Kyutai's Speech-To-Text and Text-To-Speech models based on the Delayed Streams Modeling framework.

Moshi is a speech-text foundation model and full-duplex spoken dialogue framework. It uses Mimi, a state-of-the-art streaming neural audio codec.