Computational Modeling of Assimilated Speech: Cross-Linguistic Evidence

Models of speech perception have shown divergent views concerning the mechanisms that underlie perception of assimilated speech. In the present study, we evaluated an existing probabilistic model for English place assimilation and conducted a series of simulations on voice assimilated speech in French. The model was trained on a realistic acoustic-phonetic data set of word-final assimilated stops. Our findings showed that the model accommodates the asymmetric assimilation pattern that exists in French, as well as the finding from a cross-modal priming study in French indicating strong regressive context effects for fully voiceassimilated segments. These results suggest that the perceptual system for speech learns to deal with surface variants of speech probabilistically, through experience of phonological variations that are available in the linguistic environment.