Deriving manner of articulation classes from phoneme co-occurrence frequencies

This paper looks into the possibility of deriving manner of articulation classes from phoneme cooccurrence frequencies. We show that distance measures based on phoneme co-occurrence frequencies group abstract manner of articulation classes that participate in phonological processes in spite of high acoustic and articulatory variability, e. g: obstruents, liquids, approximants, etc. The test language is Brazilian Portuguese, as represented by two large phonemicized corpora. Phonetically meaningful groups of phonemes emerge from the statistical analysis of co-occurrence frequencies of types and tokens of phoneme pairs. The techniques employed are cluster analysis and multidimensional scaling. Results show consistent groupings across corpora, counts, and statistical techniques. Besides consonants and vowels, co-occurrence frequencies satisfactorily separate the major classes, as well as some smaller manner of articulation subclasses.