Speaker Recognition Using MFCC and Combination of Deep Neural Networks

Electronics and Communication Engineering

Speaker Recognition is a process of validation of a person’s identity based on his/her voice. The two major steps in speaker recognition are: Extraction of Features and Matching of Features. A direct analysis and synthesizing of the complex voice signal is difficult as a large amount of information is contained in the signal. Several methods have been used to obtain speaker related features and matching of the features. In the following paper we combine the use of different methods to obtain a speaker recognition system. The conventional method can be combined with other methods to obtain better efficiency of the system. In the proposed approach the feature extraction is done using conventional Mel Cepstral Coefficients and a Butterfly structure Deep Neural Network (also called as Deep Autoencoder). After obtaining the coefficients the Deep Neural Network (DNN) is trained for the classification purpose. DNN can be directly used to pull out features and then classify speakers using same DNN but the MFCC and Auto-encoder are used at first for data compression and to get maximum number of features thus getting better efficiency and faster results. The Deep Autoencoder undergoes unsupervised training and DNN undergoes supervised training. Features are obtained for three different classes and these features are then used to train the Deep Neural Network. After the training phase deep neural network can be used as classifier but only for the classes that have been used during the training and for the same text or phrase used during the training time.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.