RESEARCH DOSSIER
Speech Monitoring and Error Repair
Abstract. Speech monitoring enables speakers to evaluate prepared and overt utterances, detect possible deviations from linguistic and communicative targets, and change production while speaking. This critical narrative research monograph examines the origins of self-repair research, the perceptual-loop tradition, conflict- and production-based monitoring, node activation, auditory and somatosensory feedback, and computational accounts of control. Evidence is synthesized from spontaneous repairs, experimentally induced sound slips, temporal distributions of interruption and replacement, divided-attention studies, neuropsychological dissociations in aphasia, auditory-feedback perturbation, and functional neuroimaging. The analysis separates error generation, detection, interruption, and repair success; it evaluates lexical bias, phonetic similarity, early and late detection, and the relationship between motor compensation and linguistic self-correction. The scientific conclusion is that monitoring is a family of interacting functions with different targets and timescales, for which evidence supports multiple information sources while important causal and generalization questions remain open. This publication reports a critical narrative synthesis, not a newly conducted experiment or preregistered systematic review.
Keywords: speech monitoring; speech self-monitoring; speech errors; self-repair; covert repair; overt repair; perceptual-loop theory; conflict monitoring; speech production; phonological slips; lexical bias; inner speech; auditory feedback; error-to-cutoff latency; speech motor control; aphasia; stuttering; neural monitoring; Artificial Era
Speech Monitoring as a Psychological Research Object
Speech monitoring comprises the processes through which a speaker evaluates an intended, prepared, or articulated utterance against linguistic, communicative, and motor requirements, detects possible deviations, and adjusts production. An error can involve a mistaken word, misplaced segment, disordered sequence, inappropriate formulation, or mismatch between predicted and heard sound. A speaker can intercept an error before articulation, interrupt an ongoing utterance, or revise it afterward. These events support distinct inferences: an interruption establishes detection of trouble but not necessarily identification of the erroneous representation; a successful repair shows that an alternative production could be selected Levelt (1983) Postma (2000) Nozari (2025).
The scientific problem has at least four separable components: how speakers represent the intended target, which signals indicate that the current plan is unreliable, how those signals trigger control, and how a replacement expression is prepared. Detection may be implicit, reflected in latency or hesitation, or accompanied by explicit awareness and verbal identification. Repair includes stopping, restarting, replacing a segment or phrase, and reformulating an intended message. Inferences about an internal monitor require measures that distinguish each component; repair frequency alone confounds error generation, opportunity for detection, willingness to interrupt, and success of replacement Hartsuiker & Kolk (2001) Nozari et al. (2011) Nooteboom & Quené (2017).
This dossier examines self-monitoring within speech production across conceptual planning, lexical retrieval, sound-form encoding, articulatory control, and overt speech. It treats listener-initiated clarification and collaboratively negotiated conversational repair as neighboring phenomena with their own mechanisms. Its approach is a critical narrative integration of experimentally induced slips, spontaneous self-repairs, neuropsychological dissociations, feedback perturbation, electrophysiology, imaging, and computational modeling. The dossier does not claim a new experiment, comprehensive systematic literature search, or quantitative meta-analysis Postma (2000) Gauvin et al. (2016) Nozari (2025).
What Counts as an Error and What Counts as Repair
Research requires a target definition against which an observed performance can be judged. An error of lexical selection replaces the intended lexical item; a phonological slip exchanges, anticipates, or perseverates segments; an articulatory deviation changes an acoustic realization; and an appropriateness revision changes an expression whose linguistic form may be fully well formed. These categories depend upon the intended message, temporal context, and communicative goal. For observational studies, coders ordinarily infer intent from a correction, the prior context, or a constrained elicitation task, and that inference has uncertainty Levelt (1983) Levelt et al. (1999) Postma (2000).
A self-repair contains a trouble source, a possible interruption or editing expression, and a repaired segment. The original utterance need not contain an unequivocal speech error: speakers also revise information or improve wording. Conversely, a clearly erroneous output may remain uncorrected because the speaker did not notice it, judged correction unnecessary, or failed to retrieve the desired replacement. The analytic distinction between detection, interruption, and repair prevents an observed correction from being mistaken for a complete measure of monitoring accuracy Levelt (1983) Hartsuiker & Kolk (2001) Nooteboom & Quené (2020).
Covert repair designates a theoretical event in which a problematic plan is corrected before its erroneous form becomes audible. A hesitation or restart may be compatible with covert repair, but it is not unique evidence for that explanation: lexical competition, respiratory timing, attention shifts, and ordinary planning difficulty can produce similar surfaces. Overt repair provides a visible sequence of attempted utterance and correction, yet its timing may reflect earlier internal detection. Strong tests therefore combine controlled error opportunities, acoustic timing, and direct manipulations of the monitoring channel rather than assigning a psychological cause to each disfluency by appearance alone Postma & Kolk (1993) Nooteboom & Quené (2017) Quené & Nooteboom (2024).
Historical Origins: Slips, Self-Correction, and the Speaking System
Speech errors became scientifically productive because their systematic patterns can reveal intermediate representations in an otherwise rapid process. The traditions of observing spontaneous slips and experimentally inducing errors meet in a common inferential problem: a distortion in overt output can arise from competition at several stages before articulation. The historical transition from cataloging error types to manipulating phonological and semantic conditions transformed slips into tests of hypotheses concerning lexical selection, segmental planning, and prearticulatory editing Motley & Baars (1976) Levelt et al. (1999) Postma (2000).
In a foundational observational study, Levelt analyzed 959 spontaneous self-repairs and separated monitoring and interruption, editing expressions, and the repair proper. The temporal and syntactic organization of the corrections supplied evidence that speakers preserve relations between an abandoned phrase and its replacement. His account assigned a central role to language understanding in evaluating a speaker's own planned or produced speech, a proposal that became the perceptual-loop tradition. It was historically influential because it linked linguistic architecture, temporal behavior, and functional self-control in one research program Levelt (1983) Hartsuiker & Kolk (2001).
Controlled spoonerism induction supplied a complementary method. Motley and Baars manipulated phonological precursors and semantic properties of potential error outcomes, finding that error incidence was sensitive to more than the phonological form of a target. Such results were interpreted as evidence of filtering or editing before articulation; they also raised the still-active question whether observed lexical and semantic biases arise through an independent monitor or through feedback and competition within the production system. Later studies compared the survival of word-form and nonword-form slips and explicitly modeled internal and overt detection Motley & Baars (1976) Nooteboom & Quené (2008) Nooteboom (2005).
