Dynamicity in Speech Perception
Emily J. Clare · TSpace (University of Toronto) · 2019
Human listeners dynamically adjust their perceptual category boundaries and cue weights following brief exposure to idiosyncratic speech (Norris et al., 2003; Idemaru and Holt, 2011). This dissertation presents five experiments, each exploring a different facet of when and how listeners recalibrate. The first three experiments use the classical perceptual learning paradigm, where listeners are exposed to speech exhibiting an idiosyncratic pattern, such as an atypically high VOT boundary between /t/ and /d/, and then tested on whether they adjusted their boundary according to that training. Experiment 1 used weaker evidence to train listeners on the pattern than previous studies, by using word-initial targets in training. Experiment 2 exposed listeners to a more complex pattern than previous studies, by using a second-order (i.e. context-dependent) idiosyncrasy. In both experiments, listeners succeeded at learning. Experiment 3 tested whether listeners make use of phonological category knowledge in learning these patterns. There was no evidence of learning. However, some methodological weaknesses made it difficult to conclude that listeners could not learn this type of pattern; further work is needed to assess this. The last two experiments looked at cue reweighting -- reliance on different acoustic cues to the same contrast. Experiment 4 tested the effects of noise on cue weighting. It confirmed previous work showing that noise affects cue weighting, and expanded beyond previous work in showing that different types of noise affect cue weighting differently. Noise especially suppressed reliance on cues incongruent in periodicity with the noise type, suggesting listeners may be sensitive to spectral characteristics of background noise in adaptive cue reweighting. Experiment 5 was the first test of unsupervised learning of cue reweighting based on exposure. Listeners were exposed to a contrast varying on along two acoustic dimensions where one was a more reliable cue than the other. The results suggested that listeners learned to reweight perceptual cues based on their exposure, supporting the hypothesis that listeners implicitly learn multi-dimensional statistical distributions in speech. Together, these experiments provide new data and contribute novel methodologies to further our understanding of the dynamic nature of adaptation in human speech perception.