Towards Better Dynamics in Full-Duplex Spoken Language Models
Invited Talk, UT Austin, Texas, US
Slides: Towards Better Dynamics in Full-Duplex Spoken Language Models
Related Survey: Speech-Trident
This talk starts from our survey paper: “On the Landscape of Spoken Language Models” (TMLR 2025) to introduce spoken language models (SLMs) and the development of full-duplex models that can listen and speak simultaneously. It then covers two contributions: (1) the Game-Time Benchmark (ICASSP 2026), which evaluates temporal dynamics in SLMs such as reaction timing, tempo adherence, and silence management, and (2) Full-Duplex-Bench-v2, a multi-turn evaluation framework with an automated examiner for assessing full-duplex SLMs.