REAL TIME FUZZY CLASSIFICATION OF CEPSTRAL COEFFICIENTS FOR SPEAKER IDENTIFICATION

Abstract

Speaker Identification refers to using utterances from a speaker, in order to determine who the speaker is out of a set of known speakers. Speaker Identification is a widely popular but a complex problem. It is difficult to design accurate algorithms capable of extracting salient features and matching them in a robust way. A novel system for speaker identification based on fuzzy logic has been proposed in this research. Mel-frequency cepstral coefficients (MFCC), which are commonly used for voice-based biometric recognition, form the voice features. In this thesis, a fuzzy nearest neighbour classifier is built for text independent speaker identification using MFCC features. The fuzzy classifier is based on the Gaussian membership function. Speaker identification experiments using the fuzzy nearest neighbour classifier have been carried on a well known publicly available voice dataset. The performance of the proposed classifier has been compared against some of the other commonly known techniques used for speaker identification. The results obtained by the different techniques have been compared on a number of parameters like the efficiency, the time complexity and the space complexity of each algorithm. The obtained results are very promising and verify our claim that the proposed scheme gives better performance than the already existing techniques for speaker identification and that it has the potential to attain a reasonable real-time performance.