Fact · Voice and AI
The Moshi paper says that when text is the intermediate step in dialogue, non-linguistic information that changes meaning is lost in the interaction.
Checked against its source on October 17, 2026
In the source's own words
“text being the intermediate modality for dialogue, non-linguistic information that modifies meaning ... is lost in the interaction.”
Moshi: a speech-text foundation model for real-time dialogue, Alexandre Défossez and and colleagues at Kyutai, arXiv (Kyutai), 2024, Abstract, second problem listed.
Open the original- What it measures
- The paper names emotion and non-speech sounds as examples of the non-linguistic information it has in mind.
- What it does not show
- This is the authors' argument for a speech-to-speech design. It is a claim about information loss in a design, not a measured error rate for a product.