A Novel Approach for Text-Dependent Speaker Identification System

Abstract

This research paper will present an Automatic Speaker Identification System using MFCC (Mel-Frequency Cepstral Coefficients) to extract speech features and BPNN (Back Propagation Neural Network) to classify speakers. This paper aims to classify 16 speakers’ patterns with acceptable classification accuracy rates and identify each registered speaker correctly while testing with new speech data without any false identification of the registered test speaker. MFCC method is used to extract the speech features from each speaker and BPNN is used to identify the registered test speaker.  The speech classifier model is built with 13 input nodes with 1 hidden layer consisting of 298 hidden neurons and 16 output nodes for 16 speaker classifications. Initially, the developed classifier model is trained for speaker’s pattern classification and later it is tested with all the registered speakers and found that it successfully identifies all the registered speakers correctly achieving a 100% correct identification rate. A scaled conjugate gradient training function will be used for training the BPNN. A speech database consisting of 16 speakers from different age groups is created in a relatively noise-free environment with the same sentence spoken once. The classification accuracy rate obtained from the classification is 92.6%. MATLAB simulation tool is used in the entire work.