"Blind Source Separation (BSS) is a classic problem that attempts to separate unknown sources from observed mixtures. A common framework of this is the Cocktail Party problem, where multiple individuals speaking in a noisy room are recorded by a configuration of microphones, with the challenge being to split the observed signals into individual conversations. This dissertation covers the evaluation of an existing algorithm known as Cascaded ICA with Intervention Alignment (CICAIA), the establishment of a new parameter for characterizing blind mixtures, and three new algorithms, of which two are capable of separating any number of sources given more microphones than sources.
Evaluating the performance of the existing algorithm CICAIA reveals that the "reverberation time" parameter most commonly used to describe a mixing environment is actually insufficient to characterize the performance of a given algorithm. This leads to the new descriptor: sparseness. Its formulation is defined, and additional evaluations show its importance in algorithm characterization, while at the same time describing the strengths and weaknesses of CICAIA-namely, it has mediocre performance of about 13 dB SIRI in good conditions, but performs better than others in high noise or low microphone spacing. Modifying the trigger for intervention alignment to a more sensitive version creates the new algorithm Cascaded ICA with Demixing Intervention (CICADI). This proves to have better performance, roughly 20 dB SIRI, in good conditions, while still performing well in adverse conditions.
With three or more sources, neither CICAIA nor CICADI produce more than 5 dB of separation. A new framework is developed to separate additional sources using redundant information present in overdetermined setups. The first approach, Inter-frequency Correlation with Microphone Diversity (ICMD), efficiently chooses a set of microphones to produce a determined mixture that provides the best separation. The microphone set is selected at one frequency, and maintained for subsequent frequencies until a new set is necessary. The second approach, ICA with Triggered Principal component analysis (ITP), extracts the principal components of the overdetermined mixture, resulting in a determined mixture with minimal noise. Both approaches utilize an efficient detection algorithm to determine when separation has failed at a given frequency, and corrects the separation at that bin before proceeding to the next. Both algorithms produce roughly 15 dB of SIRI in good conditions"--Abstract, page iii.
Grant, Steven L.
Beetner, Daryl G.
Moss, Randy Hays, 1953-
Kosbar, Kurt Louis
Grow, David E.
Electrical and Computer Engineering
Ph. D. in Electrical Engineering
Missouri University of Science and Technology
xi, 90 pages
© 2013 Christopher Thomas Paul Osterwise, All rights reserved.
Dissertation - Restricted Access
Library of Congress Subject Headings
Blind source separation -- Mathematical models
Source separation (Signal processing) -- Mathematical models
Independent component analysis
Print OCLC #
Electronic OCLC #
Link to Catalog RecordElectronic access to the full-text of this document is restricted to Missouri S&T users. Otherwise, request this publication directly from Missouri S&T Library or contact your local library.http://laurel.lso.missouri.edu:80/record=b10115760~S5
Osterwise, Christopher, "Advances in blind source separation" (2013). Doctoral Dissertations. 3.