Teamwork Quality Prediction Using Speech-Based Features

MARTIN MEZA; LARA GAUDER; LAUTARO ESTIENNE; GERMAN BARCHI; AGUSTIN GRAVANO; PABLO RIERA; LUCIANA FERRER

Dublin

Workshop; Speech, Music and Mind Workshop; 2023

ISCA

This paper describes a novel protocol for annotating teamwork quality and related variables, based only on the speech signal. Our protocol was designed to annotate a Spanish version of the Objects Games corpus, a publicly available corpus that contains dialogues of people playing a collaborative computer game. The corpus was annotated by 4 raters, who achieved an Intraclass Correlation Coefficient of 0.64 for the main teamwork quality metric. Using the resulting annotations, we developed a system for automatic prediction of the average teamwork quality across raters using features extracted from the conversations, reaching a coefficient of determination, R2 of 0.56. This result suggests that automatic prediction of teamwork quality from the speech signal of the teammates is a feasible task.