Separation and localization of acoustic signals
Kevin H. Knuth · The Journal of the Acoustical Society of America · 2001
The cocktail party problem is considered where there are several sound sources present in a free-field environment and an observer, who has the ability to record only linear mixtures of these sounds, is interested in obtaining estimates of the original source signals. Previous application of Bayesian inference to the general problem of blind source separation has demonstrated that it is possible to include additional prior information such as the transmission of the signals to the detectors and the possible locations of the sources to improve the source estimates. The derivation of prior probabilities describing such information will be discussed as will the construction of useful models and application of a general Bayesian methodology allowing one to deal with noninstantaneous inverse square transmission, noisy environments and any number of sources and detectors. Methods for dealing with environments in enclosure will also be described. [Work supported in part by NARSAD Young Investigator Award.]