Rime Prism: a fast speech-to-speech voice AI that writes before it speaks

Voice assistants usually make you pick between speed and control, and Rime says its new model doesn't. On October 28, the voice AI company launched Prism, a speech-to-speech model aimed at businesses that run phone agents.

Here is what stood out to me in the announcement:

• Fast replies. Rime says Prism answers in about 300ms, and measured 0.57 seconds to first audio on a public voice agent benchmark.
• A text checkpoint. Prism listens to audio, writes out what it plans to say, then speaks. Companies can run their own rules over that text before a single word is spoken.
• Tools while talking. Tool calls happen in parallel with speech, so the agent can look up an order without dead air.
• Fine-tuning. Rime calls Prism the first speech-to-speech model anyone can fine-tune, and it is building self-serve training APIs with beta customers.

Rime reports a score of 56.6 on Artificial Analysis's τ-Voice benchmark for customer service calls, and says Prism beats the non-reasoning realtime models from OpenAI and Google there. Those are the company's own numbers, so I'd wait for independent results.

Why it matters

The old way to build a voice bot chains three steps: speech to text, a text model, then text to speech. You can inspect every word, but the tone of the caller gets lost. Newer all-in-one models hear tone but hide what they are about to say. Prism tries to keep both, and that inspectable text step is the part I find most practical. If you run a support line, being able to block a wrong promise before it is spoken matters more than shaving off a few milliseconds.

I also liked a small detail from a customer test: after training on a restaurant's calls, Prism asked "shall I place it?" before ordering far more often. Confirming before acting is exactly what I want from any AI that touches my money.

Read the official Rime announcement.