Year: 2010
Type: dissertation
Abstract: This thesis investigates the problem of using nearest-neighbor based non-parametric methods for performing multi-class class-conditional probability estimation.The methods developed are applied to the problem of acoustic modeling for speech recognition.Neighborhood components analysis (NCA) (Goldberger et al. [2005]) serves as the departure point for this study.NCA is a non-parametric method that can be seen as providing two things: (1) low-dimensional linear projections of the feature space that allow nearest-neighbor algorithms to perform well, and (2) nearest-neighbor based class-conditional probability estimates.First, NCA is used to perform dimensionality reduction on acoustic vectors, a commonly addressed problem in speech recognition.NCA is shown to perform competitively with another commonly employed dimensionality reduction technique in speech known as heteroscedastic linear discriminant analysis (HLDA) (Kumar [1997]).Second, a nearest neighbor-based model related to NCA is created to provide a class-conditional estimate that is sensitive to the possible underlying relationship between the acoustic-phonetic labels.An embedding of the labels is learned that can be used to estimate the similarity or confusability between labels.This embedding is related to the concept of error-correcting output codes (ECOC) and therefore the proposed model is referred to as NCA-ECOC.The estimates provided by this method along with nearest neighbor information is shown to provide improvements in speech recognition performance (2.5% relative reduction in word error rate).Third, a model for calculating class-conditional probability estimates is proposed that generalizes GMM, NCA, and kernel density approaches.This model, called locally-adaptive neighborhood components analysis, LA-NCA, learns different lowdimensional projections for different parts of the space.The models exploits the fact that in different parts of the space different directions may be important for discrimination between the classes.This model is computationally intensive and prone to over-fitting, so methods for sub-selecting neighbors used for providing the classconditional estimates are explored.The estimates provided by LA-NCA are shown I would like first to thank my advisor Michael Collins who taught me how to be a researcher.He provided insight when I needed it and always knew the answers to my technical questions.I especially appreciate the freedom to explore (and fail) he gave me while carefully directing my research towards fruitful endeavors.Finally, he provided me with unfliching support when I chose to start a family, for which I will always be grateful.There are numerous other researchers who deserve thanks.Tommi and Jim both provided me with excellent feedback during the course of my research and on the thesis itself.My group taught me a lot over the years about various topics in speech and language and were excellent for bouncing around ideas.T. J. Hazen taught me a lot about the speech recognizer and Hung-An Chang helped me develop and understand various ideas for acoustic modeling.To all these people and many others (too numerous to list) whose discussions, support, and ideas helped me shape my own, I am sincerely grateful.No one supported me more during my graduate work than my husband, Nick, who took care of me and our family when we needed it and whose continued interest in my work and my ideas were was always inspiring.My son Neel provided a much needed sense of urgency to actually finish and submit my thesis by giving me additional responsibilities, so I'm grateful to him as well.I'm thankful to my parents for supporting me throughout my many years at MIT and always trying to take responsibilities off my shoulders so that I could study and explore unhindered.My sisters and brother-in-law also always supported me and were proud of me no matter what which helped me through this process.I would also like to thank numerous friends who made life fun and kept me connected to the outside world throughout these many years.Last, but not least, I'd like to thank MIT where I studied for my Bachelor's, Master's, and Ph.D.This has been an amazing place to live and study, and I've met some of the best friends of my life here
Author: Natasha Singh-Miller
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W1510457122