Insider Brief
- Rime launched Prism, a customizable speech-to-speech AI model that companies can fine-tune and run on-premises or in a private cloud.
- Prism generates text before speech, allowing businesses to apply their own policies and safety guardrails, maintain audit trails and customize vocabulary, pronunciation and response latency.
- In a Revmo trial, Rime said Prism reached 100% for order confirmations and pickup-method completions, compared with 68% for an unspecified OpenAI model, though the company did not disclose the trial size or methodology.
Rime has launched Prism, a customizable speech-to-speech AI model designed to let companies fine-tune voice systems and run them on their own infrastructure.
According to the San Francisco voice AI company, Prism can be deployed on-premises or in a private cloud, giving businesses control over customer data, latency, vocabulary, pronunciation and safety rules. Rime describes Prism as the first fine-tunable frontier speech-to-speech model.
Speech-to-speech models take spoken input and generate spoken responses without requiring companies to assemble separate speech-recognition, language-model and text-to-speech systems. Rime is going after enterprise applications such as customer service, where companies may need more control over how an AI agent responds and how conversations are recorded and reviewed.
Unlike systems that generate speech directly, Prism first produces text internally before converting the response to speech. Rime said that design allows company policies and guardrails to check the response before it reaches a customer and creates a text record for auditing.
The model can also be fine-tuned on a company’s own conversations and terminology. Rime said Prism was trained using human conversation data recorded through its own studios and additional reinforcement-learning post-training. Rime said Prism is designed to recognize conversational signals such as frustration, hesitation and confusion and use those signals when responding or routing a call.
“A generic model with a custom prompt is not your voice,” co-founder and CEO Lily Clifford said in the announcement. “It’s a rental with your logo on it. 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. Quickly, the model running your calls will be one no competitor can buy.”
Enterprise Controls
Companies can adjust pronunciation, vocabulary and response latency and establish their own safety policies. Since industries such as healthcare and financial services, where customer conversations may contain sensitive information and responses may need to follow specific policies, customers can operate Prism within infrastructure configured to meet SOC 2 and HIPAA requirements rather than sending voice data through a shared external service.
Testing Against OpenAI
Dialpad has worked with Rime as a design partner on Prism and plans to deploy the model across its agentic AI products and Rime also released results from a customer trial with Revmo comparing a fine-tuned Prism model with an OpenAI model.
According to Rime, order confirmations and pickup-method completions reached 100% with Prism, compared with 68% using OpenAI. The share of calls without hallucinations increased to 85% from 35%, while calls without periods of dead air increased to 88% from 66%.
The company did not provide additional details in the announcement about the OpenAI model used, the number of calls in the trial or the methodology behind the comparison, so the results represent Rime’s reported performance in that specific deployment.
Rime, which is backed by M13, Twilio Ventures, Corazon Capital and Unusual Ventures, was founded in 2022 by researchers and engineers with backgrounds at Stanford and Amazon Alexa. The company said its voice models currently support more than 100 million interactions a month for customers including Asurion, Mayo Clinic, Upstart and Dialpad.