Speaker Identification of Disguised Voices Using MFCC Statistical Moment And SVM Classifier

Abstract

Speaker Recognition is the most powerful scheme in which security is the major concern but misclassification rate often increases when speakers attempt to disguise their voices. This project mainly focuses on electronic voice disguise method that involves modification, deviation or twisting of the original voice of a speaker electronically via voice changing software’s. With sophisticated algorithms, electronic methods achieve better results than nonelectronic methods as they provide a vast range of choices for voice disguise. Voice disguise has been expanding in several illegal applications like annoying phone calls, ransom calls, and bomb threats etc. It has adverse effects on determining the veracity of audio evidence. It is utmost important in audio forensics to spot whether the so-called voice of a speaker is original or disguised.The proposed system presents an efficient way to identify a speaker using MATLAB tool. The system recognizes a speaker and makes a decision on the basis of information present in his/her voice that may be an original voice or disguised voice.The first and foremost task is to find the disguised voices from the original voice and then the next step to find its identity speaker. Several related studies have been accounted on Automatic Speaker Recognition (ASR) systems but not many studies on recognizing speaker from their disguised speech. Here acoustic information of each speaker is passed through various stages i.e., preprocessing, feature extraction, pattern matching, decision making. To carry out this research, MFCC based statistical moments involving mean and correlation coefficients are used as acoustic features in feature extraction stage and SVM classifier are used in feature matching stage. These features and classification technique provide useful and precise results for the identification of hidden speaker.