Speech

Abstract

Speech is a type of verbal behavior: the motor act of producing spoken language, which unfolds in stages from conceptual preparation through lexical selection and phonological encoding to articulation (Levelt, 1989). Its acoustics follow the source–filter principle, a glottal source shaped by the changing resonances of the vocal tract (Fant, 1960). Slips of the tongue expose the planning units that assemble an utterance (Fromkin, 1971), which articulatory phonology recasts as coordinated gestures (Browman & Goldstein, 1992). Neuroimaging and computational models map the feedforward commands and sensory feedback that steer the articulators toward their acoustic goals (Guenther, Ghosh, & Tourville, 2006; Hickok, 2012). This article surveys how speech is planned, produced, and controlled, and how it is classified within verbal behavior.

Keywords: speech production, source-filter theory, articulatory phonology, speech motor control

Speaking feels effortless, yet a single spoken clause is the output of one of the most intricate motor programs the human body runs. In under a second the speaker settles on a message, retrieves the right words, orders their sounds, and drives roughly a hundred muscles of the chest, larynx, and face in precise coordination to shape a stream of air into a signal another brain can decode. Speech is the behavior that accomplishes this: the production of spoken language. It is distinct from the abstract linguistic system it expresses, from the perception that recovers it, and from the voice that carries it, and it is studied as a planning problem, an acoustic problem, and a motor-control problem at once.

What Speech Is

Speech is the production of spoken language: the coordinated respiratory, laryngeal, and articulatory activity that converts an intended message into an acoustic signal. Three distinctions give the term its precision. First, speech is not language. Language is the abstract, largely amodal system of words and rules; speech is one channel through which that system is externalized, alongside writing and sign. A person can lose speech to paralysis while retaining language intact. Second, speech is not speech perception. Production and perception share representations and brain regions, but speaking and listening are different acts with different problems, and this article concerns the first. Third, speech is not the voice. The voice is the laryngeal sound source and its quality; speech is the full behavior that shapes that source into linguistic structure.

Because it is a behavior, speech is studied through its products and its breakdowns. The timing of its stages, the acoustics of its output, the structure of its errors, and the activity of the brain that drives it are all observable, and each has yielded a distinct research tradition. MeSH classifies the construct under verbal behavior, placing it with the other goal-directed verbal acts rather than with the sensory or linguistic systems it recruits.

Types of Speech

Within the Medical Subject Headings vocabulary, the descriptor Speech is a form of verbal behavior and has two narrower descriptors. They partition the construct along different axes — one acoustic, one perceptual — rather than naming subtypes of the behavior itself, so they are complementary views of speech rather than a single branching taxonomy.

The classification is an indexing convenience, not a theory of speech. MeSH is built to organize the biomedical literature, so its branches track how research is catalogued rather than how the production system is structured. Speech acoustics has no article of its own here and is listed for completeness; speech intelligibility, which has its own treatment, is the perceptual measure of how well the produced signal survives its journey to a listener.

The Architecture of Production

The most influential account of how a message becomes speech is Levelt's staged model of the speaker (Levelt, 1989). It divides production into three broad processing components that run largely in sequence but overlap in time. Conceptualization selects the message to be expressed and its information structure. Formulation translates that message into linguistic form in two sub-stages: grammatical encoding retrieves the words’ abstract lemmas and builds a syntactic frame, and phonological encoding spells out their sound forms and organizes them into syllables and a prosodic contour. Articulation then executes the resulting phonetic plan as movement.

A central claim is that lexical access is a two-step process. The speaker first selects a lemma — a word’s meaning-and-syntax representation — and only then retrieves its lexeme, the phonological form. The computational model WEAVER++ made this architecture explicit and quantitative, deriving the time course of naming from a spreading-activation network in which a lemma is selected before its phonemes are encoded and syllabified (Levelt, Roelofs, & Meyer, 1999). The two-step structure explains why a speaker can be sure of a word’s meaning and grammar while its sound hovers out of reach — the tip-of-the-tongue state, a lemma selected without its lexeme.

Conceptualization. The speaker fixes the idea to express and its information structure — a preverbal message, not yet any particular word.

Lexical access is two-step: the lemma (meaning and grammar) is selected before the lexeme (sound form) is retrieved. A lemma selected without its lexeme is the tip-of-the-tongue state — you know the word but cannot yet say it.

