Typomora

1McLuhan's deferred prediction

McLuhan's central historical claim was a reversal. The phonetic alphabet and then print pulled Western culture into visual space: linear, sequential, fixed in point of view. Electric media, he argued, would pull it back into acoustic space: simultaneous, involved, voiced. Radio and television seemed to confirm it. Then came the computer, and for roughly fifty years the defining new electric medium was operated almost entirely by typing.

Terminals, email, bulletin boards, Usenet, IRC, ICQ, HTML, SMS: the networked world of the 1970s to the 2010s was built out of keystrokes. A whole generation learned to think, argue, flirt and work in typed text. On its face, this is a counterexample to McLuhan. The newest electric medium looked more like the printing press than like the radio.

This paper argues the counterexample is only apparent. The typed era was a typomora, from Greek typos (impression, type) and Latin mora (delay). The coinage carries three senses. In Latin, mora is delay, and in Roman law specifically the delay of an obligation already owed. In phonology, a mora is the unit of spoken timing. The typomora is therefore a delay counted in units of speech: the return to voice held in arrears, not cancelled.

Text had real advantages, among them asynchrony, editability and privacy, but these do not explain why voice stayed out of the machine. The binding cause was the input layer. Until bandwidth, compute and speech recognition matured, the only economical way to put language into a networked machine was the keyboard. The keyboard brought with it a configuration inherited from literacy: language held in a visual buffer, checked by eye, composed serially and revised at leisure. As those technical limits fell, the interface began to return to voice, through dictation, voice notes, video and conversational AI. We call that return epanaphony, from Greek epan- (back, again) and phōnē (voice).

The paper builds on an earlier paper in this series, Sensory Syncrasis, which argues that McLuhan's "sense ratios" were a literal hypothesis about how the nervous system weights its channels. Typomora applies that framework to one concrete case: the layer through which people produce language for machines and for each other.

Three claims are made here, in order of confidence. First, the typed era coincided with measurable technical limits, and the turn toward voice coincided with their removal. Second, a writing medium is a physiological configuration, not only a cultural habit, so the keyboard era plausibly tuned its users in specific ways. Third, the return of voice changes the production, timing, feedback and persistence of language, not its semantic core. The framing of the typed era as a bandwidth-forced delay and the mapping onto syncrasis are this paper's own contributions. We have found no prior source that frames the typed era this way.

2The literacy precedent

If a typographic configuration is only a metaphor, the typomora is only a story. The strongest reason to treat it as physiology is the neuroscience of literacy, the best-documented case of a communication medium reshaping the brain.

What reading changes

Dehaene and colleagues (2010, Science) scanned 63 adults in Portugal and Brazil: some illiterate, some who learned to read as adults, and some schooled from childhood. Literacy produced several measurable changes. A region of left visual cortex, the visual word form area, responded more strongly to written strings as reading skill rose. Responses in early visual cortex, down to primary visual cortex, increased across all categories of images. Spoken sentences drove the brain's spoken-language network more strongly in literates, and a region of auditory cortex (planum temporale) showed enhanced responses to speech sounds. Reading changes the ear as well as the eye.

The auditory change has a behavioural counterpart. Morais and colleagues (1979) found that illiterate Portuguese adults largely could not add or remove individual speech sounds from words, while adults who had learned to read as adults could. Awareness of phonemes is a product of alphabetic literacy, not a precondition of speech. This is close to a direct physiological statement of McLuhan's claim that the phonetic alphabet re-trained the ear.

Literacy also suppresses something. People who cannot read treat mirror images as the same object, a generalisation that serves well in a natural world where a tiger facing left is still a tiger. Literate people lose this for letters, where b and d must differ, and the loss partly spreads to other images (Kolinsky and colleagues). An oriented, sequential code displaces a symmetrical, all-at-once way of seeing.

Fast, and in adults

These changes do not require childhood. Skeide and colleagues (2017, Science Advances) trained 21 illiterate women in India to read Devanagari for six months and compared them with untrained controls. The trained group showed strengthened coupling between visual cortex and deep structures (the superior colliculus and pulvinar), reaching roughly first-grade reading level. The finding is from one site and one cluster, but it shows that a medium can reconfigure an adult brain within months.

