Application of Data Mining Techniques in the Analysis of Acoustic Sound Characteristics

Abstract

In recent years, the analysis of acoustic characteristics of speech and sound has been one of the areas that data mining has found its way through. The present research study is also related to this topic which aims to detect the gender of the speaker by using the acoustic feature of his voice. In this research, the data set includes 3,168 recorded voice samples gathered from female and male speakers. Through acoustic analysis, 20 characteristics along with the desired labels were extracted and prepared for the data mining process. Finally, using Python programming language tools, 6 different techniques were used to construct an appropriate problem-solving model. These techniques were: support vector machines, logistic regression, random forest, regression and classification trees, adaptive boosting, and K-nearest neighbor. The accuracy of the models was also compared with each other. The obtained results revealed that the accuracy of all these techniques was sufficiently high (above 90%) for solving the problem and the model made by them had the necessary efficiency for classification. Moreover, the obtained model was specifically evaluated through decision tree and some principles and rules existing in the model were extracted. As a result, it was revealed that average fundamental frequency measured across the audio signal is the key characteristic of the sound for the evaluation of the voice gender to the extent that it plays a key role in data classification.