Speech Emotion Recognition using Fully Convolutional Network and Augmented RAVDESS Dataset

Speech Emotion Recognition (SER) refers to the process of determining an individual’s emotional state based on their spoken utterance. The success of SER will enhance the quality of human-machine interaction. It will enable the machine to understand and respond to the emotional content conveyed through speech. In this study, we adopted a framework which first augments the number of training samples in the dataset by a factor of 5 to increase the training data quantity and then extracts five informative features, namely Mel-Frequency Cepstral Coefficients (MFCC), Mel Spectrograms, Chroma features, Zero Crossing Rate (ZCR), and Root Mean Square (RMS) for an audio signal. A CNN- classifier with 3 dense layers (FCN) is used to accurately recognize emotions in speech. The Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) is used for conducting the experiments. Experimental results showed that augmented RAVDESS dataset, achieved an accuracy of 92%, which is approximately 30% higher than when tested on the non-augmented RAVDESS dataset. Compared with the baseline model, adopted frameworks achieved 3% higher SER accuracy.