SVM BASED TEXT-DEPENDENT SPEAKER IDENTIFICATION FOR LARGE SET OF VOICES (TuePmPO2)

Author(s) : Piotr Staroniewicz (Institute of Telecommunication and Acoustics, Wroclaw University of Technology, Poland) Wojciech Majewski (Institute of Telecommunication and Acoustics, Wroclaw University of Technology, Poland)

Abstract : The paper presents the test results of speaker identification system based on the Support Vector Machines. The usefulness of SVM classifier for large voice telephone quality database (1300 speakers) was examined. The tested database was recorded according to SpechDat(E) conditions. The SVM classifier has shown its ability of feature generalization for large sets of classes. The obtained high scores (around 90%) of speaker identification have not changed significantly with the increase of number of tested voices. At the same time, very large training sets significantly increase the amount of computation required both during training and classification.

Menu