What Influence Does the Type of Extracted Audio Features Have on Emotional States?

This paper presents the way in which different audio emotional states are classified using various feature extraction types. Our audio soundbase for robot service application, in Romanian language, alongside with four established soundbases, Crema, Ravdess, Savee and Emo-DB were used in this paper. We analyzed the influence of different feature extraction methods upon classification accuracy. We focus on Mel Frequency Cepstral Coefficients (MFCC), Linear Predictive Coding (LPC), Magnitude-based Spectral Root Cepstral Coefficients (MSRCC), Normalized Gammachirp Cepstral Coefficients (NGCC) and Gammatone Frequency Cepstral Coefficients (GFCC). One classifier, k-nearest neighbor (k-NN), is utilized to evaluate the effectiveness of each feature set in recognizing emotional states. We compared the correct classification rates to determine which features have the best results in the emotional recognition field. We show that by using $30-\mathrm{NGCC}$ and 5 -nearest neighbors we reach an overall correct classification rate of $96.46 \%$ in the testing phase. Computational time was also incorporated into the analysis, including the testing and training phases. This study makes contributions to the field of emotion recognition from audio signals, facilitating advancements in human-computer interaction.