From Stage Models to Control Architectures
The production architecture described by Levelt, Roelofs, and Meyer distinguishes conceptual preparation, lemma selection, phonological encoding, phonetic planning, and articulation. This decomposition permits researchers to ask where an error arises and which information is accessible before sound reaches a listener. A staged production theory and a monitoring theory answer different questions: the former identifies representational transformations; the latter specifies a process by which their reliability is assessed and future processing is altered. A complete explanation must allow both successful fluent production and the interruption of a problematic plan Levelt et al. (1999) Postma (2000).
The perceptual-loop model locates the primary monitor in language comprehension, using access to inner speech and auditory output. Production-based proposals instead emphasize local signals within formulation, motor prediction, or response competition. A third possibility is layered monitoring, where different error classes have different useful signals and response times. Postma's comparative review organized these alternatives around access, automaticity, available feedback, and the conditions under which control can intervene. The evidence available at that point favored a substantial role for perception-based evaluation while allowing additional automatic components Postma (2000) Hartsuiker & Kolk (2001).
Subsequent experiments and models made the older opposition more precise. Lexical conflict can provide an internal estimate of uncertainty before speech is emitted; auditory and somatosensory feedback can reveal motor deviations during articulation; communicative appropriateness involves broader conceptual and interpersonal criteria. No single detector must provide all signals with equal sensitivity. This is an integration problem rather than a proof that every candidate architecture is equivalent: a model earns support through distinct, discriminating predictions about timing, lesion patterns, feedback disruption, and repair success Nozari et al. (2011) Tourville & Guenther (2011) Nozari (2025).
The Perceptual-Loop Theory and Inner Speech
The perceptual-loop account proposes that a speaker can analyze a planned expression through processes also implicated in understanding speech from other people. An inner route allows evaluation of prepared phonological material before the corresponding sound is emitted; an outer route evaluates auditory output after articulation. An additional conceptual pathway can evaluate whether an intended message fits the situation. The crucial theoretical commitment is the involvement of comprehension mechanisms in an assessment of self-generated linguistic representations, with time needed to transform plans into monitorable form Levelt (1983) Postma (2000).
A strong interpretation predicts that impairments to speech comprehension should impair error detection when the monitor must use corresponding representations. Yet comprehension itself comprises multiple functions, including auditory discrimination, phonological recognition, lexical understanding, and explicit judgment, and clinical tests cannot treat them as a single undifferentiated capacity. Demonstrating shared functional architecture requires convergent task-specific associations or causal perturbations; overlap in measured activation alone leaves several interpretations open Nozari et al. (2011) Gauvin et al. (2016).
Hartsuiker and Kolk constructed a computational treatment of perceptual-loop monitoring that made the timing of interruption and correction explicit. In their account, repair preparation may proceed in parallel with the decision to interrupt, offering an explanation for variation in observable cutoffs. This contribution matters methodologically: a claim that inner speech is monitored should specify the representational delay, where repair starts, and what changes when a speaker lacks access to an auditory channel. Temporal structure is not an incidental property of the theory; it is part of its mechanism Hartsuiker & Kolk (2001) Nooteboom & Quené (2017).
Production-Based and Conflict-Based Monitoring
Production-based monitoring proposes that information within the generation process can disclose when a selected representation is unreliable. In the conflict formulation developed by Nozari, Dell, and Schwartz, simultaneous activation of competing lexical or phonological candidates creates a measurable signal of response conflict. A monitoring system can exploit that signal without first converting an intended word into a perceived internal utterance. A damaged production network may therefore degrade monitoring even when basic speech comprehension remains relatively preserved Nozari et al. (2011).
The model connects psycholinguistic production to control research, but the inferential step deserves care. High competition can accompany correct decisions, while some errors may result from a strongly activated incorrect candidate with little concurrent competition. A detector must therefore specify decision threshold, temporal integration, reference targets, and noise. Successful simulation of an empirical correlation shows that the proposed mechanism can generate the data; causal identification requires further experiments that selectively change conflict while holding the actual error opportunity constant Nozari et al. (2011) Nozari (2025).
The aphasia evidence reported by Nozari and colleagues is especially informative because it compared production-based predictions, comprehension measures, and observed detection. It supported the conflict account under the study's tests and challenged a simple shared-comprehension necessity claim. Its scope is the sampled patients, modeled production deficits, and applied error-detection tasks; generalization to all conversational speech, all linguistic units, or every neurological condition remains a separate empirical question. Production conflict is a credible informational source, whose explanatory reach must be established across tasks Nozari et al. (2011) Postma (2000).
Distributed Monitoring, Motor Prediction, and Multilevel Control
Monitoring at the articulatory level poses a different problem from deciding whether a lexical choice conveys the correct meaning. A speaker possesses auditory predictions about expected sound, proprioceptive information about movement, and sensory outcomes once vocalization begins. Predictive motor accounts compare intended trajectories or expected sensory consequences with estimates derived from ongoing feedback. Discrepancy signals can recruit rapid compensation even when the person cannot verbally identify a word-level error. The units of analysis here include formant patterns, pitch, timing, and articulator state Houde & Jordan (1998) Houde & Nagarajan (2011) Tourville & Guenther (2011).
The DIVA framework specifies learned auditory and somatosensory targets, feedforward command generation, and feedback-error correction within an explicit computational and neuroanatomical system. State feedback control likewise emphasizes estimates of articulatory state constructed from motor predictions and delayed sensory observations. These models are particularly informative about how speech movements remain stable under acoustic or mechanical perturbation, yet a corrected pitch trajectory should not automatically be equated with detecting a mistaken lexical item. Different representational levels require distinct target and error definitions Tourville & Guenther (2011) Houde & Nagarajan (2011).
A multilevel research program asks whether lexical, phonological, and articulatory monitoring share control infrastructure or merely interact through functional dependencies. Imaging overlap with domain-general performance monitoring may indicate coordinated recruitment rather than one centrally located error detector. Models can accommodate fast automatic adjustment alongside slower deliberate self-repair if they specify information flow and causal timescales. The strongest synthesis therefore tests cross-level transfer, double dissociations, and the circumstances under which one source compensates for reduced reliability of another Gauvin et al. (2016) Nozari (2025) Nooteboom & Quené (2017).
