Skip to content
Robotic Brain Universe

Language and speech

Where language lives in the brain, why it is not the same as thought, and how AI language models and speech implants are changing the field.

Intermediate · about 10 min · updated 2026-10-02 · awaiting clinical review

See it in 3D:

Illustrative simulation excitatory inhibitory

Broca's and Wernicke's areas, the arcuate fasciculus and the core language network; the dual-stream model, semantic maps and the speech motor cortex; development, FOXP2 and handedness; aphasia, Broca's patients revisited and the evidence for therapy; Zipf's law, surprisal and the laterality index; language models as models of the brain, semantic decoding and speech neuroprostheses.

Contents
  1. The talking brain
  2. What the language network is
  3. Why language is not the same as thought
  4. How speech is understood and produced
  5. When language develops
  6. When language fails
  7. The mathematics of language
  8. Language, AI and decoding
  9. Milestones
  10. Frontiers
  11. Check yourself

The talking brain

In 1861 the French surgeon Paul Broca described two patients who had lost the ability to speak after damage to the back of the left inferior frontal gyrus. Their case launched the idea that functions live in particular places in the brain, and 'Broca's area' became the anchor for the brain's speech centres.[1]

Today language is one of the most active meeting points of neuroscience and AI. Language models that are better at predicting the next word also better predict human brain activity during language processing, and the same predictive machinery now helps decode speech directly from the cortex of people who cannot speak.[2,3]

This reading explains where language lives in the brain (and where it does not), how speech is perceived and produced, how dominance for language is measured, what happens in aphasia, and how brain–computer interfaces are giving voices back.[4,5]

What the language network is

In the dominant hemisphere (usually left), the pars opercularis and pars triangularis of the inferior frontal gyrus form Broca's region, roughly Brodmann areas 44 and 45, whose cytoarchitectonic extent varies markedly between people. The posterior superior temporal gyrus is the core of the classical Wernicke area, whose boundaries and role are still debated.[6,7,8]

These areas are linked through the perisylvian network, including the arcuate fasciculus, which runs around the end of the sylvian fissure through the supramarginal gyrus region and connects frontal and temporal language areas.[9]

Modern imaging describes a core language network of left frontal and temporal areas that is strongly interconnected, works the same whether language is heard, read, spoken or written, is causally important for language and responds selectively to it.[4]

Key numbers

Right-hemisphere language dominance in strong right-handers
4%[7]
In ambidextrous people
15%[7]
In strong left-handers
27%[7]
People with a stroke who experience aphasia
about one third[10]
Attempted speech decoded from motor cortex
62 words per minute[3]

Why language is not the same as thought

People with global aphasia, who can barely understand or produce language, can still add and subtract, solve logic problems, think about another person's thoughts, appreciate music and find their way around. In healthy people, the language areas respond strongly to sentences but not to arithmetic, working-memory tasks or music.[11]

The language network therefore sits between two other kinds of system. Perceptual and motor mechanisms handle the surface of the signal, the sounds or letters, without caring about meaning; systems for knowledge and reasoning are sometimes engaged during language use but are not specific to language.[4]

How speech is understood and produced

Two streams. Hickok and Poeppel's dual-stream model proposes that speech is first analysed in the superior temporal lobe and then follows two routes: a ventral stream that maps sound onto meaning and a dorsal stream that maps sound onto the movements of speaking.[5]

Maps of meaning. Recording fMRI while people listened to hours of stories, Huth and colleagues mapped which concepts each patch of cortex responds to. Meaning turned out to be represented across large parts of the cortex in intricate patterns that seemed consistent across people, with most areas tuned to particular semantic domains.[12]

Speaking. Speech is one of the most complex movements we make, coordinating lips, jaw, tongue and larynx within milliseconds. Recordings from the surface of the ventral precentral and postcentral gyri show the articulators laid out somatotopically and activity patterns organised by phonetic features, switching between consonants and vowels within tens of milliseconds.[13]

