Transfer learning based feature selection for feedforward neural network for speech emotion classifier

This work discusses speech emotion recognition via custom feature engineering and feature selection techniques using mel-frequency cepstral coefficients as initial audio features. Proposed transfer learning approach consist in employing the backward-step selection algorithm for feature selection using statistical learning classifiers, the obtained subset of features than subsequently used to train feedforward neural networks. This technique allowed us to significantly reduce initial feature vector size while increasing models’ prediction quality. We used TESS and RAVDESS datasets to estimate the performance of proposed method. To evaluate the quality of the model, unweighted average recall (UAR) was used. Experimental results demonstrate promising accuracy (UAR = 82 % for TESS and UAR = 53 % for RAVDESS), showcasing the potential of this approach for applications like virtual agents, voice assistants and mental health diagnostics.