Retuning, not destruction

The original 2010 study reported a small reduction in face responses where words took over, which suggested literacy might cost something. Later work from the same Indian cohort points the other way. Hervais-Adelman and colleagues (2019) found literacy enhanced visual responses without shrinking responses to other images, and van Paridon and colleagues (2021) found face recognition improved with reading. The defensible conclusion is that literacy reallocates and retunes. It adds a specialised channel, sharpens early processing, re-trains the ear, and gives up mirror invariance for letters.

This matters for the argument. McLuhan's "sense ratio" is about weighting, not damage. The literacy evidence supports exactly that: a medium changes the relative weights of the senses and the habits of processing, measurably and in adults. One gap should be stated plainly. This evidence concerns learning to read, not habits of producing language. It shows that a writing medium can reconfigure the brain; it does not show that decades of composing by keyboard did. That step is a hypothesis, but given what reading does, it is a reasonable one rather than a leap.

3Two production systems

Reading is reception. The typomora concerns production: how people put language out. Here the brain runs two partly separate systems that share a core.

Writing and typing

A meta-analysis of brain imaging studies of written word production (Purcell, Turkeltaub, Eden & Rapp, 2011) separates two layers. A central spelling layer, in left fusiform and inferior temporal cortex, inferior frontal cortex and the intraparietal sulcus, retrieves and assembles a word's letters. A peripheral motor layer, in premotor and superior parietal cortex, turns them into strokes or keypresses. Spelling holds the word in an orthographic working memory, sometimes called the graphemic buffer, while the hand catches up (Rapp & Dufor, 2011). Typing specifically engages left intraparietal regions more than handwriting does (Higashiyama and colleagues, 2015), and isolated typing deficits after brain lesions, sometimes called dystypia, suggest the typed channel has some neural independence.

The written channel is checked by eye. The writer sees the word appear, compares it with intention, and corrects. Output persists and can be re-scanned.

Speaking

Speech runs differently. In Hickok and Poeppel's dual-stream model (2007), one pathway maps sound to meaning and another maps sound to articulation, the second running through a region at the junction of parietal and temporal cortex. Guenther's DIVA model describes speech as feedforward motor commands corrected continuously by what the speaker hears and feels in the mouth. The correction is fast and largely unconscious: when experimenters shift the pitch or vowel sounds a speaker hears of their own voice, the speaker compensates within moments, and the adjustment persists (Houde & Jordan, 1998).

The spoken channel is checked by ear and by the body. Output is transient. It cannot be re-scanned, only remembered or repeated.

Shared core, different loops

The two systems share the stages where meaning is chosen and sentences are formed. They split at the stage where language is held and checked: a visual-orthographic buffer for writing, an auditory and somatosensory loop for speech.

This is where the typomora claim lives. For decades, the keyboard made the visual-orthographic loop the default way to produce language for networked machines. No study yet shows that heavy typing weakens speech production, or that heavy speaking improves typed composition. Whether practice in one channel transfers to the other is an open question.

The typomoric configuration

In the terms of the first paper in this series, the keyboard era plausibly set the syncrasis of language production in three ways. This is interpretation, not a finding. The sense ratio of production tilted toward the eye: language was checked visually rather than by ear. The inside–outside balance tilted outward: the speaker's bodily, self-hearing loop gave way to watching words appear on a screen. And persistence became the default: language was produced as an object that stays, can be re-scanned and can be revised, rather than as an event that passes. A generation practised this configuration for hours a day, for decades.

4The cost of the keyboard

If the keyboard were cognitively free, the typomora would be a matter of style. It is not free.

Transcription takes working memory

Bourdin and Fayol (1994) asked children and adults to recall word lists either aloud or in writing. Children recalled less when writing, and slower writing speed did not explain it. The cost came from the work of transcription itself. A follow-up, titled "Even in adults, written production is still more costly than oral production"14 (2002), showed the cost persists in adults when the task gets harder. The result replicated in German children and adults.

