Speech Separation in Humans and Machines

Summary form only given. Any system for extracting information from sound must face the problem of interference between multiple overlapping sound sources. The human auditory system does this very successfully, but our computer systems have some way to go. In this talk, I will review various approaches to separating information from sound mixtures, particularly for the difficult case of target voice mixed with interfering voices. I will also review some results in human abilities to understand sound mixtures, and what these may be able to tell us about promising automatic approaches.