Examination of distinctive capacity of human voice

Biometric-based authentication systems have been begun to be widely used instead of passwords. Because voice can be captured easily by using a microphone, voice is more popular between all biometric modalities. Although the use of voice biometrics is increasing day by day, the studies about capacity of speaker verification systems are limited. Moreover, the results of these studies conflict with each other and which in turn raise doubts reliability of speaker verification systems instead of answering questions. Because of these reasons, in this thesis, the capacity of voice-based authentication systems, in other words, speaker verification systems, is investigated in terms of entropy. The subject has been examined under three main headings. Firstly, proposed approaches up to now for measuring capacity of biometric systems are examined in detail and whether these approaches are suitable for voice-based authentication systems or not was also investigated. Secondly, state-of-the-art methods used in speaker verification systems are overviewed. The features extracted from the speeches, the models used for representation of the features, and voice-based authentication methods for these models are examined separately. Thirdly, because the dataset used in speaker verification systems contains limited number of speakers and speeches, by using open sources a new dataset containing more than 20000 speakers is created. A new approach suitable with state-of-the-art speaker verification system is proposed for measuring capacity and the mathematical background of this approach is explained in detail. This approach is examined in different cases by using different datasets. Finally, new research topics on capacity estimation are mentioned.