Extending a North American English category learner to a non-standard variety: Categorizing vowels across speech styles in Glasgwegian English

Despite much research on the performance of distributional category learning models on Standard North American English (e.g., Feldman [2], deBoer and Kuhl [1], McMurray et al. [10], Vallabha et al. [16], and many others), statistical learning of vowel categories of other regional varieties remains vastly underaddressed in computational literature. This paper applies an unsupervised infinite mixture model (as developed in Feldman [2]) to vowels from a corpus of Glaswegian English sociolinguistic interviews. While originally developed for North American English vowels in carrier syllables devised by Hillenbrand et al. [4] to limit variation due to phonetic context, the distributional learner was also able to categorize vowels largely correctly across speech styles common to sociolinguistic interviews. This displays the ability of the distributional learner to operate relatively well on data with extensive overlap from running Glaswegian English speech, demonstrating that computational models of category acquisition can handle more complex inputs than minimal pair lists, and can be used with naturally-occurring speech from non-standard regional varieties.