Perspectives on Speech Separation -- A Workshop

East Bay Institute For Research And Education Inc, Martinez CA

Investigators

Abstract

The situation that may be the most detrimental to verbal communication is when a speech signal is embedded in a background of other, interfering speech, such as crowd noise or echoes of the target speech in a reverberant room. It is also known that speech understanding in such a "cocktail-party" situation is especially difficult for elderly individuals, even for those with relatively good hearing. Over the last 25 years, data collected by experimental psychologists, audiologists, and neuroscientists have unveiled some important characteristics of the process of separating speech sources, as well of its dysfunction, and developed models to explain and predict human performance in "cocktail-party" settings. Motivated by the same data, and by the fact that human listening performance under such difficult conditions has its limits, information scientists have become increasingly interested in finding computational solutions to speech separation. The last 10-15 years have generated an unprecedented rush toward the development of diverse computational schemes aimed at automatic separation of speech sources by filtering out unwanted speech and enhancing the target. Nevertheless, to the onlooker of this work there appears to be no general solution to this problem. Because the facts, models, and schemes are dispersed over different scientific fields and in specific sub-areas, it is likely that for any solution to bear fruit requires a concerted effort among scientists from these various disciplines. The objective of this multidisciplinary, international workshop is to bring together an invited panel of around twenty scientist-experts whose work represents the cutting edge of speech separation within their particular discipline. Over a two-day period, these experts will present their own and discuss each other's research, in addition to taking part in general discussions in which the current status and future trends of speech separation research will be examined. The workshop will also produce a report containing all presentations, as well as highlights of the discussions.