8–12 Sept 2026
Europe/Vienna timezone

Classifying overlappee and overlapper based on acoustic prosodic features

FA2026/84
9 Sept 2026, 15:00
3h
Messehalle (Poster+Exhibition) (Messe Congress Graz)

Messehalle (Poster+Exhibition)

Messe Congress Graz

A20 Speech A20.00 Speech

Speaker

Barbara Schuppler (Signal Processing and Speech Communication Laboratory)

Description

Conversational overlaps, i.e., instances where two or more participants speak simultaneously, are a prevalent feature of human communication, influencing conversational dynamics and speaker roles. While previous studies have mainly analysed and categorised them based on competitiveness, this paper focuses on the roles of individual speakers within these overlaps. Our study is based on instances of overlapping speech from a corpus of casual, spontaneous dyadic conversations, from which we extract acoustic features and communicative function labels that were created manually. Using XGBoost for classification, we find that overlaps as coming from an overlappee or overlapper, can be discriminated with a balanced accuracy score of approx. 75%. Our analysis of feature importances shows that the communicative function along with the intensity levels are the key indicators in identifying speaker roles during overlaps. This study contributes to our understanding of conversational dynamics and provides insights applicable to making turn-taking in human-computer interaction more natural.

Authors

Barbara Schuppler (Signal Processing and Speech Communication Laboratory) Maximilian Nothnagel (Graz University of Technology)

Presentation materials