Models of the phonetics/phonology interface have typically dealt with phonetic detail in a quite idealized way, for instance by placing exemplars in a one or two-dimensional idealized phonetic cue space. This paper explores a more phonetically realistic approach to modeling phonological cues building on a “Cue-based Features” account of perceptual cue trading and the interface between phonetic cues and phonological features. We train non-negative matrix factorization over large speech corpora to discover spectrotemporal patterns that are potential acoustic cues for phonological contrast, and then we use linear discriminant analysis to find combinations of cues that can be optimally used to specify the value of a phonological feature. We present a preliminary study of plosive voicing cues, and the resulting models reveal that the cues trade with each other across exemplars and that they encode information that is temporally spread beyond the segment’s acoustic boundaries.
Pfiffner et al. (Tue,) studied this question.