Speech Emotion Recognition Using Deep Learning Techniques and Augmented Features

This paper presents a study on speech emotion recognition using deep learning techniques. The focus is on utilizing the Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) and the Toronto Emotional Speech Set (TESS) datasets. The proposed approach involves preprocessing the audio signals by converting them into a sequence of numbers and normalizing them using max normalization. Furthermore, the signals are augmented using three methods: noise with a 0.015 noise rate, pitch adjustment, and a combination of noise with a 0.03 noise rate and pitch adjustment. The extraction of audio features includes a combination of two features, namely Mel Frequency Cepstrum Coefficients (MFCC) and Root Mean Square (RMS) energy, which provide distinct information from each other. These features are then used to train a CNN for emotion classification. Experimental results demonstrate the effectiveness of the proposed approach in accurately recognizing emotions from speech signals. The proposed approach achieves an accuracy of 93.6% on the RAVDESS dataset and 99.9% on the TESS dataset. These results highlight the potential of the proposed approach in practical applications such as speech therapy and human-computer interaction.