Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Speech Segmentation and Spoken Word Recognition

Abstract. This fundamental research monograph investigates how listeners find lexical boundaries and recognize competing word forms in a continuous acoustic signal. A critical narrative synthesis of original psycholinguistic experiments, developmental research, computational models, and neuroscientific findings reconstructs the historical emergence of the problem and analyzes its principal methods, including word spotting, gating, cross-modal priming, infant familiarization, and eye tracking. The chapters compare cohort theory, TRACE, Shortlist, Shortlist B, and distributed models; examine coarticulation, phonotactics, stress, prosody, frequency, and predictive context; and assess infant statistical learning, cross-language variation, speech in noise, adaptation, and the neural time course of recognition. The evidence supports graded, incremental competition and the integration of multiple boundary cues whose reliability changes across conditions. Model architecture, ecological generalization, developmental diversity, and mechanistic interpretation remain active research questions. This is a critical narrative monograph grounded in published literature; it reports no new experiment or meta-analysis.

Keywords: speech segmentation; spoken word recognition; lexical competition; spoken lexical access; word boundaries; speech perception; prosodic segmentation; phonotactics; coarticulation; TRACE model; Shortlist B; cohort model; statistical learning; infant language acquisition; visual-world eye tracking; incremental comprehension; speech-in-noise; phonetic detail

Scientific Definition and Explanatory Scope

Speech segmentation is the process through which a listener infers the locations and identities of word-sized units in an unfolding acoustic stream. Spoken word recognition is the activation, evaluation, and selection of lexical representations that match that stream. These operations are coupled: the identification of a plausible word helps establish its boundaries, while a proposed boundary changes which lexical candidates can be considered. The psychological object is therefore a continuously updated mapping from acoustically variable speech to temporally organized lexical hypotheses. It includes temporal competition among words beginning at the same and different positions, not merely the retrospective insertion of spaces into a transcript Norris (1994) Norris et al. (1995).

The relevant input contains overlapping articulatory gestures, reduced syllables, variable duration, coarticulation, pauses of different linguistic significance, and fluctuations in prominence. A word is a linguistic and lexical unit; an acoustic discontinuity is an observable event. Their alignment is probabilistic. Listeners can use segmental transitions, phonotactic regularities, rhythm, word frequency, and discourse constraints to infer boundaries and recognize forms, but the reliability of each source varies with language, speaker, and listening condition. Scientific explanation requires specifying the information available at a given moment rather than attributing recognition to an idealized, fully segmented sequence Cutler et al. (1986) Mattys et al. (2005).

This monograph critically synthesizes original behavioral experiments, infant learning studies, computational models, and selected neuroscientific and methodological accounts. It considers the evidence for incremental lexical activation, the coordination of boundary cues with lexical competition, and changes across development and experience. Phonemic categorization, written word recognition, syntactic parsing, and broad language comprehension enter only where they constrain the target mechanism. No experiments, participants, or pooled effects are introduced here; the study is a source-based narrative synthesis whose claims inherit the design and access limitations of the cited investigations Allopenna et al. (1998) Norris & McQueen (2008).

Why Fluent Speech Creates a Segmentation Problem

In print, white space supplies an unusually explicit visual convention for word separation. Continuous speech has no equivalent general-purpose boundary marker. The waveform is shaped by movement of the tongue, lips, jaw, larynx, and respiratory system, and movements associated with neighboring sounds overlap. Silence can reflect a phrase break, hesitation, stop closure, or environmental event, while genuine word boundaries commonly pass without silence. Consequently, a listener cannot solve segmentation merely by locating energy minima or pauses. The difficulty is amplified when a shorter word is embedded within a longer lexical form McQueen et al. (1995) Salverda et al. (2003).

Many candidate segmentations are locally compatible with partial input. At the beginning of an utterance, the listener may activate one set of words sharing an onset; later sound can favor a different word or reveal that an apparent onset belonged inside a previous word. Lexical candidates can also overlap in time, including sequences that correspond to the end of one word and the beginning of another. This creates a competition problem whose solution may depend on global coverage of the utterance and the plausibility of residual material. A fixed left-to-right commitment at the first familiar syllable would frequently be premature Norris (1994) Norris et al. (1997).

The distinction between possible and actual word boundaries is central. A listener can exploit regularities associated with word onsets without claiming that every onset-like acoustic feature marks an independent word. Likewise, hearing a familiar sound sequence does not show that it has been selected as a lexical unit. Identification must survive competing alternatives and subsequent evidence. Experiments that compare embedded words with matched acoustic contexts expose these differences. Salverda and colleagues manipulated the origin of ostensibly identical phonemic sequences and found effects of prosodic realization on lexical interpretation, making boundary inference an empirical matter of graded cue use Salverda et al. (2003).

Historical Formation of the Research Problem

The contemporary program developed as psycholinguists moved from recognizing isolated words toward explaining fluent comprehension. Cross-linguistic work in the 1980s showed that speakers of different languages may rely on different rhythmic segmentation routines. Cutler, Mehler, Norris, and Segui compared French and English listeners and found evidence for syllabic segmentation among French speakers that did not generalize to English listeners. That result made the structure of the acquired language an explanatory variable rather than treating segmentation as one universal auditory routine Cutler et al. (1986).

Computational work then transformed lexical access from a serial dictionary lookup into a dynamic problem. The TRACE model of McClelland and Elman in 1986 used interacting units to represent evolving phonetic and lexical interpretations. Marslen-Wilson's 1987 treatment distinguished access, selection, and integration while refining cohort-based accounts. Both lines acknowledged that information arriving over time changes the status of multiple candidates, but they differed in representational commitments and the permitted influence of higher-level knowledge on early selection McClelland & Elman (1986) Marslen-Wilson (1987).

During the 1990s, segmentation and recognition became increasingly inseparable in formal models. Norris's Shortlist explicitly addressed continuous speech and overlap among lexical hypotheses; subsequent experiments examined metrical segmentation, lexical competition, and the Possible-Word Constraint. Eye-movement research later made it possible to observe candidate activation while listeners processed spoken instructions. Developmental studies established another research path, showing how infants can begin discovering word forms without a mature lexicon. The modern question concerns how these complementary sources of constraint are learned and combined during real-time listening Norris (1994) Norris et al. (1997) Allopenna et al. (1998) Saffran et al. (1996).

Constructs, Time Scales, and Causal Questions

Three constructs require operational separation. Boundary evidence concerns the probability that a location in the sound stream begins or ends a lexical unit. Candidate activation concerns the degree to which a lexical representation fits information available so far. Lexical selection concerns the comparative advantage of one candidate over alternatives as processing proceeds. These constructs may covary, but they do not have identical observable signatures. A listener can register an onset cue before selecting a word, and several words can remain active after one has become the most plausible interpretation Allopenna et al. (1998) Norris et al. (1995).

Acoustic time, processing time, and developmental time must also be distinguished. Acoustic time tracks physical information in a recorded signal. Processing time concerns when an experimental response begins to reflect that information, including delays associated with planning an eye movement or a motor response. Developmental time concerns changes in learned knowledge across months or years. An infant's successful familiarization to a sound sequence demonstrates sensitivity to specified relations under that procedure; it does not automatically imply the rich semantic recognition exhibited by fluent adults Saffran et al. (1996) Jusczyk et al. (1999) Fernald et al. (2008).

