Generative AI · advanced · concept 177 of 311 · Updated Nov 4, 2026

🎙️Speech-to-Speech and Realtime Voice Models

A speech-to-speech model takes audio in and produces audio out directly, instead of chaining speech recognition, a text model and text-to-speech. Removing the text step lowers latency and keeps tone, pauses and interruptions that a transcript drops. Kyutai's open Moshi model reports about 200 ms of latency in practice and can listen while it speaks. OpenAI's Realtime API and Google's Gemini Live API expose the same kind of model.