How Does the Brain Process Language?

Last reviewed on October 2, 2026.

Language is a defining feature of human cognition. Within about half a second of hearing a word, you have usually recognised it, retrieved its meaning and started fitting it into the sentence around it — while already predicting what comes next. Psycholinguistics studies the mental processes involved, and neurolinguistics studies how they are implemented in the brain. This guide walks through the whole pipeline, from sound waves to meaning, then covers speaking, the brain networks involved, reading, how children acquire language, second-language learning and what large language models do and do not tell us.

What Is Language Processing?

Language processing is the set of cognitive operations that let us understand and produce spoken, written and signed language. It is usually described at several levels of linguistic structure:

• Phonology: the sound system of a language — phonemes (the distinct speech sounds that change meaning, like /p/ and /b/ in pat and bat) and the rules for combining them.
• Morphology: morphemes, the smallest units of meaning, and how they combine (un-break-able, walk-ed).
• Syntax: the rules that govern how words combine into phrases and sentences.
• Semantics: the meaning of words and sentences.
• Pragmatics and discourse: how context, shared knowledge and the speaker’s intentions shape interpretation across sentences and conversations.

These levels are a useful description of language, but the brain does not finish one level before starting the next. Processing is incremental (we interpret word by word, not sentence by sentence) and interactive (meaning and context feed back to help recognise sounds and resolve ambiguity). Cognitive psychology contributes the experimental methods and models of memory and attention; linguistics contributes precise descriptions of what has to be computed. Psycholinguistics sits at the intersection, which is why language is one of the core disciplines of cognitive science.

The Language Comprehension Pipeline: From Sound to Meaning

1. Speech perception

Speech is a continuous acoustic stream with no reliable pauses between words. Each sound is also shaped by its neighbours (coarticulation): the /d/ in dee and doo is acoustically quite different. Listeners nonetheless hear stable categories — a phenomenon called categorical perception. Perception is also multisensory: in the McGurk effect (McGurk & MacDonald, 1976), seeing lips say “ga” while hearing “ba” often produces the percept “da”. Early acoustic analysis happens in primary auditory cortex (Heschl’s gyrus) and the surrounding superior temporal gyrus, largely in both hemispheres.

2. Word recognition and lexical access

Recognising a word means matching the input against a mental dictionary (the lexicon) of tens of thousands of entries. The cohort model (Marslen-Wilson) proposes that the first sounds of a word activate all candidates that begin that way — cap, captain, capital — and the set narrows as more input arrives, often identifying the word before it ends. The connectionist TRACE model (McClelland & Elman, 1986) captures the same competition with interacting layers of feature, phoneme and word units. Frequent words are recognised faster, and ambiguous words briefly activate several meanings: in Swinney’s (1979) cross-modal priming studies, “bugs” momentarily primed both ant and spy, even when context favoured one.

3. Parsing: building sentence structure

To understand “The dog that chased the cat was tired”, you must work out that the dog, not the cat, was tired. This structure-building is called parsing. Because we parse incrementally, we sometimes commit to the wrong structure, as in the classic garden-path sentence “The horse raced past the barn fell” (Bever, 1970). Two families of theory explain this. Garden-path models (Frazier) propose that the parser first builds the simplest structure using syntactic principles and revises later. Constraint-based models (MacDonald, Pearlmutter & Seidenberg; Trueswell and Tanenhaus) propose that syntax, word frequency, meaning and context all compete at once. Eye-tracking studies in the “visual world” paradigm (Tanenhaus et al., 1995) showed that a visual scene can change how listeners parse an ambiguous instruction within a few hundred milliseconds, supporting rapid use of context.

4. Semantics: combining meanings

Word meanings are combined according to the sentence structure to compute who did what to whom. Comprehension is also strongly predictive: readers and listeners anticipate likely upcoming words, and unexpected words cost more processing (see the N400 below). This fits the wider idea that the brain constantly generates predictions, explored in our article on predictive processing. Much semantic knowledge appears to be stored in distributed networks, with the anterior temporal lobes acting as a hub that binds features into concepts; some meaning also seems grounded in perceptual and motor systems, a theme of embodied cognition.

