Speech Emotion Classification Using Deep Learning

Over the last several years, the fields of affective computing and human-computer interaction have shown a great deal of interest in the topic of speech emotion recognition studies. In the study of speech emotion classification, numerous strategies, many of which are well-established methodologies, have been utilized to extract emotions from speech signals. These strategies include speech analysis and classification methods. Deep learning techniques have recently been proposed as a potential alternative to more traditional methods for the classification of the emotions conveyed in speech. In this study, we used MFCC features for emotion classification and compared the performance of DNN, DCNN, and CNN-LSTM models on RAVDESS, EMODB, and TESS datasets. In the proposed models, DNN achieved an average accuracy of 76.38% on RAVDESS, 82.83% on EMODB, and 100% on TESS. The overall accuracy of DCNN was 74.65% when it was tested on RAVDESS, 77% when it was tested on EMODB, and 100% when it was tested on TESS. On RAVDESS, CNN-LSTM achieved an average accuracy of 70.27%, while on EMODB it achieved 71.8%, and on TESS it achieved 99%.