Rime Prism

Prism is a speech-to-speech (s2s) model for enterprise voice agents from Rime (Rime Labs), launched on October 28, 2026 by CEO Lily Clifford. Rime calls it "the first composable speech-to-speech model": it reasons over the caller's real audio, writes its reply as text that your guardrails can check, and then speaks while still listening.

Announcement: Introducing Prism · Product page: rime.ai/prism · Launch post: Lily Clifford on X

Key points

1. The problem it targets: big-lab s2s models fuse text and audio at runtime, so you can't inspect or rewrite what the agent is about to say. Classic cascades (speech-to-text, LLM, text-to-speech) give control but throw away tone, hesitation, laughter and background sound.
2. How it works: one end-to-end model that listens, writes what it will say, and keeps hearing the user while it speaks. The text is a checkpoint: the decision is made over audio, your rules read every word before it becomes speech, and the speech still carries the caller's emotion.
3. Extras at the checkpoint: rewritable text for action and content tags, phonemic conditioning for names and account terms, and tool management for composable agent flows. Tool calls run in parallel with speech.
4. Speed: about 300ms response time per the launch post, and 0.57s to first audio on the Artificial Analysis harness, about half the wait of GPT-Live-1, GPT-Realtime-2.1 and Gemini 3.8 Live.
5. Accuracy: 56.6 on τ-Voice (airline, retail and telecom customer-service calls) without a reasoning step, ahead of every non-reasoning realtime model from OpenAI and Google by 17 to 30 points. In telecom it scores 82.5 vs 45.6 for GPT-Live-1.
6. Ownership: Rime says Prism is the first s2s model anyone can fine-tune, including with reinforcement learning on your own calls, and it can run on-premises, in your VPC or in the cloud. Dialpad is a design partner.

How to try it

1. Sign up on rime.ai/prism ("Try Prism for free"). The launch post offered the first 100 sign-ups one month free.
2. Connect it to your voice-agent stack and add your policy checks at the text checkpoint, so every reply is checked before the caller hears it.
3. For production, fine-tune on your own call data and measure against business outcomes such as resolution rate or confirmed orders.

Good fits

• Contact centers and customer-service voice agents
• Regulated flows that need required disclosures and guardrails on every spoken line
• Teams that want to own and deploy their voice model on their own infrastructure

Related

Rime's text-to-speech models Coda (flagship) and Mist v3 (lowest latency) are for cascade systems.