Augmented audio data in improving speech emotion classification tasks

To achieve high performance and classification accuracy, classification of emotions from audio or speech signals requires large quantities of data. Big datasets, however, are not always readily accessible. A good solution to this issue is to increase the data and augment it to construct a larger dataset for the classifier’s training. This paper proposes a unimodal approach that focuses on two main concepts: (1) augmenting speech signals to generate additional data samples; and (2) constructing classification models to identify emotion expressed through speech. In addition, three classifiers (Convolutional Neural Network (CNN), Naïve Bayes (NB) and K-Nearest Neighbor (kNN)) were further tested in order to decide which of the classifiers had the best results. We used augmented audio data from a dataset (SAVEE) in the proposed method to conduct training (50%), and testing (50%) was executed using the original data. The best performance of approximately 83% was found to be a mixture of augmentation strategies using the CNN classifier. Our proposed augmentation approach together with appropriate classification model enhances the efficiency of voice emotion recognition.