Variable print quality

In the literature, much research work has been done in the area of speaker verification. The developments include: different types of speaker verification techniques, methods for feature extraction, measures for telephone channel compensation, system robustness etc. In contrast, the problem of acoustic feature selection for speaker verification has been relatively neglected. Hence our aim is to study the effect of feature selection in speaker verification. In this thesis, a text-dependent Dynamic Time Warping speaker verification system is developed. Different feature selection approaches are then studied and compared using this verification system. Feature selection is performed on voice and lip-based input features. The results show that feature selection never degrades the performance of the system, but may improve it while reducing the cardinality of input features. Recently, the emphasis in personal identity verification based on biometric data has shifted to the use of multiple modalities in order to improve the recognition performance. In this context, we investigated the possibility of combining outcomes of voice-based verification system with lip-based system. The different combination rules used for this are Sum, Product, Max and Min. It has been shown that this fusion improves the performance of the system. We also studied an application of speaker verification to a video annotation problem. For this, we propose an automatic speaker-based segmentation of multimedia material using very few samples for training. The segmentation process involved is unsupervised, hence segmentation error achieved is not satisfactory. Therefore, in order to improve this segmentation error, correlation of audio with video shot detection is proposed and implemented.