5. Discourse and pragmatics

Real understanding goes beyond single sentences. Listeners link pronouns to their referents, draw inferences that were never stated, and interpret what the speaker means rather than only what they say. If someone asks “Is there any coffee left?” and you reply “There’s a café on the corner”, they infer the answer is no. Philosopher Paul Grice explained such implicatures through the assumption that speakers cooperate by being informative, truthful, relevant and clear. Pragmatic processing draws on theory of mind, working memory and general knowledge, and recruits regions well beyond the classic language areas, including the right hemisphere.

How the Brain Produces Speech: Levelt’s Model

Speaking reverses the comprehension pipeline. The most influential account is Willem Levelt’s blueprint of the speaker (Speaking, 1989; later developed with Ardi Roelofs and Antje Meyer):

1. Conceptualisation: deciding what to say — a preverbal message.
2. Formulation: selecting words (first an abstract lemma with its meaning and grammatical properties, then its sound form), building syntactic structure and preparing a phonological plan.
3. Articulation: turning that plan into coordinated movements of the lungs, larynx, tongue and lips.
4. Self-monitoring: checking the output, which is why we can catch and repair errors mid-word.

Evidence for separate lemma and sound-form stages comes from tip-of-the-tongue states, in which people know a word’s meaning and often its grammatical gender or first letter but cannot retrieve the whole form, and from speech errors such as sound swaps (“heft lemisphere”) versus word swaps. Fluent speakers typically produce two to three words per second, choosing each from a vocabulary of tens of thousands of words, with remarkably few errors.

Language Processing in the Brain: Beyond Broca and Wernicke

The classic model

In 1861 Paul Broca described a patient, Louis Leborgne, who could produce little more than the syllable “tan” but understood much of what was said to him; his autopsy showed damage to the left inferior frontal lobe. In 1874 Carl Wernicke described patients with fluent but meaningless speech and poor comprehension after damage to the left posterior superior temporal lobe. Wernicke and later Ludwig Lichtheim and Norman Geschwind proposed a simple circuit: Wernicke’s area for comprehension, Broca’s area for production, and a fibre bundle, the arcuate fasciculus, connecting them.

Why the classic model is incomplete

Modern imaging and lesion studies have refined this picture considerably. When Nina Dronkers and colleagues (2007) re-scanned the preserved brains of Broca’s original patients, the damage extended well beyond “Broca’s area” into deeper white matter. Damage confined to Broca’s area often produces only temporary problems, and Broca’s area itself is involved in comprehension of complex syntax, not just production. “Wernicke’s area” is also defined inconsistently across studies. Language is now seen as a distributed network.

The dual-stream model (Hickok & Poeppel)

Gregory Hickok and David Poeppel’s dual-stream model (2007) is the most widely cited modern framework, drawing an analogy with the “what” and “where/how” streams of vision:

• Ventral stream (sound to meaning): runs through the superior and middle temporal gyri toward the anterior temporal lobe. It is largely bilateral, which helps explain why comprehension of single words often survives damage to one hemisphere.
• Dorsal stream (sound to articulation): runs from the posterior superior temporal lobe through a region at the parietal–temporal boundary (area Spt) to frontal motor and premotor areas, including Broca’s area. It is strongly left-dominant and supports repeating words, learning new words and verbal working memory.

Angela Friederici and others similarly describe dorsal and ventral white-matter pathways, with the dorsal arcuate fasciculus/superior longitudinal fasciculus particularly important for complex syntax. Evelina Fedorenko’s work has further identified a left frontotemporal “language network” that responds strongly to sentences but very little to arithmetic, music or general problem solving — suggesting language relies on dedicated circuitry even though it interacts with other systems.

Left-lateralisation

For the large majority of right-handed people (around 95% in many estimates), core language functions are left-hemisphere dominant; most left-handers are left-dominant too, though right-hemisphere or bilateral language is more common among them. The right hemisphere still contributes prosody (the melody of speech), emotional tone, metaphor, humour and discourse-level integration.

