In 1929 the German psychologist Wolfgang Köhler showed people two shapes — one rounded and blob-like, one spiky and star-shaped — and told them that in an invented language one was called takete and the other maluma. Almost nobody hesitated. The spiky one was takete. The round one was maluma. Ramachandran and Hubbard repeated the test with the now-famous pair bouba and kiki and reported that about 95 per cent of respondents gave the same answer, in the same direction, without any instruction.

This should not work. The founding assumption of modern linguistics, inherited from Saussure, is that the link between a word's sound and its meaning is arbitrary — that there is nothing dog-like about the sounds in dog, which is why the same animal is hond in Dutch and inu in Japanese. Arbitrariness is what makes language powerful: any sound sequence can carry any meaning, so vocabulary is unconstrained.

And yet the spiky shape is kiki, reliably, across cultures and ages. That fact has a practical consequence for anyone learning vocabulary. Some words in your target language are not arbitrary at all, and those words are measurably easier to learn — but only if you notice why, and only if you know where the pattern stops working, because it stops in ways that will quietly teach you the wrong meaning.

The Test You Cannot Fail

The bouba/kiki effect is the most replicated result in this literature, and it has been pushed well beyond European undergraduates. In 2022 a team tested speakers of 25 languages across nine language families and ten writing systems, and found the effect everywhere: the estimated proportion of bouba/kiki-consistent matches was 72 per cent overall, with a 95 per cent credible interval from 56 to 82. Participants who used Roman script were only marginally more likely to show the effect, and scripts whose letter shapes happen to be rounder for bouba did not produce stronger effects. The bias is in the ear and the mouth, not the page.

The strongest single test of the cultural explanation came from the Himba of northern Namibia, a remote population with little exposure to Western media and no written language. Bremner and colleagues (2013) found that Himba participants made shape–sound matches in the same direction as Westerners. They differed on shape–taste matches, which suggests culture shapes some cross-modal associations and not others — but the sound–shape mapping survived.

Köhler's original demonstration is worth keeping in mind for a practical reason. The effect is not a subtle statistical tendency that requires a lab. It is a bias strong enough that a room full of people who have never heard the words will converge on the same answer, and it is already doing work inside every language you might study.

Four Things People Mean by "Sound Symbolism"

The phrase covers at least four different phenomena, and mixing them up produces bad study advice. The taxonomy that follows is the one used in Dingemanse and colleagues' 2015 review, which distinguishes arbitrariness, iconicity and systematicity as separate properties of a lexicon.

  • Onomatopoeia — direct imitation of a sound: buzz, hiss, crash, Japanese wanwan for a bark. The word copies the acoustic event.
  • Ideophones — a whole word class, not a curiosity. Ideophones depict sensory events: sounds, manner of motion, texture, inner states. Japanese has the largest and most studied inventory, split into giongo (sounds), gitaigo (manner and state) and gijōgo (feelings). Korean has 의성어 and 의태어. Dingemanse's cross-linguistic work sampled 208 mimetics from five languages — Japanese, Korean, Semai, Siwu and Ewe — and found both shared and language-specific sound–meaning links inside a single semantic domain.
  • Phonaesthemes — a sub-morphemic sound cluster that carries a meaning without being a morpheme. English gl- clusters around light and vision: glow, gleam, glimmer, glisten, glitter, glare, gloom. sn- clusters around the nose: sniff, snore, snort, snout, sneeze. -ump clusters around blunt mass: bump, lump, hump, thump, dump. Bergen (2004) showed that English speakers process these clusters as meaningful units, with priming between words that share them.
  • Cross-modal correspondences — the synaesthetic layer: high front vowels like /i/ read as small, low vowels like /a/ as large; high pitch as small and bright, low pitch as large and dark. Sapir demonstrated the size effect in 1929 with the nonce words mil and mal: English speakers matched mil to a small table and mal to a large one.

The distinction that matters most for study design is between iconicity and systematicity. Iconicity is a resemblance between form and meaning that a learner can feel: kiki is spiky because the sound is spiky. Systematicity is a statistical regularity in a language that need not be feelable at all — the -ung nouns in German are feminine, and nothing about the sound of -ung means "feminine". Iconicity is the part that gives you a free first encoding. Systematicity is the part you have to be taught, and it belongs in a rule deck, not a mnemonic.

The Evidence That This Is Not Just a European Party Trick

The obvious objection to bouba/kiki is that English speakers have been primed by a lifetime of English words. The best answer is a study that ignored meaning entirely and looked at the shape of basic vocabulary at global scale.

Blasi, Wichmann, Hammarström, Stadler and Christiansen (2016), published in PNAS, took word lists covering nearly two-thirds of the world's languages and tested 100 basic vocabulary items for associations with particular speech sounds. They kept only associations attested in at least ten different lineages and three different macro-areas, with no area running in the opposite direction. 74 sound–meaning associations survived. Among the strongest: "small" tends to contain /i/; "nose" tends to contain a nasal; "tongue" tends to contain /l/; "full" tends to contain /p/ or /b/. The areal and historical distribution suggested these associations arose independently rather than being inherited from a common ancestor.