Node Structure, Feedback, and Lexical Bias
Node-structure approaches and interactive activation models link errors to activation dynamics in the production network. Candidate segments and words may continue to compete as an utterance is prepared, and feedback can strengthen actual lexical possibilities. Under these conditions, lexical bias—the greater occurrence or survival of certain word-form errors compared with nonword errors—can reflect both the process that generates errors and the process that filters them. Observed distributions cannot uniquely identify a monitoring mechanism unless the competing generation account is independently constrained Postma (2000) Nooteboom & Quené (2008).
Nooteboom and Quené directly examined whether lexical bias should be attributed to self-monitoring of inner speech, immediate interactive feedback, or their combination. Their experimental tests highlight a general identification difficulty: a category-dependent change in completed slips may originate from selecting a candidate, intercepting an invalid one, or repairing a detected error. Monitoring models should predict the entire sequence of error opportunities, cutoffs, completed slips, and corrections, not just the proportion of lexical outcomes Nooteboom & Quené (2008) Nooteboom (2005).
An ecologically realistic account also acknowledges that lexical status, familiarity, semantic acceptability, phonological neighborhood, and communicative context vary together. Experiments must manipulate these variables separately where possible and describe the exact stimulus distribution. The literature is most informative when it supplies independent evidence about the latent candidate set, temporal dynamics, and detection process. Such constraints allow interactive production and monitoring hypotheses to be compared rather than resolved by assigning any one observed bias to a favorite component Motley & Baars (1976) Nooteboom & Quené (2020) Nozari (2025).
Spontaneous Self-Repairs and Their Temporal Anatomy
Spontaneous speech supplies phenomena that laboratory designs can miss: revisions of communicative appropriateness, syntactic retreat, abandoning an utterance, and repair without audible error. Levelt's analysis distinguished the original trouble source from interruption, editing term, and replacement, and showed that the structural relation between a truncated utterance and its repair is systematically organized. This made spontaneous data evidence about grammar and control as well as speech performance. The corresponding weakness is inferential: researchers often cannot observe the speaker's original intention independently of the correction itself Levelt (1983).
Temporal measures help divide the inference into testable stages. Error-to-cutoff latency starts with the first acoustically identifiable erroneous material and ends when the speaker interrupts; cutoff-to-repair latency measures a subsequent interval before the replacement begins. Their distributions depend on segment type, position, prosodic continuation, speech rate, and access to alternative words. A short error-to-cutoff interval supports early detection under some conditions, yet rapid interruption can also reflect an unusually salient external signal. Timing alone must be interpreted jointly with experimental manipulations Hartsuiker & Kolk (2001) Nooteboom & Quené (2017) Nooteboom & Quené (2020).
Production may continue after detection because a motor plan is already committed or because interrupting inside a particular unit is costly. Conversely, speakers can pause while searching for an alternative after having located the problem. These alternatives make it essential to avoid translating all pauses into monitoring events. In naturalistic annotation, independent coders need operational criteria for errors, repairs, false starts, editing terms, and incomplete constituents, with a record of disagreements and examples from the speech community being studied Levelt (1983) Postma (2000) Nozari (2025).
Laboratory-Induced Slips and Phonological Selection
The SLIP family of tasks manipulates the phonological environment before a target utterance, increasing the probability of selected speech-sound exchanges. This design offers a known intended form and a planned error alternative, making it possible to compare detected and completed errors under controlled cue conditions. The original semantic-bias studies showed that outcome properties beyond segmental overlap can influence slip frequencies. They established experimentally that errors expose interactions between planning constraints and the linguistic acceptability of possible outputs Motley & Baars (1976).
Induced errors are valuable precisely because their rates can exceed those in fluent ordinary production, yet the induced population may not resemble the full natural error distribution. Repeating cue pairs can create strategic attention to initial sounds, and instructions may change what participants consider an error. Moreover, a self-reported error can be missed, and a difficult word can elicit a hesitation without an actual segment substitution. Each experiment needs separate counts of elicitation opportunities, unintended outputs, cutoffs, and fully corrected responses Nooteboom (2005) Nooteboom & Quené (2017).
Careful factorial designs distinguish lexical from phonological biases by comparing word-word and nonword-nonword candidate outcomes, phonetic similarity, target position, and articulation timing. They should avoid interpreting the relatively low number of nonword slips as automatic proof of prearticulatory editing, because interactive activation could also bias which competitors emerge. Studies that manipulate monitoring or feedback separately from induction are most discriminating. Nooteboom and Quené's sequence of experiments demonstrates why visible self-correction and completed-error frequencies need to be modeled jointly Nooteboom & Quené (2008) Nooteboom & Quené (2020).
The Lexical Bias and Phonetic Similarity Problems
Whether an error is detected depends partly on its perceptual and representational distance from the intended target. A substitution that changes a word into a highly similar phonetic neighbor might evade notice even when the produced form is incorrect; a large acoustic contrast can provide a stronger signal. Lexicality and familiarity also shape the plausibility of the candidate and the speaker's expectation of what should be uttered. These variables interact across production and monitoring and can create apparent differences in editing rates Nooteboom & Quené (2008) Quené & Nooteboom (2024).
In their 2024 reanalysis, Quené and Nooteboom reported distinct classes of early and late detection and related them partly to the size of the phonetic discrepancy between error and target. Their interpretation distinguishes an earlier preparation stage involving auditory target representations from a later stage concerning motor commands. Since the study reanalyzed previously collected experimental responses, its findings refine an existing evidence base rather than constitute a wholly independent replication. The proposed representational stages remain hypotheses to test using additional paradigms Quené & Nooteboom (2024).
The correct comparison is between models that jointly predict error occurrence and error awareness. A larger target-to-output contrast may improve detection while reducing production errors through more distinct phonological selection; observed frequencies then cannot isolate perception. An experiment can address the confound by controlling the probability of errors, parametrically altering auditory or phonetic contrasts, and measuring detection sensitivity conditional on observed errors. Such approaches connect linguistic error patterns to signal-detection logic and improve the interpretability of apparent lexical-bias effects Nozari et al. (2011) Nooteboom & Quené (2017) Quené & Nooteboom (2024).
Attentional Load, Time Pressure, and Production Fluency
Speech planning and evaluation occur while the speaker allocates attention to content, timing, interlocutors, and environmental demands. Divided attention can affect message formulation, candidate competition, the amount of time available for internal checking, and the decision to interrupt. Oomen and Postma experimentally combined picture-story narration with a tactile-form recognition task in eighteen nonstuttering adults; filled pauses and repetitions increased under divided attention. That result establishes an effect on surface fluency under the tested conditions, not a one-to-one mapping from every pause to a covertly corrected error Oomen & Postma (2001a).
Time-pressure studies likewise highlight the difficulty of isolating a monitoring mechanism from the speed of upstream encoding. Faster required speech may increase error generation, compress the interval for intervention, alter attention allocation, or change the number of acceptable restarts. A suitable design collects separate measures of error frequency, detected errors, repair latency, and output quality, with pressure and task complexity manipulated independently. Oomen and Postma's work motivates this distinction but cannot settle which pathway dominates every speaking condition Oomen & Postma (2001b) Postma (2000).
Attentional constraints connect speech monitoring with executive control. Theoretical proposals involving domain-general conflict signals suggest that the allocation of control to speech may depend upon task priorities, while learned articulatory corrections can proceed with little focal attention. The prediction is a dissociation between deliberate editing of words and automatic compensation to perturbation, although the exact boundary depends on the paradigm. Carefully matched dual-task experiments should distinguish workload costs on generation from workload costs on detection Nozari et al. (2011) Houde & Nagarajan (2011) Nozari (2025).
The Timing of Detection: Early and Late Pathways
Self-repair timing provides one of the most direct behavioral windows onto internal speech control. Nooteboom and Quené experimentally induced segmental slips and compared error detection and repair in conditions with and without normal auditory feedback. They reported a bimodal distribution of error-to-cutoff intervals, consistent with earlier and later pathways, along with differences in repair preparation time. The inference is more informative than counting the number of corrections: a distribution can reveal distinct temporal regimes even when the mean cutoff masks them Nooteboom & Quené (2017).
The same findings place a constraint on the assumption that a late correction requires hearing the erroneous sound. Under the conditions of their experiments, detection remained possible without normal audition, suggesting that postarticulatory information includes additional pathways or representations. Removing auditory feedback is not equivalent to eliminating somatosensory input, internal prediction, or knowledge of the planned motor sequence. The result therefore qualifies a purely auditory interpretation of external monitoring while supporting a richer account of access to the speaker's own actions Nooteboom & Quené (2017) Tourville & Guenther (2011).
Early and late detection labels are theoretical summaries of observed temporal classes. Their interpretation depends on reliable onset coding, the distribution of actual errors, timing correction for phonetic differences, and a model of motor execution latency. A bimodal distribution does not alone prove two anatomically separate detectors; heterogeneous decisions or processing durations can also produce multimodality. Independent perturbations of prediction, hearing, and articulatory feedback are required to distinguish true processing pathways from descriptive classes Nooteboom & Quené (2017) Quené & Nooteboom (2024) Nozari (2025).
Repair Planning, Competitor Persistence, and Interruption
Detection of an error and preparation of a replacement are conceptually distinct. Once a speaker detects an unsuitable segment, the intended competitor may remain available in the activation landscape; alternatively, the correct word may require renewed retrieval. Nooteboom and Quené proposed that competitors surviving from earlier preparation can supply rapid repairs, with their accessibility diminishing over time. Their analyses of induced slips mainly supported some proposed patterns but did not establish the account uniformly across error categories Nooteboom & Quené (2020).
A timing model must represent overlap between monitoring, stopping, and replacement planning. If the speaker prepares a correction before halting, a short cutoff-to-repair delay need not imply instantaneous lexical retrieval. If interruption precedes discovery of a replacement, an extended pause may follow. Observed pauses are further shaped by segment completion, respiratory timing, and the syntactic requirements of re-entering a phrase. Hartsuiker and Kolk's parallel-interruption proposal illustrates the need to specify these dependencies rather than treating self-repair as a serial three-box process Hartsuiker & Kolk (2001) Levelt (1983).
The quality of a repair supplies an independent constraint. A replacement can be correct, partially improved, another error, or a strategic circumlocution. Models in which correction is selected from a competition set predict systematic relations among the error's phonological form, candidate availability, repair latency, and final outcome. Tests should retain failed repairs and abandoned attempts in the denominator, because restricting analysis to successful self-corrections overstates the efficiency of monitoring and obscures the candidate dynamics that may matter most Nooteboom (2005) Nooteboom & Quené (2020) Quené & Nooteboom (2024).
Auditory Feedback Perturbation and Sensorimotor Compensation
Experiments that alter the sound returned to a speaker reveal how speech control responds to an externally induced discrepancy. In a landmark study, Houde and Jordan shifted auditory feedback during production and showed compensatory changes consistent with sensorimotor adaptation. Such adaptation indicates that the nervous system updates relationships between action and expected sensory consequences. It does not, by itself, establish that the speaker consciously detected a lexical or phonological substitution, nor that the same mechanism causes a speaker to restart a sentence after choosing the wrong word Houde & Jordan (1998).
Pitch-shift and formant-shift manipulations can separate rapid feedback corrections from longer-term recalibration. A perturbation changes a sensory target or an expected outcome while leaving the intended word constant, making it a strong test of motor feedback control. The acoustic system, however, includes delays, masking, bone conduction, somatosensory feedback, and adaptation to repeated disturbances. Researchers should report perturbation size, onset, duration, latency, microphone and earphone characteristics, and relevant baseline conditions, because changes in these parameters can alter both measured compensation and participant awareness Tourville & Guenther (2011) Houde & Nagarajan (2011).
Studies of speaking-induced auditory suppression indicate that self-generated sound is processed differently from playback. Heinks-Maldonado and colleagues observed a smaller auditory N100 response during unchanged self-voice feedback than during altered or substituted feedback. That pattern is consistent with a precise prediction signal reducing expected sensory responses, but electrophysiological attenuation does not uniquely identify a linguistic error detector. It belongs to the auditory-motor control evidence layer, which must be linked experimentally to conscious detection and lexical repair before a single-mechanism claim is warranted Heinks-Maldonado et al. (2005) Houde & Nagarajan (2011).
Inner Monitoring Without Normal Audition
The availability of internal monitoring is indicated by interrupted or repaired utterances before a completed audible error, yet the interpretation of such events is delicate. A speaker can assess a phonological plan before articulating it, and can also experience conflict or uncertainty in lexical retrieval. Experiments using noise masking or reduced auditory feedback seek to determine whether self-monitoring survives loss of reliable acoustic evidence. The survival of detection supports internal information sources, while the failure of a particular task may reflect changed attention or the difficulty of hearing instructions Nooteboom & Quené (2017) Gauvin et al. (2016).
Noise masking does not make speech production acoustically silent: vibrations, tissue-borne sensations, and motor predictions remain, and the masker can itself change speech loudness or effort. Accordingly, the claim that error detection is entirely independent of auditory processes requires convergent evidence that separates central speech-perception representations from peripheral hearing. The experiments of Nooteboom and Quené establish a meaningful dissociation in their task, but cannot by themselves rank every internal pathway in all forms of speech Nooteboom & Quené (2017) Quené & Nooteboom (2024).
The practical research strategy is to couple masking with measures of the exact error contrast and temporal outcome, then compare performance across internal speech judgments, overt articulation, and listening to another voice. Adding articulatory interference or targeted feedback delays offers further constraints. This design helps resolve whether an early correction uses a planned phonological representation, a motor efference copy, conflict among alternatives, or a combination. It also guards against equating a subjective report of inner speech with direct observation of a specific intermediate neural code Postma (2000) Nozari et al. (2011) Nozari (2025).
Neural Correlates of Detecting One's Own Speech Errors
Neuroimaging tests address whether internal error detection recruits language-perception cortex, domain-general control systems, or specialized motor-monitoring circuits. Gauvin, De Baene, Brass, and Hartsuiker compared internally detected errors during a tongue-twister production task with speech perception judgments. Their fMRI analysis implicated an overlapping network including pre-supplementary motor area, dorsal anterior cingulate cortex, anterior insula, and inferior frontal gyrus. This is consistent with shared task-control resources, while the inference from overlap to identical computation remains underdetermined Gauvin et al. (2016).
Imaging contrasts matter as much as activated regions. Error trials can differ from correct trials in difficulty, effort, awareness, head movement, speech output, surprise, or likelihood of later repair. An activation contrast therefore requires well-matched baselines and an explicit statement about whether the participant detected an error, produced an error, or both. Separating internal monitoring from auditory perception can help, although masking, button reports, and production demands introduce their own costs. The anatomical distribution of responses provides constraints on models rather than a complete functional localization of monitoring Gauvin et al. (2016) Nozari et al. (2011).
Domain-general performance monitoring offers a plausible functional analogy: conflicting candidates create uncertainty and recruit greater control, while successful detection may modify subsequent policy. The analogy should be tested with trialwise conflict estimators and individual differences in performance rather than inferred solely from the presence of anterior cingulate activity. Different neuronal populations can occupy the same macroscopic fMRI region, and a widely activated region can support several operations. Strong neural claims combine temporal measurements, lesion or stimulation evidence, and predictive modeling Nozari et al. (2011) Nozari (2025) Gauvin et al. (2016).
Auditory–Motor Networks and Predicted Sensory Consequences
The neural control of speech articulation depends on sensorimotor loops operating across cortex and connected circuits. Tourville and Guenther's DIVA model identifies mappings among learned sound targets, motor commands, sensory feedback, and error correction. It is a specified theory that links computational functions to candidate neural systems; correspondence between modeled activity and brain imaging is an empirical test of its proposals, not automatic verification of every anatomic assignment Tourville & Guenther (2011).
Magnetoencephalographic work on feedback perturbations offers temporal information unavailable from static region lists. In pitch-shift paradigms, responses related to feedback modification have been reported in auditory, premotor, and temporoparietal regions, while activity associated with expected self-produced sounds can be suppressed. These patterns are compatible with prediction and correction during vocal control. However, the size and latency of a neural response cannot be read directly as a measure of word-level self-awareness Heinks-Maldonado et al. (2005) Houde & Nagarajan (2011).
The auditory system can compare expected acoustic consequences with actual outcomes while somatosensory information estimates articulatory state. The balance of those signals plausibly varies with context and perturbation. A motor adjustment can occur without any lexical substitution, and a lexical correction can occur without a manipulated acoustic error. An integrated neural account therefore needs explicit routes between phonological target selection, sensory prediction, discrepancy estimation, and language-specific repair decisions Tourville & Guenther (2011) Houde & Nagarajan (2011) Nozari (2025).
Aphasia and the Dissociation of Production, Comprehension, and Detection
Aphasia creates a natural but heterogeneous test of monitoring models. Damage can disrupt speech production, auditory word comprehension, phonological selection, language-specific working memory, and general executive capacity in different combinations. A simple perceptual-loop theory predicts some dependence of self-detection on language comprehension. Nozari, Dell, and Schwartz used computationally characterized word production and aphasic performance to show that an account based on production conflict predicted detection better than the comprehension measures included in their design Nozari et al. (2011).
The study is theoretically important because it does not merely correlate overall language impairment with error monitoring. Its explanatory proposal connects the particular production state to the likelihood of an internal error signal. Yet patient groups vary in lesions, chronicity, compensation, task demands, and response criterion. A null association with a comprehension test is not equivalent to proof that comprehension never contributes, and performance in a constrained naming task does not automatically generalize to conversational repair, sentence editing, or emotional self-censorship Nozari et al. (2011) Postma (2000).
A strong clinical research design analyzes errors conditional on production opportunities and distinguishes production severity from detection sensitivity. It should characterize auditory comprehension at the relevant representational level, describe whether a judgment is explicit or inferred from spontaneous correction, and account for motor and executive limitations. Longitudinal change could then test whether improvements in underlying production reliability predict monitoring independently of comprehension training. Such tests would advance causal explanation while treating clinical heterogeneity as part of the research object Nozari et al. (2011) Nozari (2025).
Disfluency, Stuttering, and the Limits of Covert-Repair Accounts
The covert repair hypothesis proposed by Postma and Kolk connects some disfluencies to interruption and reconstruction of prearticulatory plans. Under this view, repeated phonological difficulties can provoke internal attempts at correction whose side effects become audible as repetitions, blocks, or delays. The hypothesis was historically influential because it assigned a proposed mechanism to otherwise heterogeneous speech interruptions and yielded predictions about when problems in planning might surface Postma & Kolk (1993).
Clinical disfluency has multiple relevant dimensions, including speech motor timing, planning load, learned responses, anticipation, communication context, and individual developmental history. The existence of covert correction in fluent speakers does not establish that stuttering is generally caused by accumulated undetected phonological errors. To evaluate the hypothesis, studies need independent measures of planning-error prevalence, timing and accuracy of internal corrections, and disfluency phenotypes, ideally with designs that distinguish initiation and stopping deficits from error-detection mechanisms Postma & Kolk (1993) Nozari (2025).
Disfluency should not be scored as an error merely because speech is interrupted, and an interrupted error should not be classified as a clinical stuttering event without relevant behavioral criteria. This distinction matters for both scientific validity and responsible clinical interpretation. Current integrative approaches favor assessing the functions of monitoring, control, motor coordination, and communicative demand separately. The present monograph draws theoretical and methodological implications from published studies and makes no individualized diagnostic or treatment recommendations Oomen & Postma (2001a) Postma & Kolk (1993) Nozari (2025).
Individual Differences, Expertise, and Language Experience
Monitoring accuracy varies with the representations available to the speaker and the demands of the task. A familiar word provides a more stable target for comparison than an unfamiliar form; technical language, rapid naming, or rare phonological sequences can place greater demands on selection and control. The same person can generate more errors while maintaining a high probability of detecting each one, so raw error rates cannot establish superior or inferior monitoring. Individual differences are best described through separate measures of generation, detection, criterion, correction latency, and final intelligibility Levelt et al. (1999) Nozari et al. (2011) Nozari (2025).
Expertise may alter both anticipatory planning and the cost of intervening. A trained speaker can anticipate risky phonological sequences or restructure discourse before an error arises, making overt correction less frequent without necessarily strengthening an internal detector. Conversely, a practiced articulatory movement can compensate rapidly for perturbation with minimal explicit deliberation. Studies of expertise should therefore include matched tasks with calibrated difficulty, assess speech rate, and report whether the benefit reflects preparation, reliable execution, earlier detection, or a more effective recovery strategy Tourville & Guenther (2011) Houde & Nagarajan (2011).
Cross-language comparisons must account for phonotactics, morphological complexity, lexical frequency, syllable structure, and conventions for restarting an utterance. A repair classified as a phoneme substitution in one language may involve a morpheme boundary or different phonological representation in another. Such differences are scientifically useful if coding is explicit, but they complicate the transfer of error rates and cutoff timings from a small set of well-studied languages. Claims about linguistic universality require corresponding evidence across language communities Postma (2000) Nozari (2025).
Communicative Goals, Semantic Editing, and Social Consequences
Speakers sometimes revise utterances because a message is factually wrong, socially inappropriate, ambiguous, or insufficiently specific. These revisions involve a broader target than phonological accuracy: they compare planned meaning and expected consequences with current communicative goals. Levelt's distinction between error repair and appropriateness repair remains valuable because a sentence can be grammatical and flawlessly articulated while still requiring revision. The scientific task is to identify what criterion the speaker used and when it became available Levelt (1983) Postma (2000).
Semantic constraints can shape the population of produced slips. Motley and Baars demonstrated that properties of possible error outcomes can affect induced errors, highlighting the importance of meaning even in a phonologically constrained task. The result does not warrant universal psychoanalytic readings of verbal slips; experimental semantic priming and contextual set provide plausible mechanisms without assuming that every unintended word reveals a hidden personal wish. Interpretation requires controlled alternatives and an explicit causal account Motley & Baars (1976) Nooteboom & Quené (2008).
Communicative repair at the interpersonal level extends beyond the current research object's focus. A listener may request clarification, infer intended meaning, or modify their understanding following a speaker's correction. Those processes feed back into the speaker's decisions but also involve pragmatic common ground and turn-taking. This dossier concentrates on what the speaker monitors and changes within an utterance, using the social outcome as relevant context rather than treating the entire dialogue as an internal error-monitoring process Levelt (1983) Nozari (2025).
Development, Aging, and the Problem of Generalization
Developmental research on speech monitoring must distinguish the emergence of linguistic targets from the emergence of metacognitive access to one's own errors. A child may improve a word after hearing a correction, interrupt a planned utterance, or recognize a discrepancy after articulation; these observations recruit different learning and control capacities. Age comparisons also vary with vocabulary growth, motor maturation, and task instruction comprehension. The present evidence corpus supports a mechanism-centered framework for studying these changes but does not establish a universal age of emergence for any single detector Postma (2000) Tourville & Guenther (2011).
Studies across adulthood face similar measurement problems. Changes in hearing, lexical retrieval, speaking rate, attention, and articulatory reliability can move error prevalence and repair behavior in opposite directions. A slower correction in an older participant may reflect delayed retrieval of a replacement rather than failure to detect the error. Longitudinal designs with stable calibration and independently measured sensory access would help distinguish changes in generation, monitoring, and repair efficacy Nozari et al. (2011) Nozari (2025).
Generalization also requires attention to the task ecology. Reading isolated words, naming pictures, producing induced spoonerisms, and telling autobiographical narratives require different kinds of planning and expose different opportunities for repair. An experiment that elicits phoneme substitutions under strong phonological priming is informative about those conditions; its results do not automatically determine how speakers regulate factual statements in lengthy dialogue. The most useful cross-population evidence compares mechanisms while specifying the representational and situational demands Levelt (1983) Nooteboom & Quené (2017).
Speech Monitoring and Human–Artificial Language Interaction
Artificial language systems introduce a related but distinct control problem: generated verbal output can be checked against intended constraints, detected contradictions, recorded context, and externally provided feedback. These properties make mechanisms of monitoring and repair a useful functional comparison, especially when investigating revision policies, uncertainty estimates, and response verification. The comparison requires operational variables, such as the presence of an independently specified target, an observable error signal, a correction trigger, and measurable improvement after revision Nozari (2025).
Human speech integrates auditory, motor, linguistic, and social processes, whereas different artificial language architectures can operate through token prediction, search, validators, and external tools. Similar outward corrections therefore permit a comparative research question, not an inference of identical neural or experiential mechanisms. Tests should distinguish errors detected during generation from errors identified by a later verification pass, and should evaluate false alarms, successful repairs, computational costs, and robustness under altered context Nozari et al. (2011) Tourville & Guenther (2011).
Within the Artificial Era research horizon, this dossier situates monitoring as a general problem of maintaining the reliability of evolving linguistic action while preserving the specifically psychological evidence about human speaking. Human–Artificial dialogue can provide new behavioral settings for studying how people accept external corrections and redistribute checking effort. Such extensions belong to testable comparative and interaction research; the findings from speech-error experiments do not by themselves establish claims about artificial consciousness, lived inner speech, or a shared psychology of all language-producing systems Nozari (2025) Postma (2000).
Evidence Appraisal: What Each Method Can Establish
Speech-error corpora establish the structure, distribution, and contextual embedding of spontaneous corrections, but their error denominators are imperfect because intended messages are often inferred retrospectively. Laboratory induction provides better control over target and competitor, while changing strategy and natural incidence. Auditory perturbation establishes sensitivity of motor output to feedback mismatch, yet does not isolate lexical self-monitoring. Neuropsychological dissociations can constrain which components are necessary for certain tasks, and neuroimaging can locate candidate networks while leaving the underlying computational operation partly unresolved Levelt (1983) Motley & Baars (1976) Nozari et al. (2011) Gauvin et al. (2016).
Computational models serve an additional role by turning verbal proposals into explicit predictions. A model of conflict must specify candidate activation and decision signals; a perceptual-loop model must specify encoding and inspection latency; a sensorimotor model must specify predictions and correction gains. A model can fit observed errors without proving that brains implement its mechanism, because distinct internal architectures sometimes yield the same output. Valid evaluation tests generalization to new materials, perturbations, patients, and time-resolved measurements, and compares alternative parameterizations under shared constraints Hartsuiker & Kolk (2001) Tourville & Guenther (2011) Nozari (2025).
Evidence quality depends on sample characteristics, transparent inclusion criteria, acoustic annotation reliability, planned statistical comparisons, and the reporting of uncertainty. Particularly important are the rates of undetected errors and false alarms, since a detector that flags every utterance could appear exceptionally sensitive but impair fluency. The present monograph identifies convergent findings and tensions in original studies and scholarly syntheses; it is a critical narrative review without an exhaustive registered search, formal study-level risk-of-bias instrument, pooled effect size, or original experimental data Postma (2000) Nooteboom & Quené (2017) Nozari (2025).
Measurement Design and Reproducible Tests of Monitoring
A rigorous experiment should define an error opportunity before measuring detection. In a picture-naming task, investigators can specify the intended lexical target, record every produced response, and distinguish semantic, phonological, phonetic, and mixed deviations. Two independent raters can code onset and cutoff times from audio with documented resolution and adjudication procedures; a separate response can capture whether the speaker recognized an error without requiring that every recognized event result in repair. Preserving the raw recording, consent constraints, annotations, and exclusions supports later reanalysis Nozari et al. (2011) Nooteboom & Quené (2017).
Signal-detection measures improve interpretation when trials include both errors and correct productions. Hits are genuine errors detected, misses are genuine errors overlooked, false alarms are correct productions rejected, and correct rejections are accepted correct productions. Sensitivity and response criterion can then be separated under a justified measurement model. This framework remains incomplete for continuous speech, where errors can be graded and intention is uncertain, but it clarifies why self-correction count is not a sufficient proxy for monitoring ability Postma (2000) Nozari (2025).
Experimental tests should preregister primary contrasts when feasible, report participant and item variability, examine alternative acoustic alignments, and make model-comparison choices explicit. Repeated measures across lexicality, sound similarity, masking, time pressure, and confidence could estimate how detection changes with signal quality. Clinical analyses should report lesion and language profiles and avoid treating a diagnostic category as a uniform mechanism. Cross-laboratory protocols would be particularly valuable for resolving the extent to which the early and late repair patterns generalize Nozari et al. (2011) Nooteboom & Quené (2017) Quené & Nooteboom (2024).
Unresolved Theoretical Questions and Research Priorities
The first major unresolved issue concerns the information actually compared during internal monitoring. One model requires an interpreted inner speech representation, another detects competition before the word is perceived, and motor architectures compare predicted and observed sensory outcomes. These hypotheses are partly complementary but retain distinct predictions. Experiments that dissociate target confidence, acoustic discrepancy, and lexical competition could determine which signal controls particular classes of interruption and when multiple cues are integrated Postma (2000) Nozari et al. (2011) Tourville & Guenther (2011).
The second issue is how error awareness becomes a control action. Conflict may trigger caution without identifying a specific mistaken segment; auditory mismatch may cause articulatory compensation without a verbal report; and explicit recognition may fail to produce correction when the alternative cannot be retrieved. A more complete science should link subjective detection, motor stopping, candidate reactivation, and listener-facing reconstruction through temporally resolved tasks. Studies of repair candidate persistence and cutoff latency make that connection empirically tractable Hartsuiker & Kolk (2001) Nooteboom & Quené (2020) Quené & Nooteboom (2024).
The third issue is generalizability: from sound slips to conceptual editing, from isolated tasks to real conversation, and from neurologically typical adult speakers to patients, children, multilingual groups, and speakers with altered sensory feedback. Each extension changes the target, error prevalence, and available evidence. Comparative artificial-system research further raises the question of which functional properties of monitoring can be studied independently of human biological implementation. These are productive research directions when constrained by observable hypotheses and explicit domain boundaries Levelt (1983) Nozari (2025).
Conclusion
Speech monitoring is best characterized as an organized family of processes through which speakers evaluate intended, prepared, and executed speech and adjust production when a discrepancy matters. Spontaneous repair reveals linguistic and communicative structure; experimentally induced slips constrain phonological and lexical competition; clinical dissociations test proposed functional dependencies; and feedback perturbations illuminate rapid auditory-motor control. The combined evidence supports multiple useful signals and timescales while leaving the relative contribution of each to particular error classes open Levelt (1983) Postma (2000) Nozari (2025).
The most influential research debate concerns the role of comprehension-based perceptual loops compared with internal production conflict and motor prediction. Aphasia and simulation results support a production-based source of monitoring information under specified tasks, while imaging and sensorimotor findings identify shared and specialized neural networks. Timing studies show that error detection and repair preparation are more complex than a uniform serial sequence and that auditory feedback is not the only resource available to speakers. These contributions should be integrated through explicit models rather than reduced to a single universal monitor Nozari et al. (2011) Gauvin et al. (2016) Nooteboom & Quené (2017).
