An Algorithm for Calculating the Contribution of Acoustic Features in Speaker Recognition

Abstract

AbstractIn order to solve the problem that the model performance will be different if different acoustic features are selected in the speaker recognition field, an algorithm for calculating the contribution degree of acoustic features in speaker recognition is proposed. First of all, the use of increase or decrease in weight method to calculate contribution of each dimension, and then, with a fisher ratio than the acoustic features of each dimension fisher ratio calculation, to undertake the corresponding percentage of fusion after can get the characteristics in the contribution of speaker recognition model. Taking the acoustic feature GFCC and the speaker model i-vector as an example, this paper calculates the contribution rate of each component of GFCC under the speaker recognition by using the proposed new algorithm, analyzes the differences of the three methods, and provides a new method for subsequent feature processing.KeywordsContributionSpeaker recognitionIncrease or decrease component method