ارشيا اموري سرابي

Feature selection plays an important role in creating a successful speech emotion recognition system. In this research, a feature selection approach that modifies the initial population generation stage of meta-heuristic search algorithms with the help of ReliefF algorithm is proposed. This approach is eva‎luated on the meta-heuristic flame-butterfly (MFO) search algorithm, in the context of speech emotion recognition using the Interactive Emotional Binary Motion Capture (IEMOCAP) database. The results show that the proposed feature selection algorithm significantly reduces the number of features and is still effective for the classification of emotions from speech. Specifically, for the IEMOCAP database, an accuracy of 64.07% is obtained using the selected features (142), which is only less than 10% of the total features (1582) using the SVM classifier. The results obtained compared to the calculation of all the features have a 10% improvement in accuracy, and compared to similar methods such as the cuckoo search algorithm and genetics, each of which selected 787 and 815 features, respectively, and reached 62.44% and 66.18% accuracy. They found that our proposed method achieved similar and even better accuracy than cuckoo search with a much smaller number of features.

انتخاب ويژگي , تشخيص احساسات , الگوريتم هاي فراابتكاري , بهينه ساز شعله-پروانه

Feature selection , Emotion recognition , Meta-heuristic Algorithms , moth-flame optimizer

Arshiya Omoori

Dr. Vahidi