Using the Distributional Statistics of Speech Sounds for Weighting and Integrating Acoustic Cues
Joseph Christopher Toscano, Bob McMurray · eScholarship (California Digital Library) · 2008
A great deal of behavioral evidence suggests that infants can use distributional statistics to learn speech sound categories.Recently, a number of computational approaches have demonstrated the feasibility of statistical learning by showing that the distributional statistics of linguistically-relevant acoustic cues can be learned in an unsupervised way.However, speakers and listeners use a large number of acoustic cues to distinguish phonetic categories, and it is not clear how multiple cues are combined during perception.We propose a model of speech sound category acquisition that learns the distributions of multiple cues that lie along the same dimension and combines them.We demonstrate that the model is able to account for trading relations between cues (an indicator of the size of the effect of each cue) for word-initial voicing contrasts in English.