Fast Gaussian Evaluations in Large Vocabulary Continuous Speech Recognition

Rapid advances in speech recognition theory, as well as computing hardware, have led to the development of machines that can take human speech as input, decode the information content of the speech, and respond accordingly. Real-time performance of such systems is often dominated by the evaluation of likelihoods in the statistical modeling component of the system. Statistical models are typically implemented using Gaussian mixture distributions. The primary objective of this thesis was to develop an extension of the Bucket Box Intersection algorithm in which the dimension with the optimal number of splits can be selected when multiple minima are present. The effects of normalization of mixture weights and Gaussian clipping have also been investigated. We show that the Extended BBI algorithm (EBBI) reduces run-time by 21% without introducing any approximation error. EBBI also produced a 12% lower word error rate than Gaussian clipping for the same computational complexity. These approaches were evaluated on a wide variety of tasks including conversational speech. I would like to dedicate this thesis to my parents and my husband for their constant support, encouragement and sacrifices throughout my education and career development. DEDICATION i ACKNOWLEDGMENTS First of all, I want to thank Joe Picone for introducing me to the area of speech recognition. He not only created my interest in this area but also inspired me to do my graduate research in speech recognition. He never got tired of my incessantly argumentative discussions and I always learned something valuable from those discussions. Having him as my mentor for my graduate studies has been a memorable experience for me. I would also like to thank Jon Hamaker for all the valuable suggestions and ideas he gave me for this research. I owe Jurgen Fritsch for answering any questions I had regarding the Bucket-Box Intersection algorithm which has been used as a baseline for this research. The environment at ISIP was very conducive to serious research and that has played a big role in my work. Excellent computing resources provided in the lab helped me a lot with the experiments. My years at ISIP will always be a sweet memory for me not only for the knowledge I gained there but also for giving me an opportunity to meet some very wonderful people. I deeply thank all my colleagues there for making those years so enjoyable in spite of the heavy work load we always had. Finally …