Prism targets roughly 500 ms enterprise voice responses

Rime launched Prism on October 28 as an enterprise speech-to-speech model. The company says Prism responds in roughly 500 milliseconds and emits an inspectable text layer before synthesizing speech, so policy and safety checks can review every word. Tool calls can run in parallel with spoken output, reducing the pause created by waiting for an action to finish before the agent speaks. Prism can also be fine-tuned for industry vocabulary, pronunciation and emotional cues, and deployed inside customer-controlled infrastructure for data and runtime control. Its central proposition is not merely human-like audio, but real-time conversation that remains composable and reviewable. The 500 ms figure, performance claims and human-perception results come from Rime; broad independent benchmarks are not yet available. Buyers still need to test multilingual accuracy, noise, interruptions, hallucinations, end-to-end cost and reliability in their own call flows.