Models of writing explain why. Planning, translating ideas into sentences, and reviewing all compete for the same limited working memory (Kellogg, 1996; McCutchen, 1996). Transcription, whether by pen or key, takes a share until it becomes automatic. Practised adults automate much of it, but not all, and the share matters most when ideas are complex or the writer is under strain.

Dictation was viable decades early

The most striking evidence comes from IBM in the late 1970s and early 1980s. John Gould found that after a few hours of practice, adults could dictate one-page letters as well as they could write them, judged both by time and by raters (Gould, 1978). Authors at first rated their own dictated letters worse, but later they and the recipients rated them the same (Gould & Boies, 1978). The bias against speaking was in the author's head, not in the letters. Composing on an early computer text editor produced letters of equal quality but took about 50% longer than writing by hand (Gould, 1981).

Then Gould, Conti and Hovanyecz (1983) simulated a "listening typewriter": a speech recogniser operated by a hidden human typist, set to mimic different machine limits. When the simulated recogniser forced speakers into isolated words or a limited vocabulary, performance fell. An imperfect recogniser was no better than ordinary dictation.

Read together, these studies say something unusually direct. A decade before the web, researchers had shown that people could compose by voice as well as by hand, and that the binding constraint was the recognition technology. That is close to a primary-source demonstration that the typed era was a hardware delay rather than a human preference.

The scribe, before the machine

Dictation is not new. It was a privilege. Cicero dictated to his secretary Tiro, whose shorthand system, the Tironian notes, survived for centuries. Milton, blind, dictated Paradise Lost. Dostoevsky dictated The Gambler in 1866 to a stenographer, Anna Snitkina, to meet a deadline; he later married her. Henry James, from the late 1890s, dictated his novels to a typist; his secretary Theodora Bosanquet described the practice in Henry James at Work (1924).

James is the closest thing to a historical experiment in production mode. Critics have long linked the elaborate, parenthetical, conversational quality of his late style to the switch from pen to voice. The link is debated, and age and ambition changed too. But the case suggests that a change in how language is produced can show up in the language itself.

For most of history, only the powerful, the wealthy or the disabled with help could compose by voice. Everyone else wrote. Speech recognition and conversational AI make the scribe universal.

When the constraint lifted

By 2016, the constraint had largely gone. Ruan and colleagues tested 32 people entering text on an iPhone: English speech recognition was 2.93 times faster than the on-screen keyboard (153 versus 52 words per minute), with 20.4% fewer errors. Mandarin showed a similar gap. That study measured short phrases, not composition, but it marks the crossover. On entry speed and accuracy, speaking had overtaken typing on the phone. Skilled typists on a full keyboard are faster than phone typists, but few reach 150 words per minute.

Modern speech-to-text also appears to draw on a partly different skill from handwriting and typing. In a 2025 study of 64 eleven-year-olds, speech-to-text produced the most output in copying tasks, but children's handwriting and keyboard scores correlated with each other more than with speech-to-text. Voice is not simply faster typing. It is a different route.

5The typomora as history

The technical thresholds

The typomora has a measurable end. Two thresholds mark it.

The first is speech recognition. On the standard Switchboard benchmark of conversational telephone speech, word error rates stalled around 15% through the mid-2000s, then fell sharply with deep learning. In 2016 and 2017, Microsoft reported error rates between 5.1% and 5.9%, which it claimed matched professional human transcribers (Xiong and colleagues). IBM disputed where human parity sat, measuring careful human transcription at 5.1% (Saon and colleagues, 2017). The disagreement is about where the finish line is, not about the trend. In 2022, OpenAI's Whisper, trained on 680,000 hours of speech in many languages, extended this robustness to noisy, accented and everyday speech.

The second is bandwidth. In early networks, text was the only content cheap enough to send at scale. That changed through the 2010s. Cisco estimated video passed half of consumer internet traffic by the end of 2016 and reached 73% in 2017. Sandvine's measurements put video at about 58% of downstream traffic in 2018 and 65% of all internet traffic in 2022. These are industry figures, and traffic is counted in bytes, which flatters video. Even so, the direction is not in doubt.

The timing matters. Usable voice input and video-heavy networks arrived within a few years of each other, around 2016 to 2022. The typomora ends where the hardware ran out of excuses.

Media theory already saw the symptoms

Media theorists noticed that networked text did not feel like print. Walter Ong called radio and television "secondary orality," an orality built on top of writing, and suggested computerised text was a kind of "secondary literacy." Robert Logan described email, chat and blogs as "tertiary" or "digital orality": oral in spirit, written in form. Oren Soffer (2010, Communication Theory) called texting and chat "silent digital orality."

These accounts describe the symptoms of the typomora exactly. People wrote in typed text the way they would talk: fast, informal, phatic, conversational. What these accounts do not offer is a cause. Soffer's own word, silent, points straight to it. The orality was silent because the channel could not yet carry voice. The style was oral; the production was typographic. The typomora is the name for the gap between the two, and bandwidth and recognition limits explain why it existed and why it is closing.

The idea of the "Gutenberg Parenthesis" (Sauerberg; Pettitt) treats the print era as a bounded interlude between oral cultures. Typomora proposes a second, much shorter bracket inside the electric age: a keyboard bracket, opened by the terminal and now closing.

Kittler and the typewriter

The theorist who took the typewriter most seriously was Friedrich Kittler. In Gramophone, Film, Typewriter (1986), he argued that around 1900 three machines split apart the streams of human expression: the gramophone stored sound, film stored images, and the typewriter stored text, separating writing from the hand and the body that had produced it.

Kittler's emblem was Nietzsche. Half-blind, Nietzsche bought a Malling-Hansen writing ball, an early typewriter, in 1882. In a typed letter to his friend Heinrich Köselitz at the end of February that year, he wrote: "Unser Schreibzeug arbeitet mit an unseren Gedanken"33, our writing tools are also working on our thoughts. Kittler called him the first mechanised philosopher and read a shift toward compressed, aphoristic style into the work that followed. That reading is contested, and Nietzsche gave up the machine within weeks. But the sentence itself is the typomora thesis in miniature, written on the device it describes.

Kittler also foresaw that digitisation would erase the differences between media, merging sound, image and text into one stream of data. He was right about storage and transmission. Within the network, voice, video and text did converge. But the input layer lagged behind. People went on putting language into a converged medium through a nineteenth-century device. The typomora names that lag: the gap between the convergence of channels and the convergence of input.

6What returns and what does not

A strong version of the thesis would say that returning to voice changes how people think. The evidence does not support that, and the paper does not claim it.

Meaning is largely modality-invariant

Deniz and colleagues (2019, Journal of Neuroscience) had nine people read and listen to the same stories while their brains were scanned, and built models of how each part of cortex represents meaning. The semantic maps for reading and listening were almost identical. Later work found that the brain's tuning to different timescales of language is also shared across reading and listening beyond early sensory areas (Chen and colleagues, 2024), and that semantic decoders transfer across people and across modalities (Tang & Huth, 2025).

Behaviour agrees. Rogowsky, Calhoun and Tallal (2016) randomly assigned 91 adults to an audiobook, an e-text, or both at once. Comprehension did not differ, immediately or two weeks later. A review of audiobook research (Singh & Alexander, 2022) found listening helped younger and struggling readers but found few differences for skilled adults.

For skilled adults, then, the meaning a person takes from language is about the same whether it arrives by eye or by ear. Any claim that voice makes people understand differently is not supported.

What does differ

The claim moves to four places where modality demonstrably matters.

