In 2024, Kyutai released Moshi, a full-duplex voice-to-voice model with 200 ms practical latency. This note documents the diagrams we drew while studying Moshi and Mimi, with pseudocode and equations for readers who want the extra detail.
Moshi for Mere Mortals
Explore the articleTongKe Xue, Rohit Swamy, Charles Niu
Moshi for Mere Mortals
Keep reading.
All articlesWhat is full-duplex? How does half-duplex imitate it?
A guide to overlapping speech, interruptions, backchannels, and the design of full-duplex and half-duplex systems.
Grafting a Speech Head onto Gemma 4 E4B
An architecture walkthrough and early experiment mapping Gemma hidden states to Mimi audio tokens with a trainable speech head.
