Speech Emotion Recognition Using Deep Learning

Speech emotion recognition can be used in many applications, mainly in the field of mental health and human–robot interaction. SER can be used to monitor anxiety, depression, and post-traumatic stress disorder, among other mental health disorders. In this work, we have developed deep learning models such as CNN, DANN, and TCN to recognize emotional states from speech signals. Each model is trained with different datasets with different feature extraction techniques such as MFCC, etc., to recognize various emotions. The emotional states of a person can be classified based on factors like pitch, tone, intensity, and dimensions of emotion such as arousal and valence. We have used four different datasets for training and evaluating the model. This work used CNN, GRU, DANN, and TCN with various feature extraction techniques, among that TCN performs better in large datasets (MFCC 58 features) with 93.66% accuracy and with eight emotion classes (Angry, Calm, Disgust, Fear, Happy, Neutral, Sad, Surprise).