Multimodal speech emotion recognition optimization using genetic algorithm

Abstract

Speech emotion recognition (SER) is a technology that can detect emotions in speech. Various methods have been used in developing SER, such as convolutional neural networks (CNNs), long short-term memory (LSTM), and multilayer perceptron. However, sometimes in addition to model selection, other techniques are still needed to improve SER performance, namely optimization methods. This paper compares manual hyperparameter tuning using grid search (GS) and hyperparameter tuning using genetic algorithm (GA) on the LSTM model to prove the performance increase in the multimodal SER model after optimization. The accuracy, precision, recall, and F1 score improvement obtained by hyperparameter tuning using GA (HTGA) is 2.83%, 0.02, 0.05, and 0.04, respectively. Thus, HTGA obtains better results than the baseline hyperparameter tuning method using a GS.