The principal causal questions ask what cues listeners can use, how those cues are weighted, how competing word hypotheses are resolved, and which learning experiences create the relevant sensitivities. Experimental control can isolate a feature such as segment duration or word stress, but generalization requires a broader sample of words, voices, phonetic environments, and communicative settings. A computational model may fit behavioral responses by several parameter arrangements, so identifying a plausible algorithm does not by itself establish its neural realization. Throughout the monograph, observed task effects and proposed internal mechanisms remain separately attributed Shatzman & McQueen (2006) Norris & McQueen (2008).

Experimental Paradigms and Interpretive Limits

Word spotting asks participants to detect familiar words embedded in longer sequences and measures accuracy or detection latency. Its strength lies in exposing whether candidate words survive competition and boundary constraints. Its limitation lies in the strategic search demanded by the instruction itself. Gating progressively reveals a longer portion of recorded speech, allowing researchers to estimate when particular interpretations become possible, but repeated fragments and explicit judgments can alter listening strategy. Neither paradigm offers a literal readout of the entire unconscious candidate set Norris et al. (1997) Allopenna et al. (1998).

Cross-modal priming tests whether an auditory fragment influences subsequent decisions about a visually presented word. It can reveal lexical activation under a specified delay, though effects depend on visual decision demands and prime preparation. Eye tracking in the visual-world paradigm follows fixation probabilities to pictured referents during auditory instructions. Allopenna, Magnuson, and Tanenhaus observed transient attention to both onset-sharing and rhyme-related competitors; those patterns supported continuous mapping, but display composition and oculomotor latency remain important controls. Visual search environments provide useful proxies for lexical dynamics rather than a complete model of unrestricted conversation Allopenna et al. (1998) Dahan et al. (2001).

Infant head-turn preference, familiarization, and looking-while-listening procedures add developmentally appropriate measurements. Preferential listening can establish discrimination or recognition of trained sound sequences, while looking patterns offer a measure of online referential interpretation once words are associated with objects. Computational simulation contributes a separate test by making explicit predictions about competition and word boundary recovery. Convergence is strongest when experiments manipulate the same hypothesized cue in different tasks, and when models predict held-out input rather than explaining only the original stimuli Jusczyk et al. (1999) Fernald et al. (2008) Norris & McQueen (2008).

Coarticulation, Segment Duration, and Subphonemic Evidence

Coarticulation distributes information about one segment across neighboring intervals. Anticipatory gestures can begin before an intended boundary, and perseverative effects can persist beyond it. The resulting signal contains systematic local detail rather than a sequence of independently produced phonemes. Segmentation therefore depends in part on how the listener interprets such temporal relationships. In Mattys's cross-modal fragment priming experiments, coarticulatory continuity affected segmentation in clear speech, whereas stress information became more consequential when noise changed access to the detailed signal Mattys (2004).

Duration can distinguish interpretations even where phoneme labels coincide. Salverda, Dahan, and McQueen contrasted embedded monosyllables with the matching beginnings of longer words and found that prosodic realization modulated fleeting recognition of the short competitor. Shatzman and McQueen showed that listeners learning novel word forms used syllable duration in deciding between monosyllabic and longer alternatives. These findings indicate that stored lexical forms and general expectations about word duration participate in recognition before the utterance has been fully heard; neither result implies an invariant temporal boundary threshold Salverda et al. (2003) Shatzman & McQueen (2006).

Fine-grained segment evidence also modulates lexical competition. In an eye-tracking study with Dutch stimuli, longer and shorter realizations of a preceding fricative shifted gaze toward different lexical alternatives. That manipulation connects continuous acoustic variation to candidate activation without requiring categorical changes in transcription. An experimental design must nevertheless consider neighboring vowels, speaking rate, prosodic position, and the particular lexical sets on display. The scientifically relevant unit is a joint probability of form and context, and the same number of milliseconds can signal different structures across speakers or speaking styles Shatzman & McQueen (2006) Dahan et al. (2001).

Phonotactic Regularities and the Possible-Word Constraint

Languages restrict which sound sequences occur as syllables and words, though those restrictions vary in strength, dialect, and productivity. Such phonotactic knowledge can help listeners evaluate whether a proposed leftover portion of speech could form a word. A candidate that accounts for one stretch of input may become less plausible if accepting it strands an impossible residue. Norris, McQueen, Cutler, and Butterfield formalized this intuition as the Possible-Word Constraint and tested it by embedding real words within sequences that left viable versus nonviable residual material Norris et al. (1997).

The Possible-Word Constraint is especially informative because it separates finding a familiar sequence from obtaining a coherent segmentation of the larger stream. Participants were less successful at spotting an embedded target when its selection left a lone consonant than when the neighboring residue could constitute an allowable word. The model therefore treated the viability of the surrounding material as a penalty on candidates, a form of constraint on competition. The experiment does not establish that the human mind runs an explicit dictionary check on every residue; it establishes a pattern that a candidate mechanism must explain Norris et al. (1997).

Phonotactic expectations interact with acquired vocabulary. An unfamiliar sequence may be legal without being familiar, and word frequency changes what candidates become prominent once the listener entertains it. Conversely, a well-learned word can survive partial mismatch when the context makes the mismatch explicable. The general issue is how constraints influence activation at the moment relevant information becomes available. Overreliance on permissible sound patterns alone would underpredict recognition of rare proper names and borrowed words, while a system that ignores them would retain too many incompatible segmentations Norris & McQueen (2008) Gow (2001).

Rhythm, Stress, and Language-Specific Segmentation

Rhythm supplies probabilistic expectations about where words begin. Classic French–English comparisons found that French listeners displayed syllable-linked segmentation behavior, whereas English listeners did not show the same procedure under the tested conditions. These observations challenged accounts that assume one fixed sublexical unit for all languages. A listener's segmentation strategy emerges from regularities in the language or languages experienced, including the relationship between syllables, stress, and likely lexical onsets Cutler et al. (1986) Cutler et al. (1992).

In English, strong syllables often provide useful evidence for word onsets, but stress alone cannot delimit every word. English includes weak-syllable initial words, unstressed grammatical words, and contexts in which acoustic reduction obscures textbook stress patterns. Jusczyk, Houston, and Newsome found that English-learning infants initially performed better on strong–weak word patterns and later improved at recovering weak–strong words. The developmental progression supports learning to combine multiple sources of boundary evidence rather than a permanent hardwired rule that all words start on stressed syllables Jusczyk et al. (1999).

Bilingual findings add complexity to the relation between language experience and segmentation. Cutler and collaborators investigated bilingual listeners whose speech experience involved different rhythmic patterns, emphasizing that segmentation routines could remain strongly language dependent. Those experiments belong to a particular historical debate and should be interpreted with attention to bilingual proficiency, testing language, and stimulus selection. More broadly, multilingual listeners need to infer which expectations fit the currently heard language while managing overlapping phonological and lexical possibilities. The evidence supports flexible language-conditioned constraints, with considerable room for direct tests of switching and exposure effects Cutler et al. (1992) Christiansen et al. (1998).

Prosodic Boundaries and Multiword Grouping

Prosody organizes spoken language through patterns of pitch, timing, prominence, and phrasing. Stronger phrase boundaries may coincide with word boundaries, while many individual word boundaries lack a separate prosodic break. A plausible segmentation account must therefore assign context-sensitive evidential weight to prosody. Shukla, Nespor, and Mehler showed that syllable-transition regularities and prosodic grouping can interact: sequences with favorable statistics were less likely to be recognized as wordlike units when they crossed prosodic constituents Shukla et al. (2007).

Prosodic evidence also changes the status of embedded lexical alternatives. When an acoustic sequence could denote a short word or the beginning of a longer word, its duration and phrasing can increase one interpretation's relative plausibility. Eye movements reveal these effects as transient changes in attention to competitors. Such findings connect lexical representations with phonetic realization and learned expectations about the shapes of utterances. They do not imply that prosodic phrasing uniquely determines lexical identity; the same phrase contour accommodates many distinct word sequences Salverda et al. (2003) Shatzman & McQueen (2006).

Researchers must distinguish prosodic structure from grammatical analysis. An intonational phrase may aid segmentation while also reflecting information structure, communicative intention, or syntactic planning. If a manipulation simultaneously changes duration, pitch, pauses, and semantic plausibility, an observed effect cannot be attributed to one isolated cue. Controlled resynthesis, matched recordings, and factorial designs can determine which dimensions carry relevant information. The explanatory value of prosody lies in its predictable relationship to the emerging speech interpretation, which becomes more or less useful as competing cues and task demands change Mattys et al. (2005) Shukla et al. (2007).

Hierarchy, Reliability, and Integration of Boundary Cues

Connected speech offers several partly redundant and sometimes conflicting sources of information. Lexical knowledge can suggest which sound sequences form words; local acoustic detail may signal coarticulation across or within a boundary; stress and prosodic grouping supply broader structural expectations. Mattys, White, and Melhorn directly examined conflicts among cues and reported a hierarchical organization in which lexical information generally had priority under their conditions, with segmental and prosodic information receiving different weights. The hierarchy was sensitive to the interpretive circumstances and quality of the input Mattys et al. (2005).

A hierarchy should be interpreted as a pattern of experimentally measured preference, not an immutable serial processing pipeline. Acoustic degradation can reduce the reliability of fine detail and increase the relative usefulness of rhythmic or contextual knowledge. Conversely, a rare unfamiliar word may lack strong lexical support while revealing a clear phonotactic boundary. This distinction matters because a description such as bottom-up versus top-down can conceal differences between the arrival of information, its statistical reliability, and the control exerted by instructions or expectations. These variables should be independently manipulated wherever feasible Mattys (2004) Mattys et al. (2012).

The integrated inference problem can be expressed without committing to one implementation: at each time point, the listener evaluates alternative lexical segmentations in light of the accumulating signal and prior learned constraints. The research program then asks how quickly probabilities change, which alternatives inhibit one another, and whether abstract phoneme categories or distributed acoustic representations are required. Statistical models may formalize uncertainty, while activation models formalize temporal competition. The central empirical criterion is prediction of new listening behavior under controlled cue conflict and variation, beyond the fitting dataset Norris & McQueen (2008) Christiansen et al. (1998).

Cohort Activation and the Temporal Uniqueness Problem

An early word onset can be compatible with many members of the mental lexicon. Cohort theory emphasized the set of words supported by the initial sound sequence and the way subsequent evidence narrows that set. Marslen-Wilson distinguished initial access, selection among viable candidates, and integration of a recognized form with syntactic and semantic context. The approach made temporal uniqueness an explanatory concern: some candidates can be distinguished before their acoustic offset, whereas others remain ambiguous until later portions of an utterance Marslen-Wilson (1987).

Uniqueness is not equivalent to certainty. Initial fragments can be degraded, assimilated, or insufficiently informative, while a listener may anticipate a word from context before all diagnostic segments arrive. The cohort framework remains useful for formulating questions about the duration of lexical alternatives, but its specific predictions depend on whether initial mismatches automatically exclude candidates and how later context is permitted to act. Experimental gating and eye movement procedures provide different windows on these predictions because one asks for explicit identification and the other monitors graded competition during a task Marslen-Wilson (1987) Allopenna et al. (1998).

Within an integrated segmentation problem, cohorts can begin at several possible times. Selecting a plausible onset can favor a group of words, but there is no guarantee that the proposed onset marks a real lexical boundary. Additional simultaneous cohorts must be considered when long words contain short embedded words and when alternative word sequences cover overlapping acoustic material. A theory restricted to choosing among candidates at one preassigned onset addresses only a portion of continuous speech recognition. Competition across candidate onsets is the natural extension Norris (1994) McQueen et al. (1995).

TRACE and Interactive Activation

McClelland and Elman's TRACE represented speech perception through mutually constraining activation among processing units. Patterns of activation are updated as auditory evidence arrives, and competition can occur among phonemic and lexical hypotheses. The model was influential because it generated temporally specific predictions instead of treating lexical access as a single yes-or-no event. Its interactive design allowed activation at one representational level to influence another, making it a major reference point for questions about feedback, context, and perceptual interpretation McClelland & Elman (1986).

Interactive activation must be evaluated against concrete phenomena rather than a verbal claim that perception is contextual. The model can be assessed by comparing predicted competitor activation with eye movement trajectories, errors, or identification latencies. Allopenna, Magnuson, and Tanenhaus reported gaze patterns consistent with multiple partially activated candidates, including competitors sharing beginnings and those sharing endings. Those observations support graded temporal mapping but do not, by themselves, prove the model's particular architecture or require lexical feedback into sublexical representations Allopenna et al. (1998) McClelland & Elman (1986).

TRACE also carries computational and representational commitments that demand scrutiny. A simulation can explain a stimulus set while relying on input coding, timing choices, or network size that do not generalize to a large vocabulary. Repeated temporal representations may simplify aspects of the physical signal, and fit to laboratory behavior need not establish how neural circuits encode the same operations. The right comparison concerns which model features are indispensable for independent empirical predictions and which are convenient engineering assumptions. Competing architectures should be tested using matched conditions and transparent parameter variation Norris (1994) Norris & McQueen (2008).

Shortlist, Lexical Competition, and Continuous Input

Shortlist was designed to address lexical access from continuous speech without beginning from known word boundaries. Norris proposed that the incoming phonological evidence activates a manageable set of lexical candidates, among which competition helps determine compatible interpretations. Importantly, plausible words can compete even when their onsets are temporally offset. The explanatory target is the overall pattern of segmental coverage, making the model directly relevant to situations where a familiar word occurs within another word or multiple word sequences fit local material Norris (1994).

Competition can produce segmentation indirectly. If an early candidate strongly matches a span, it may suppress partially overlapping alternatives; if subsequent input undermines that candidate, previously less favored interpretations may gain ground. Norris, McQueen, and Cutler used experimental results and simulations to argue that competition must be considered alongside prosodic information. Their observations indicated that stress-related segmentation effects were especially visible when many lexical candidates competed for the same speech material. Thus cue use depends on the active lexical environment, not solely on acoustic form Norris et al. (1995).

Models of this type retain important empirical questions. The relevant competition network could reflect inhibitory links, normalized posterior evaluation, or distributed interference among representations. Behavioral suppression is compatible with several mechanisms, and computational plausibility alone does not settle which mechanism is implemented in the human brain. Word frequency, candidate neighborhood density, and timing of bottom-up evidence affect model predictions and offer points of possible falsification. Robust model comparison requires all alternatives to predict the same stimulus sets with clear constraints on free parameters Norris & McQueen (2008) Gaskell & Marslen-Wilson (1997).

Shortlist B and Probabilistic Lexical Evaluation

Shortlist B recast continuous speech recognition in Bayesian terms. Norris and McQueen retained competing lexical hypotheses and a segmentation procedure sensitive to possible-word structure, while using probabilistic evidence to evaluate matches. An especially important feature was its representation of uncertain phoneme information rather than an error-free string of recognized phonemes. The model made explicit how probabilities over incoming evidence contribute to the relative support for alternative lexical interpretations as a continuous utterance unfolds Norris & McQueen (2008).

A Bayesian statement does not mean that listeners consciously perform arithmetic or retain every candidate in numerical form. It specifies a computational-level account: possible word sequences receive support according to the interaction of prior constraints and the likelihood of the observed signal. Different representations and algorithms can approximate such inference. The empirical questions concern whether the chosen likelihood structure predicts human behavior, whether a candidate's support changes with expected timing, and how segmentation constraints influence posterior competition under uncertain acoustic evidence Norris & McQueen (2008) Norris et al. (1997).

Comparison with interactive activation should distinguish abstract inference from processing architecture. Two models can sometimes predict similar response distributions while disagreeing about feedback between phonetic and lexical representations. The existence of a good Bayesian fit cannot rule out neural feedback, just as a successful interactive activation fit cannot establish that feedback is necessary for a given experiment. Discriminating evidence requires manipulating acoustic ambiguity, learning, and temporal context in ways that make the competing mechanisms predict observably different outcomes. Shortlist B supplies one formal benchmark for that continuing program Norris & McQueen (2008) McClelland & Elman (1986).

Distributed Representations and Alternative Implementations

Gaskell and Marslen-Wilson developed a distributed approach in which phonological and lexical information are integrated within learned representations. Competition can emerge through interference among patterns rather than only through explicitly named word units inhibiting one another. This broadens the mechanistic space: separate dictionary entries with inhibitory connections, distributed attractor-like structures, and probabilistic sequence evaluation can all implement aspects of lexical selection. Their relative merit depends on predictions for phonetic mismatch, partial input, generalization, and novel vocabulary Gaskell & Marslen-Wilson (1997).

Representational choice is consequential for continuous speech. An abstract phoneme-coded model must explain why subtle duration differences affect lexical competition; a richly acoustic model must explain generalization across speakers and pronunciations. Distributed systems can learn relations from correlated patterns but may generalize incorrectly to unseen forms without appropriate constraints. In turn, explicit lexical architectures offer interpretable candidate lists but risk imposing boundaries that the listener actually has to infer. The best evidence concerns behavior under carefully controlled changes in spoken form and contextual probability Shatzman & McQueen (2006) Gow (2001).

These model families agree on several findings without settling the deepest architectural debates. Spoken candidates become available before the entire word is heard; alternatives coexist and influence responses; boundary information and lexical support are coordinated over time. Disagreements concern the format of input, feedback, competition, and learning rather than the basic existence of incremental recognition. A mature research synthesis therefore treats models as explicit explanatory instruments with predictions and failure conditions, avoiding the assumption that one successful fit licenses a unique account of the human recognition system Allopenna et al. (1998) Dahan et al. (2001) Norris & McQueen (2008).

Infant Statistical Learning and the Discovery of Candidate Words

Early word discovery presents a striking version of the segmentation problem because learners initially have limited lexical knowledge. Saffran, Aslin, and Newport exposed eight-month-old infants to continuous sequences organized by statistical relationships among neighboring syllables. The infants discriminated patterns consistent with the artificial stream's recurrent units after a brief exposure. This provided influential evidence that distributional learning can support discovery of candidate word forms before a mature vocabulary supplies lexical boundaries Saffran et al. (1996).

The central statistic is often described as a conditional transition between adjacent syllables: given one syllable, how predictable is the next? Relatively strong within-unit transitions and weaker between-unit transitions can provide boundary information under the constructed input conditions. This does not show that learners always calculate one particular formula, nor that the same contrast is sufficient for segmentation in natural conversational speech. Real input includes variable phonemes, morphology, prosody, competing topics, and acoustic reduction. An experiment demonstrating learning of controlled sequences establishes an important capacity whose ecological contribution requires further tests Saffran et al. (1996) Christiansen et al. (1998).

The extension from speech to tonal sequences helps specify the scope of learning. Saffran, Johnson, Aslin, and Newport investigated statistical learning with tone sequences in infants and adults, suggesting that the extraction of sequential structure has applications beyond the lexical domain. General sensitivity to transitional information may supply one resource for speech learning, while language-specific representations and communicative contexts determine how that resource supports actual words. A domain-general learning result is therefore compatible with specialized knowledge about phonology, stress, and lexical meaning Saffran et al. (1999) Jusczyk et al. (1999).

Natural-Language Development of Word Segmentation

Evidence from natural English learning demonstrates that the statistical capacity is joined by language-conditioned boundary expectations. Jusczyk, Houston, and Newsome examined infants' recognition of bisyllabic forms in connected speech across a series of experiments. At approximately seven and a half months, infants were more successful with words following a frequent strong–weak pattern; weak–strong words presented greater difficulty. By approximately ten and a half months, performance on these harder forms had improved. The finding ties segmentation development to changes in how multiple cues are utilized Jusczyk et al. (1999).

These laboratory tasks do not measure the full content of infant vocabulary. A child may respond differently to familiarized sound sequences without demonstrating adult-like semantic reference or stable word identification across all speakers. Nevertheless, successful recovery of a recurring form across sentences is a necessary learning achievement. Natural-language studies thereby connect momentary perceptual discrimination with the longer process of constructing lexical representations. The developmental schedule depends on the tested language and task; ages from one set of English-learning samples should not be treated as a fixed timetable for every linguistic community Jusczyk et al. (1999) Jusczyk (1999).

Longitudinal associations add another level of evidence. Newman and colleagues examined whether early ability to segment conversational speech predicted later language development. Such a relationship is important because it links laboratory segmentation behavior with a subsequent outcome, but a retrospective association cannot alone establish that better segmentation caused later linguistic performance. Shared auditory capacities, attention, exposure, and caregiver interaction are possible contributors. Prospective designs using repeated measures and appropriate covariates are needed to evaluate how early sensitivity, accumulating vocabulary, and language experience influence one another over time Newman et al. (2006).

Learning Multiple Cues from Speech Experience

Infants encounter transitional statistics, stress patterns, phonotactics, and recurring word forms simultaneously. A learning theory must account for whether these sources can be combined without explicit instruction about boundaries. Christiansen, Allen, and Seidenberg trained a recurrent network on child-directed speech with information about phonemes, stress, and utterance edges. Their model demonstrated that multiple individually imperfect cues could support discovery of lexical divisions. The computational result establishes a proof of concept for cue integration, subject to the usual questions about training corpus, architecture, and correspondence with infant behavior Christiansen et al. (1998).

Learning also has a feedback structure across development. An emerging lexicon gives evidence about likely sound sequences and boundaries, which can in turn improve segmentation of unfamiliar utterances. More successful segmentation creates additional candidate words and opportunities for learning. The relation need not be uniform across all languages: stress-based expectations may be especially informative for an English learner, whereas a different combination of syllabic, morphological, and prosodic patterns may support another learner. This reciprocity links online word recognition to longer-term vocabulary formation Jusczyk et al. (1999) Cutler et al. (1986).

A complete account must confront the difference between segmentation in an artificial language and segmentation in a communicative environment. Artificial-language experiments isolate a few statistical variables and make causal interpretation cleaner. Natural speech adds reference, familiar voices, gesture, discourse structure, and strong fluctuations in signal quality. If a model is trained on idealized phoneme strings, its apparent success may not survive reduced vowels or uncertain consonants. Convergent developmental evidence requires both controlled manipulations and measurements of learning from representative speech samples Saffran et al. (1996) Christiansen et al. (1998) Fernald et al. (2008).

Lexical Frequency, Familiarity, and Vocabulary Structure

Familiarity changes which candidates are expected during recognition. A frequent lexical form can receive support earlier or more strongly than a competing rare form under otherwise similar acoustic evidence. Dahan, Magnuson, and Tanenhaus used eye movements during spoken instructions to investigate the time course of frequency effects, supplying evidence that lexical frequency affects evolving interpretation rather than simply a final button press. The theoretical issue is how such frequency information is represented and how it interacts with detailed bottom-up support Dahan et al. (2001).

Neighborhood structure adds competition beyond raw frequency. A word with many phonologically similar neighbors may face a different identification problem from an equally frequent word with few close alternatives. Embedding is particularly relevant: a short word can be the initial portion of a longer word, and the larger vocabulary contains many possible overlapping forms. McQueen, Cutler, Briscoe, and Norris documented the importance of lexical embeddings in English word structures, motivating competition mechanisms that operate across possible onsets and offsets. The size and composition of the candidate set therefore constrain the task McQueen et al. (1995).

Experimental comparisons must avoid conflating frequency, phonotactic probability, length, semantic familiarity, and phonetic predictability. Frequency norms depend on corpus and speech register; a participant's actual exposure may differ from aggregate counts. High-frequency words may be pronounced with more reduction, producing a tradeoff between lexical familiarity and signal clarity. Better models distinguish priors over forms from likelihoods based on the incoming acoustics, while behavioral designs test the two through matched stimulus materials and manipulation of expectation. The same word can be easy in one utterance and difficult in another Norris & McQueen (2008) Mattys et al. (2012).

Adaptation, Perceptual Learning, and Newly Acquired Words

Speech recognition operates on a learned sound system that remains adaptable. Norris, McQueen, and Cutler demonstrated lexically guided perceptual learning with acoustically ambiguous fricatives. Listeners' knowledge of which word was likely to have been spoken shaped later categorization of ambiguous sounds. This result establishes that lexical knowledge can guide the adaptation of phonetic interpretation across exposure, even if specific online-feedback mechanisms remain contested. Adaptation can influence future segmentation because shifts in categorization change the evidence available to competing lexical candidates Norris et al. (2003).

Word learning also creates new competitors. Shatzman and McQueen trained participants on novel spoken words associated with unfamiliar objects, including pairs in which a short learned form was embedded at the beginning of a longer learned form. The subsequent influence of duration on fixations demonstrated that newly acquired lexical items could be recognized using expectations about prosody. The result matters because it links the acquisition of lexical representations to general patterns of spoken recognition without requiring every acoustic exemplar to have been heard previously Shatzman & McQueen (2006).

Adaptation raises questions about generalization and stability. A shift learned with one speaker or segment context may or may not transfer to new voices, novel lexical environments, or differently degraded recordings. Short-term recalibration is not necessarily equivalent to permanent phonological restructuring. Developmental experience and adult learning may rely on overlapping mechanisms but operate under different constraints. Strong evidence therefore requires separating the acquisition of a word form, the reassignment of ambiguous acoustic input, and the updating of expectations about likely boundaries. Each changes recognition through potentially different pathways Norris et al. (2003) Gow (2001).

Connected-Speech Variation and Phonological Assimilation

Real speech changes with neighboring sounds, speaking rate, and phonological environment. A listener who expects every lexical entry to match an invariant dictionary pronunciation would frequently reject the correct word. Gow examined English place assimilation and recognition of words whose surface form differed from a canonical representation. Such work asks whether the listener tolerates mismatches because the surrounding context renders them predictable, or because lexical representations encode a broader range of phonetic possibilities. The evidence favors sensitivity to contextual structure, while the interpretation of the underlying architecture remains contested Gow (2001).

Variability can affect two stages at once. A reduced consonant may weaken the evidence for a particular lexical candidate, while the articulatory transition can supply information about an upcoming segment or a boundary. These effects are experimentally separable only when recordings and comparisons carefully control for contextual phonetics. Dahan and colleagues manipulated subcategorical mismatches and found consequences for target fixation time courses, revealing sensitivity to lexical alternatives beyond fully categorized phonemic identity. Recognition is therefore incremental at a granularity finer than conventional phoneme transcription Dahan et al. (2001).

Acoustic tolerance should not be mistaken for unrestricted guessing. A mismatch that fits an ordinary assimilation pattern may be handled differently from one incompatible with the acoustic context. Listeners combine learned expectations with the observed signal and can revise their interpretation as later speech arrives. A strong model must predict when apparently deviant forms remain recognizable and when the same amount of deviation defeats recognition. Generalization across dialect, accent, speaker, and rate requires diversified training and testing materials rather than an assumption that one corpus represents every form of spoken English Gow (2001) Kapnoula et al. (2024).

Segmentation and Recognition under Adverse Listening Conditions

Speech understanding can be degraded at the source, during transmission, or through properties of the receiver. Background noise masks acoustic distinctions; reverberation blurs temporal cues; an unfamiliar accent can change correspondence between expected and realized forms. Mattys, Davis, Bradlow, and Scott reviewed adverse speech-recognition conditions and their consequences for auditory, linguistic, cognitive, and neural processing. For segmentation, the critical prediction is that the reliability of boundary cues changes with the source of difficulty rather than declining uniformly with signal quality Mattys et al. (2012).

When coarticulation becomes difficult to detect, listeners may rely more heavily on stress or lexical plausibility. Mattys's experimental comparison of clear and noisy materials demonstrated a change in the relative influence of stress and coarticulatory cues under the tested conditions. This result is a direct challenge to any theory assigning the same fixed weight to every cue. Context may improve performance while simultaneously producing stronger expectations that occasionally favor the wrong candidate. Performance must be analyzed for both correct recovery and systematic error patterns Mattys (2004) Mattys et al. (2005).

Recent work examined whether clearer speaking styles aid segmentation itself. Guo and Smiljanić used eye tracking to compare conversational and clearly articulated sentences containing a short-word versus longer-word ambiguity. Clear speech increased relative target fixation for a short target under both quiet and noisy conditions, with an asymmetric pattern when the longer form was intended. The design suggests that speaking style can alter access to word-boundary evidence, but its scope is the manipulated materials and listener population. It does not establish an equal benefit for every speech rate, language, or hearing profile Guo & Smiljanić (2024).

Multilingual Experience and Individual Differences

The phonological probabilities, stress patterns, and typical word shapes learned in one language shape how unfamiliar streams are interpreted. Cross-language comparisons provide evidence that segmentation routines adapt to language structure, while also warning against treating a language label as an explanation by itself. Differences in age of acquisition, proficiency, current exposure, literacy, and familiarity with the recorded accent may all influence performance. Historical findings on French and English segmentation therefore serve as a foundation for carefully controlled multilingual research rather than a universal account of bilingual cognition Cutler et al. (1986) Cutler et al. (1992).

An individual's recognition system also changes across the life span. Kapnoula, Jevtović, and Magnuson reviewed plasticity in spoken word recognition, including influences of development, linguistic experience, and changing sensory or cognitive capacities. Their perspective broadens the object from the often-studied idealized adult monolingual listener to a more variable system. A key issue is whether experience changes the content of lexical representations, the dynamics of competition, or both. Models tested on a single demographic group cannot automatically settle that question Kapnoula et al. (2024).

Developmental comparisons increasingly use real-time measures rather than relying only on overall accuracy. Kutlu and colleagues investigated changes in spoken and written word recognition across development and argued that variation in underlying ability can account for changes often interpreted as maturation. The study gives a useful modern comparator, but spoken and written recognition have distinct temporal and representational properties. For the present subject, the practical lesson is to measure vocabulary, listening skill, and task performance alongside age, and to avoid treating chronological age as a complete causal mechanism Kutlu et al. (2024).

Neural Timing and the Architecture of Spoken Comprehension

Neural evidence can constrain accounts of how sounds become lexical interpretations, but its spatial and temporal measures answer different questions. Hemodynamic imaging identifies activity associated with processing conditions while averaging over relatively slow changes. Electrophysiology and magnetoencephalography provide finer temporal information about responses to unfolding input. Davis and Johnsrude reviewed the influence of top-down processes on speech perception and its auditory interface, offering a framework for comparing sensory evidence with linguistic expectations without reducing observed activity to a single algorithm Davis & Johnsrude (2007).

Sohoglu, Beckers, and Davis combined magnetoencephalographic and functional magnetic resonance measurements to examine prediction-related activity for spoken words whose initial syllables varied in how strongly they anticipated the remainder. Their investigation of superior temporal responses provides a contemporary example of testing computational predictions with convergent neural methods. These responses constrain theories about prediction and recognition, but neural correlations do not directly reveal a consciously computed word boundary or establish one unique way of linking candidate words to phonetic evidence Sohoglu et al. (2024).

Gwilliams and colleagues argued for studying the computational architecture of speech comprehension by separating representations, operations, and information flow. Their synthesis emphasizes parallel, distributed spatiotemporal processing rather than simple one-region-per-stage assignments. For segmentation, the research program should connect the timing of neural sensitivity to boundary cues with behavioral evidence for lexical competition. Such integration requires stimuli whose acoustic and lexical properties are quantified, explicit temporal alignment, and independent validation. Neural decoding of a hypothesized category remains distinct from demonstrating its causal necessity Gwilliams et al. (2025).

Prediction, Semantic Context, and Lexical Selection

Sentence context changes how plausible particular words are before they are heard. This can make recognition faster or more accurate when the predicted form fits incoming speech, but a predictive advantage may operate at several stages: attentional allocation, candidate prior probability, postlexical integration, or a later decision. Marslen-Wilson's treatment of lexical access distinguished form-based access and selection from integration, exposing the need to specify where contextual effects occur. A behavioral improvement associated with context is insufficient evidence for a particular neural feedback pathway Marslen-Wilson (1987) Davis & Johnsrude (2007).

Semantic and visual context likewise modulate what eye movements appear to reveal. A pictured target can attract gaze because the spoken form becomes probable, because its meaning fits the utterance, or because a response strategy favors one visible object. Duta and Plunkett modeled the temporal movement from early phonological competition to later semantic or visual preferences in visual-world tasks. Their results illustrate how an unfolding phonological-to-semantic mapping can produce patterns sometimes attributed to explicit top-down control, strengthening the case for mechanism-comparative modeling Duta & Plunkett (2021).

Prediction should therefore be represented as a testable family of hypotheses rather than a blanket explanation of speech comprehension. The relevant questions concern which statistical regularities are learned, how rapidly they constrain lexical candidates, and what happens when unexpected sounds arrive. Laboratory manipulations of lexical predictability are useful when they also match or model acoustic realization. Natural discourse adds pragmatics, changing speakers, and long-range topical structure. Models that explain local acoustic segmentation should be evaluated again when applied to those richer conditions Norris & McQueen (2008) Sohoglu et al. (2024).

Ecological Validity and Generalization beyond Laboratory Stimuli

Laboratory segmentation research derives much of its inferential strength from stimuli designed to isolate one candidate explanation. That strength creates a generalization problem: recorded nonsense sequences, four-picture displays, and repeated short fragments differ from unplanned conversations. Spontaneous speech contains repairs, reductions, overlapping turns, irregular pauses, unfamiliar names, and referential information outside the acoustic channel. A mechanism that works on controlled word pairs must be tested again in these conditions. The outcome of a laboratory experiment should therefore be described at the level of the manipulation actually performed Allopenna et al. (1998) Mattys et al. (2012).

Ecologically richer research can preserve causal testing through nested designs. Investigators may compare a controlled acoustic manipulation across isolated fragments, sentence contexts, and naturalistic passages while retaining common measurements of the predicted competitor. Speech corpora can estimate the distribution of stress, phonotactic structure, and acoustic reductions in real interactions, while experimental recordings test their separate effects. When a naturalistic corpus contains several correlated sources of information, observational model comparison alone cannot determine which source caused a response. Controlled intervention remains necessary Christiansen et al. (1998) Guo & Smiljanić (2024).

Cross-population generalization deserves the same attention. A theory based on highly proficient young adult speakers may depend on language-specific vocabulary, hearing acuity, and sustained attention. Developmental, multilingual, and sensory-diverse groups can exhibit different candidate activation patterns without representing the failure of lexical competition as a broad principle. A valid generalization argument describes the measured population, the nature of the speech materials, and the robustness of the effect across new speakers and items. Variation becomes a property to explain rather than an afterthought Kapnoula et al. (2024) Kutlu et al. (2024).

Reproducibility, Statistical Inference, and Evidence Quality

Most influential segmentation experiments are based on complex relationships among items, participants, phonetic realizations, and analysis windows. The unit of inference must match that structure. A pattern visible for a few highly controlled words can be driven by their idiosyncratic acoustic properties; a participant-level average can obscure differences among word types. Modern analysis should use item-sensitive designs, prespecified hypotheses where possible, and uncertainty estimates appropriate to nested observations. Reporting response times without examining accuracy or trials excluded by preprocessing can distort the interpretation of an apparent segmentation advantage Mattys et al. (2005) Allopenna et al. (1998).

Eye-movement evidence introduces decisions about saccade latency, baseline windows, time binning, regions of interest, and correction for multiple comparisons. A gaze preference emerging after an auditory event may reflect earlier processing because planning a movement requires time. Researchers should therefore provide the temporal alignment of the speech signal, the analysis definition, and alternatives to a preferred window when a major theoretical conclusion depends on millisecond-scale differences. A result's replicability also depends on sharing usable stimuli, transparent exclusions, and sufficient information to reimplement the analysis Fernald et al. (2008) Dahan et al. (2001).

This monograph is a critical narrative synthesis, with sources selected for their direct relationship to its research questions and verifiable publication records. It does not implement a registered database search with reproducible inclusion screening, and it does not compute a pooled effect size or grade evidence using a formal systematic-review protocol. Accordingly, no quantitative estimate of the average segmentation effect is asserted. The most reliable conclusions are those supported by convergence among different manipulations and methods; disputes about representational feedback, relative cue weighting, and generalization to natural listening retain their status as open empirical questions Norris & McQueen (2008) Kapnoula et al. (2024).

Computational Tests and Mechanism Discrimination

A model of spoken recognition should predict an unfolding probability distribution or other explicit representation over alternative word hypotheses, rather than only a correct final transcription. Its evaluation can include segmentation boundaries, lexical identity, temporal eye movements, and changes under acoustic degradation. The strongest comparisons hold the training and evaluation material separate, limit flexible parameter adjustment, and examine whether a model predicts phenomena beyond those that motivated its creation. In these respects, TRACE, Shortlist, Bayesian approaches, and distributed networks are valuable competing research instruments McClelland & Elman (1986) Norris (1994) Norris & McQueen (2008) Gaskell & Marslen-Wilson (1997).

Several kinds of model success need to remain distinct. An algorithm can recover word boundaries from idealized syllables while failing on a continuous waveform; it can identify phonological word forms without understanding their meanings; it can approximate listener behavior without reproducing the same learning history or neural mechanism. Conversely, a model trained end to end on rich speech may achieve useful recognition accuracy while revealing little about which boundary cues account for human errors. Explanatory research requires interventions that differentiate possible mechanisms and measurements that link model variables to independently observed behavior Christiansen et al. (1998) Duta & Plunkett (2021).

A productive comparison also involves deliberately adverse cases: a short word embedded in a longer one, a misleading stress pattern, assimilation across a boundary, an unfamiliar pronunciation, and a low-frequency form competing with a common neighbor. The interaction among these cases is often more revealing than success on a single test. A unified mechanism should predict when lexical competition strengthens reliance on prosody and when detailed coarticulation overrides rhythmic expectations. Future research can treat these combined manipulations as a common benchmark across model families and languages Norris et al. (1995) Mattys (2004) Gow (2001).

Relations to Artificial Speech Systems and the Artificial Era

Automatic speech recognition and human spoken word recognition share a practical requirement to transform a continuous, noisy acoustic input into structured linguistic output. In artificial systems, learned acoustic encoders and sequence decoders can produce word sequences without explicit preliminary placement of all boundaries. This makes machine recognition useful as a computational comparator for how uncertainty and temporal evidence may be integrated. A machine transcript's accuracy, however, supplies only a task-level measurement and does not demonstrate that the system uses a human lexical architecture or follows the same developmental pathway Gwilliams et al. (2025) Norris & McQueen (2008).

Psychology of Artificial can use speech segmentation as a well-defined case for cross-system functional comparison. The shared research object is the resolution of competing candidate units under streaming uncertainty. Comparisons can manipulate masking noise, unfamiliar vocabulary, alternative prosodic grouping, and domain shifts, measuring whether internal or behavioral uncertainty develops in ways that predict recognition errors. Such experiments test organizational analogies rather than asserting identical subjective experience or identical neural substrates. Their value lies in operational precision and the possibility of falsifying specific processing hypotheses Duta & Plunkett (2021) Mattys et al. (2012).

In the historical horizon of the Artificial Era, spoken interaction is also a site of human–artificial coordination. Errors in detecting names, commands, corrections, or turn boundaries can change the practical outcome of a dialogue. Human listeners and speech-based artificial systems can therefore be studied as different processing organizations exchanging signals in a shared environment. Research on this coordination benefits from separating acoustic segmentation, lexical identification, discourse inference, and action selection, then measuring failures and recovery at each interface. The present dossier supplies the human psycholinguistic foundation for that broader interdisciplinary comparison Kapnoula et al. (2024) Gwilliams et al. (2025).

Unresolved Questions and a Research Agenda

The first unresolved problem is the representation of uncertainty. Behavioral data establish that candidate interpretations compete and can change after additional speech, yet they leave open whether this competition is best expressed through interactive activation, probabilistic ranking, or distributed interference. Direct comparisons require the same acoustic materials, explicit model parameter constraints, and methods that distinguish early lexical activity from later decision strategies. Neural data may add timing and location constraints but must be linked to computationally discriminating predictions McClelland & Elman (1986) Norris & McQueen (2008) Sohoglu et al. (2024).

The second problem concerns development and variation. Segmentation capacities are present in infancy, and their organization changes with exposure to lexical, phonotactic, and prosodic patterns. Questions remain about how learners combine cues under conflicting evidence, how vocabulary feedback changes their relative importance, and how those changes vary across languages and populations. Longitudinal designs, cross-language replication, representative speech corpora, and developmental modeling can test whether different listeners learn the same computation with different data or acquire substantively different recognition routines Saffran et al. (1996) Jusczyk et al. (1999) Kapnoula et al. (2024) Kutlu et al. (2024).

The third problem concerns realistic listening. Most established findings derive from carefully restricted contexts, while genuine conversation contains competing talkers, variable pronunciation, rapid topic shifts, and partially predictable meanings. Research should connect acoustic measures, moment-by-moment candidate activation, neural responses, and final communicative success within matched experiments. Uncertainty about how these levels interact is a scientific reason to design stronger discriminating studies. The likely future of the field is a more explicit account of how segmentation mechanisms adapt while preserving the capacity to recover lexical structure from imperfect speech Guo & Smiljanić (2024) Mattys et al. (2012) Gwilliams et al. (2025).

Conclusion

Speech segmentation and spoken word recognition are a coupled psychological process in which unfolding acoustic evidence supports multiple possible lexical organizations. The central research finding is that listeners use more information than a sequence of discretely identified phonemes: coarticulation, duration, phonotactics, rhythmic structure, lexical familiarity, and broader contextual probabilities can all influence which words are recognized and where their boundaries are placed. The relative influence of these sources changes with acoustic conditions, learned language structure, and competing lexical alternatives Mattys et al. (2005) Norris et al. (1995) Salverda et al. (2003).

Historical and experimental research supplies complementary explanations. Cohort approaches emphasize temporal access and candidate narrowing; TRACE models interactive activation; Shortlist and Shortlist B formalize continuous lexical competition and possible-word constraints; distributed models offer another route to graded lexical selection. Eye movements, priming, gating, and word spotting converge on incremental recognition while differing in their inferential limits. Infant experiments demonstrate early statistical sensitivity and the subsequent development of language-conditioned segmentation skills. Contemporary neural and developmental work extends the explanatory agenda without establishing one definitive processing architecture Marslen-Wilson (1987) McClelland & Elman (1986) Norris & McQueen (2008) Saffran et al. (1996) Kutlu et al. (2024).

The resulting scientific synthesis identifies speech segmentation as a dynamic inference problem situated at the interface between acoustic perception, learned lexical organization, and temporal prediction. Its strongest empirical claims concern sensitivity to graded evidence, the coexistence of alternatives, and the usefulness of multiple boundary cues. The remaining program concerns implementation, comparative generalization, developmental plasticity, and the coordination of human and artificial speech systems. A defensible theory must predict both successful word recognition and structured failure under realistic changes of signal, listener, and context Kapnoula et al. (2024) Guo & Smiljanić (2024) Gwilliams et al. (2025).

