Learning segments from waveforms: A statistical model of early phonological acquisition

Abstract

We consider the problem of early phonological acquisition within a statistical learning framework. Given a set of waveforms, our goal is to output a set of models for phonological categories as well as representations of the utterances in terms of these categories. The basic assumption of our model is that the learning of sound categories neither precedes nor follows segmentation of the waveform, but improves together with segmentation in an iterative manner. In statistical learning terms, segmentation is regarded as ‘‘missing data,’’ while category models are treated as parameters of interest. Our algorithm iterates over two basic steps: First, given a set of category models, we compute a mixture of segmentations for each utterance; Second, we improve the category models with unlabeled segments using unsupervised learning. Starting from an acoustic segmentation, each iteration produces an initial estimate of the parameters for the next iteration, until the algorithm converges to a set of sublexical models and a set of segmentations for each utterance. The results of running this algorithm on TIMIT and motherese data suggest that it approximately identifies segment-sized units and their associated categories in a completely unsupervised manner.