Spoken conversational AI is expected to become the next major interface for human–computer interaction, enabling natural, real-time communication beyond conventional text-based systems. However, current speech-based conversational agents still face fundamental challenges, including realistic turn-taking, limited availability of high-quality spoken dialogue data, and increased latency caused by complex reasoning. The presentation reviewed recent advances addressing these challenges and discussed future directions toward personalized spoken conversational AI. First, it examined self-supervised approaches for modeling conversational turn-taking, highlighting how voice activity prediction enables the learning of interaction dynamics without costly manual annotations. Next, it introduced generative spoken dialogue language models that directly synthesize natural spoken conversations while preserving conversational rhythms such as pauses, overlaps, and backchannels. The presentation also discussed recent efforts to reduce response latency through interleaving listening and reasoning, allowing speech language models to begin reasoning before user utterances were fully completed while maintaining response quality. Building upon these developments, the speaker presented ongoing research on personalized spoken interaction, which aims to model speaker- and context-specific conversational behaviors rather than assuming uniform interaction patterns. The proposed framework focuses on controllable turn-taking decisions by balancing the confidence of the current understanding with the expected benefit of waiting for additional user input. Ultimately, the presentation highlighted how advances in turn-taking, efficient reasoning, and personalization can contribute to the development of more natural, efficient, and human-centered spoken conversational AI systems.

· Moderator: 강필성 교수 pilsung_kang@snu.ac.kr