On the Impact of Voice Encoding and Transmission on the Predictions of Speaker Warmth and Attractiveness

Abstract

Modern human-computer interaction systems may not only be based on interpreting natural language but also on detecting speaker interpersonal characteristics in order to determine dialog strategies. This may be of high interest in different fields such as telephone marketing or automatic voice-based interactive services. However, when such systems encounter signals transmitted over a communication network instead of clean speech, e.g., in call centers, the speaker characterization accuracy might be impaired by the degradations caused in the speech signal by the encoding and communication processes. This article addresses a binary classification of high versus low warm--attractive speakers over different channel and encoding conditions. The ground truth is derived from ratings given to clean speech extracted from an extensive subjective test. Our results show that, under the considered conditions, the AMR-WB+ codec permits good levels of classification accuracy, comparable to the classification with clean, non-degraded speech. This is especially notable for the case of a Random Forest-based classifier, which presents the best performance among the set of evaluated algorithms. The impact of different packet loss rates has been examined, whereas jitter effects have been found to be negligible.