The Acoustics of Speech

Articulation produces sound, and the sound obeys a principle that organizes the entire acoustic study of speech: the source–filter theory (Fant, 1960). Voiced speech begins with a source — the vocal folds chopping the airstream into a periodic buzz rich in harmonics, whose repetition rate is the fundamental frequency heard as pitch. That buzz then passes through a filter: the vocal tract from larynx to lips, an air-filled tube whose resonances amplify certain frequencies and damp others. The resonant peaks the filter imposes are the formants, and their frequencies — chiefly the lowest two — are what distinguish one vowel from another.

The theory’s power is that source and filter are largely independent. Changing the rate of vocal-fold vibration moves the pitch without moving the formants; changing the shape of the tract — raising the tongue, rounding the lips — moves the formants without changing the pitch. The same vowel can be sung at any note, and the same note can carry any vowel, because the speaker controls the two systems separately. This separation is why a vowel’s identity survives changes in intonation and why the worked example below can compute formant frequencies from the tract alone.

The blue lines are the harmonics of the source, spaced at multiples of the pitch; the gold curve is the formant envelope set by the tract. Move the pitch slider and the harmonics slide while the formant peaks stay put; move the tract-length slider and the formants shift while the harmonics stay put. That independence is the whole content of the source–filter theory.

Planning Units and the Evidence from Errors

If production assembles an utterance from smaller pieces, the pieces should occasionally be assembled wrongly, and they are. Fromkin’s analysis of spontaneous speech errors established that slips of the tongue are not random noise but lawful rearrangements that expose the units and rules of the production system (Fromkin, 1971). When a speaker says heft lemisphere for left hemisphere, the exchanged elements are single phonemes; when word onsets swap in a spoonerism, whole consonant clusters move as units; when a reading list becomes a leading list, a feature has anticipated. The errors respect structure: a vowel exchanges with a vowel, an onset with an onset, and a stressed syllable with another stressed syllable. Those regularities are direct evidence that phonemes, syllable positions, and prosodic frames are real stages in planning. The separation of frame from content has a proposed evolutionary basis: MacNeilage’s frame/content theory argues that the syllabic frame derives from the rhythmic open–close cycle of the mandible seen in ingestive and communicative behavior, into which segmental content is later inserted — tying the architecture of speech planning to the motor system it grew out of (MacNeilage, 1998).

A later tradition reinterpreted the units themselves. Articulatory phonology argued that the primitives of speech are not static phoneme symbols but gestures — dynamically defined constriction actions of the lips, tongue, and glottis, coordinated in time (Browman & Goldstein, 1992). On this view, much of what looks like phonological rule — assimilation, deletion, the blurring of sounds in fast speech — falls out of how gestures overlap and reduce, making the gap between abstract phonology and physical articulation smaller than the symbol-based account assumed.

The Neural Control of Speech

Speaking is a control problem: the nervous system must drive the articulators to acoustic and somatosensory targets it can only partly foresee, correcting as it goes. The DIVA model makes this explicit, casting speech production as the interaction of a feedforward controller that issues learned motor commands with feedback controllers that compare the resulting sound and feel against expected targets and correct any mismatch (Guenther, Ghosh, & Tourville, 2006). The model grounds each component in specific cortical and subcortical regions and reproduces behaviors such as the compensation a speaker makes when auditory feedback is artificially perturbed. Guenther later synthesized two decades of this work into a full neurocomputational account of the speaking brain (Guenther, 2016).

Converging evidence comes from two directions. A meta-analysis of neuroimaging studies traced the spatial and temporal signatures of the word-production components — from conceptual preparation to articulation — onto a reproducible left-lateralized network, giving the staged model a neural time course (Indefrey & Levelt, 2004). In parallel, a dual-stream framework distinguished a ventral stream mapping sound to meaning from a dorsal stream mapping sound to articulation, and argued that the dorsal sensorimotor interface is what lets the brain translate an intended sound into the movements that produce it (Hickok, 2012).

Anticipation — planning unit: single phoneme. A later sound is produced early: the /l/ of list anticipates and replaces the /r/ of reading. The planning unit is a single phoneme.

Errors are not random: a vowel exchanges with a vowel, an onset with an onset, a stressed syllable with another stressed syllable. The regularities are direct evidence that phonemes, syllable positions, and prosodic frames are real stages in the plan.

