The History and Future of CASA

1 INTRODUCTION In this chapter I briefly review the history and the future of computational auditory scene analysis (CASA). Auditory scene analysis describes the process we use to understand the world around us. Our two ears hear a cacophony of sounds and understand that the periodic tic-toc comes from a clock, the singing voice comes from a radio and the steady hum is coming from the refrigerator. The field of computational auditory scene analysis crystallized with the publication of Al Bregman's book " Auditory Scene Analysis " (Bregman 1990). The commonly understood goal is to listen to a cacophony of sounds and separate the sounds from the mixture, just as humans do. I would like to argue that this is not what people do. In this review, I will describe some progress to date towards modeling human sound separation, and review why this is the wrong direction for those of us interested in modeling human perception. Instead, we should be thinking about sound understanding. This is clearly a much harder problem, but should provide a better model of human sound separation abilities. In particular, this paper makes two related points. 1) We need to consider a richer model of sound processing in the brain, and 2) human sound separation work should not strive to generate acoustic waveforms of the separated signals. Towards this goal, this paper reviews the use of a correlogram for modeling perception and understanding sounds, the success at inverting the correlogram representation and turning it back into sound, and then summarizes recent work that questions the ability of humans to isolate separate representations of each sound object.