Abstract

Whilst there is no doubt about the immense practical benefits that could be derived from the automated analysis of audio scenes, it is not clear that the research community has yet developed a sufficiently sophisticated theoretical framework to realise its full potential. Recent years have seen measurable progress in computational approaches to information extraction by applying the latest machine learning techniques to annotated (or even unannotated) data, but most of the focus has been on classifying the surface phenomena associated with acoustic events. Little attention has been given to interpreting the underlying ‘intentional’states that are unique to living organisms and which drive the physical actions that are performed (particularly communicative behaviour such as speech). Of course, if there was a simple one-to-one relationship between internal intentional states and the consequent surface behaviour, then interpretation would be relatively straightforward. However, in reality there is significant ‘coupling’(ie, dependencies) between objects, agents and their environment, and this means that interpreting what is happening in an acoustic scene requires a yet-to-bedefined unified computational modelling approach which is capable of integrating the relevant contingencies. This stimulus talk illuminated these issues and raised the following issues for discussion: How important is it that we acknowledge that the world contains intentional agents? Can we envisage a unified computational modelling approach which is capable of integrating the relevant contingencies? What are the implications of modelling self (recursion, context dependency …

Date

2014

Authors

Meinard Müller, Shrikanth S Narayanan, Björn Schuller

Journal

Computational Audio Analysis

Pages

9