1. 1Time structure. Speech is transient, serial and paced by the speaker. Text persists, can be re-scanned in any order, and is paced by the reader. This is McLuhan's contrast between acoustic and visual space, restated as a claim about persistence and control of sequence.
2. 2The production loop. Speech is monitored by ear and body, written language by eye (section 3). Shifting the main channel shifts which loop gets daily practice.
3. 3Prosody and presence. Voice carries tone, emphasis, hesitation and emotion that text strips out or must encode with punctuation and emoji.
4. 4Transcription cost. Voice removes the working-memory cost of transcription (section 4).

The hybrid configuration

The present moment has a feature neither McLuhan nor Ong anticipated. When people dictate, leave voice notes, or talk to an AI, their speech is usually turned into text and kept. Production is oral; storage is typographic. Every utterance becomes searchable, editable, persistent text.

The epanaphony is therefore not a simple return. It retrieves the voice while keeping the archive. In the terms of the syncrasis framework, the typographic configuration migrates from the production side to the storage side. Whether that hybrid changes how people revise, remember and retrieve their own ideas is still open.

There is a known cost to this kind of storage. When people photograph objects in a museum, they later remember those objects less well than ones they only looked at (Henkel, 2014), a finding that has since replicated (Soares & Storm, 2018). Saving something externally seems to license the mind to let it go, although saving can also free memory for what comes next (Storm & Stone, 2015). Total transcription may do the same to speech. People may remember less of what they said, because the transcript remembers it for them. Whether that is a loss or a reallocation is the question every extension raises, and for this one it is still open.

7Epanaphony: voice, AI and the speaking self

The shift is under way, unevenly

Voice input is common but not yet dominant. Survey data from GWI, published through DataReportal, put weekly use of voice assistants at 22.5% of internet users worldwide in late 2025, and 28.8% in the US. Smart-speaker ownership in the US has sat at roughly a third of adults for four years (Edison Research). The older voice-assistant wave has plateaued. The newer growth is in dictation tools and voice conversation with AI models, for which no rigorous public survey data yet exist. Measuring it is one of the paper's open tasks.

What has changed most is what voice can do. A voice assistant answers a narrow question. A conversational model can take a rambling spoken explanation, ask follow-up questions, and turn the result into a document, a plan or working software. Voice has become a way to think and build, not only to command.

Why speaking to an AI might change the speaker

Two well-supported mechanisms suggest that talking ideas through, including with an AI, could improve both thinking and speaking. Neither has been tested in that exact setting.

The first is self-explanation. In Chi and colleagues' classic study (1994), students prompted to explain a text aloud to themselves understood it more deeply than students who read it twice. A meta-analysis of 69 effect sizes found a moderate average benefit (g = 0.55; Bisra and colleagues, 2018). Explaining to a listener, even an artificial one, is a form of spoken generative elaboration.

The second is exposure. Psychological treatments for fear of public speaking reduce it substantially, with an average effect of g = 0.74 across 62 interventions (Ebrahimi and colleagues, 2019), and exposure-based treatments show larger effects (Reeves and colleagues, 2022). An AI listener removes social judgement, so it may work as a first, low-stakes rung of exposure. Meanwhile, each utterance runs the auditory-motor loop described in section 3, so sheer volume of speaking is practice in the motor sense.

Vygotsky in reverse

There is a deeper reading of what happens when people think aloud with an AI. It comes from Lev Vygotsky's account of how thought develops (Thought and Language, 1934). For Vygotsky, thinking begins as social speech: a child talks with others. It then becomes private speech, the child talking aloud to themselves while working something out. Finally it goes silent and becomes inner speech, which is compressed and abbreviated, carrying only the parts the thinker needs. Modern research broadly supports this sequence and the role of inner speech in planning and self-regulation (Alderson-Day & Fernyhough, 2015).

Talking ideas through with an AI runs that path backwards. Inner speech is spoken aloud again, and this time to a partner who answers, questions and asks for more. Thinking regains its social form. The compression of inner speech has to be undone, because a listener needs the full version. That expansion may be part of why explaining aloud improves understanding: the speaker discovers what the abbreviated inner version left out.

This is the paper's own synthesis; no study has tested it. But it suggests the epanaphony returns more than the voice. It may return the social origin of thinking. For someone who found speaking hard, a patient and non-judging listener may be a route by which thought becomes speech again, which is a different and more specific mechanism than practice or exposure alone.

