AbstractsLanguage, Literature & Linguistics

Structural and activity relationship of thienopyrimidines analogs as inhibitors of the human farnesyl pyrophosphate synthase

by Aanchan Mohan




Institution: McGill University
Department:
Year: 2016
Keywords: Electrical and Computer Engineering
Posted: 02/05/2017
Record ID: 2066377
Full text PDF: http://digitool.library.mcgill.ca/thesisfile141326.pdf


Abstract

The objective of this thesis is to develop efficient methods for the transfer of knowledge between languages and speakers by sharing acoustic model parameters for automatic speech recognition (ASR). Knowledge transfer between languages is often useful when only a limited amount of transcribed data is available for ASR system development in a target language. Additionally, boot-strapping acoustic phonetic knowledge is also seen to improve ASR performance when adequate training data is available. These scenarios are used as examples to study issues in acoustic-phonetic knowledge-transfer for ASR. Furthermore, the parameters that characterize speaker variability could often be thought to lie in a low-dimensional subspace or a manifold. Parameters for a new test speaker are often estimated with knowledge transfer from training speaker information that is parametrized as a set of subspace vectors or low-dimensional embeddings on a manifold. The technical contributions in this thesis are as follows. First, acoustic mismatch due to different recording instruments and background conditions poses a problem when training a single multi-lingual statistical model on data from multiple languages. The subspace Gaussian mixture model (SGMM), which allows for natural sharing of model parameters between acoustic-phonetic units of different languages is used in this study. A two-stage procedure is proposed to compensate for speaker variability and environmental variability, prior to multi-lingual acoustic model training. As a result of this compensation procedure, ASR performance improvements are observed for all languages used in multi-lingual acoustic model training. Experimental results are presented on Hindi and Marathi speech data on a small-vocabulary agricultural commodities task. With only one hour of available Hindi data, multi-lingual acoustic model training with Marathi is seen to improve Hindi language ASR performance significantly compared to mono-lingual training. Second, to reduce the number of context-dependent errors in Hindi, an algorithm for borrowing state-level SGMM parameters from Marathi in the multi-lingual SGMM acoustic model is proposed. A statistically significant improvement is observed in Hindi language ASR. Furthermore, in order to reduce the number of parameters in the Hindi-Marathi multi-lingual acoustic model, the use of semi-tied covariance (STC) instead of full-covariance matrices is proposed. With a reduction of a factor of five relative to full-covariance parameters, similar ASR accuracy is maintained through the use of STCs. Third, the use of multi-task training for multi-lingual neural network acoustic models is studied. The use of multi-task training provides state of the art results on a well-known large vocabulary read speech task. Experiments on cross-language adaptation when only a limited amount of target language data is available are also presented. To reduce space and time-complexity to train these networks the impact of low-rank matrix factorization of the weight matrix in the final… Advisors/Committee Members: Milica Popovich (Supervisor2), Richard Rose (Supervisor1).