Automatic Text Independent Speaker Recognition Using Voice and Lip Movement

Abstract

This thesis is focussed on the implication of combining multiple biometric traits on the newlineperformance of Text-Independent Automatic Speaker Recognition. A multimodal biometric newlinesystem combines more than one biometric trait acquired from a person which could be newlinephysiological or behavioural in nature. Besides the biometric traits, the sources of information newlineincluded in a multimodal biometric system can involve the multiple sensing devices employed, newlinemultiple instances of data collection and multiple algorithms for multiple traits. In this era of newlinedigitization, mobile devices have become affordable and offer ease of use. Additionally, the newlineavailability of cutting edge sensor technology has simplified the acquisition of biometric traits newlineand the focus of all transactions has moved to a digital platform. Implementing a secure and a newlinereliable biometric system which can identify, recognize, or verify an individual within newlineminimum time while safeguarding the privacy of an individual is challenging and remains an newlinearea of research. newlineAfter conducting an exhaustive literature survey, two major conclusions were drawn. Unimodal newlinesystem have limitations such as failure to enrol, limited flexibility, reduced security to spoof attacks amongst others, making multimodal system the need of the hour. The additional biometric trait to be employed to reinforce the voice parameter for speaker recognition was the other concern to be addressed. Voice acquisition being a non-invasive and easy to acquire trait newlineis supplemented with lip movement features to add liveness detection for implementing a newlinerobust speaker recognition system. Lip movement of a speaker being an extension of sound in the process of speech production besides being unique for every individual, made this as the choice for another biometric parameter.A standard audio-video database - VidTIMIT, which is derived from the N-TIMIT corpus, is chosen for this research to explore a text -independent speaker recognition system along with another database of 72 speakers created in the college.