The central scientific conclusion is methodological as well as theoretical. Valid explanations distinguish errors that occur from errors that are noticed, noticed errors from interrupted speech, and interruption from successful replacement. They specify the target being evaluated, the source of discrepancy information, the temporal window for control, and the observational criterion for repair. This structure makes future cross-language, clinical, neural, and human–artificial comparisons testable while preserving the empirical specificity of human speech production Nooteboom & Quené (2020) Quené & Nooteboom (2024) Nozari (2025).
An integrated research program should measure monitoring at several analytically distinct moments: preparation before articulation, the onset of an observable disruption, the decision to interrupt an utterance, and the success of the resulting correction. Each moment calls for different evidence. Speech-error induction can reveal the conditions under which competing linguistic representations become active; spontaneous conversation can reveal how speakers preserve mutual understanding; neurophysiological measures can constrain the timing of evaluation; and altered auditory feedback can isolate selected components of sensorimotor adjustment Levelt (1983) Nozari et al. (2011) Tourville & Guenther (2011). These measurements should be combined within designs that record the acoustic signal, task context, participant characteristics, and opportunities for repair. A further priority is to compare predictions made before data collection, including predictions about errors that go undetected or corrections that fail. Such designs can clarify when monitoring relies on comprehension, production conflict, or forward estimates, and when several information sources contribute jointly. Their explanatory strength will depend on replication across languages and communicative settings, transparent reporting of excluded responses, and a clear distinction between empirical findings and the broader theoretical model used to interpret them Postma (2000) Gauvin et al. (2016).
References
- Gauvin, H. S., De Baene, W., Brass, M., & Hartsuiker, R. J. (2016). Conflict monitoring in speech processing: An fMRI study of error detection in speech production and perception. NeuroImage, 126, 96–105. https://doi.org/10.1016/j.neuroimage.2015.11.037
- Hartsuiker, R. J., & Kolk, H. H. J. (2001). Error monitoring in speech production: A computational test of the perceptual loop theory. Cognitive Psychology, 42(2), 113–157. https://doi.org/10.1006/cogp.2000.0744
- Heinks-Maldonado, T. H., Mathalon, D. H., Gray, M., & Ford, J. M. (2005). Fine-tuning of auditory cortex during speech production. Psychophysiology, 42(2), 180–190. https://doi.org/10.1111/j.1469-8986.2005.00272.x
- Houde, J. F., & Jordan, M. I. (1998). Sensorimotor adaptation in speech production. Science, 279(5354), 1213–1216. https://doi.org/10.1126/science.279.5354.1213
- Houde, J. F., & Nagarajan, S. S. (2011). Speech production as state feedback control. Frontiers in Human Neuroscience, 5, Article 82. https://doi.org/10.3389/fnhum.2011.00082
- Levelt, W. J. M. (1983). Monitoring and self-repair in speech. Cognition, 14(1), 41–104. https://doi.org/10.1016/0010-0277(83)90026-4
- Levelt, W. J. M., Roelofs, A., & Meyer, A. S. (1999). A theory of lexical access in speech production. Behavioral and Brain Sciences, 22(1), 1–38. https://doi.org/10.1017/S0140525X99001776
- Motley, M. T., & Baars, B. J. (1976). Semantic bias effects on the outcomes of verbal slips. Cognition, 4(2), 177–187. https://doi.org/10.1016/0010-0277(76)90003-2
- Nooteboom, S. G. (2005). Lexical bias revisited: Detecting, rejecting and repairing speech errors in inner speech. Speech Communication, 47(1–2), 43–58. https://doi.org/10.1016/j.specom.2005.02.003
- Nooteboom, S. G., & Quené, H. (2008). Self-monitoring and feedback: A new attempt to find the main cause of lexical bias in phonological speech errors. Journal of Memory and Language, 58(3), 837–861. https://doi.org/10.1016/j.jml.2007.05.003
- Nooteboom, S. G., & Quené, H. (2017). Self-monitoring for speech errors: Two-stage detection and repair with and without auditory feedback. Journal of Memory and Language, 95, 19–35. https://doi.org/10.1016/j.jml.2017.01.007
- Nooteboom, S. G., & Quené, H. (2020). Repairing speech errors: Competition as a source of repairs. Journal of Memory and Language, 111, Article 104069. https://doi.org/10.1016/j.jml.2019.104069
- Nozari, N. (2025). Monitoring, control and repair in word production. Nature Reviews Psychology, 4, 222–238. https://doi.org/10.1038/s44159-025-00417-1
- Nozari, N., Dell, G. S., & Schwartz, M. F. (2011). Is comprehension necessary for error detection? A conflict-based account of monitoring in speech production. Cognitive Psychology, 63(1), 1–33. https://doi.org/10.1016/j.cogpsych.2011.05.001
- Oomen, C. C. E., & Postma, A. (2001a). Effects of divided attention on the production of filled pauses and repetitions. Journal of Speech, Language, and Hearing Research, 44(5), 997–1004. https://doi.org/10.1044/1092-4388(2001/078)
- Oomen, C. C. E., & Postma, A. (2001b). Effects of time pressure on mechanisms of speech production and self-monitoring. Journal of Psycholinguistic Research, 30(2), 163–184. https://doi.org/10.1023/A:1010377828778
- Postma, A. (2000). Detection of errors during speech production: A review of speech monitoring models. Cognition, 77(2), 97–132. https://doi.org/10.1016/S0010-0277(00)00090-1
- Postma, A., & Kolk, H. (1993). The covert repair hypothesis: Prearticulatory repair processes in normal and stuttered disfluencies. Journal of Speech and Hearing Research, 36(3), 472–487. https://doi.org/10.1044/jshr.3603.472
- Quené, H., & Nooteboom, S. G. (2024). Some properties of mental speech preparation as revealed by self-monitoring. Speech Communication, 158, Article 103043. https://doi.org/10.1016/j.specom.2024.103043
- Tourville, J. A., & Guenther, F. H. (2011). The DIVA model: A neural theory of speech acquisition and production. Language and Cognitive Processes, 26(7), 952–981. https://doi.org/10.1080/01690960903498424