Real-time comprehensive sociometrics for two-person dialogs

Abstract

A real-time system is proposed to quantitatively assess speaking mannerisms and social behavior from audio recordings of two-person dialogs. Speaking mannerisms are quantitatively assessed by low-level speech metrics such as volume, rate, and pitch of speech. The social behavior is quantified by sociometrics including level of interest, agreement, and dominance. Such quantitative measures can be used to provide real-time feedback to the speakers, for instance, to alarm to speaker when the voice is too strong (speaking mannerism), or when the conversation is not proceeding well due to disagreements or numerous interruptions (social behavior). In the proposed approach, machine learning algorithms are designed to compute the sociometrics (level of interest, agreement, and dominance) in real-time from combinations of low-level speech metrics. To this end, a corpus of 150 brief two-person dialogs in English was collected. Several experts assessed the sociometrics for each of those dialogs. Next, the resulting annotated dialogs are used to train the machine learning algorithms in a supervised manner. Through this training procedure, the algorithms learn how the sociometrics depend on the low-level speech metrics, and consequently, are able to compute the sociometrics from speech recordings in an automated fashion, without further help of experts. Numerical tests through leave-one-out cross-validation indicate that the accuracy of the algorithms for inferring the sociometrics is in the range of 80-90%. In future, those reliable predictions can be the key to real-time sociofeedback, where speakers will be provided feedback in real-time about their behavior in an ongoing discussion. Such technology may be helpful in many contexts, for instance in group meetings, counseling, or executive training.