Prism is a speech-to-speech model that processes audio input and produces spoken responses with roughly 300 millisecond latency. It creates an intermediate text layer for applying guardrails and supports fine-tuning through reinforcement learning on custom data.