ITR: Blind Identification of Multivariate Systems

University Of California-Riverside, Riverside CA

Investigators

Abstract

Blind identification of multivariate systems concerns with modelling, estimation and detection of multivariate systems driven by unknown sources. It is an emerging area of fundamental importance to applications such as wireless communications, human-computer interface, and video surveillance. It provides a foundation for, as well as a unification of many application-specific views arid techniques. In particular, it brings a bridge between the field of space-time coding for wireless communications and the field of speech recognition for human-computer interface. This project continues a systematic study previously conducted by the P1 in the past a few years. A primary focus is to understand the limits of blind identification of convolutive multiple-input-multiple-output (MIMO) systems driven by nonwhite sources. This is known to be a challenging problem. Prior work in this area mainly concerns with single-input-multiple-output (SIMO) systems. instantaneous MIMO systems, MIMO systems driven by white sources, or MIMO systems driven by modulated sources. Preliminary discoveries on convolutive MIMO systems driven by nonwhite sources have been made recently by the P1. and more are yet to be discovered. Great efforts will be made to draw connections between the generic identifiability conditions associated with unknown sources and those associated with encoded and/or modulated sources. The results of this work will be a complete understanding of the identifiabilty of MIMO systems. a complete taxonomy of identification algorithms for various conditions of MIMO systems with various coding schemes, and a complete evaluation of performance bounds of MIMO systems driven by unknown sources. This project will also explore a key application in speech enhancement. The acoustic channel in a common environment (such as offices) is known to have a convolutive distortion that severely hampers the performance of today's best speech recognition systems. Blind deconvolution is of great importance at the very front end of a speech recognition system. The flexibility or friendliness of future human-computer interface depends on how well blind deconvolution can be carried out and consequently how well speech recognition can be performed. To some degree, research has either neglected the structural model of acoustic channels or ignored the hidden models in speech signals. This project will cross-fertilize between the field of sensor arrays and the field of speech recognition. By exploiting multiple microphones, the structural details of acoustic channels as well as the hidden Markov model of speech signals, we are expecting a significantly improved speech recognition system by the end of this project.