A metaheuristic-based hybrid approach for feature selection in emotional classification from speech

Throughout human history, people have been communicating with each other through speech. People communicate by speaking and can also express their feelings to the person they are talking to. Speech consists of signals that the air we breathe passes through certain organs in the human body and emits through the vocal cords. We can guess the emotional state of the person we are talking to by the tone we receive from these signals. Today, when human-computer interaction is used intensively, studies on emotion recognition from voice are also frequently encountered. In the development of today's technology, we often see examples inspired by living things. Especially in the tools and equipment used, systems are developed by imitating living things found in nature both in terms of design and operating principle. In the field of artificial intelligence, the study of imitating living things appears as metaheuristic algorithms. These algorithms are algorithms developed by being inspired by the food search and finding behaviors of living things. In our study, a hybrid metaheuristic algorithm (HGWSSO) design was carried out inspired by the GWO and SS algorithms used in the literature. The HGWSSO algorithm was used for emotional classification of speech in our study. For this purpose, EmoDB, SAVEE, eNTERFACE and EMOVA data sets, which are frequently used in the literature, were used. First, the audio files in this data set went through the data preprocessing stage and then the feature extraction process was performed with OpenSmile software. With the obtained features, feature selection process was performed with both GWO and SS algorithms based on our algorithm and HGWSSO algorithm. Then, with the obtained features, the accuracy of HGWSSO algorithm was tested using SVM and KNN classifiers. In the obtained results, it was observed that SVM classifier gave better results compared to KNN classifier. The highest accuracy rate with KNN classifier was obtained with HGWSSO algorithm with 586 features as 91.92% in EmoDB data set. In the results obtained with the SVM classifier, the HGWSSO algorithm achieved better results than other datasets with 137 features and 98.14% accuracy rate in the EmoDB dataset. As a result, it was seen that the HGWSSO algorithm we developed gave better results especially with the SVM classifier compared to the GWO and SS algorithms we took as basis.