Biometrics

DECISION THRESHOLD MODEL ESTIMATION AND QUALITY EVALUATION TECHNIQUES FOR SPEAKER VERIFICATION.

Author: RODRIGUEZ SAETA JAVIER.
Year: 2004.
University: POLITÉCNICA DE CATALUÑA.
Place of defense: DEPARTAMENTO DE TEORIA DEL SENYAL I COMUNICACIONS.
Place of preparation: EDIFICI D4 CAMPUS NORD.

Summary: The number of biometric applications has experienced a boom in recent years. Concern about security is becoming increasingly evident and it is in this context where the automatic recognition of individuals by some of its characteristic features such as fingerprints, faces, voices or iris, among others, plays a dominant role. A growing number of users who are demanding this type of applications at a time when the technology is beginning to be already sufficiently mature. At the same time it seeks security, accuracy and low cost, other factors related to biometric applications start to grow in parallel importance. The degree of intrusiveness is certainly worth booming in deciding what biometric technology is most suitable for the application you want to perform. And that's when the recognition of broadcasters is an attractive choice for the use of voice, the natural method of communication between people, their ability to act remotely and its low cost. Automatic recognition of broadcasters has great usefulness as a method of recognition over the phone but also can be used as an access control application witness or forensic analysis. In applications of verification and identification of broadcasters can distinguish various etapas.En First we find the phase parameterization of the speech signal, where the signal is processed to be modeled and compared. Secondly we have the learning stage models, if the training is being conducted, or the stage of decision, if you want to get the result of a comparison. This dissertation focuses on the stages of training and decision of a verification system for broadcasters. In such systems, the result of the comparison is determined by the existence of a threshold decision. The score obtained by comparing the speech signal with a specific model will lead to positive verification if the score is above the threshold, or negative, whichever is the lowest. Moreover, the quality of the samples with those made by the process of training a decisive influence on the performance of the system. The detection of low-quality samples is also under consideration of this thesis. In real applications usually have little data for the estimation of the model and calculation of the threshold, a further additional complexity lies in the difficulty of obtaining data from imposters. Another negative factor linked to the scarcity of data is that the influence of samples with small noises or low quality influence in a very incisive in benefits. This thesis proposes a new system of calculating the threshold decision based strictly dependent announcer broadcasters customers in the application, and a method of detecting those sequences voice that negatively affect the calculation of the threshold. In addition, also suggests new methods to determine the quality of the samples of a model. One of the most interesting proposals is to determine the quality of the samples of a model 'on-line' during training, so that if they detect any sample that fail to meet the minimum quality requirements, it could be replaced by another new instantly. To demonstrate the validity of these proposals has been recording a database in Spanish multisession of 184 broadcasters, specially designed for recognizing announcer called BioTech. Finally, it presents the case of an actual implementation of a verification system for broadcasters 8 in the q 329 ue are implemented some of the techniques developed during this thesis. This application is the remote revocation of certificates through the voice.

BIOMETRICS