Speech Emotion Recognition using K-means Apriori Feature Selection Algorithm

Abstract

The novelty of this paper lies in the extraction of an effective feature vector in classifying speech emotions. Observation shows the spectral features extracted over the entire range of frequencies remain noise-sensitive with a distorted power spectrum. Thus, the focus is to extract the high frequency, low noisy spectral, and voice quality components for a possible improvement in classification accuracy. The extracted low noisy feature vectors are high-dimensional, containing redundant data. To alleviate the issue, this work further investigates the K-means apriori feature selection (KAFS) algorithm to derive a novel reduced feature vector for a better result. While the K-means algorithm has clustered the raw feature vectors, the apriori algorithm fetches only the relevant features with the desired outcome. The efficient Decision Tree (DT) and the Random Forest (RF) classifiers have been simulated to validate the derived feature vectors for their efficacy. The KAFS-based optimized feature sets are more reliable with an average accuracy of 64.89% with RF and 53.17% with DT. On the contrary, the corresponding accuracy, using the traditional baseline feature vector has been 64.21% with RF and 52.57% with DT.