A filtering approach for speech emotion recognition using wavelet approximation coefficient

The understanding of emotions from human speech signals is one of the most attractive areas of research. However, it is also challenging due to the influence of language and accent on speech. This framework presents a solution for speech emotion recognition based on wavelet approximation method. In this work, the Wavelet Approximation Speech Emotion Recognition (WaSER) model is presemted. The ‘db4′ wavelet is explored to calculate the approximation coefficient. At the initial stage, silent segments of the raw speech signal are eliminated using wavelet approximation and energy thresholding, resulting in a filtered signal. For the wavelet selection and threshold value optimization, various experiments are performed. A set of prosodic and spectral features is then extracted from the filtered signal. The potential features are selected using the ReliefF algorithm, and these optimal features are utilized for further processing. An ablation experiment is also presented to demonstrate the potential of the proposed WaSER model. The WaSER model was tested on various language-based datasets, including RAVDESS, EMOVO, Emo-DB and ShEMO, attaining a highest accuracy of 94.6% on the ShEMO dataset. The ablation experiment shows that filtering the speech significantly enhanced accuracy across all datasets and maintaining a balance between accuracy and sensitivity. Also, the proposed model drastically reduces the response time compared to the existing state of art.