Aphasias: what brain damage reveals

Aphasia is a language impairment caused by brain damage, most often stroke. Broca’s (non-fluent) aphasia involves effortful, telegraphic speech with relatively preserved comprehension of simple sentences. Wernicke’s (fluent) aphasia involves fluent speech that is often empty or full of wrong words, with poor comprehension. Conduction aphasia leaves comprehension and fluent speech fairly intact but severely impairs repetition — classically attributed to arcuate fasciculus damage, now also linked to cortical damage around area Spt. Primary progressive aphasias, caused by neurodegeneration, have helped reveal the role of the anterior temporal lobes in word meaning. Most real patients show mixed profiles, which is one reason clinicians increasingly describe specific deficits rather than relying only on classic syndrome labels.

Brain regions involved in language and their roles

How Fast Does the Brain Process Language? ERP Evidence

fMRI shows where language is processed, but it is too slow to track word-by-word processing. EEG and MEG, which measure electrical and magnetic brain activity with millisecond precision, show when. Researchers average the brain’s responses to many words to obtain event-related potentials (ERPs). Two components are especially well known:

• N400: a negative wave peaking roughly 400 ms after a word appears, discovered by Marta Kutas and Steven Hillyard (1980). It is larger for words that are hard to integrate or unexpected in context — “He spread the warm bread with socks” produces a large N400 — and smaller for predictable words. It is widely treated as an index of semantic access and prediction.
• P600: a positive wave around 500–900 ms, reported by Lee Osterhout and Phillip Holcomb (1992) for grammatical violations and garden-path sentences (“The child throw the toy”). It is often linked to syntactic reanalysis or repair, although some “semantic” anomalies also produce it, so its interpretation is debated.

Together with eye-tracking, these measures show that the brain begins interpreting a word within a few hundred milliseconds and uses context and prediction continuously. More on these tools is in our guide to research methods in cognitive science.

Reading: How the Brain Processes Written Text

Writing is only about 5,000 years old — far too recent for the brain to have evolved a dedicated reading organ. Instead, learning to read recycles existing visual and language circuits, a proposal developed by Stanislas Dehaene. With literacy, a patch of the left fusiform gyrus becomes specialised for letter strings: the visual word form area. It responds to familiar scripts in readers but not in people who have not learned to read.

Skilled readers do not read letter by letter in a smooth sweep. The eyes make brief fixations (typically around 200–250 ms) separated by rapid jumps called saccades, taking in only a small window of text each time; short, predictable words are often skipped.

The dual-route model of reading

The dual-route cascaded model (Coltheart and colleagues, 2001) proposes two ways to read a word aloud. The lexical route looks up familiar whole words in the mental lexicon and is needed for irregular words like yacht or colonel. The sublexical route converts letters into sounds using spelling-to-sound rules and is needed for new words and pronounceable nonwords like blint. Acquired dyslexias fit this split: people with surface dyslexia regularise irregular words (“pint” to rhyme with “mint”), while people with phonological dyslexia read familiar words but struggle with nonwords. Connectionist “triangle” models (Seidenberg & McClelland, 1989; Plaut et al., 1996) offer a competing single-mechanism account in which both kinds of word are handled by one learning network — see our article on connectionism.

What type of cognition comes from reading text?

The process of building meaning from written text is usually called reading comprehension (or text comprehension), and the broader set of skills and knowledge that reading develops is literacy. According to the simple view of reading (Gough & Tunmer, 1986), reading comprehension depends on two things: decoding (recognising written words) and language comprehension (understanding the language those words express). Psychologists describe the outcome of successful comprehension as a situation model (van Dijk & Kintsch, 1983; Zwaan & Radvansky, 1998): a mental representation not of the words themselves but of the situation they describe — the characters, places, time, causes and goals. Readers generally remember this situation model far longer than the exact wording. Reading also builds vocabulary and background knowledge, which in turn makes later comprehension easier; this accumulated, language-based knowledge is close to what intelligence researchers call verbal or crystallised ability.

