The Sequential GMM: A Gaussian Mixture Model Based Speaker Verification System that Captures Sequential Information

1 Introduction This report presents a novel speaker verification system that generates a new feature set that captures long duration speaker identifying characteristics while taking advantage of the well-established and well-studied Gaussian Mixture Model system (GMM). Much of the innovation in the system is contained in the intelligent exploitation of traditional cepstral features such that temporal aspects of speech, which are otherwise disregarded in traditional GMM frameworks, can be explicitly modeled. The system consists of a collection of independent GMMs, one for each phoneme, built on these long duration feature vectors. The outputs of these GMMs are then combined at the score level using a neural network. Despite using traditional tools with respect to the GMM and the front-end feature extraction, combining this system with a run-of-the-mill GMM system dramatically reduced both the equal error rate (EER) and the minimum value of the decision cost function (DCF) on a standard speaker verification test set, in comparison to the GMM system alone. This improvement indicates that the long duration features are capturing speaker characterizing information that the regular GMM ignores. The min DCF fell by nearly 65% and the EER fell by approximately 36%. Moreover, the new system's performance, when operating in isolation, approached that of the state-of-the-art GMM. Speaker recognition is a task that is familiar to everyone. When answering the telephone, people often know immediately who is on the other end of the line. Unfortunately, speaker recognition is not such a simple task for computers. Part of the problem is that it is difficult for humans to determine what characteristics they use in identifying speakers. Perhaps they recognize a phrase the person commonly uses or maybe just the way the person laughs. Human based speaker recognition can be studied, and has been to some extent [1]. However, perhaps humans are not the optimal system; perhaps machines can do much better. There are a number of distinguishing speech characteristics that can be utilized, such as acoustic qualities, prosodic patterns, pronunciation preferences, and word usage, to name a few. The sources of these different pieces of information depend on factors ranging from the shape of the nasal passage to where the person was raised [2]. The aspiration for speaker recognition systems is to use all of the above-mentioned sources of information; simply stated, the goal is to capture every piece of information that reveals the identity of the speaker. The difficulty, …