Worked Example

The source–filter theory makes formant frequencies calculable from the vocal tract treated as a simple acoustic tube. For the neutral vowel schwa, the tract is roughly uniform in cross-section and behaves as a tube closed at the glottis and open at the lips — a quarter-wave resonator. Such a tube resonates at odd multiples of its lowest resonance, with formant n given by

F_n = (2n − 1) × c / (4L),

where c is the speed of sound (about 35,000 cm/s) and L is the tract length. Take an adult tract of L = 17.5 cm:

- F₁ = 1 × 35,000 / (4 × 17.5) = 35,000 / 70 = 500 Hz - F₂ = 3 × 35,000 / 70 = 105,000 / 70 = 1,500 Hz - F₃ = 5 × 35,000 / 70 = 175,000 / 70 = 2,500 Hz

The model predicts formants near 500, 1,500, and 2,500 Hz for schwa — close to the values actually measured for a neutral vowel, and a striking return for so crude a tube. The spacing of 1,000 Hz is a direct consequence of the quarter-wave geometry. Now lengthen the tract to L = 20 cm, as for a larger speaker or rounded lips: every formant scales down by the factor 17.5 / 20, giving 437.5, 1,312.5, and 2,187.5 Hz. Because the source is untouched, the pitch is unchanged; only the filter has moved, lowering the formants and shifting the perceived vowel. The arithmetic shows in miniature why formants track the tract and pitch tracks the source — the independence that source–filter theory asserts.

Discussion

The study of speech has advanced by treating one behavior as three problems. The planning problem asks how a message becomes an ordered string of linguistic units, and the staged and spreading-activation models answer it with a lemma-then-lexeme architecture whose predictions survive both chronometric experiments and the structure of natural errors. The acoustic problem asks how articulation produces a decodable signal, and source–filter theory answers it with an independence between source and filter that organizes a century of phonetics. The control problem asks how the brain drives the articulators accurately in real time, and feedforward–feedback models answer it with a loop that leans on prediction and corrects against the senses.

These are not rival accounts; they describe different cross-sections of the same act, and the field’s progress has come from connecting them. Articulatory phonology narrows the gap between the planning units and the articulatory movements. The DIVA framework ties the control loop to the acoustic goals that source–filter theory defines. Neuroimaging maps the planning stages onto the network the control models implement. The remaining tension is the one every stage shares: speech is at once a discrete, symbolic, linguistic object and a continuous, physical, motor one, and no single description has yet captured both faces without seams.

Current Directions

The most active contemporary work formalizes speech production as control theory. A comparative review laid out the modern model landscape — DIVA, state-feedback control, task-dynamic and optimal-control accounts — in a common architecture-and-properties framework, clarifying where the models agree, where they diverge, and which behavioral and neural data discriminate them (Parrell, Lammert, Ciccarelli, & Quatieri, 2019). The shared move is to treat the speaker as an internal-model controller that predicts the sensory consequences of its own commands and uses the prediction, not just the raw feedback, to stay on target.

A closely related line probes exactly what sensory feedback does. By modeling how auditory and somatosensory signals are weighted and integrated during speaking and during motor learning, this work asks when the system trusts its feedforward plan and when it defers to feedback, and how the balance shifts as a new articulation is learned (Parrell & Houde, 2019). The open problems are to unify these controllers with the linguistic planning stages above them, and to extend single-utterance models to connected, conversational speech — the register in which speech is actually used.

Common Misconceptions

Speech and language are the same thing.

They are separable. Language is the abstract system of words and rules; speech is one way of externalizing it. The dissociation is clinical fact: a person can lose fluent articulation while retaining language, and command language fully through writing or sign (Levelt, 1989).

Vowel pitch and vowel identity are produced by the same adjustment.

They are controlled independently. Pitch is set by the rate of vocal-fold vibration (the source); vowel identity is set by the tract resonances, the formants (the filter). The same vowel can be sung at any pitch (Fant, 1960).

Slips of the tongue are random mistakes.

They are highly structured. Sounds exchange with like sounds and onsets with onsets, respecting the units of planning; that lawfulness is precisely why errors are evidence about the production system (Fromkin, 1971).

Fluent speech is driven moment to moment by feedback.

