Increase Noise Resistance of the Automatic Speaker Recognition System of Critical Use

Abstract

The relevant speaker recognition systems in which i-vector/PLDA modeling is applied to the description of soundtracks synthesize the generalized PLDA model with average parameters on all soundtracks base without their segregation on the noise level. As a result such systems provide the acceptable level of reliability only in the presence of the large training selection, both by quantity, and on duration of soundtracks. Authors suggest to synthesize separate PLDA models for the description of soundtracks with the determined levels the relation signal / noise (RSN) therefore factors which characterize specific features of a speaker’s voice, will be concentrated in the most changeable areas of i-vector space. It is assumed that statistical analysis of the parameters of such variability regions for phonograms with a signal-to-noise ratio determinants will determine the factors that are stable to the noise level in the signal and informative for the speaker's identity recognition. The statistical analysis of parameters of such areas of variability for soundtracks with the determined RSN level allowed to define noise resistant and informative for speaker recognition factors. For the solution of this task analytical expression for PLDA model which parameters are defined only by values of i-vectors, into which it is entered the parameters describing the RSN levels is received. Criterion functions and stages EM-algorithm of training RSN depended PLDA mixture are also synthesized and check of efficiency of the offered models by their comparison with results which show RSN independed PLDA mixture for a certain base of the speaker’s soundtracks is carried out. For complex testing of the proposed theoretical results, the authors formed two test samples of phonograms that differed in the way of making noise into a signal. Experimental results show that the RSN depended PLDA model allows for better results than the RSN independed PLDA model for almost all test data variants, when phonograms from the first set were used for training models. However, when the training of models occurred according to data from the second set, the situation turns out to be the opposite. This can be explained by the fact that the use of phonograms with the three levels of the RSN the formation of the first set of training data provides greater informativity than the second way of obtaining training data.