Rime is launching Prism, a speech-to-speech model for enterprise voice agents on live calls. The model hears the caller's audio, writes the reply as text so a company's rules can check every word before anyone hears it, then speaks, and can call tools while it is still talking. It is built for teams with high call volume that need to review what an agent says and run the model in the cloud, a private VPC, or on premises, including healthcare and financial services. On the harness Artificial Analysis uses for speech-to-speech models, Rime reports a tau-Voice score of 56.6 with no reasoning step before speech, ahead of realtime models from OpenAI and Google that also skip that step, and ahead of GPT-Realtime-2.1 even with reasoning turned on. Time to first audio was 0.57 seconds, about half what GPT-Live-1, GPT-Realtime-2.1, and Gemini 3.8 Live took to start talking. Rime says Prism can be fine-tuned on a customer's own calls and is building self-serve training APIs with beta customers. On about ten hours of Revmo ordering calls, a trained Prism model got a yes before placing an order 94 percent of the time, against 57 percent for the OpenAI model already on those calls. At launch Prism is English only and ships with four voices: aurora, wawona, brando, and astra. CEO Lily Clifford, a Stanford NLP PhD dropout, linguist Brooke Larson, who previously worked on Amazon Alexa, and engineer Ares Geovanos founded the San Francisco company in 2022 and built a studio to record spontaneous conversation rather than scrape audio from the web. Rime already supplies voices to customers including Mayo Clinic, Dialpad, Upstart, and Asurion. The launch follows a $24 million Series A led by M13 in July 2026. The launch video puts Prism on a fake dating show with five men who are not told their date is a voice model.