Mostly it is not. Feedback loops are too slow to steer every gesture; fluent speech relies on accurate feedforward commands, with feedback dominating only when the plan is new or the signal is perturbed (Guenther, Ghosh, & Tourville, 2006).

Glossary

Articulation.

The execution stage of speech, in which the phonetic plan is carried out as coordinated movement of the respiratory, laryngeal, and supralaryngeal structures.

Articulatory phonology.

A theory in which the primitive units of speech are dynamically defined articulatory gestures coordinated in time, rather than static phoneme symbols.

Conceptualization.

The first stage of production, in which the speaker selects the message to be expressed and its information structure before any linguistic form is chosen.

Feedforward command.

A learned, predictive motor command issued to the articulators in advance of any feedback, responsible for most of fluent speech.

Formant.

A resonant peak in the speech spectrum imposed by the vocal tract; the lowest two formants chiefly determine vowel identity.

Formulation.

The stage that converts a message into linguistic form, comprising grammatical encoding (lemmas and syntax) and phonological encoding (sound forms and syllables).

Fundamental frequency.

The repetition rate of vocal-fold vibration, heard as the pitch of the voice; a property of the source, independent of the formants.

Gesture.

In articulatory phonology, a goal-directed constriction action of an articulator, defined over time, serving as the basic unit of speech production.

Lemma.

A word’s abstract representation of meaning and syntax, selected in the first step of lexical access before its sound form is retrieved.

Lexeme.

A word’s phonological form, retrieved in the second step of lexical access once its lemma has been selected.

Phoneme.

The smallest contrastive sound unit of a language; a frequent unit of exchange in speech errors.

Source–filter theory.

The account of speech acoustics in which a glottal sound source is independently shaped by the resonant filter of the vocal tract.

Speech error.

An unintended rearrangement of speech units, such as an exchange or anticipation, whose lawful patterns reveal the planning units of production.

Spoonerism.

A speech error in which the onsets of two words exchange, as in heft lemisphere for left hemisphere.

Vocal tract.

The air-filled passage from the larynx to the lips and nostrils whose changing shape forms the acoustic filter of speech.

Key Researchers

Gunnar Fant

(1919–2009). Swedish speech scientist whose Acoustic Theory of Speech Production (1960) gave the source–filter theory its quantitative form, deriving vowel formants from the geometry of the vocal tract and founding modern acoustic phonetics. Wikipedia - Wikidata

Victoria Fromkin

(1923–2000). American linguist at UCLA whose analysis of spontaneous speech errors demonstrated that slips of the tongue are rule-governed, using them to reveal the phonemes, syllables, and frames that serve as the planning units of production. Wikipedia - Wikidata

Louis M. Goldstein

(living). American linguist, co-developer with Catherine Browman of articulatory phonology, which recast the units of speech as dynamically coordinated gestures and narrowed the gap between abstract phonology and physical articulation. Google Scholar - Faculty page - Wikipedia - Wikidata

Frank H. Guenther

(living). American computational neuroscientist at Boston University whose DIVA model cast speech production as feedforward and feedback control grounded in specific brain regions, synthesized in Neural Control of Speech (2016). Google Scholar - Faculty page - Wikipedia - Wikidata

Gregory Hickok

(living). American cognitive neuroscientist at the University of California, Irvine, whose dual-stream framework distinguished the ventral sound-to-meaning and dorsal sound-to-articulation pathways and clarified the sensorimotor interface for speech. Google Scholar

Willem J. M. Levelt

(living). Dutch psycholinguist, founding director of the Max Planck Institute for Psycholinguistics, whose book Speaking (1989) and the WEAVER++ model set the standard staged, two-step architecture of lexical access in production. Google Scholar - Faculty page - Wikipedia - Wikidata

Peter F. MacNeilage

(living). American psycholinguist at the University of Texas at Austin whose frame/content theory proposed that the syllabic frame of speech evolved from the rhythmic mandibular cycle of ingestive behavior, linking speech motor control to its evolutionary origins. Faculty page

Frequently Asked Questions

What is the difference between speech and language?

Language is the abstract system of words, meanings, and grammatical rules; speech is the motor behavior that externalizes that system as sound. The two are separable: a person can lose the ability to articulate while fully retaining language and expressing it through writing or sign, and speech perception can be impaired while production is intact (Levelt, 1989).