A first-person case

The author is a case of the typomora. Online from the late 1990s, he became fluent in typed text, building a career on the web while finding speech difficult. Email and writing let him say clearly what he struggled to say aloud. In recent years his main interface has shifted to voice: recording ideas on a phone, dictating, describing experience, talking projects through with AI, and directing software builds by speaking. Over the same period, his spoken fluency has improved.

One case proves nothing, and the paper does not treat it as evidence. It is offered the way McLuhan and Ong used their own experience: as a probe that generates hypotheses.

A caution

Concern that AI weakens thinking, so-called "cognitive debt," rests mainly on a 2025 MIT preprint (Kosmyna and colleagues) that measured typed use of ChatGPT for essays, has not been peer reviewed, and has drawn a published methodological critique. Offloading concerns apply to AI in any modality. Whether speaking to an AI makes people think more or less is an open empirical question, not a settled one.

8Objections and replies

"Early online writing was already oral." Logan and Soffer argue that chat and email were oral in style. This paper agrees, and treats it as evidence for the typomora rather than against it. An oral bias was being forced through a typographic channel. Soffer's term "silent" orality names the constraint precisely. The style was oral; the production physiology, with its visual buffer, eye-checking and keystrokes, was typographic. The typomora is the gap between the two.

"Text never declined." True in absolute terms, and the paper does not claim otherwise. McLuhan's claim was about ratios, not totals. The thesis concerns the main input interface and where growth is happening. Voice with transcription also produces text, so total text can rise while the share of text keyed by hand falls. The useful measure is keyed versus spoken input, which needs better data than currently exist.

"People still prefer text." Many people, especially younger people, avoid phone calls and prefer messaging. The preference is real, but it is mostly about asynchrony and control: a message can be composed in private, revised and answered later. It is not a preference for the keyboard as such. Voice notes and dictation now offer the same asynchrony and editability by voice, and their spread suggests the preference was for asynchrony, not for typing. The typomora concerns the input layer, not the synchronous phone call.

"Meaning is the same either way." Conceded in full (section 6). The claim is relocated to production, timing, feedback, prosody and persistence, where modality demonstrably matters.

"Literacy's costs are contested." Also conceded. The literacy precedent is used for reconfiguration, not damage. Mirror invariance and phoneme awareness are cited as documented reallocations; the face-competition claim is not relied on.

"This is technological determinism."16 The opposite. The typomora framing is explicitly contingent: it ties timing to measurable thresholds in recognition error, entry speed and bandwidth. Gould's listening-typewriter studies show the demand for voice existed in the early 1980s and the technology was what held it back.

"Voice AI may weaken thinking." The main evidence concerns typed use and is a contested preprint. Self-explanation research suggests speaking ideas through may be more generative, not less. This is treated as an open question.

"It is built on one person's story." The author's case generates hypotheses; it does not support them. The argument rests on the literacy neuroscience, the production research and the technical history.

9Conclusion: the tetrad of the voice interface

McLuhan was not wrong about electric media and voice. He was early. The return to acoustic space was deferred for half a century by the one part of the system he did not analyse: the input layer. The keyboard kept networked language inside a typographic configuration long after the medium's other channels had gone electric. That deferral, the typomora, is now ending.

McLuhan's tetrad asks four questions of any medium. Applied to the voice interface:

That last reversal is the most consequential and the least understood. The typomora tuned a generation to compose by eye. The epanaphony may tune the next to speak into a permanent record. What that does to how people talk, revise and remember is exactly the kind of question the syncrasis framework was built to ask.

