Estimating uncertainty models for speech source localization in real-world environments

Abstract

This thesis develops improved solutions to the problems of audio source localization and speech source separation in real reverberant environments. For source localization, it develops a new time- and frequency-dependent weighting function for the generalized cross-correlation framework for time delay estimation. This weighting function is derived from the speech spectrogram as the result of a transformation designed to optimally predict localization cue accuracy. By structuring the problem in this way, we take advantage of the nonstationarity of speech in a way that is similar to the psychoacoustics of the precedence effect. For source separation, we use the same weighting function as part of a simple probabilistic generative model of localization cues. We combine this localization cue model with a mixture model of speech log-spectra and use this combined model to do speech source separation. For both source localization and source separation, we show significantly performance improvements over existing techniques on both real and simulated data in a range of acoustic environments.