What are the stages of speech production?

Production runs through conceptualization, formulation, and articulation. Conceptualization selects the message; formulation encodes it grammatically (retrieving word lemmas and building syntax) and phonologically (spelling out sounds and syllables); articulation executes the phonetic plan as movement. The stages overlap in time but follow this order (Levelt, 1989; Indefrey & Levelt, 2004).

What is source-filter theory?

It is the account of how speech sounds are generated: a source, the buzz produced by the vibrating vocal folds, is shaped by a filter, the resonances of the vocal tract. The filter imposes formant peaks that distinguish speech sounds, and because source and filter are largely independent, pitch and vowel identity can be controlled separately (Fant, 1960).

What are formants?

Formants are the resonant frequency peaks the vocal tract imposes on the speech signal. The frequencies of the lowest two formants chiefly determine which vowel a listener hears. Speakers change formants by reshaping the tract — moving the tongue and lips — without necessarily changing the pitch of the voice (Fant, 1960).

Why do speech errors matter to psychologists?

Because they are lawful. Sounds exchange with similar sounds, word onsets swap as units, and stressed syllables trade with stressed syllables. These regularities reveal the units and ordering rules the production system uses, making errors a natural experiment on the architecture of planning (Fromkin, 1971).

What is articulatory phonology?

A theory that the basic units of speech are not static phoneme symbols but gestures — goal-directed constriction actions of the lips, tongue, and glottis, defined over time and coordinated with one another. It explains much of connected-speech variation as the natural overlap and reduction of gestures (Browman & Goldstein, 1992).

How does the brain control speaking?

Through a combination of feedforward and feedback control. The brain issues learned motor commands that predict the sounds they will produce, then compares the actual auditory and somatosensory results against the expected targets and corrects any mismatch. Fluent speech leans mainly on the feedforward commands (Guenther, Ghosh, & Tourville, 2006; Hickok, 2012).

Is speech production still an active research area?

Yes. Current work formalizes speaking as control theory, comparing models such as DIVA and state-feedback control within a common framework and probing how auditory and somatosensory feedback are weighted during speaking and motor learning. Open problems include joining these controllers to linguistic planning and extending them to conversational speech (Parrell, Lammert, Ciccarelli, & Quatieri, 2019; Parrell & Houde, 2019).

References

Browman, C. P., & Goldstein, L. (1992). Articulatory phonology: An overview. Phonetica, 49(3–4), 155–180. https://doi.org/10.1159/000261913

Fant, G. (1960). Acoustic theory of speech production. Mouton.

Fromkin, V. A. (1971). The non-anomalous nature of anomalous utterances. Language, 47(1), 27–52. https://doi.org/10.2307/412187

Guenther, F. H. (2016). Neural control of speech. MIT Press.

Guenther, F. H., Ghosh, S. S., & Tourville, J. A. (2006). Neural modeling and imaging of the cortical interactions underlying syllable production. Brain and Language, 96(3), 280–301. https://doi.org/10.1016/j.bandl.2005.06.001

Hickok, G. (2012). Computational neuroanatomy of speech production. Nature Reviews Neuroscience, 13(2), 135–145. https://doi.org/10.1038/nrn3158

Indefrey, P., & Levelt, W. J. M. (2004). The spatial and temporal signatures of word production components. Cognition, 92(1–2), 101–144. https://doi.org/10.1016/j.cognition.2002.06.001

Levelt, W. J. M. (1989). Speaking: From intention to articulation. MIT Press.

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

MacNeilage, P. F. (1998). The frame/content theory of evolution of speech production. Behavioral and Brain Sciences, 21(4), 499–511. https://doi.org/10.1017/S0140525X98001265

Parrell, B., Lammert, A. C., Ciccarelli, G., & Quatieri, T. F. (2019). Current models of speech motor control: A control-theoretic overview of architectures and properties. The Journal of the Acoustical Society of America, 145(3), 1456–1481. https://doi.org/10.1121/1.5092807

Parrell, B., & Houde, J. (2019). Modeling the role of sensory feedback in speech motor control and learning. Journal of Speech, Language, and Hearing Research, 62(8S), 2963–2985. https://doi.org/10.1044/2019_JSLHR-S-CSMC7-18-0127