This work presents a phonologically and phonetically based description for Portuguese Beira Interior regional speech. In particular, the performance of a post-stressed vowel system is shown. We also describe the solutions adopted to build a phonological-phonetic inventory. We explore the view that Quantal Theory and Optimality Theory can support phonological representation sets of phone inventories and language models for speech processing could claim its frequency distribution providing a feasible alternative to the standard front-ends.
An interesting feature of our pattern models is that they relate phonological categories to the structure of phonetic variation along assured parameters. The sounds are considered by the fact that speech is perceived and produced in distinctive features terms and optimal centre of gravity regions are correlates of distinctive features. In our approach, based on models for speech categorical perception, those discrete-units correspond to categorical prototype regions and they tend to be maximally combined by Quantal Theory and feature a hierarchy issue too. The syllabic position constitutes another finer-grained phonological category. In fact, patterns of vowel variability depend on the contextual factors involved. Those patterns are accounted for assuming that vowels’ uninformed variability is ruled by the care involved in achieving a specific articulatory target, and that contextual variability is determined by both the post-stressed vowel articulatory requirements and the relative compatibility between the articulatory gestures for adjacent phones or silences.
The resulting phone inventory is supported by statistical measures obtained from the analysis of an oral corpus. The major goal is to present the allophones which are more or less perceived in speech continuum in order to give a code with which an algorithm will match in a language model of Portuguese speech processing, including regional variety. Our belief is that formalised knowledge of pronunciation variants will also be acquired by analysing very large amounts of data in the future and will ultimately contribute to improving Portuguese multi-pronunciation modelling, taking minority speech into account.