Performance analysis of text-dependent speaker recognition system based on template model based classifiers

Speaker recognition technology has found its usage in varied different fields including security systems, banking operations and forensic laboratories etc. due to its easy implementation, flexibility and security. Each of these applications demand high performance i.e. high speaker recognition rate. In this paper, MFCC based text-dependent SR system is presented. The speaker voice is modeled by Mel spaced frequency spectrum. The performance of the system has been evaluated using our own recorded subject database and compared on the basis of two template model based classifiers - Nearest Neighbor (NN) and Vector Quantization (VQ), in terms of parameters such as recognition accuracy rate, system processing time, number of users and size of the training database. Based on the experiments, it was found that VQ performed better than NN for all the parameters. Although both the classifiers use Euclidean distance as a similarity determination function between training and testing data but NN being an instance based algorithm (lazy algorithm) do not build any model for training data and hence lags behind VQ in performance whereas the VQ builds code vectors of different users by clustering as soon as the training data is fed.