A Multibiometric Speaker Authentication System with SVM Audio Reliability Indicator

Performances of biometric speaker authentication systems are good in clean conditions but their reliability drops severely in noisy environments. Implementation of multibiometric systems using audio and visual experts is one of the solutions to this limitation. In this study, weighting for fusing the audio and visual expert scores is proposed to be adapted corresponding to the current environment. Frequent approach uses fixed weighting but this is inappropriate if the systems are executed in uncertain conditions. In this study, we propose a novel approach by introducing Support Vector Machine (SVM) as indicator system for audio reliability estimation. This approach directly validate the quality of the incoming (claimant) speech signal so as to adaptively change the weighting factor for fusion of both subsystems scores. It is important to priory check the speech signal quality because unreliable speech data give incorrect scores hence affect the accuracy of the total scores of the fusion systems. The effectiveness of this approach has been experimented to a multibiometric authentication system that employs lipreading images as visual features. This system uses SVM as a classifier for both subsystems. Principle Component Analysis (PCA) technique is executed for visual features extraction while for the audio feature extraction; Linear Predictive Coding (LPC) technique has been utilized. In this study, we found that the SVM indicator system is able to determine the quality of the speech signal up to 99.66%. For comparison, EER percentages at 10dB are observed as 51.13% for audio only system, 9.3% for fixed weighting system and 0.27% for adaptive weighting system.