Moshi is a speech model from the AI lab Kyutai that listens and talks at the same time, generating speech directly for real-time spoken conversations instead of chaining separate steps.
In plain words
Most voice assistants chain separate steps: detect speech, transcribe it, write a text reply, then read it aloud. That adds seconds of delay and loses tone and emotion. Moshi treats conversation as speech in, speech out (it predicts matching text tokens along the way), and models its own voice and the user's as two parallel streams, so it can handle interruptions. It uses a streaming audio codec called Mimi and reaches about 200 milliseconds of latency in practice.
A quick example
Imagine chatting out loud and cutting in with "wait, say that again" halfway through a reply. Because Moshi listens while it speaks, it can react to the interruption straight away rather than waiting for its turn to end.
Where you'll meet it
- Real-time voice assistants and spoken dialogue research.
- Base for related Kyutai speech models, such as live translation.
Related
Sources
- PaperMoshi: a speech-text foundation model for real-time dialogue, Défossez et al. (arXiv) · read 29 Sept 2026
- Docsmoshi, Kyutai (GitHub) · read 29 Sept 2026
Checked by WhatIsAI against the sources above · 29 Sept 2026. Spotted an error? Report it.