Speaker identification using fuzzy i-vector tree

Article type: Research Article

Authors: Gałka, Jakub^{; *} | Jaciów, Paweł

Affiliations: AGH University of Science and Technology, Department of Electronics, Kraków, Poland

Correspondence: [*] Corresponding author. Jakub Gałka, AGH University of Science and Technology, Department of Electronics, Kraków, Poland. E-mail: [email protected].

Abstract: In this paper, we propose a new method for concurrent accuracy and computational efficiency optimization using a fuzzy clusters tree for i-vector speaker identification. The design assumptions and an algorithm for a new type of fuzzy i-vector tree construction were introduced. The obtained solution was evaluated using the NIST 2014 i-Vector Speaker Recognition Machine Learning Challenge dataset. A 15% relative equal error rate reduction for a 74% reduction in computation time was achieved when compared to the baseline with only a 5.5% relative identification rate loss for discussed tree configurations.

Keywords: speaker recognition, fast identification, i-vector, fuzzy clustering, decision trees

DOI: 10.3233/JIFS-181359

Journal: Journal of Intelligent & Fuzzy Systems, vol. 37, no. 4, pp. 4937-4949, 2019