and can be fine-tuned on your own calls

Tool Interface

Interactive tool will be available soon

Features

✓ Rime reports about 300ms response to caller audio, with tool calls running in parallel with speech instead of waiting for the sentence to finish
✓ Text is a checkpoint: guardrails can read every word before speech, and the same layer supports rewritable action and content tags plus phonemic conditioning for names and account terms
✓ One end-to-end system does three jobs: it listens, it writes what it will say, and it keeps listening while speaking, so tone and background noise are not thrown away
✓ Rime's own chart puts τ-Voice task completion at 56.6% with 0.57s time to first audio, against GPT-Live-1 (Sol, low) at 59.3% and 1.24s and Gemini 3.8 Live at 30.1% and 1.18s
✓ Fine-tunable and self-hostable across cloud, on-prem and co-located endpoints, which is the differentiator Rime leans on: the model becomes yours rather than a rental

How to Use

Classify your call flows first: high-stakes flows such as refunds or contract changes are where a text checkpoint pays off, while low-risk flows are cheaper on a lighter model
Write guardrails as rules that run on the text layer: banned phrasing, mandatory disclosures, amount and date checks, all enforced before speech
Fine-tune on real call recordings, especially place names, account terms and product names that get mispronounced, and use phonemic conditioning for them
Before launch measure three things: time to first audio, task completion, and behaviour under interruption. Whether it sounds human is the least useful metric

FAQ

When did it launch?

28 October 2026. Product page https://www.rime.ai/prism and launch post https://www.rime.ai/resources/prism-first-speech-to-speech-for-enterprise .

Does writing text first add latency?

Rime's answer is that the text is a checkpoint, not a cascade: the decision happens on audio and the text exists to run guardrails and rewrites. Rime quotes about 300ms response and 0.57s to first audio. Real latency still depends on how many guardrail rules you stack.

How is this different from ASR plus LLM plus TTS?

Rime's argument is that a cascade loses tone, hesitation, laughter and background noise at transcription, while an end-to-end model keeps them. The cascade's advantage of inspectable control is what the text checkpoint restores.

Can I self-host it?

Rime lists cloud, on-prem and co-located endpoint deployments plus fine-tuning on your own calls. Licensing and commercial terms come from Rime directly.

Is 56.6% on τ-Voice the top score?

No. Rime's own chart shows GPT-Live-1 (Sol, low) at 59.3%, but with 1.24s time to first audio and reasoning enabled. Prism trades a little task completion for 0.57s, so the pitch is speed rather than the absolute best completion rate.