Automatic detection of turn-taking events in continuous EEG data from spontaneous dialogue

BRUSCO P; KAMIENKOWSKI JE; GRAVANO A

San Francisco

Workshop; 1st Workshop on Speech Engineering and Computational Neuroscience of Speech; 2016

This paper describes a series of machine learning experiments for automatically classifying the type of turn-taking transition based on features extracted from the EEG signal. These were conducted on a corpus of unrestricted dialogues between pairs of subjects, with simultaneous recordings of speech and EEG from both subjects. Our preliminary results indicate that the listener?s EEG signal contains useful information for predicting whether the current speaker will either yield the conversational turn or continue talking. In other words, we show it is possible to detect the listener?s perception of turn-yielding cues present in the speaker?s speech. Similarly, from the current speaker?s EEG signal we can extract useful information for detecting whether the speaker will yield the turn or continue talking. These results may lead to new tools valuable for thedevelopment of brain-computer interfaces.