Current Research Projects

Data on the perceptual magnet effect (Kuhl, 1991) indicate that perceptual space is shrunk in the neighborhood of vowel prototypes and expanded near category boundaries. We are modeling this pattern using a statistical model in which listeners need to recover the phonetic detail of a speaker's target production through a noisy speech signal. Because experienced listeners know that speakers are more likely to produce sounds near the centers of phonetic categories, they should bias their perception toward category centers for optimal accuracy. We have shown a close correspondence between model predictions and data from Iverson & Kuhl (1995), and we have also collected data showing that as predicted, listeners use category information to different degrees depending on the amount of noise in the speech signal.

Feldman and Griffiths (2007) - conference paper introducing the Bayesian model
Shi, Feldman, and Griffiths (2008) - conference paper describing an exemplar-based implementation of the model
Feldman, Griffiths, and Morgan (2009) - paper with a full description of the Bayesian model and the noise experiment

Infants begin to segment words from fluent speech at about the same time they acquire native language phonetic categories (between 6 and 12 months). A learner that simultaneously learns about both levels of structure might be able to use information about which sounds occur together in words when deciding whether sounds belong to the same or different phonetic categories. We are exploring how this type of information can provide useful constraints to guide phonetic category acquisition. Our simulations suggest that information from words can help disambiguate overlapping categories in cases where distributional information alone would not be sufficient. Experiments are currently underway to test whether adults and infants use information from words to constrain their interpretation of phonetic variability.

Feldman, Griffiths, and Morgan (2009) - conference paper describing the model