UC Santa Cruz

Eliminativist usage-based and classical generative speech production models differ fundamentally in two respects: (i) the nature of speech representations and (ii) the locus of linguistic knowledge. In this dissertation, I ask what levels of abstraction and detail are present in speech representations and whether a lexicon of detailed speech memories is sufficient to explain a speaker’s linguistic knowledge of their language. I investigate these questions using phonetic, phonological, experimental, and computational evidence from two understudied Pakistani languages, Punjabi and Mankiyali. First, I present evidence from Punjabi that lexical representations can be completely covert, containing information that is not present in any experienced realization of a lexical item. I then demonstrate that, given an appropriately structured learning architecture, a learner can acquire covert speech representations efficiently. At the same time, a pilot study provides preliminary evidence that talker-specific phonetic and socially indexed information from recent linguistic experience is retained and reflected in subsequent production. Crucially, the same lexical items used to motivate covert representations also appear to be represented at a richly detailed episodic level. Turning to the locus of linguistic knowledge, I present acoustic data from a production experiment in Mankiyali, showing that the neutralization of a phonological contrast can be phonetically complete. I then use simulations of exemplar accumulation to examine whether a lexicon of detailed speech memories can account for the stable realization of such patterns. In these simulations, once the production and similarity biases define a stable target for an acoustic parameter, introducing exemplars with different values cannot by itself shift that target. A persistent shift in a category's optimal acoustic realization instead seems to require recalibration of the biases shaping production, motivating higher-order linguistic knowledge that operates beyond stored speech memories. Taken together, these findings challenge both eliminativist usage-based models, which reduce linguistic knowledge to accumulated exemplars, and classical generative models, which exclude fine-grained phonetic and social detail from linguistic competence. I argue for a hybrid architecture in which speakers retain detailed episodic memories, maintain abstract categories that may diverge substantially from experience, and possess higher-order knowledge that regulates their joint influence on speech production.