Порівняння ефективності класифікаторів машинного навчання у контексті голосової біометрії

• Date: 2019
• Main Authors: Danilov, Valery Ya., Grushko, Yaroslav G.
• Format: Article
• Language: Ukrainian
• Published: The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" 2019
• Subjects: голосова біометрія; MFCC; порівняння класифікаторів; к-найближчих сусідів; машинне навчання; штучний інтелект
• Online Access: http://journal.iasa.kpi.ua/article/view/188562
• Journal Title: System research and information technologies

The purpose of this work was to compare the seven popular classifiers of scikit-learn python-based library in the context of the performance of the voice biometrics system. The MFCCs (Mel-Frequency Cepstral Coefficients) method was used to compute the feature vectors of the person's voice undergoing verification. The classifiers involved in this study are the following: K-NN (K-Nearest neighbors classifier), MLP (Multilayer perceptron), SVM (Support vector machine), DTC (Decision tree classifier), GNB (Gaussian Naive Bayes classifier), ABC (AdaBoost classifier), RFC (Random forest classifier). As the data, we used voice samples from 40 individuals with an average duration of 9 minutes per person. The performance criteria of the classifiers were dictated by the needs of voice biometrics systems. Thus, in the framework of this work, the fraud simulation was conducted during authentication. The most effective in voice recognition was the K-NN classifier, which, with zero number of incorrectly admitted persons, provided 3-85% better accuracy of verification than other classifiers.