Events Calendar

Manifold Constrained Acoustic Modeling for Automatic Speech Recognition

Speaker: Richard Rose, McGill University

Talk Title: Manifold Constrained Acoustic Modeling for Automatic Speech Recognition

Abstract: This presentation investigates the application of manifold learning approaches to automatic speech recognition (ASR). All of the approaches considered rely on very high dimensional feature representations for speech while at the same time assuming that speech features are constrained to lie on a low dimensional embedded manifold. Discriminative manifold based linear projections are investigated as dimensionality reducing feature space transformations. These techniques attempt to preserve local within-class relationships along a nonlinear manifold while maximizing separability between classes. The ASR word error rates obtained from these techniques are compared to those obtained using more well known discriminative dimensionality reducing linear transformations on multiple speech in noise tasks. The high computational complexity associated with computing the Laplacian matrices for these techniques is reduced by an order of magnitude through the use of locality sensitive hashing (LSH) algorithms. As time permits, a discussion of additional applications of manifold based constraints to speech processing will be presented. These include manifold based constraints for regularizing training for speaker adaptation transformations, regularized least squares classifiers for spoken term detection, and manifold regularization for training deep neural networks.

Biography: Richard Rose is an Associate Professor and Graduate Program Director of Electrical and Computer Engineering at McGill University in Montreal, Quebec, Canada. His major area of research is in speech and language processing. His recent research contributions have been in acoustic modeling for speech recognition, computer aided human language translation, and computer aided speech therapy. Over his career, he has published over 130 articles in refereed international journals and conference proceedings. He has served as Adjunct Research Scientist at the Human Language Technology Center of Excellence in Baltimore and as Adjunct Professor of ECE at Johns Hopkins University. Prof. Rose is an IEEE Fellow. Before coming to McGill in 2004, Prof. Rose was a senior member of technical staff at AT&T Labs Research where he contributed to AT&T's speech enabled services and was inventor or co-inventor on twelve patents. His professional service has included General Chair of the IEEE Automatic Speech Recognition and Understanding Workshop, membership in the IEEE Speech Technical Committee, elected membership on the IEEE Signal Processing Society Board of Governors, associate editor of the IEEE Transactions on Speech and Audio Processing, associate editor of the IEEE Transactions on Audio, Speech, and Language Processing, and founding editor of the IEEE Speech Technical Committee Newsletter. Prof. Rose is a member of Tau Beta Pi, Eta Kappa Nu,
and Phi Kappa Phi.


Host: Prof. Shrikanth Narayanan & Alexandros Potamianos

Location: Hughes Aircraft Electrical Engineering Center (EEB) - 132

Audiences: Everyone Is Invited

Contact: Talyia Veal