Investigating Matching Pursuit Decompositions of Non-Noisy Speech Signals

Previous

Students

Bob Sturm, Elec & Comp Engineering

Faculty Advisors

Jerry Gibson, Elec & Comp Engineering
Curtis Roads, Media Arts & Tech

Next

Abstract

The relatively recent signal time-decomposition method "matching pursuits" (MP) may provide an interesting tool for high-resolution time-frequency distributions, signal denoising, and audio coding. This method essentially finds good approximations of signals using a linear combination of "atoms" selected from a non-orthogonal set of elementary waveforms, or a "dictionary." A signal can thus be represented by functions that are well-localized in time and frequency, or are better correlated to the signal than orthogonal bases such as a Fourier basis.

Others have shown that care must be taken when designing a dictionary, but none have studied in detail the effects of the algorithm and dictionary on the decomposition of different types of signals. How good of a representation does MP provide? For what kinds of signals does it work well? To what extents do the types of waveforms used in the decomposition affect the results? We have investigated these questions for synthetic and speech signals and have found that the resulting decompositions can be inaccurate and misleading no matter what is in the dictionary. For purposes of signal analysis this is a disturbing flaw; but for other uses, such as signal modification for artistic purposes, it is less problematic.