Spoken dialogue alignment from live user interactions

What changed

Constructed content- and timing-related preference pairs from multi-turn spoken interactions using transcription, AI feedback and speech resynthesis, then applied offline preference alignment to full-duplex speech models.

Scientific scope

Constructed content- and timing-related preference pairs from multi-turn spoken interactions using transcription, AI feedback and speech resynthesis, then applied offline preference alignment to full-duplex speech models. Inherits Moshi and offline preference learning. Resynthesized preference data is not an unchanged archive of user audio; benchmark and human-study improvements do not guarantee factuality or safety in all conversations.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Inherits Moshi and offline preference learning. Resynthesized preference data is not an unchanged archive of user audio; benchmark and human-study improvements do not guarantee factuality or safety in all conversations.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• Aligning Spoken Dialogue Models from User Interactions
• Kyutai scientific publications index

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction