Preparing words in speech production is normally a fast and accurate process. We generate them two or three per second in fluent conversation; and overtly naming a clear picture of an object can easily be initiated within 600 msec after picture onset. The underlying process, however, is exceedingly complex. The theory reviewed in this target article analyzes this process as staged and feedforward. After a first stage of conceptual preparation, word generation proceeds through lexical selection, morphological and phonological encoding, phonetic encoding, and articulation itself. In addition, the speaker exerts some degree of output control, by monitoring of self-produced internal and overt speech. The core of the theory, ranging from lexical selection to the initiation of phonetic encoding, is captured in a computational model, called weaver++. Both the theory and the computational model have been developed in interaction with reaction time experiments, particularly in picture naming or related word production paradigms, with the aim of accounting for the real-time processing in normal word production. A comprehensive review of theory, model, and experiments is presented. The model can handle some of the main observations in the domain of speech errors (the major empirical domain for most other theories of lexical access), and the theory opens new ways of approaching the cerebral organization of speech production by way of high-temporal-resolution imaging.

Levelt et al. describe a model of speech production in which lemma access is achieved via input from nondecompositional conceptual representations. They claim that existing decompositional theories are unable to account for lexical retrieval because of the so-called hyperonym problem. However, existing decompositional models have solved a formally equivalent problem.

Though weaver has knowledge that gets activated by words and pictures, it is incapable of responding appropriately to these words and pictures as task demands are varied. This is because it has a most severe case of attention deficit disorder. Indeed, it has no attention at all. I discuss the very complex attention demands of the tasks given to weaver.

Language production and comprehension are intimately interrelated; and models of production and comprehension should, we argue, be constrained by common architectural guidelines. Levelt et al.'s target article adopts as guiding principle Ockham's razor: the best model of production is the simplest one. We recommend adoption of the same principle in comprehension, with consequent simplification of some well-known types of models.

Levelt, Roelofs & Meyer present a comprehensive and sophisticated theory of lexical access in production, but we question its reliance on binding-by-checking as opposed to binding-by-timing and we discuss how the timing of retrieval events is a major factor in both correct and errorful production.

Things should be made as simple as possible, but not simpler. –Albert Einstein

The theory of lexical access in speech production reported by Levelt, Roelofs & Meyer is exciting, well-described and well-organized, but because it relies mainly on the principle of simplicity (Ockham's razor), I argue that it might not be true. In particular, I suggest that overapplying this principle is wrong.

Any complete theory of lexical access in production must address how words are produced in prosodic contexts. Levelt, Roelofs & Meyer make some progress on this point: for example, they discuss resyllabification in multiword utterances. I present work demonstrating that word articulation takes into account overall prosodic context. This research supports Levelt et al.'s hypothesized separation between metrical and segmental information.

The theory of lexical selection presented by Levelt, Roelofs & Meyer addresses the mechanisms of semantic activation that lead to the selection of isolated words. The theory does not appear to extend naturally to the referential use of words (particularly pronouns) in coherent discourse. A more complete theory of lexical selection has to consider the semantics of discourse as well as lexical semantics.

I examine four core aspects of weaver++. The necessity for lemmas is often overstated. A model can incorporate interaction between levels without feedback connections between them. There is some evidence supporting the absence of inhibition in the model. Connectionist modelling avoids the necessity of a nondecompositional semantics apparently required by the hypernym problem.

The use of lexical concepts in Levelt et al.'s model requires further refinement with regard to syntactic factors in lexical choice, the prevention of pleonasm, and the representation of near-synonyms within and across languages.

Levelt et al. attempt to “model their theory” with weaver++. Modeling theories requires a model theory. The time is ripe for a methodology for building, testing, and evaluating computational models. We propose a tentative, five-step framework for tackling this problem, within which we discuss the potential strengths and weaknesses of Levelt et al.'s modeling approach.

According to the theory of lexical access presented by Levelt, Roelofs & Meyer, processing of semantic–syntactic information (i.e., lemma information) and phonological information (i.e., lexeme information) proceeds in a strictly discrete, serial manner. We will evaluate this claim in light of recent evidence from the literature and unpublished findings from our laboratory.

The weaver++ model discussed by Levelt et al. assumes incremental encoding and articulation following complete encoding. However, many of the response latency results can also be accounted for by assuming incremental articulation. Another temporal variable, initial segment duration, can distinguish weaver++'s incremental encoding account from the incremental articulation account.

Lexemes supposedly represent phonological but not grammatical information. Phonological word substitutions pose problems for this account because the target and error almost always come from the same grammatical class. This grammatical congruency effect can be explained within the Levelt et al. lexicon given that (1) lexemes are organized according to phonological similarity and (2) lexemes from the same grammatical category share phonological properties.

(1) Reaction time (RT) studies give only a partial picture of language processing, hence it may be risky to use the output of the computational model to inspire neurophysiological investigations instead of seeking further neurophysiological data to adjust the RT based theory. (2) There is neurophysiological evidence for differences in the cortical representation of different word categories; this could be integrated into a future version of the Levelt model. (3) EEG/MEG coherence analysis allows the monitoring of synchronous electrical activity in large groups of neurons in the cortex; this is especially interesting for activation based network models.

The insistence on strict seriality and the proscription of feedback in phonological encoding place counterproductive limitations on the theory and weaver++ model. Parsimony cannot be stipulated as a property of the language system itself. Lifting the methodological prohibition on feedback would allow free exploration of its functionality in relation to the substantial content of this major research program.

The following questions are addressed concerning how a theory of lexical access can be realized in the brain: (1) Can a brainlike device function without inhibitory mechanisms? (2) Where in the brain can one expect to find processes underlying access to word semantics, syntactic word properties, phonological word forms, and their phonetic gestures? (3) If large neuron ensembles are the basis of such processes, how can one expect these populations to be connected? (4) In particular, how could one-way, reciprocal, and numbered connections be realized? and, (5) How can a neuroscientific approach for multiple access to the same word in the course of the production of a sentence?

We discuss our concerns associated with three assumptions upon which the model of Levelt, Roelofs & Meyer is based: assumed generalisability of decontextualised experimental programs, assumed highly modular architecture of the language production systems, and assumed symbolic computations within the language production system. We suggest that these assumptions are problematic and require further justification.

Recent aphasiological findings, not mentioned in the target article, have been accounted for by Levelt et al.'s theory and have, in turn, provided it with empirical support and new leads. This interaction is especially promising in the domain of complex word retrieval. Examples of particular categories of compounds are discussed.

Despite several positive features, such as extensive theoretical and empirical scope, aspects of Levelt, Roelofs & Meyer's theory can be challenged on theoretical grounds (inconsistent principles for phonetic versus phonological syllables, use of sophisticated homunculi, underspecification, and lack of principled motivation) and empirical grounds (failed predictions regarding effects of syllable frequency and incompatibility with observed effects of syllable structure).