Prism
Rime's speech-to-speech model for enterprise voice agents that writes inspectable text before speaking and can be fine-tuned.
Overview
Prism is Rime's speech-to-speech model, announced October 28, 2026. It is trained end-to-end to listen, write the text it will say, and speak while still hearing the caller, so guardrails can inspect or rewrite the text before it becomes audio, and tool calls run in parallel with speech. Rime reports 0.57 seconds to first audio and a 56.6 overall score on the tau-Voice benchmark with no reasoning before speaking, ahead of non-reasoning OpenAI and Google realtime models by 17 to 30 points (company-reported). Access is through a Rime API key with Python and Node.js SDKs or a WebSocket endpoint; self-serve RL and SFT fine-tuning APIs are in beta with selected customers.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- Inspectable text layer between listening and speaking, with guardrail hooks
- Parallel tool calls while speech is being generated
- Reported 0.57 s to first audio and 56.6 on tau-Voice (company-reported)
- Phonemic conditioning for names and account terms; content and action tags
- Fine-tuning via RL and SFT APIs, in beta with customers
Use cases
- Customer-service and telecom voice agents needing auditable responses
- Healthcare and financial voice workflows with compliance review of agent text
- Voice agents that must call backend tools mid-conversation
Pricing
Not published in the announcement; free trial and demo request offered. SDK and WebSocket access via a Rime API key.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 28 Oct 2026
- Record updated
- 2 Nov 2026
- Record last checked
- 2 Nov 2026