Text vs. Speech Analysis – Detecting Sentiment of Customer Calls
(A2023-114205)

AUTHORS

Manuel Weber, WHU - Otto Beisheim School of Management; Christian Schlereth, WHU - Otto Beisheim School of Management

ABSTRACT

In recent years, sentiment analysis has been adopted in customer service to better address customer needs during call center calls. We aim to learn about the value of text-based vs. speech-based sentiment analysis to understand customer satisfaction with a call center call. We apply a pre-trained transformer model to classify customer sentiment of mock call transcripts. Then, we train a convolutional neural network on a public speech dataset and use it to classify the sentiment of call recordings. We find that the “simple” text-based model is more accurate in predicting customer sentiment than the speech-based model (~84% vs. ~53%). While the further predicts sentiment best using the entire call transcripts, the latter predicts sentiment best using just the beginning of call recordings. For some calls, the speech-based model detects sentiment more accurately, which indicates that both approaches could complement each other.