Language Acquisition: How Children Learn Language

Children acquire their first language without formal teaching, and the main milestones are remarkably consistent across languages (individual ages vary widely):

• Before birth and at birth: newborns already prefer their mother’s voice and the rhythm of their native language, heard through the womb.
• Around 6–10 months: canonical babbling (“bababa”) begins. Infants start to tune into the sounds of their own language, and by about 10–12 months they become much worse at distinguishing some non-native sound contrasts (Werker & Tees, 1984).
• Around 8 months: infants can use statistical regularities — which syllables tend to follow which — to find word boundaries in continuous speech (Saffran, Aslin & Newport, 1996).
• Around 12 months: first words, usually understood some months before they are produced.
• Around 18–24 months: vocabulary growth accelerates and two-word combinations (“more juice”) appear.
• Around 3–5 years: grammatical morphemes, questions and complex sentences emerge; overregularisations such as “goed” show that children are extracting rules, not just copying.

How this is possible is one of cognitive science’s longest-running debates. Noam Chomsky argued that the input children hear is too limited to explain what they learn, so humans must have an innate language faculty (“universal grammar”). Usage-based and statistical-learning accounts (Michael Tomasello and others) argue that powerful general learning mechanisms, combined with rich social interaction, can do much of the work. The debate was central to the cognitive revolution and continues today. For the wider picture of how children’s thinking develops, see cognitive development.

Bilingualism and Second-Language Learning

More than half of the world’s population is often estimated to use two or more languages. Studies of bilinguals show that both languages are active at once: when a Dutch–English bilingual reads an English word, similar-looking Dutch words are activated too, and the speaker must constantly select the intended language.

Is there a critical period for language?

Eric Lenneberg (1967) proposed a critical period for language ending around puberty. The evidence supports a sensitive period with gradual decline rather than a sharp cut-off, and the timing differs by component:

• Accent and phonology are most sensitive to age; people who start a language after early childhood rarely sound fully native, though many come close.
• Grammar: Johnson and Newport (1989) found that immigrants who arrived in the US at younger ages performed better on English grammar. A study of around 670,000 people by Hartshorne, Tenenbaum and Pinker (2018) estimated that the ability to learn grammar to a native-like level stays high until about age 17–18 and then declines — later than many had assumed — though reaching native-like levels still requires starting much earlier because learning takes years.
• Vocabulary can be learned well throughout life.

Adults are not simply worse learners: they often learn faster at first, using explicit strategies and existing knowledge. Children tend to reach higher final levels mainly because they get more years of immersive input, have more implicit learning opportunities and face less interference from an established first language. The amount and quality of input, motivation and opportunities to use the language matter a great deal at every age.

Claims that bilingualism produces a general “executive function advantage” are contested: early positive results have often not replicated in larger studies, so it is safest to say that any such cognitive benefit is small or context-dependent. The practical benefits of knowing more than one language are not in doubt.

Language Processing and Large Language Models

Large language models (LLMs) such as GPT-style transformers learn to predict the next word from enormous amounts of text, and their internal activations turn out to predict human brain responses to sentences surprisingly well (for example, Schrimpf et al., 2021). This supports the idea that prediction is central to human comprehension. But important differences remain. LLMs are trained on vastly more text than any child ever hears, they learn without the social interaction, perception and action in which children acquire language, and they can produce fluent language without the grounded understanding or communicative intentions that human speakers have. Whether, and how, they can serve as models of human language processing is an active research question — explored further in our article on AI and cognitive science.

Why Language Processing Research Matters

Understanding language processing informs speech and language therapy after stroke, early identification and support for developmental language disorder and dyslexia, evidence-based reading instruction (the strong evidence for systematic phonics comes directly from research on decoding), second-language teaching and the design of speech and language technology. Language also links to other parts of cognition: holding a sentence in mind relies on working memory (see memory and learning), and following a conversation in a noisy room depends on attention and perception. If you want to work in this area, psycholinguistics and speech technology are covered in our guide to cognitive science careers.