Gender interruption spotting: A proof-of-concept using audio machine learning to detect gender dynamics of simultaneous speech events in recorded conversation

There is a long history of research interest in gender and conversation interruptions, especially about what it means for equality and power. Despite the extensive literature, there is wide disagreement on gender patterns due to methodological differences and the constraints of qualitative research designs. Only a recent few studies have begun to use computational methods, which have the potential to improve validity and generalizability to studies about interruptions. This paper is a proof-of-concept that documents the first attempt to investigate how audio machine learning using a neural network with a transformer architecture on large audio datasets can add to research on gender-based interruption patterns, especially compared to existing text-based methods. Audio classifiers trained on a novel dataset of 3,076 examples of interruptions and non-interruptions show promising results, with classification metrics from 0.66 to 0.87. This suggests that it is possible to detect interruptions including the gender dynamic of the speakers based on the audio signature of the audio events. A rule-based text classifier tested on automatically generated transcripts performed much worse, underscoring the unique potential of audio methods. Future developments on this method (larger training dataset, modified operationalization, and larger models) could create a reliable classifier for detecting interruptions with their gender dynamic, enabling a greatly expanded range of research opportunities into this phenomenon.