The dual-stream model of speechAuditory cortexsound analysis on the superiortemporal gyrusSuperior temporal lobespeech sounds (phonology)Ventral streammiddle and inferior temporal:sound to meaningMeaningconcepts across the cortexDorsal streamtemporo-parietal junction: soundto articulationInferior frontal and motorcortexplanning and articulating speecharcuate fasciculus
The dual-stream model of speech. Speech is analysed in the superior temporal lobe and then splits into a ventral 'what does it mean' route and a dorsal 'how do I say it' route, which connects to frontal speech areas through the perisylvian pathways.[5,9,12]
Text version of the diagram
  1. Auditory cortex: sound analysis on the superior temporal gyrus. Leads to Superior temporal lobe.
  2. Superior temporal lobe: speech sounds (phonology). Leads to Ventral stream; Dorsal stream.
  3. Ventral stream: middle and inferior temporal: sound to meaning. Leads to Meaning.
  4. Meaning: concepts across the cortex.
  5. Dorsal stream: temporo-parietal junction: sound to articulation. Leads to Inferior frontal and motor cortex (arcuate fasciculus).
  6. Inferior frontal and motor cortex: planning and articulating speech.

When language develops

Language areas are active long before a baby speaks. In 3-month-old infants listening to speech, functional MRI showed left-lateralised activity in regions similar to those of adults, including the superior temporal and angular gyri.[14]

Genes play a part. In the KE family, a severe speech and language disorder is inherited as a single dominant trait; the cause is a mutation in FOXP2, a gene for a transcription factor involved in the developmental process that culminates in speech and language.[15]

Dominance is linked to handedness. Measuring blood-flow changes during word generation in 326 healthy people, Knecht and colleagues found right-hemisphere language dominance in 4% of strong right-handers, rising linearly to 27% of strong left-handers.[7]

When language fails

Aphasia is an acquired language impairment after brain damage that can affect speaking, understanding, reading and writing; about a third of people who have a stroke experience it.[10]

Broca's patients revisited. Re-imaging the preserved brains of Broca's two patients showed that their lesions extended well into the medial brain beyond the surface damage he described, and that what is now called Broca's area does not exactly match the region he identified. The classical deficits are not explained by one surface area alone.[1]

Variable sites. Stimulation mapping during surgery shows that the location of essential language sites varies widely between people, which is why resections near language areas of the dominant hemisphere (usually left) are guided by awake mapping. Damage to the posterior superior temporal gyrus is associated with disturbed phonological processing, such as phonemic paraphasias.[8,16,17]

Therapy helps. In a Cochrane review of 57 randomised trials with 3,002 participants, speech and language therapy improved functional communication, reading, writing and expressive language compared with no therapy, and therapy at higher intensity, dose or duration may be more beneficial.[10]

The mathematics of language

Some of the most robust laws in cognitive science are about words: how often they occur, how predictable they are, and how much effort they cost the brain.[20,21]

Zipf's law of word frequency[20]
f(r)∝1rα,α≈1f(r) \propto \frac{1}{r^{\alpha}}, \qquad \alpha \approx 1

Rank the words of a language by how often they occur: the frequency of the word at rank rr falls roughly as 1/r1/r, so the second most common word appears about half as often as the first. Language follows this simple form approximately, with reliable structure beyond it that no single explanation yet accounts for.

Symbols in Zipf's law of word frequency
SymbolMeaningUnit
rrfrequency rank of a word (1 = most common)—
f(r)f(r)how often the word of rank r occurs—
α\alphaexponent, close to 1 for word frequencies—
Surprisal[2,21]
S(wt)=−log⁡2P ⁣(wt∣w1,…,wt−1)S(w_t) = -\log_2 P\!\left(w_t \mid w_1, \dots, w_{t-1}\right)

How unexpected a word is given the words before it, in bits. Smith and Levy found that the time people spend reading a word grows with the logarithm of its predictability, linearly in surprisal, over six orders of magnitude. Language models compute exactly these probabilities, which is one reason they can predict brain responses.

Symbols in Surprisal
SymbolMeaningUnit
wtw_tthe word at position t—
P(wt∣⋅)P(w_t \mid \cdot)probability of the word given its context—
SSsurprisalbits
Laterality index[19]
LI=QL−QRQL+QR\mathrm{LI} = \frac{Q_L - Q_R}{Q_L + Q_R}

A single number for hemispheric dominance in functional MRI: +1 means all activity on the left, −1 all on the right. Its value depends on choices such as the regions compared and the statistical threshold, so these must be reported with it.

Symbols in Laterality index
SymbolMeaningUnit
QL,QRQ_L, Q_Ramount of language-related activation in the left and right hemisphere (for example, the number of active voxels)—
Word error rate[3,22]
WER=S+D+IN\mathrm{WER} = \frac{S + D + I}{N}

The standard accuracy measure for speech decoders: substitutions, deletions and insertions needed to correct the output, divided by the number of words intended.

Symbols in Word error rate
SymbolMeaningUnit
S,D,IS, D, Isubstituted, deleted and inserted words—
NNnumber of words intended—

Language, AI and decoding

Models of the brain. Schrimpf and colleagues compared many artificial language models with human recordings and found that the most powerful ones predict neural and behavioural responses up to the noise level of the data; models better at next-word prediction were better brain models. Caucheteux and King, testing 102 people on 400 sentences with fMRI and MEG, found that brain-likeness depends mainly on a model's ability to predict words from context.[2,23]

Decoding meaning. Combining fMRI with a language model, Tang and colleagues reconstructed word sequences that captured the meaning of stories people heard or imagined, but only with the person's cooperation, both to train and to use the decoder.[24]

Giving voices back. Speech neuroprostheses decode attempted speech from the cortex of people who cannot speak: 15.2 words per minute in 2021 for a man with anarthria after a brainstem stroke, then 62 and 78 words per minute in 2023 with large vocabularies, and synthesised speech and a talking avatar for one participant.[3,22,25]

Reading language from the brain[3,22,24,25]
ApproachRecordingWhat it decodes
Semantic decoder (2023)Functional MRI, no surgeryThe gist of heard or imagined speech
ECoG speech neuroprosthesis (2021, 2023)Electrode grid on the speech cortexWords and sentences, synthesised voice, avatar
Intracortical speech BCI (2023)Microelectrode arrays in motor cortexAttempted speech from a 125,000-word vocabulary

Milestones

From Broca to brain-to-text

  1. 1861Broca describes two patients who lost speech after damage to the left inferior frontal gyrus.[1]
  2. 1960The intracarotid amytal (Wada) test lateralises speech dominance.[18]
  3. 1989Stimulation mapping in 117 patients shows how widely essential language sites vary.[16]
  4. 1999Broca's region is mapped cytoarchitectonically, revealing large differences between people.[6]
  5. 2000Language dominance is shown to vary with handedness in healthy people.[7]
  6. 2001FOXP2 is identified in a family with an inherited speech disorder.[15]
  7. 2002Left-lateralised language areas are active in 3-month-old infants.[14]
  8. 2005Tractography describes the perisylvian language networks.[9]
  9. 2007Broca's original patients are re-imaged with MRI; the dual-stream model of speech is set out.[1,5]
  10. 2013The speech articulators are mapped on the ventral sensorimotor cortex.[13]
  11. 2016Semantic maps tile the human cortex.[12]
  12. 2021Next-word prediction is linked to brain responses to language; sentences are decoded in a person with anarthria.[2,22]
  13. 2023Speech neuroprostheses reach 62 and 78 words per minute; non-invasive semantic decoding.[3,24,25]

Frontiers

The language network is now seen as a 'natural kind': a dedicated system that stores knowledge of words and constructions and combines them to interpret and generate messages, working closely with, but distinct from, perception, motor control and general reasoning.[4]

Prediction is the common thread linking brains and machines. Models trained only to predict the next word converge partially on brain-like representations, and how well they predict words from context is the main factor in how brain-like they are.[2,23]

Check yourself

Check yourself

  1. Which gyri make up Broca's region?
    Show answer

    The pars opercularis and pars triangularis of the inferior frontal gyrus, in the dominant hemisphere (usually left).

  2. What did MRI of Broca's original patients show?
    Show answer

    Their lesions extended deep into medial regions beyond the surface damage Broca described.

  3. What is the evidence that language and thought are distinct?
    Show answer

    People with global aphasia can still do arithmetic, logic and social reasoning, and language areas do not respond to non-linguistic tasks in healthy people.

