Array signal processing under model errors with application to speech separation
Victor C. Soon, Yih-Fang Huang · 1992
In recent years there has been increasing interest in eigenstructure based array signal processing techniques especially in the areas of radar and sonar processing and in signal separation applications as found in communications or speech acquisition/enhancement. However, model errors are not accounted for in these methods. Unfortunately, the problem of model error is a ubiquitous part of real systems: they arise from such phenomena as non-ideal sensor characteristics (e.g., sensor calibration errors in phase and gain, etc.), unknown source characteristics and locations, etc. This dissertation is broadly divided into two parts. The first part examines the effects of model errors on the performance of existing eigenstructure based methods in array signal processing. A statistical analysis of the ESPRIT (Estimation of Signal Parameters via Invariance Techniques) algorithm under random model errors is performed. The analysis provides interesting insight into the sensitivity of ESPRIT to model errors: in particular, for uniform linear arrays of sensors, it is found that the mean square error of DOA (direction of arrival) estimates found using ESPRIT is almost totally dependent on errors in sensor phases and not that of sensor gains. The second part of the dissertation deals with the development of algorithms that take model errors into account. Two approaches are proposed. The first is based on a signal subspace constraint on the possible set of model parameters as determined from the array covariance. This constraint is incorporated into an iterative procedure which calibrates the array under model errors and unknown source signals. The second approach incorporates blind identification and clustering for the source estimation problem under model errors or uncertainties. The proposed approach is shown to be robust to the simultaneous presence of such uncertainties such as unknown sensor and channel gains, unknown combinations of near-field and far-field sources, unknown combinations of narrowband and wideband sources, unknown source spectral characteristics and unknown number of sources.