Abstract: Speech research in several past decades has been defined by two major themes. The first one is the search for invariance, that is, a set of physical attributes that map to phonological entities (phonemes or distinctive features) independent of the context these entities occur in (examples are the works by Stevens, Liberman, Fowler, Perkell, among many others). The relative failure to do so brought into prominence the pervasive variability of speech, which is the other issue that has been devoted much attention to. The two issues are inextricably linked: speech variation has been seen as an obstacle to defining invariant properties of phonological entities. This even prompted a position that the goal of finding invariance in speech is misguided and no invariant features can be located in the physical signal (Lindblom, 1990). On the other hand, there have also been attempts to harness variability as an aid in categorical perception (McMurray and Jongman, 2011). There is, therefore, a continued need to address the apparent disparity between the search for invariance and inherent variability of speech. The question I ask is whether and how contrastive features of place and voicing are maintained under co-variation of acoustic cues of duration and amplitude in American English stops. Drawing from a large spontaneous speech database, the patterns of co-variation of closure and burst cues are examined, as they are expressed in the direction and magnitude of slope of the linear regression line that defines the relationship between a pair of acoustic cues. One of the findings is that the differences between values of contrastive features (voiced vs voiceless, or labial vs coronal vs velar) are often reflected in either the direction of the association between two cues or its degree. This suggests that quantitative indices of the regression line between two cues may serve as a relational invariant for phonological properties of place or voicing. This is akin to the role locus equations have been proposed to play in consonantal place identification (Sussman et al. 1991, Iskarous et al. 2010). These findings suggest that that, despite previous inconclusive attempts at invariance, the speech signal may yet hold information needed for recovery of contrastive features in the form of relational invariants.
The intriguing possibility is that variability, instead of being an obstacle to defining invariant attributes, may be the key to this search.