Classification of Emotion with Audio Analysis

Abstract

Classification is an important technique used to predict which group the samples in the data belong to. In this study it is aimed to classify according to emotions by extracting audio features. The study was performed using audio data from 2 male and 2 female individuals that spoke in four different emotions amused, angry, neutral, and sleepy. “MFCC” as a Cepstral feature, “Centroid, Flatness, Skewness, Crest, Flux, Slope, Decrease, Kurtosis, Spread, Entropy, roll-off point” as Spectral Feature, “Pitch, Harmonic ratio” as Periodicity Features were used in voices. All classification algorithms that were located in the classification learner toolbox in Matlab were applied to the data, and the algorithm providing the highest accuracy was asked to estimate the emotion. 26 inputs and one output were calculated, and the performance results had compared to each other. According to the results, it has been seen that the support vector machine algorithm provides the highest accuracy performance. Considering the performances obtained, this study reveals that it is possible to distinguish and classify sounds using sentimental data and sound feature parameters.