A non-parametric Bayesian prior for causal inference of auditory streaming

Human perceptual grouping of sequential auditory cues has traditionally been modeled using a mechanistic approach. The problem however is essentially one of source inference – a problem that has recently been tackled using statistical Bayesian models in visual and auditory-visual modalities. Usually the models are restricted to performing inference over just one or two possible sources, but human perceptual systems have to deal with much more complex scenarios. To characterize human perception we have developed a Bayesian inference model that allows an unlimited number of signal sources to be considered: it is general enough to allow any discrete sequential cues, from any modality. The model uses a non-parametric prior, hence increased complexity of the signal does not necessitate more parameters. The model not only determines the most likely number of sources, but also specifies the source that each signal is associated with. The model gives an excellent fit to data from an auditory stream segregation experiment in which the pitch and presentation rate of pure tones determined the perceived number of sources.