Much of the sound interpretation world uses Mel-frequency cepstral coefficients (MFCCs) for coding sound for interpretation. While such coding is highly compressed, it loses all information about the fine time-structure of the sound, and this is particularly important both for locating sound sources and for separating foreground sound from the background. We describe a simple spike (event) based coding technique that maintains fine time structure, and report on how we have used it for sound direction finding, for segmenting sounds, and for determining the type of musical instrument that generated a sound. We discuss why this type of representation can be useful for sound interpretation in realistic environments, and discuss possible ways forward for this type of approach.
Leslie S. Smith: After graduating in Mathematics in 1973, went off to work in the computing industry, but then returned to do a PhD in computing at Glasgow, graduating in 1981. Initially he worked on parallel computing, and taught at Glasgow, before arriving at Stirling University in 1984 and becoming involved in neural networks and computational neuroscience. More recently, he has worked on neuroinformatics, including helping create the CARMEN e-Science portal, and on biologically inspired early auditory processing. Currently, he is Professor of Computing at Stirling University, and, having stepped down two years ago from being Head of Department, has a little more time to devote to research.