Social RoboticsNagoya University releases J-Moshi, a Japanese AI that talks and listens at once
The full-duplex dialogue system can produce the short listener responses typical of Japanese conversation.
Nagoya University's Higashinaka Laboratory adapted Kyutai's Moshi model to Japanese, training it on tens of thousands of hours of speech, and made it publicly available. Unlike turn-taking voice assistants, J-Moshi can speak and listen simultaneously and produce aizuchi, the brief acknowledgements Japanese speakers use to show they are listening. The work was accepted to Interspeech 2025. Natural overlapping speech is one of the missing pieces for conversational robots, particularly in languages and cultures where back-channelling is expected.
Source: Tech Xplore (Nagoya University)
This archive entry summarises reporting published elsewhere; follow the source link for the full story. The summary was drafted with AI from that reporting.