Effect of Reverberation in Speech-based Emotion Recognition

In room environment, echo, reverberation, interference and additive noise cast the major challenges for emotional speech recognition due to degradation in quality and reliability of recorded speech signals. In this paper, we investigate effects of reverberation and noise on speech-based emotion recognition by comparing clean speech signal, adding simulated reverberant data, de-reverberant data and signal with added noise. First, we develop an emotional speech corpus of these four kinds of emotional speech data sources. Then we apply GMM-UBM framework to evaluate the performance of emotion recognition based on them. Results show that reverberation reduces emotion recognition accuracy by 5.87%, and a process of de-reverberation can largely cover this reduction.