Speaker Recognition

Abstract

As the technology is escalating in all the directions we are intending to move one step ahead of using user name and passwords for our identification. Now a days it is obvious that speakers can be identified from their voices. First step is to start with the basics of speaker identification which mainly includes DSP concepts.DSP definitions have been reviewed which are involved in the speaker identification. After that the step involved is feature extraction. It is a method through which we can acquire the unique features of a particular speaker which are then used to identify them. Many features have been analyzed like cepstrums, LPCC (Linear predictive cepstral coefficient), MFCC (Mel Frequency Cepstrum Coefficients) and finally used MFCC for the identification. The Overall system is mainly divided into two parts, Text dependent identification and Text Independent identification. For both the identifications MFCC is used as the extracted features .In text dependent identification many matching methods are analyzed like Vector Quantization, DTW (Dynamic Time Warping), HMM (Hidden Markov Model) and performed pattern matching using method (DTW) after which the system is giving the accuracy of 91.30% .For Text independent identification VQ (Vector Quantization) has been used for feature modeling and then feature matching is done using Euclidean distances. Some advancement is done in the VQ method by calculating the weights of the patterns as per the unique information contained in them about a particular speaker which increases the accuracy. The text independent recognition is giving the accuracy of 86% when VQ is used and when the advancement is done its accuracy increases to 90%.So, it can be said that this thesis work provides different approaches for the speaker recognition.