Nonparametric modeling for speaker recognition.

摘要

Recognition of people based on their voices is an innate property of all humans. It is a natural phenomenon which almost happens without any effort. However, recognition of people by a machine (computer)---known as speaker recognition---is a challenging task. In this dissertation research, a novel approach to speaker recognition, based on nonparametric probability density function estimation is developed. Speaker recognition has a wide variety of applications ranging from creating user friendly interactive voice response systems to detection of criminal and terrorist activities over the telephone. Characterizing a person's voice and creating a representation is known as speaker modeling, and this challenging task is the central theme of this research. Firstly, speaker modeling is identified as a process of estimating the long-term statistical properties, which is in turn obtained by estimating the probability density function of characteristic features (which preserve speaker information) extracted from speech data. Secondly, it is observed that the conventional methods to develop speaker models are based on inappropriate (or even incorrect) assumptions on the density functions of features. A class of statistical models, known as nonparametric models, as against to the traditional parametric models, are introduced to model speakers. The use of such models make no a priori assumptions on the underlying distributions of features, and thus determine the speaker model rather than enforcing a functional form. Novel strategies to perform speaker recognition tasks, such as speaker identification and speaker verification were developed. Experimental evaluation of nonparametric methods were performed under various real-life scenarios, and the nonparametric system was found to have higher performance as well as have less sensitivity to noise, as compared to the parametric system. Methods to analyze conversational speech data, which consists of voice data from one or more speakers are also introduced. It was found that the analysis of conversational data becomes extremely difficult when the speech is recorded over a telephone, where very limited amount of data is available. Simple algorithms to perform speaker clustering and speaker counting were developed. Evaluation of these methods were performed and results indicate that the proposed method has the potential to perform well in real-world conditions.