THE ESSENTIALS IN 30 SECONDS

  • Prism handles audio in and audio out in one model, but drafts each reply as text that guardrails can read and edit before it becomes speech.
  • Rime reports 0.57 seconds to first audio and a τ-Voice score of 56.6, ahead of non-reasoning realtime models from OpenAI and Google; these are company figures.
  • A free trial is open; self-serve fine-tuning APIs are still being built with beta customers.

Rime, a San Francisco company that supplies synthetic voices for customer phone lines, released Prism on October 28. It is the company's first speech-to-speech model: a single system that listens to a caller and answers out loud, instead of chaining separate transcription, language and voice models. Rime is aiming it squarely at enterprises, which typically need to control what an automated agent tells their customers.

A text checkpoint inside the model

Prism's distinguishing feature is that, although trained end to end, it writes a draft of each reply as text before turning it into audio, and keeps listening while it talks. Rime says this written layer serves as a checkpoint. A company's guardrail rules can inspect every word, the text can be rewritten to trigger actions or add content tags, and pronunciation can be steered for names and account terms. Tool calls run alongside speech rather than before it, and the model does not reason before answering. Dealroom adds that the text gives companies an audit trail and that Prism can run in a private cloud or on premises.

Speed and benchmark claims

According to Rime, Prism needs about 0.57 seconds to produce its first audio in Artificial Analysis's test setup, roughly half the delay of GPT-Live-1, GPT-Realtime-2.1 and Gemini 3.8 Live. On τ-Voice, a benchmark of tool use and instruction following in simulated customer-service calls, it scores 56.6, which the company says puts it 17 to 30 points ahead of OpenAI's and Google's non-reasoning realtime models. Two reasoning models score higher but take 1.24 to 1.35 seconds to begin speaking. In a trial with Revmo, which runs a phone ordering agent, using ten hours of its production calls, Rime says Prism got callers to confirm before placing 94% of orders, against 57% for the OpenAI-based agent Revmo uses today. None of these figures has been independently checked.

Availability and context

Prism can be tried for free. Rime calls it the first speech-to-speech model anyone can fine-tune, but its self-serve reinforcement learning and supervised fine-tuning APIs are still being built with beta customers. Pricing and supported languages were not stated. The launch follows a $24 million Series A in July led by M13; Rime says it handles nearly 100 million calls a month for clients such as Mayo Clinic and Dialpad.