Optimal Classifier Selection in Turkish Speech Emotion Detection

Emotion detection comprises various signal processing steps since emotion perception is subjective in speech. Signals used in emotion detection would be extracted using acoustic, visual and textual data. The quality of data acquisition and labelling has positive effects on emotion classification. Emotion extraction varies depending on factors such as spoken language, emotion polarity, age, and gender. Thus, it is crucial to use relevant features of spoken language in aural sentiment detection. In this study, acoustic feature extraction methods; MFCC, LFCC, PLP-RASTA, LPC, Mel-Spectrogram are used. Feature selection is performed with Principal Component Analysis method. The architecture of artificial neural network models is applied on emotion classification using Turkish speech datasets (TurES and TurEV-DB). The benchmark analysis is studied through different features and comparative results are obtained. Grid Search and Randomized Search are used to choose the best parameters for models.