Behavior-SD
dataset Your tags
Your notes
NAACL 2025 Senior Area Chair Award (Speech/SLU): 108K full-duplex spoken dialogues (2,164 hours) with per-utterance annotation of conversational behaviors — backchannels, interruptions, filler words — plus BeDLM, a behavior-conditioned full-duplex dialogue LM trained on it. From Gunhee Kim's Vision & Learning Lab; a rare large-scale resource for naturalistic speech-to-speech modeling.
Paper
Venue NAACL 2025 (SAC Award)
Dataset
Size 108K dialogues (2,164 h)