Hamid Satar

Title: Partial Tracking of Music Signals using Robust Kalman Filtering

Date: Thursday, September 21, 2006
Time: 3:00 PM - 5:00 PM
Location: 442 Dana Research Center

Abstract:

In this project we discuss the problem of partial tracking as applied to music signals, and propose tracking algorithms based on Kalman filtering. We first introduce a novel technique for detection of peaks in spectral representations of music signals. We also introduce different evolution models in the form of linear state-space equations and based on the shape of frequency and power partials in different classes of melodic instruments. Parameters of these models are estimated using a large database of music signals. These parameters are frequency-dependant and for different frequency regions, different evolution models are constructed. A major part of this project is the tracking of harmonic partials through separate time instances of spectral data. We formulate a conventional Kalman tracker and analyze the performance of this tracker through a comparison with another method and also by observing its effectiveness in critical situations. We also propose novel data association techniques and rejection rules based on quality factors for individual components of a track and also for individual tracks. The second major area of this research is the development of robust partial trackers based on the concept of robust Kalman filter. Compared with the conventional Kalman filter, this class of filtering has less dependency on the accuracy of model, in which the underlying parameters are treated as bounded uncertainties. We develop two robust trackers and compare their performances. These robust methods provide improved tracking capabilities, while preserving the useful properties of the conventional Kalman tracker.

Thesis Committee:
Prof. Bahram Shafai (advisor)
Prof. Dana H. Brooks
Prof. Jeffrey A. Hopwood
Prof. Patrick J. Wolfe (Harvard University)