Text Dependent Speaker Recognition using Deep Neural Networks

Speaker Recognition is used for identification of a person depending on the characteristics contained in the speech signal. In this paper we propose the use of Deep Neural Network (DNN) for text dependent speaker Recognition system (SRS). Mel Frequency Cepstral Coefficients (MFCC) and Auto-encoder (Butterfly Structure Neural Network) are used to extract the features of speech signal at the initial stage. The previously obtained coefficients are then used to train the DNN to later classify the speakers. DNN can be directly used to extract features and classify speakers but the MFCC and Auto-encoder are used initially for data compression and maximum number of feature extraction thus aiming to get better efficiency and faster results.