Language-independent hyperparameter optimization based speech emotion recognition system

Speech emotion recognition is challenging due to substantially overlapping regions of emotions. Extracting desired features that influence emotions in a speech and categorizing these emotions is a tedious task. We intend to develop an effective and robust speech emotion recognition system capable of classifying ambiguous and overlapping emotions through this manuscript. Three feature sets Spectral, Prosodic, and Discrete Wavelet Transform are extracted and further processed to reduce the required combination of features. The use of hyper-parameter optimization in the machine learning model has been done to tune the support vector machine classifier parameter for the Speech emotion recognition system. The suggested model is also verified with two different language datasets: ‘SAVEE’ and ‘EmoDB’ resulting in a language-independent emotion recognition system from speech. The performance result achieved by employing the proposed technique in EmoDB with 535 samples and SAVEE with 480 samples in seven different emotion types is 90.02% and 71.66%, respectively. © 2022, The Author(s), under exclusive licence to Bharati Vidyapeeth's Institute of Computer Applications and Management.