Someone reads out a phone number and you run it through twice in your head before you find a pen. You hit a tangled sentence and your lips move a little. At midnight you replay a remark from the afternoon, deliver the reply you should have given, and then supply the other person's next line too.
All of this passes for "thinking". But the three have something in common that is slightly uncomfortable to state: they are all actions — you are speaking without making a sound. So take that voice apart three ways: what job it does, what shape it has, and what happens to people who do not have it.
Do this first. Look at these seven letters, cover them, and write them back in order: B C D G P T V. Most people scramble the order or drop one. Try F K L Q R Y S instead — same seven, distinctly easier. The only difference is how they sound: the first set all end in the same vowel, the second set does not. Conrad and Hull reported this in 1964, and it is called the phonological similarity effect. The crucial part is that you read the letters with your eyes and never said them out loud, yet the set you could not hold was the one whose sounds smear together: storing them, you converted them into sound. The effect is not particular to English — try it yourself with 四 (sì, "four"), 十 (shí, "ten"), 是 (shì, "is"), 事 (shì, "matter"), 时 (shí, "time"), 使 (shǐ, "to make") against 猫 (māo, "cat"), 车 (chē, "car"), 树 (shù, "tree"), 灯 (dēng, "lamp"), 门 (mén, "door"), 伞 (sǎn, "umbrella"), and see which list slips out of order.
The second demonstration is more direct. Have someone mutter "the the the the" continuously while holding that string of letters: span falls, and the phonological similarity effect vanishes with it — the two lists become equally hard. The technique is called articulatory suppression. What it occupies is not your attention but the act of saying things over silently.
Put those together and you have Baddeley's phonological loop: a small store of sound that fades on its own (roughly two seconds' worth), plus a rehearsal mechanism that keeps it alive by saying it again. At this point the voice in your head gets a concrete job description for the first time. It is not an accompaniment to thinking; it is refreshing something that is about to fade.
If that is right, then a language whose words take longer to say should let you refresh fewer of them — a checkable prediction. Ellis and Hennelly checked it in 1980 with Welsh–English bilinguals: Welsh number words take longer to articulate on average (four is pedwar, two syllables, against the single syllable of English four), and the same people's digit span in Welsh was clearly shorter than in English. Their own purpose was to argue against reading digit span as a measure of intelligence: what differs is not the brain but how many items per second you can refresh.
The ground is not firm, though. Stigler, Lee and Stevenson found in 1986 that Chinese children recalled at least two more digits on average than American or Japanese peers, and Chinese number words from one to ten are all monosyllabic — 一 (yī, "one"), 二 (èr, "two"), 三 (sān, "three") — which looks like a clean fit. But when Mattys, Baddeley and Trenkic measured bilinguals' articulation speed directly in 2018, it could not carry the gap: the correlation between span and articulation speed held for English natives and was much weaker for Mandarin natives. So no language here is "better for arithmetic" either. Digit span is one measure among others, and the within-bilingual comparison (the Welsh one) is far stronger evidence than a cross-country one.
Hurlburt took a plodding approach: give people an earpiece that beeps at random moments, have them note what was in awareness in the last instant before the beep, then interview them about each moment the next day. This is descriptive experience sampling. Heavey and Hurlburt's 2008 result: 10 random moments from each of 30 people, and inner speech — that soundless talking in your head — turned up in about a quarter of them.
Several other categories came in at roughly a quarter each: inner pictures, feelings, sensory awareness, and one called unsymbolized thinking, where you are definitely thinking something definite while neither words nor images are present. Individual variation was wide: some people were talking to themselves at nearly every beep, others at none. The evidence is thin in the usual ways — small sample, self-report, an interview method that is itself disputed — but it is enough to overturn one thing. The impression that you are talking to yourself all the time is not reliable; what you remember are the moments you noticed.
