Rime launches Prism, a fine-tunable speech-to-speech model that checks its words before it speaks

Rime launched Prism on October 28, 2026, a speech-to-speech model companies can fine-tune on their own calls and run on their own infrastructure — and it writes out its reply as text first, so a company's guardrails can vet every word before a caller hears it.

• Who: Rime, a San Francisco enterprise voice AI company founded in 2022.
• What: Prism, billed as the industry's first fine-tunable frontier speech-to-speech model.
• When: Launched October 28, 2026; available to try at rime.ai/prism.
• How it works: Spoken input becomes a spoken reply in one model; text is generated before audio so guardrails can vet it, and the model runs on-premises or in a private cloud.
• Speed: Rime lists 570 milliseconds to first audio on its homepage.
• Trial numbers (Rime-reported, Revmo): Completions 100% vs 68%; calls without hallucinations 85% vs 35%; calls without dead air 88% vs 66% — against an unspecified OpenAI model, with size and methodology undisclosed.
• Scale: More than 100 million interactions a month; customers include Mayo Clinic, Asurion, Upstart, and Dialpad.

A voice model enterprises can own

San Francisco voice AI company Rime launched Prism on October 28, 2026, billing it as the industry's first fine-tunable frontier speech-to-speech model. Unlike the text-to-speech and transcription tools Rime already sells, Prism takes spoken input and produces spoken replies in a single model, with no need to bolt separate speech-recognition, language-model, and speech-synthesis systems together. It is built for the calls companies say they cannot get wrong: customer service, healthcare scheduling, and financial services, where a wrong answer or a leaked recording carries real cost.

The pitch is ownership. Companies can fine-tune Prism on their own conversations and terminology, set its latency and guardrails, and run it entirely on their own infrastructure — on-premises or in a private cloud — rather than sending customer calls through a shared outside service. “A generic model with a custom prompt is not your voice. It's a rental with your logo on it,” said Lily Clifford, Rime's co-founder and CEO. “Prism is the first speech-to-speech model that becomes yours: it learns from your conversations, lets you set the latency and guardrails, and lives on your infrastructure.”

Text first, audio second

Prism's defining design choice is ordering: it generates text before it generates speech. That intermediate text step is what makes the rest of the product possible. A company's own rules, policies, and AI guardrails can review every word before the caller hears it, and the text itself becomes an instant audit trail — a strict requirement in regulated sectors such as finance and healthcare, where Rime says customers can run Prism on infrastructure configured to meet SOC 2 and HIPAA requirements.

The same control extends to how the model sounds. Companies can customize vocabulary and pronunciation — so product names come out right — and tune response latency. Rime lists 570 milliseconds to first audio on its homepage, a figure that matters on live calls, where even a short pause reads as a machine thinking.

Trained on real, messy conversation

Rime did not build Prism's base model by scraping the web. The company recorded thousands of hours of real, channel-separated human conversations in its own studios — the interruptions, hesitations, and overlapping dialogue that make phone calls messy — and then applied reinforcement-learning post-training. The result, Rime says, is a model that reads conversational signals such as frustration, hesitation, and confusion, and uses them to cut dead air and route calls better, rather than just sounding fluent.

The Revmo numbers — and their limits

Rime's boldest evidence comes from a real-world enterprise trial with Revmo, in which a fine-tuned Prism model went up against an OpenAI model. Rime reports that order confirmations and pickup-method completions rose from 68% with OpenAI to 100% with Prism; calls completed without hallucinations rose from 35% to 85%; and calls without dead air rose from 66% to 88%. Dialpad, which says it has been a design partner on Prism from the start, plans to roll the model out across its agentic products, according to Shezan Kazi, its VP of AI Products.

Those numbers deserve a clear caveat: Rime did not disclose which OpenAI model was used, how many calls the trial covered, or its methodology, so the comparison is Rime's reported result from one deployment, not an independent benchmark. What is independently checkable is the company's footing: founded in 2022 by voice researchers and engineers from Stanford and Amazon Alexa, backed by M13, Twilio Ventures, Corazon Capital, and Unusual Ventures, Rime says its voice models already power more than 100 million interactions a month for customers including the Mayo Clinic, Asurion, Upstart, and Dialpad.

Why this launch lands now

Prism arrives in a crowded week for enterprise voice: on the same day, NVIDIA opened a 40-language transcription model aimed at meetings and medical transcription. The two launches frame the same question from opposite sides — whether companies want a bigger general model, or a smaller one they can tune, audit, and host themselves. Rime is betting the second answer wins the call center, where the cost of a hallucinated return policy or a mishandled patient call dwarfs the cost of the model. Prism is available to try at rime.ai/prism.