Too few interruptions? Using data augmentation to improve offline automatic turn-taking annotation

Abstract

Offline turn-taking annotation consists in classifying all transitions in a spoken conversation into several categories, such as smooth switches, backchannels, and interruptions. Previous research has reported low accuracy in identifying interruptions, possibly due to their infrequent occurrence in spontaneous spoken dialogue, resulting in a scarcity of data to effectively train machine-learning models. In this study, we explore three strategies to increase the number of interruptions available in existing corpora: 1) create copies of actual interruptions and subtly alter their acoustic-prosodic characteristics; 2) generate artificial interruptions at hold transitions, which are known to be prosodically similar to the speech preceding interruptions; and 3) combine the first two strategies. We report promising improvements in classification performance when using these data augmentation techniques.