References

1. 1Dehaene et al. (2010), How learning to read changes the cortical networks for vision and language, Science 330: 1359–1364, doi:10.1126/science.1194140
2. 2Dehaene, Cohen, Morais & Kolinsky (2015), Illiterate to literate, Nature Reviews Neuroscience 16: 234–244
3. 3Skeide et al. (2017), Learning to read alters cortico-subcortical cross-talk, Science Advances 3: e1602612
4. 4Hervais-Adelman et al. (2019), Learning to read recycles visual cortical networks without destruction, Science Advances 5: eaax0262
5. 5van Paridon et al. (2021), Does neuronal recycling result in destructive competition?, Psychological Science, doi:10.1177/0956797620971652
6. 6Morais, Cary, Alegria & Bertelson (1979), Cognition 7: 323–331
7. 7Kolinsky et al. (2011), mirror-image discrimination and literacy, J Exp Psychol: General
8. 8Purcell, Turkeltaub, Eden & Rapp (2011), Examining the central and peripheral processes of written word production, Frontiers in Psychology 2: 239
9. 9Rapp & Dufor (2011), J Cognitive Neuroscience 23: 1180–1197
10. 10Higashiyama et al. (2015), The neural basis of typewriting, PLoS ONE
11. 11Hickok & Poeppel (2007), Nature Reviews Neuroscience 8: 393–402
12. 12Guenther (2016), Neural Control of Speech (DIVA model)
13. 13Houde & Jordan (1998), Science 279: 1213
14. 14Bourdin & Fayol (1994), Int J Psychology 29: 591–620; (2002) Int J Psychology 37: 219–227
15. 15Kellogg (1996); McCutchen (1996), Educational Psychology Review 8: 299
16. 16Gould (1978), How experts dictate, JEP: Human Perception and Performance 4: 648–661
17. 17Gould & Boies (1978), Human Factors 20: 495–505
18. 18Gould (1981), Composing letters with computer-based text editors, Human Factors 23: 593–606
19. 19Gould, Conti & Hovanyecz (1983), Composing letters with a simulated listening typewriter, Communications of the ACM 26: 295–308
20. 20Ruan et al. (2016), Speech is 3x faster than typing, arXiv:1608.07323
21. 21Does composition mode matter? (2025), Reading and Writing, doi:10.1007/s11145-025-10725-w
22. 22Xiong et al. (2017), The Microsoft 2017 conversational speech recognition system, arXiv:1708.06073; Saon et al. (2017), Interspeech
23. 23Radford et al. (2022), Whisper, arXiv:2212.04356
24. 24Cisco Visual Networking Index (2018); Sandvine Global Internet Phenomena (2018, 2023)
25. 25Ong (1982), Orality and Literacy; Logan (2007); Soffer (2010), Silent orality, Communication Theory
26. 26Deniz et al. (2019), Journal of Neuroscience 39: 7722–7736
27. 27Chen et al. (2024), shared language timescales; Tang & Huth (2025), Current Biology 35: 1023–1032
28. 28Rogowsky, Calhoun & Tallal (2016), SAGE Open 6(3); Singh & Alexander (2022), Educational Psychology Review
29. 29Chi et al. (1994), Cognitive Science 18: 439–477; Bisra et al. (2018), Educational Psychology Review 30: 703–725
30. 30Ebrahimi et al. (2019), Frontiers in Psychology 10: 488; Reeves et al. (2022), Behavior Modification
31. 31DataReportal/GWI (Q4 2025); Edison Research, Infinite Dial 2025
32. 32Kosmyna et al. (2025), arXiv:2506.08872, and published critique (arXiv:2601.00856)
33. 33Kittler (1986; English 1999), Gramophone, Film, Typewriter, Stanford University Press
34. 34Nietzsche to Heinrich Köselitz, typescript, end of February 1882, Briefwechsel: Kritische Gesamtausgabe III-1
35. 35Bosanquet (1924), Henry James at Work, Hogarth Press
36. 36Vygotsky (1934; English 1986), Thought and Language, MIT Press
37. 37Alderson-Day & Fernyhough (2015), Inner speech, Psychological Bulletin 141(5): 931–965
38. 38Henkel (2014), Psychological Science; Soares & Storm (2018), J Applied Research in Memory and Cognition 7: 154–160
39. 39Storm & Stone (2015), saving-enhanced memory, Psychological Science