Preprint
Moshi: a speech-text foundation model for real-time dialogue
Alexandre Défossez and and colleagues at Kyutai, arXiv (Kyutai), 2024
Read the originalRead on October 17, 2026
- What it is
- A research paper from the Kyutai lab describing Moshi, a speech model that listens and speaks at the same time, built on a text language model. The authors report its latency for speech in and speech out.
- How much weight it carries
- Peer review status not shown on the abstract page. The latency is the authors own measurement in a research setup, so it describes the model, not a deployed phone agent.
Facts taken from this source
- Moshi, a full-duplex speech-to-speech model from Kyutai, has a theoretical latency of 160 ms and about 200 ms in practice.
“with a theoretical latency of 160ms, 200ms in practice” (Abstract)
- The Moshi paper says that the complexity of current spoken dialogue pipelines induces a latency of several seconds between interactions.
“First, their complexity induces a latency of several seconds between interactions.” (Abstract, first problem listed)
- The Moshi paper says current spoken dialogue systems rely on pipelines of independent components, and it casts spoken dialogue as speech-to-speech generation instead.
“Current systems for spoken dialogue rely on pipelines of independent components, namely voice activity detection, speech recognition, textual dialogue and text-to-speech. ... Moshi solves these independent issues altogether by casting spoken dialogue as speech-to-speech generation.” (Abstract)
- The Moshi paper says that when text is the intermediate step in dialogue, non-linguistic information that changes meaning is lost in the interaction.
“text being the intermediate modality for dialogue, non-linguistic information that modifies meaning ... is lost in the interaction.” (Abstract, second problem listed)