BLIND SEPARATION OF MIXED SIGNALS FROM CONVOLVED SOURCES

Scientists at UCSF have developed a practical method for signal deconvolution called Dynamic Component Analysis (DCA). DCA can resolve a complex signal comprised of mixed signals from different sources into its component parts traced to their points of origin. The practicality of earlier methods of signal processing has been limited by frequency distortion to 2 or 3 sources.

In situations where multiple sensors each receive a different mixture of source signals, signal processing is complicated by the fact that each source signal is delayed and attenuated by different amounts on its way to the different sensors. Furthermore, a given source signal may reach a given sensor through more than one path - a situation known as "convolutive mixing." The DCA algorithm receives such mixed signals as inputs and produces original source signals as outputs. Based on statistical estimation techniques, DCA can be implemented with a simple neural network and applied to a wide range of problems.

One problem that can be addressed with DCA is the processing of speech signals in reverberant, noisy, or multiple-speaker environments - the "cocktail party" problem. A signal processor using DCA might significantly enhance the performance of hearing aids or speech recognition devices.

Additional work is underway to apply DCA to the analysis of electro-encephalographic (EEG) and magneto-encephalographic (MEG) multi-electrode recordings. This work may lead to the development of new diagnostic tools for neurologic disorders such as epilepsy.

A simulation is available for evaluation. The hardware specific to a particular application has yet to be developed.

If you would like to receive further information about this technology and potential licensing opportunities, please contact:

Joel B. Kirschbaum, Ph.D.
Director & Senior Technology Portfolio Manager
(415) 353-4462 phone
(415) 348-1579 fax
Joel Kirschbaum, Ph.D.

Reference: OTM Case #SF97-093