  4. What are the two streams in the dual-stream model?
    Show answer

    A ventral stream mapping sound to meaning and a dorsal stream mapping sound to articulation.

  5. How common is right-hemisphere language dominance in strong left-handers?
    Show answer

    About 27%, against 4% in strong right-handers.

  6. What is surprisal, and why does it matter for the brain?
    Show answer

    Minus the log probability of a word given its context; reading time grows linearly with it, and language models compute it.

  7. What did the Cochrane review conclude about speech and language therapy?
    Show answer

    It improves functional communication, reading, writing and expressive language compared with no therapy.

Glossary[6,8,9,10,15,19,21]

Broca's region
Pars opercularis and pars triangularis of the dominant inferior frontal gyrus, roughly Brodmann areas 44 and 45.
Wernicke area
Classically the posterior superior temporal gyrus of the dominant hemisphere; its boundaries are debated.
Arcuate fasciculus
A white-matter tract arching around the sylvian fissure that links frontal and temporal language areas.
Aphasia
An acquired language impairment after brain damage.
Anarthria
Loss of the ability to articulate speech.
Phonemic paraphasia
Producing a word with wrong, added or swapped speech sounds.
Language dominance
The hemisphere that is principally responsible for language, usually the left.
Laterality index
A measure of hemispheric dominance from −1 (right) to +1 (left).
Surprisal
The information carried by a word: minus the log of its probability in context.
FOXP2
A gene whose mutation causes an inherited speech and language disorder.

