Modeling Phonetic Category Learning from Natural Acoustic Data

Languages make different distinctions regarding what constitutes a meaningful difference in sound, yet children quickly learn to distinguish between these sounds in their native language. A major question in the field of language acquisition regards what strategies children use to acquire knowledge of their language from linguistic input. Empirical research shows that children are able to track some statistical properties of their input (Saffran et al., 1996; Maye et al., 2002). One way to convert these statistical observations into linguistic knowledge is by tracking distributional regularities of the input, such as regularities in the frequency or duration of sounds in the child’s environment, a process known as distributional learning. This distribution of sounds is hypothesized to be important for phonetic category learning (Maye et al., 2008). Computational models help us explore how well linguistic structures can be learned from input given a set of statistical tools. Perhaps we wish to test how completely the phonetic categories of some language can be learned given information about the distribution of sounds in natural speech. Provided with this input, a model could be programmed with a learning algorithm taking advantage of these data, and its output (the categories recovered by the model) could be compared to the categories existing in the language. In practice, however, the input for such models is often simplified compared to the input available to children: acoustic vowel information is sampled from unrealistically parametric distributions whose parameters are taken from carefully enunciated words produced in laboratory settings, and lexical information providing context for these sounds is transcribed phonemically, rather than phonetically. These simplifications occur both due to technical limitations of the models and because realistic annotated data are expensive. Because the purpose of modeling language acquisition is to test how linguistic structures can be learned from input, it is important to ensure that these simplifications are not impacting model results in unexpected ways.