The study deals with the emotion recognition problem in Turkish single word pronunciations through the image recognition perspective. The CNN approach allowed us to use spectrograms as the training material. The resulting model has a feasible predictive power and shares trends with the human judges: it is robust to the changes in the sound signal at higher frequencies and generally performed better than the judges except for the 500-8000 hertz band where the human judges did not lose any significant predictive power. The model proved to be a feasible explanation of human assessment of emotions by failing at 500-8000 hertz band where the humans remained robust.

Date: 27.06.2019 13:30   Place: A-108