Study of speaker recognition system based on Feed Forward deep neural networks exploring text-dependent mode

Abstract

We aim by this work to follow the significant progress in speaker recognition systems getting the benefits of the advancement in the artificial intelligence (AI). Indeed, the deep learning algorithms have proved a real performance in the recognition and classification data. In this contest, we present a study of three different speaker recognition system based in Feed Forward neural networks. The first one is the logic regression, the second one is the Multilayer Perceptron (MLP) and the third one is the Stacked Denoising Autoencodeurs (SDA). We evaluated these recognition rates using the parameterization technique Mel Frequency Cepstral Coefficients (MFCC). To find the best results and to better optimize automatic recognition algorithms, we tested our speaker recognition system under the text-dependent database RSR2015. We studied the recognition rates by varying the values of neural networks parameters, number of neurons and number of hidden layers…etc. We discussed the different results obtained and we selected best parameter values which lead the minimum rate error of recognition.