Toward Predicting Communication Effectiveness

摘要

Voice plays an important role in effective communication. This paper investigates communication effectiveness using call center audio records, where each record contains a conversation between a sales representative and a customer. The outcome is considered positive if the conversation ended with an appointment being made and negative otherwise. Two methods are proposed to predict the outcomes. The first method uses traditional machine learning algorithms on a small set of commonly-used, well-known acoustic features of voice in psychology research (i.e. prosodic features). A data processing pipeline has been developed that consists of three major steps: 1) an audio record is split into segments by speaker turns; 2) acoustic features are extracted from each segment; and 3) classification models are trained on the acoustic features to predict conversation outcomes. The second method uses deep learning on a larger set of vocal features. Each audio segment is split into fixed-interval pieces, from which acoustic feature vectors are computed and are stacked in temporal order to form a feature matrix as input to deep convolution neural networks. Experimental results based on real call center data shows that acoustic features, such as Mel-frequency cepstral coefficients (MFCCs), timbre and Chroma features, are good indicators of call success, much better than prosodic features. Temporal information in an audio record can be captured by deep convolutional neural networks for improved prediction accuracy, up to 82.6%.