Cocktail Party Solutions: Mixing techniques

The human auditory system has an unparalleled ability to solve what is known as the cocktail problem: the task of perceptually separating superimposed acoustic sources in a noisy environment. Information about the auditory system can inform our choices of representation and computation. Recent progress has been made from several diier-ent approaches to the acoustic source separation problem. Computational auditory scene analysis (CASA) and independent components analysis (ICA) represent two extremes in a spectrum deened by the dependence of the technique upon prior knowledge versus the richness of the input data. I suggest that some of these approaches can be combined and that the combination will require information about the intrinsic qualities of sources as well as about their relationships with each other. Framing these techniques in terms of probability will allow us to learn a network of modules based on intrinsic properties of natural sounds. Interaction between modules at diierent levels of analysis is necessary for negotiating a consistent grouping policy. This negotiation process requires representations that can accomodate ambiguity among potential grouping hypotheses. These grouping hypotheses should be continuously variable to allow for negotiation between different constraints. Negotiating between source-separation possibilities requires complexity to be minimized using regularization techniques.