LARIGALDIE, NATHANAEL,CHRISTOPHE,RODO (2021) A nonparametric Bayesian clustering approach to auditory perception. Doctoral thesis, Durham University.
|PDF - Accepted Version|
Models of perceptual grouping are usually using verbal, poorly accurate predictions. As of late, probabilistic models are being used more and more to create more stringent descriptions of the underlying mechanisms, and quantitative predictions. This thesis presents a nonparametric Bayesian clustering algorithm applied to Auditory Scene Analysis (ASA), along with several validations from the classical literature on the subject, and experiments using a new paradigm in different experimental settings. Grouping/segregation processes in ASA, and therefore in the model, follow similar Gestalt principles as in the more studied visual field: the more tones are similar, the more they tend to be clustered in a single auditory stream, and conversely. The first study focuses on a mathematical description of the clustering algorithm and on its validation on well-known studies from the field. A new paradigm has been used to create situations where 3 simultaneous streams could be reached by increasing the distance in frequencies between rapidly played tones, as predicted by the classical ASA model and our own. Results were in line with the hypotheses. The second study expands on the first one and uses qualitative predictions from the clustering algorithm to observe stream segregations using increasing differences in several dimensions at once in two experiments using the same paradigm: namely, frequency and spatial distance in the first one, frequency and timbre in the second one. Results presented an unexpected pattern, suggesting a stronger influence of attention as initially supposed. The third study explored the influence of attention on the stream formation process in the same paradigm, by adding specific attentional instructions to participants. Results suggest a possible limitation to 2 simultaneous attentional streams: the foreground stream, and the background one where all tones are clustered. Overall, while the model was only used to create qualitative predictions, those were useful enough to guide experiments with impactful results.
|Item Type:||Thesis (Doctoral)|
|Award:||Doctor of Philosophy|
|Keywords:||perception; gestalt psychology; auditory scene analysis; causal inference; bayesian inference|
|Faculty and Department:||Faculty of Science > Psychology, Department of|
|Copyright:||Copyright of this thesis is held by the author|
|Deposited On:||05 May 2021 09:45|