References

  1. Dronkers NF, Plaisant O, Iba-Zizen MT, Cabanis EA. Paul Broca's historic cases: high resolution MR imaging of the brains of Leborgne and Lelong. Brain 2007;130(5):1432-1441. doi:10.1093/brain/awm042
  2. Schrimpf M, Blank IA, Tuckute G, Kauf C, Hosseini EA, Kanwisher N, Tenenbaum JB, Fedorenko E. The neural architecture of language: integrative modeling converges on predictive processing. Proceedings of the National Academy of Sciences of the USA 2021;118(45):e2105646118. doi:10.1073/pnas.2105646118
  3. Willett FR, Kunz EM, Fan C, Avansino DT, Wilson GH, Choi EY, et al.. A high-performance speech neuroprosthesis. Nature 2023;620(7976):1031-1036. doi:10.1038/s41586-023-06377-x
  4. Fedorenko E, Ivanova AA, Regev TI. The language network as a natural kind within the broader landscape of the human brain. Nature Reviews Neuroscience 2024;25(5):289-312. doi:10.1038/s41583-024-00802-4
  5. Hickok G, Poeppel D. The cortical organization of speech processing. Nature Reviews Neuroscience 2007;8(5):393-402. doi:10.1038/nrn2113
  6. Amunts K, Schleicher A, B�rgel U, Mohlberg H, Uylings HBM, Zilles K. Broca's region revisited: Cytoarchitecture and intersubject variability. The Journal of Comparative Neurology 1999;412(2):319-341. doi:10.1002/(SICI)1096-9861(19990920)412:2<319::AID-CNE10>3.0.CO;2-7
  7. Knecht S. Handedness and hemispheric language dominance in healthy humans. Brain 2000;123(12):2512-2518. doi:10.1093/brain/123.12.2512
  8. Binder JR. The Wernicke area. Neurology 2015;85(24):2170-2175. doi:10.1212/WNL.0000000000002219
  9. Catani M, Jones DK, ffytche DH. Perisylvian language networks of the human brain. Annals of Neurology 2005;57(1):8-16. doi:10.1002/ana.20319
  10. Brady MC, Kelly H, Godwin J, Enderby P, Campbell P. Speech and language therapy for aphasia following stroke. Cochrane Database of Systematic Reviews 2016(6):CD000425. doi:10.1002/14651858.CD000425.pub4
  11. Fedorenko E, Varley R. Language and thought are not the same thing: evidence from neuroimaging and neurological patients. Annals of the New York Academy of Sciences 2016;1369(1):132-153. doi:10.1111/nyas.13046
  12. Huth AG, de Heer WA, Griffiths TL, Theunissen FE, Gallant JL. Natural speech reveals the semantic maps that tile human cerebral cortex. Nature 2016;532(7600):453-458. doi:10.1038/nature17637
  13. Bouchard KE, Mesgarani N, Johnson K, Chang EF. Functional organization of human sensorimotor cortex for speech articulation. Nature 2013;495(7441):327-332. doi:10.1038/nature11911
  14. Dehaene-Lambertz G, Dehaene S, Hertz-Pannier L. Functional neuroimaging of speech perception in infants. Science 2002;298(5600):2013-2015. doi:10.1126/science.1077066
  15. Lai CSL, Fisher SE, Hurst JA, Vargha-Khadem F, Monaco AP. A forkhead-domain gene is mutated in a severe speech and language disorder. Nature 2001;413(6855):519-523. doi:10.1038/35097076
  16. Ojemann G, Ojemann J, Lettich E, Berger M. Cortical language localization in left, dominant hemisphere. Journal of Neurosurgery 1989;71(3):316-326. doi:10.3171/jns.1989.71.3.0316
  17. Sanai N, Mirzadeh Z, Berger MS. Functional Outcome after Language Mapping for Glioma Resection. New England Journal of Medicine 2008;358(1):18-27. doi:10.1056/NEJMoa067819
  18. Wada J, Rasmussen T. Intracarotid injection of sodium amytal for the lateralization of cerebral speech dominance. Journal of Neurosurgery 1960;17(2):266-282. doi:10.3171/jns.1960.17.2.0266
  19. Seghier ML. Laterality index in functional MRI: methodological issues. Magnetic Resonance Imaging 2008;26(5):594-601. doi:10.1016/j.mri.2007.10.010
  20. Piantadosi ST. Zipf's word frequency law in natural language: a critical review and future directions. Psychonomic Bulletin & Review 2014;21(5):1112-1130. doi:10.3758/s13423-014-0585-6
  21. Smith NJ, Levy R. The effect of word predictability on reading time is logarithmic. Cognition 2013;128(3):302-319. doi:10.1016/j.cognition.2013.02.013
  22. Moses DA, Metzger SL, Liu JR, Anumanchipalli GK, Makin JG, Sun PF, et al.. Neuroprosthesis for decoding speech in a paralyzed person with anarthria. New England Journal of Medicine 2021;385(3):217-227. doi:10.1056/NEJMoa2027540
  23. Caucheteux C, King JR. Brains and algorithms partially converge in natural language processing. Communications Biology 2022;5:134. doi:10.1038/s42003-022-03036-1
  24. Tang J, LeBel A, Jain S, Huth AG. Semantic reconstruction of continuous language from non-invasive brain recordings. Nature Neuroscience 2023;26(5):858-866. doi:10.1038/s41593-023-01304-9
  25. Metzger SL, Littlejohn KT, Silva AB, Moses DA, Seaton MP, Wang R, et al.. A high-performance neuroprosthesis for speech decoding and avatar control. Nature 2023;620(7976):1037-1046. doi:10.1038/s41586-023-06443-4

Related readings

  • Brain–computer interfaces

    How implants like BrainGate and Neuralink read intention from motor cortex, the maths of decoding, and the race to restore speech and movement.

    Intermediate

  • Neural networks, biological and artificial

    How brains and machines learn from experience, the maths both share, and why AI and neuroscience keep borrowing from each other.

    Intermediate

  • The motor system

    How the brain turns intention into movement, from the redrawn motor homunculus to the laws of reaching and the technology that restores it.

    Intermediate

  • The cerebral cortex and its maps

    What the folded outer layer of the brain is made of, why it folds, and how scientists map it from Brodmann's microscope to petabyte connectomes.

    Introductory

  • The frontal lobe and executive control

    How the front of the brain holds goals in mind, controls thought and action, weighs the future, and what Phineas Gage, lobotomy and AI reveal about it.

    Intermediate

  • Hearing and the temporal lobe

    How the cochlea analyses sound, how auditory cortex maps frequency, speech and music, what goes wrong, and the implants and gene therapies that restore hearing.

    Intermediate

Template anatomy for education. Not patient-specific. Not for clinical decision-making.