Reproduce “Neither Silence nor Overlap Is Failure: Intent-Conditioned Evaluation of Turn-Taking in Full-Duplex Spoken Dialogue Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kian Shamsaie, Iman Modarressi
TACT evaluates conversational timing by conditioning human reference distributions on inferred intent and speaker-specific profiles, rather than judging responses against fixed windows. It scores timing, length, and overlap across six intents. In the reported study, gpt-realtime-2 scores 0.47 versus 0.86 for the human topline; validation correlation is 0.81, but profile swaps barely change systems while humans shift.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.