References

  • Allopenna, P. D., Magnuson, J. S., & Tanenhaus, M. K. (1998). Tracking the time course of spoken word recognition using eye movements: Evidence for continuous mapping models. Journal of Memory and Language, 38(4), 419–439. https://doi.org/10.1006/jmla.1997.2558
  • Christiansen, M. H., Allen, J., & Seidenberg, M. S. (1998). Learning to segment speech using multiple cues: A connectionist model. Language and Cognitive Processes, 13(2–3), 221–268. https://doi.org/10.1080/016909698386528
  • Cutler, A., Mehler, J., Norris, D., & Segui, J. (1986). The syllable's differing role in the segmentation of French and English. Journal of Memory and Language, 25(4), 385–400. https://doi.org/10.1016/0749-596X(86)90033-1
  • Cutler, A., Mehler, J., Norris, D., & Segui, J. (1992). The monolingual nature of speech segmentation by bilinguals. Cognitive Psychology, 24(3), 381–410. https://doi.org/10.1016/0010-0285(92)90012-Q
  • Dahan, D., Magnuson, J. S., & Tanenhaus, M. K. (2001). Time course of frequency effects in spoken-word recognition: Evidence from eye movements. Cognitive Psychology, 42(4), 317–367. https://doi.org/10.1006/cogp.2001.0750
  • Dahan, D., Magnuson, J. S., Tanenhaus, M. K., & Hogan, E. M. (2001). Subcategorical mismatches and the time course of lexical access: Evidence for lexical competition. Language and Cognitive Processes, 16(5–6), 507–534. https://doi.org/10.1080/01690960143000074
  • Davis, M. H., & Johnsrude, I. S. (2007). Hearing speech sounds: Top-down influences on the interface between audition and speech perception. Hearing Research, 229(1–2), 132–147. https://doi.org/10.1016/j.heares.2007.01.014
  • Duta, M., & Plunkett, K. (2021). A neural network model of lexical-semantic competition during spoken word recognition. Frontiers in Human Neuroscience, 15, Article 700281. https://doi.org/10.3389/fnhum.2021.700281
  • Fernald, A., Zangl, R., Portillo, A. L., & Marchman, V. A. (2008). Looking while listening: Using eye movements to monitor spoken language comprehension by infants and young children. In I. A. Sekerina, E. M. Fernández, & H. Clahsen (Eds.), Developmental psycholinguistics: On-line methods in children's language processing (pp. 97–135). John Benjamins. https://doi.org/10.1075/lald.44.06fer
  • Gaskell, M. G., & Marslen-Wilson, W. D. (1997). Integrating form and meaning: A distributed model of speech perception. Language and Cognitive Processes, 12(5–6), 613–656. https://doi.org/10.1080/016909697386646
  • Gow, D. W., Jr. (2001). Assimilation and anticipation in continuous spoken word recognition. Journal of Memory and Language, 45(1), 133–159. https://doi.org/10.1006/jmla.2000.2764
  • Guo, Z.-C., & Smiljanić, R. (2024). Ham or hamster? Eye-tracking evidence of a clear speech benefit for word segmentation in quiet and in noise. Language, Cognition and Neuroscience, 39(5), 609–631. https://doi.org/10.1080/23273798.2024.2345300
  • Gwilliams, L., Bhaya-Grossman, I., Zhang, Y., Scott, T., Harper, S., & Levy, D. (2025). Computational architecture of speech comprehension in the human brain. Annual Review of Linguistics, 11, 209–226. https://doi.org/10.1146/annurev-linguistics-031120-111245
  • Jusczyk, P. W. (1999). How infants begin to extract words from speech. Trends in Cognitive Sciences, 3(9), 323–328. https://doi.org/10.1016/S1364-6613(99)01363-7
  • Jusczyk, P. W., Houston, D. M., & Newsome, M. (1999). The beginnings of word segmentation in English-learning infants. Cognitive Psychology, 39(3–4), 159–207. https://doi.org/10.1006/cogp.1999.0716
  • Kapnoula, E. C., Jevtović, M., & Magnuson, J. S. (2024). Spoken word recognition: A focus on plasticity. Annual Review of Linguistics, 10, 233–256. https://doi.org/10.1146/annurev-linguistics-031422-113507
  • Kutlu, E., Klein-Packard, J., Jeppsen, C., Tomblin, J. B., & McMurray, B. (2024). The development of real-time spoken and written word recognition derives from changes in ability, not maturation. Cognition, 251, Article 105899. https://doi.org/10.1016/j.cognition.2024.105899
  • Marslen-Wilson, W. D. (1987). Functional parallelism in spoken word-recognition. Cognition, 25(1–2), 71–102. https://doi.org/10.1016/0010-0277(87)90005-9
  • Mattys, S. L. (2004). Stress versus coarticulation: Toward an integrated approach to explicit speech segmentation. Journal of Experimental Psychology: Human Perception and Performance, 30(2), 397–408. https://doi.org/10.1037/0096-1523.30.2.397
  • Mattys, S. L., White, L., & Melhorn, J. F. (2005). Integration of multiple speech segmentation cues: A hierarchical framework. Journal of Experimental Psychology: General, 134(4), 477–500. https://doi.org/10.1037/0096-3445.134.4.477
  • Mattys, S. L., Davis, M. H., Bradlow, A. R., & Scott, S. K. (2012). Speech recognition in adverse conditions: A review. Language and Cognitive Processes, 27(7–8), 953–978. https://doi.org/10.1080/01690965.2012.705006
  • McClelland, J. L., & Elman, J. L. (1986). The TRACE model of speech perception. Cognitive Psychology, 18(1), 1–86. https://doi.org/10.1016/0010-0285(86)90015-0
  • McQueen, J. M., Cutler, A., Briscoe, T., & Norris, D. (1995). Models of continuous speech recognition and the contents of the vocabulary. Language and Cognitive Processes, 10(3–4), 309–331. https://doi.org/10.1080/01690969508407098
  • Newman, R., Ratner, N. B., Jusczyk, A. M., Jusczyk, P. W., & Dow, K. A. (2006). Infants' early ability to segment the conversational speech signal predicts later language development: A retrospective analysis. Developmental Psychology, 42(4), 643–655. https://doi.org/10.1037/0012-1649.42.4.643
  • Norris, D. (1994). Shortlist: A connectionist model of continuous speech recognition. Cognition, 52(3), 189–234. https://doi.org/10.1016/0010-0277(94)90043-4
  • Norris, D., McQueen, J. M., & Cutler, A. (1995). Competition and segmentation in spoken-word recognition. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21(5), 1209–1228. https://doi.org/10.1037/0278-7393.21.5.1209
  • Norris, D., McQueen, J. M., Cutler, A., & Butterfield, S. (1997). The possible-word constraint in the segmentation of continuous speech. Cognitive Psychology, 34(3), 191–243. https://doi.org/10.1006/cogp.1997.0671
  • Norris, D., McQueen, J. M., & Cutler, A. (2003). Perceptual learning in speech. Cognitive Psychology, 47(2), 204–238. https://doi.org/10.1016/S0010-0285(03)00006-9
  • Norris, D., & McQueen, J. M. (2008). Shortlist B: A Bayesian model of continuous speech recognition. Psychological Review, 115(2), 357–395. https://doi.org/10.1037/0033-295X.115.2.357
  • Saffran, J. R., Aslin, R. N., & Newport, E. L. (1996). Statistical learning by 8-month-old infants. Science, 274(5294), 1926–1928. https://doi.org/10.1126/science.274.5294.1926
  • Saffran, J. R., Johnson, E. K., Aslin, R. N., & Newport, E. L. (1999). Statistical learning of tone sequences by human infants and adults. Cognition, 70(1), 27–52. https://doi.org/10.1016/S0010-0277(98)00075-4
  • Salverda, A. P., Dahan, D., & McQueen, J. M. (2003). The role of prosodic boundaries in the resolution of lexical embedding in speech comprehension. Cognition, 90(1), 51–89. https://doi.org/10.1016/S0010-0277(03)00139-2
  • Shatzman, K. B., & McQueen, J. M. (2006). The modulation of lexical competition by segment duration. Psychonomic Bulletin & Review, 13(6), 966–971. https://doi.org/10.3758/BF03213910
  • Shatzman, K. B., & McQueen, J. M. (2006). Prosodic knowledge affects the recognition of newly acquired words. Psychological Science, 17(5), 372–377. https://doi.org/10.1111/j.1467-9280.2006.01714.x
  • Shukla, M., Nespor, M., & Mehler, J. (2007). An interaction between prosody and statistics in the segmentation of fluent speech. Cognitive Psychology, 54(1), 1–32. https://doi.org/10.1016/j.cogpsych.2006.04.002
  • Sohoglu, E., Beckers, L., & Davis, M. H. (2024). Convergent neural signatures of speech prediction error are a biological marker for spoken word recognition. Nature Communications, 15, Article 9984. https://doi.org/10.1038/s41467-024-53782-5