Rime has introduced Prism, a speech-to-speech system whose replies pass through readable text before being spoken. Developers can apply content rules to this checkpoint, reading and changing the proposed words while the model continues listening. The company is also working with beta customers on customization interfaces; it has not announced general availability for those interfaces.
Rime, a developer of voice AI systems, introduced Prism on 28 October with a text checkpoint between its response decision and spoken reply. Developers can read or rewrite the intended words and apply content rules before speech begins. Rime describes a system trained to listen and compose a reply while continuing to hear the caller during its own speech.[1]
The company’s design aims to preserve information such as hesitation, laughter and background noise when deciding how to respond. A conventional chain turns speech into text, passes it through a language model and synthesizes a new voice; Rime says that process can discard audio information. Prism’s checkpoint also accommodates pronunciation instructions for names and account terms, as well as action and content tags.[1]
Tool calls can run alongside a spoken reply. Rime is also working with beta customers on interfaces for self-service training, using reinforcement learning or supervised fine-tuning, rather than announcing general availability for those interfaces. Its evaluation used simulated customer-service calls and was run by the company. The published tests therefore describe Rime’s own evaluation of Prism, with no independent measurement of its claimed advantage.[1]
