Master of Science Thesis Defense by Shivesh Ranjan

ABSTRACT:
Model-based approaches to achieve Single Channel Source Separation (SCSS) have been reasonably successful at separating two sources. However, most of the currently used model-based approaches require pre-trained speaker specific models in order to perform the separation. Often, insufficient or no prior training data may be available to develop such speaker specific models, necessitating the use of a speaker independent approach to SCSS. This report proposes a speaker independent approach to SCSS using sinusoidal features. The algorithm develops speaker models for novel speakers from the speech mixtures under test, using prior training data available from other speakers. An iterative scheme improves the models with respect to the novel speakers present in the test mixtures. Experimental results indicate improved separation performance as measured by the Perceptual Evaluation of Speech Quality (PESQ) scores of the separated sources. A novel Successive Mixtures Codebook Replacement (SMCR) algorithm is presented, which uses the codebooks developed after separating one mixture, as the initial codebooks for separating the next mixture, and is able to improve separation performance, combined with a decrease in computational cost by a factor of five. Another novel Gender Restricted Eigenweights Search (GRES) algorithm is also presented, that reduces the computational expense (by a factor of eight) of separating both cross gender and same gender test mixtures without using any prior gender specific information, and without compromising separation performance. To the best of our knowledge, the GRES algorithm is the first attempt towards reducing computational complexity of SCSS, by employing a gender classifier as a pre-processing step.

NOTE: All ECE Graduate Students are ENCOURAGED to attend.
All interested parties are invited to attend.
Open to the public.

Committee Members: Dr. John Buck and Dr. Joel MacAuslan

Advisor: Dr. Karen L. Payton