Abstract
Speech emotion recognition (SER), which has gained greater attention in recent years, is a key aspect of the human-computer interaction process. However, a wide range of strategies has been offered in SER, and these approaches have yet to increase performance. In this study, a deep neural network model for classifying voice emotions is suggested. It is divided into three stages: feature extraction, normalization, and emotion recognition. The Librosa Python Toolkit is used to acquire the MFCC, Mel-Spectrogram Frequency, Chroma, and Poly Features during feature extraction. Data augmentation for the minority class using SMOTE (synthetic minority oversampling technique) and the Min-Max scaler for the normalization process were used. The model was evaluated on three frequently used languages: German, English, and French, using the Berlin Emotional Speech Database (EMODB), Surrey Audio-Visual Expressed Emotion Dataset (SAVEE), and the Canadian French Emotional (CaFE) speech datasets. The recognition rates of unweighted accuracy of 95% on EMODB, 90% on SAVEE, and 92% on CaFE are gained in speaker-dependent experiments. The results show that the suggested method is capable of efficiently recognizing emotions and outperformed the other approaches utilized for comparison in terms of performance indicators.