Modeling binaural processing: What next?

Abstract

Models of binaural processing are traditionally based on signal-processing in a bottom-up (signal-driven) architecture. Such models are sufficient for a number of technologically important applications, such as perceptual coding, sound-source identification and localization, dereverberation and decoloration, but fail when applications require cognition. Future models will thus include symbol processing in addition to signals processing, will have inherent knowledge bases and employ top-down (hypothesis-driven) strategies in addition to bottom-up ones. Some of these features are known from automatic speech recognition and may be generalized for broader application, e.g., blackboard structures. With the new models more sophisticated applications may be approached, for instance, quality evaluation and assessment on the basis of internal references, such as needed to determine estimates of the quality of performance spaces and/or audio systems. Further, to enable autonomous learning, future models will employ feed-back loops to realize active exploratory actions. Some of these features can be imported from recent research in robot audition. In our contribution, we shall, among other things, report on ideas and concepts as currently discussed in AABBA, an international grouping of 14 laboratories in Europe and the US that are dealing with auditory assessment by means of binaural algorithms.