Mayfest 2026: Production

Many current psycholinguistic theories of speech production agree on two general principles. Consistent with foundational work in the field, theories postulate the existence of distinct processes mapping between distinct types of linguistic representations (e.g., semantic representations are used to retrieve lexical representations; these lexical representations are used to retrieve speech sound representations from memory). After considerable debate in the 1990s, there is emerging consensus that there is cascading activation: a process can generate multiple co-activated representations as output, such that each of these representations simultaneously influence subsequent processes (e.g., multiple lexical items can be retrieved, and each of these representations simultaneously impacts the retrieval of speech sounds from memory). Understanding this general class of theories has proven to be quite complex. Depending on detailed assumptions about the structure of cognitive processes, cascading activation can yield quantitatively and qualitatively different behavioral effects. I'll critically examine how cascading activation has been deployed to explain phenomena at multiple levels of structure, and lay out the key challenges for developing such accounts.

Language production is often described as the process of converting a preverbal message into an utterance plan that guides speech. In practice, however, speakers do not always have a well-formed message before utterance planning begins, and time pressures can lead them to start speaking before the message is fully specified. In this talk, I explore the role of message uncertainty in shaping real-time production strategies. I present a line of experiments investigating how speakers use incremental planning and flexible word order to manage uncertainty, and examine how cross-linguistic differences influence the strategies they adopt and the time course of production. Together, this work highlights speakers’ flexibility in adapting planning strategies to the production context, and carries implications for interactive theories of cognition and language production.

In the classical approach, prosodic structure influences speech production through patterns of structure-dependent differences in phonological and/or phonetic targets. Many other grammatical and ostensibly extra-grammatical factors influence post-lexical sound patterns, including speech rate, probability/informativity dynamics, and the usage frequency of particular morphemes. This has led to suggestions that all of these influences could be unified as affecting sound production through the medium of speech planning: the greater the difficulty of retrieving or planning the implementation of upcoming material, the longer it takes to do so, and this has consequences for context-dependent realizations of sounds. This talk calls attention to several properties of post-lexical sound patterns worth considering in this ongoing debate. First, many of the factors enumerated above influence speech behavior specifically through the medium of changing the temporal separation between onsets of units in the speech stream; some of these factors may influence speech behavior *only* through that medium. Second, the sound patterns generated by this mosaic of factors will tend to be optimized for perceptual chunking through Gestalt principles: low-level acoustic cues to grouping in speech production will tend to converge on the units necessary for a listener to infer the structure generated by a speaker. Third, neither Gestalt-compliant production patterns nor their perceptual consequences are specific to speech, to language, or to any perceptual modality: production and/or perception display analogous properties in music, general audition, vision, and primate limb movement, for instance. I outline some parallels and differences between musical and prosodic grouping in this regard, as well as observations on the representations that figure in setting linguistic objects to musical structures. One conclusion is that any notion of 'planning window' intended to subsume the effects of prosodic structure will need to be both specific enough and stable enough to feed memory, perception, and cross-domain coordination in much the way that traditional descriptions of prosody predict.

While discussing the goals of theoretical phonology, John McCarthy once wrote, "if the representations are right, then the rules will follow." This suggests that understanding how speech sounds are represented precedes any understanding of the computations that are performed on said representations. In this talk, I take a different approach where I argue that the relationship between these two aspects of the speech production "program" are instead interdependent.

As a primary case study, I compare the "data structures" and "algorithms" used in symbolic theories of phonology/phonetics inspired by classic computation with gestural theories of phonology/phonetics inspired by the dynamical systems framework. In the data structure space, I use techniques from model theory to show how the multi-linear gesture-based representations used in the latter can be freely translated into the symbolic string structures used in the former, and back, with limited computational resources and without loss of information. In the algorithms space, I show how the general structure of a phonetics-phonology interface function can be described in terms of the formal types of its subfunctions. As a consequence of this symbolic approach to the interface, certain arguments in favor of gestural representations are weakened. Overall, these findings suggest that our speech production programs can be made more simple by understanding the two-way relationship between data structures and algorithms.

Several factors create interference in language production, and such interference is thought to be resolved through cognitive control. However, the mechanisms underlying the implementation of control in  language production have remained elusive. In this talk, I will present a comprehensive computational model of interference and its resolution in language production, and show that the model is compatible with behavioral, electrophysiological, neuroimaging, and neuropsychological data.

Aphasia provides a powerful window into the nature of human language and its neurobiology. The syndrome of agrammatism has been studied intensively for decades and dominates the literature on syntactic deficits in aphasia. However, its counterpart in fluent aphasia, paragrammatism, has long been identified yet little is known. Here we provide an overview of the current status of paragrammatism, including arguments against common objections to paragrammatism as an independent syntactic disorder in aphasia. Understanding paragrammatism as the primary aphasic disorder involving impaired hierarchical syntactic processing converges with psycholinguistic models of sentence production and an emerging consensus regarding the middle-posterior temporal lobe as the main hub for hierarchical syntactic representations in the brain.
 

Everyone has limited cognitive resources.  Working memory is an obvious limitation.  Everyone also has limited linguistic resources. Vocabulary is an obvious limitation. Cognitive and linguistic limitations seem to be clearest when individuals are very young, or very old, or speaking another language.  To what extent are the limitations syntactic?  And how can we decide?  I present production and comprehension data to suggest that people seem to know pretty much everything no later than 40 months of age and maybe no later than 24 months.

At the top of the hierarchy of speech production, Levelt posits a conceptualizer, a component responsible for feeding the production system with messages “tuned to the target language and to the momentary informational needs of the addressee” (Levelt, 1996). Levelt expressed some sympathy for Slobin’s (1987) notion of thinking for speaking, according to which speakers translate pre‑linguistic thoughts, assumed to be in a universal format, into the categories provided by their particular languages. However, it remains unclear how such a translation process could actually work.  

In this talk, I argue that this process is challenging, to say the least, and that positing a distinct stage of thinking for speaking does little to illuminate how the conceptualizer prepares its messages. Once thinking for speaking is set aside, two remaining options present themselves: (a) speakers move directly from a Language‑of‑Thought (LoT) representation to speech, or (b) speakers already think in the categories of their own languages. I argue that the second option is preferable. I further explain how it is possible for subjects to think in lexical concepts—that is, concepts expressed by monomorphemic forms in their languages. In particular, I suggest that subjects think in words.