Comparison of the effects of mel coefficients and spectrogram images via deep learning in emotion classification

Abstract

In the present paper, an approach was developed for emotion recognition from speech data using deep learning algorithms, a problem that has gained importance in recent years. Feature extraction manually and feature selection steps were more important in traditional methods for speech emotion recognition. In spite of this, deep learning algorithms were applied to data without any data reduction. The study implemented the triple emotion groups of EmoDB emotion data: Boredom, Neutral, and Sadness-BNS; and Anger, Happiness, and Fear-AHF. Firstly, the spectrogram images resulting from the signal data after pre-processing were classified using AlexNET. Secondly, the results formed from the Mel-Frequency Cepstrum Coefficients (MFCC) extracted by feature extraction methods to Deep Neural Networks (DNN) were compared. The importance and necessity of using manual feature extraction in deep learning was investigated, which remains a very important part of emotion recognition. The experimental results show that emotion recognition through the implementation of the AlexNet architecture to the spectrogram images was more discriminative than that through the implementation of DNN to manually extracted features.