Turn-Taking Estimation Model Based on Joint Embedding of Lexical and Prosodic Contents

A natural conversation involves rapid exchanges of turns while talking. Taking turns at appropriate timing or intervals is a requisite feature for a dialog system as a conversation partner. This paper proposes a model that estimates the timing of turn-taking during verbal interactions. Unlike previous studies, our proposed model does not rely on a silence region between sentences since a dialog system must respond without large gaps or overlaps. We propose a Recurrent Neural Network (RNN) based model that takes the joint embedding of lexical and prosodic contents as its input to classify utterances into turn-taking related classes and estimates the turn-taking timing. To this end, we trained a neural network to embed the lexical contents, the fundamental frequencies, and the speech power into a joint embedding space. To learn meaningful embedding spaces, the prosodic features from each single utterance are pretrained using RNN and combined with utterance lexical embedding as the input of our proposed model. We tested this model on a spontaneous conversation dataset and confirmed that it outperformed the use of word embedding-based features.