Analysis Kernel and Feature: Impact on Classification Performance on Speech Emotion Using Machine Learning

Abstract

The main objective of this study is to test the machine learning kernel's selection against the characteristics of the data set used, resulting in good classification performance. The goal of speech emotion recognition is to improve computers' ability to detect and process human emotions in order to improve their ability to respond to interactions between people and computers. It can be applied to feedback on talks, including sentimental or emotional content, as well as the detection of human mental health. One field of data mining work is Speech Emotion Recognition. One of the important things in data mining research is to determine the selection of the kernel Classifier, know the characteristics of datasets, perform Engineering Features and combine features and Corpus Datasets to obtain high accuracy. The research uses analysis and comparison methods using private and public datasets to detect speech emotions. Experimental analysis was done on the characteristics of datasets, selection of kernel classifiers, pre-processing, feature and corpus datasets fusion. Understanding the selection of a classifier kernel that matches the characteristics of the dataset, engineering features and the merger of features and datasets are the contributions of this investigation to improving the accuracy of the classification of speech emotion data. For models with the selection of kernels that match the characteristics of their datasets, this study gave an increase in accuracy of 12.30% for the private dataset and 14.80% for the public dataset, with accuracies of 100.00% and 74.80% respectively. Combining features and public datasets provides an increase in accuracy of 33.62% with an accuracy of 73.95%.