The shape is not like reading aloud, either. Fernyhough's questionnaire work splits inner speech into dimensions: condensed or expanded, monologue or dialogue, other people's voices present or absent, evaluative or not. Condensation is the easiest to verify on yourself: the inner sentence is often a bare predicate with subject and background stripped out — "blown it", "a bit more to the left" — and forcing it into a full sentence takes effort. The dialogic dimension is the foundation of rumination: you say the other person's line for them, so the exchange can keep going by itself. Children's out-loud self-talk works the same way — private speech increases as the task gets harder, and Vygotsky's account is that inner speech is that self-talk moved inside.
The most interesting case is signers. If the voice were really a voice, deaf signers ought not to have inner speech; in fact they report signing to themselves internally. Wilson and Emmorey transposed the phonological similarity effect to another modality in 1997: shown lists of signs, deaf signers recalled lists of handshape-similar signs worse than handshape-dissimilar ones. In 1998 they found a sign length effect as well — signs with longer movement paths are harder to recall, the counterpart of the spoken word-length effect. Same loop; the rehearsal is done with the hands rather than the mouth.
So "that voice in your head" is only half right. The more accurate version: the forms of your language are being rehearsed — by mouth or by hand, depending on which language you have.
Nedergaard and Lupyan ran the first systematic comparison in 2024: a self-report questionnaire split people into a high-inner-speech group of 47 and a low one of 46, then four tasks. Two showed a difference. Holding a list of similar-sounding words (the phonological similarity job from above) was worse in the low group, and so was judging whether two pictured objects' names rhyme — a cat and a hat.
Two tasks showed nothing: task switching, where you flip between two rule sets on cue and which is usually thought to rely on holding the current rule in words, and categorical effects on perceptual judgements. The authors gave the absence of inner-speech experience a name: anendophasia.
Evidence strength: one study, 93 people, grouped by self-report. Lind replied in the same journal in 2025 that the results are a real contribution, but that "some people have no inner speech at all" is not adequately supported — the scale asks how much you experience, which is not how much you use. The claim that holds up is narrower: how much inner speech people report correlates with performance on certain verbal tasks.
The harder blow comes from aphasia. Varley, Klessinger, Romanowski and Siegal reported in 2005 on three men with large lesions across the left-hemisphere language regions: grammar severely impaired, reading and writing number words difficult — and all the basic computational procedures intact, including the structure of arithmetic and which step has to be taken first. Earlier work from the same group found that severe grammatical impairment did not prevent passing false-belief tasks, where you judge what someone holding a mistaken view will do next. The grammar machinery broke; the skeleton of arithmetic and reasoning did not break with it.
So the debate now sits roughly here. The strong version — no language, no propositional thought — has few remaining defenders. What survives are two weaker branches. One says language is an enhancer that makes certain things easy: holding an ordered list, issuing yourself an instruction, pinning attention to a category; take it away and what you lose is exactly those, not thinking itself. The other says something stronger, that some conceptual operations cannot run at all without linguistic labels. The anendophasia results land between the two branches: the verbal tasks dropped, task switching and the categorical effect did not.
One last thing, said plainly: reporting no inner monologue is neither a deficit nor a gift — this field is still arguing about how to measure it.
At least three situations get compressed into that one sentence: the rehearsal loop does not run; it runs but produces no reportable experience; it runs and is experienced, but this person's threshold for what counts as "talking" is stricter than other people's. A self-report scale cannot separate them. Separating them means going around self-report — the silent-plus-aloud auditory damping task, say, or simply how much articulatory suppression costs them: if someone with "no inner speech" still loses span under suppression, the loop is running.
Not that far. They lost grammar after a lifetime of having it, and that arithmetic ability may well have been built with language in the first place; scaffolding can be removed once the structure stands. Telling apart "needed to run it" from "needed to build it" would require people who grew up with very little language input, and such cases are rare, bundled with other deprivation, and impossible to design ethically. So this question will probably be stuck with imperfect evidence for a long time.
The skeleton of the loop matches: similar forms mask each other, longer forms are harder to hold, and occupying rehearsal costs span. At least one thing does not match: signing uses two hands, the face and spatial locations at once, so a single instant can carry more components than a string of syllables, which means the two-second constant and the conversion into "how many items" should not be carried over from speech. Applying spoken-language span norms to signers is simply wrong.