The size dimension is the most robust. Shinohara and Kawahara ran a rating experiment on nonce words in four languages — Chinese, English, Japanese and Korean — asking speakers to rate how large the invented word felt. Vowel backness produced the strongest effect (back vowels read as larger), vowel height a weaker one, and obstruent voicing a strong one. Crucially, Korean gives the pattern a structural home. Its ideophones split vowels into a "bright" class — ㅏ and ㅗ — and a "dark" class, and brightness encodes lightness, smallness and quickness. Which is why a trickle is 졸졸 and a stream is 줄줄, why a little sparkle is 반짝 and a big flash is 번쩍, and why a small hop is 폴짝 while a big leap is 풀쩍.

Does It Actually Help You Learn?

All of the above establishes that the bias exists. It does not establish that it helps a learner, and that is the only question that matters for a deck. Three experiments answer it directly.

The learning advantage in real words

Nygaard, Cook and Namy (2009) taught monolingual English speakers the meanings of Japanese antonym pairs. Each word was heard alongside either its true translation, its antonym, or an unrelated word. At test, participants were faster and more accurate when the word had been learned with its correct meaning than with an unrelated one. The finding has since been replicated and extended, and it established the basic paradigm: sound–meaning correspondence is not just a matching task, it is a learning advantage.

The verb-learning result in children

Imai, Kita, Nagumo and Okada (2008) is the cleanest demonstration of the mechanism, because it separates sound symbolism from every other cue. They built novel verbs, confirmed by Japanese and English adults, that either matched or did not match the action they described, then taught them to Japanese three-year-olds. Children matched novel mimetics to actions well above chance at 25 months (65.7 per cent) and at three years (75 per cent). In the learning task, three-year-olds generalised a novel sound-symbolic verb to a new instance of the same action at 82 per cent, and failed on a plain novel verb at 54 per cent — chance. When the mimetic word did not match the action it described, performance fell to 44 per cent. The sound was not decoration. It was the scaffold that let the child isolate the manner of the action from the actor and the object.

The ideophone result in adult learners

Lockwood, Dingemanse and Hagoort (2016) ran the version that speaks most directly to flashcard practice. They taught Dutch speakers with no Japanese two sets of Japanese ideophones: one set with the real Dutch meanings, one set with the opposite meanings. Participants learned the real pairings significantly better than the reversed ones — and, strikingly, they could still guess the real meanings in a two-alternative forced-choice test afterwards, even when told about the manipulation. Then the authors repeated the whole thing with ordinary Japanese adjectives such as taisetsu ("important") and abunai ("dangerous"). No difference.

That null result is the most important part of the study. It rules out the explanation that foreign words are simply more memorable in a novel-sounding way, or that the effect comes from task interest. The advantage belongs to words whose form depicts their meaning.

Why It Works

Three mechanisms are usually proposed, and they are not mutually exclusive.

  • Constrained hypothesis space. Learning an arbitrary word means choosing one meaning from everything the world contains. A word that sounds like its meaning rules out most of the space before you start. This is the bootstrapping argument, and Imai's verb data are its clearest support: the sound told the child where to look.
  • Cross-modal mapping. Size and pitch, shape and articulation, brightness and vowel height — the brain already links these dimensions. A word that rides an existing cross-modal link needs one less arbitrary association to store.
  • Motor resonance. Producing kiki involves a narrow, constricted vocal tract; producing bouba involves a rounded one. The articulation mirrors the shape. Kita's work on Japanese speakers showed that mimetics are produced with matching iconic hand gestures almost automatically, which suggests the sound is bound to a bodily representation rather than to an abstract code.

For a flashcard learner, all three point the same way: an iconic word gets extra retrieval routes for free, and a deck that strips the sound out of the card throws them away.

Where the Map Breaks

This is the section that most popular articles skip, and the one that decides whether this knowledge helps or hurts you.

The effect is real and small

Monaghan, Shillcock, Christiansen and Kirby (2014) analysed the English vocabulary at corpus scale and found that sound–meaning mappings are more systematic than chance, robustly, across every control they ran — including monomorphemic words with no shared etymology. Then they reported the size of the effect: r² below 0.002. Sound explains well under one per cent of the variance in meaning across the vocabulary. Their more interesting finding is developmental: systematicity is stronger for words acquired early and declines for later-acquired words. The vocabulary appears structured to give children a foothold early, then to become arbitrary as it grows. If you are an adult learning abstract, low-frequency vocabulary — policy, assume, nevertheless — sound symbolism will do almost nothing for you.

The same feature can point in opposite directions

Saji and colleagues (2019) asked Japanese and English speakers to invent sound-symbolic words for 70 locomotion videos, then analysed the phonetic features statistically. Both languages used voicing as a primary cue — and mapped it in opposite directions. In Japanese, voiced consonants read as heavy and big; voiceless as light and small. In English, voiceless consonants read as fast and energetic. The feature is shared. The direction is not.

This is the trap in its purest form: a learner who has internalised "voiced equals heavy" from Japanese will guess wrong in English, and vice versa, while feeling confident. And a handful of published studies report null or reversed bouba/kiki results altogether (Rogers and Ross, 1975; Oberman and Ramachandran, 2008; Styles and Gawne, 2017), which means even the flagship effect is a strong tendency rather than a law.

Ideophones are a minority of any lexicon

A standard dictionary of Japanese mimetics — Giongo/Gitaigo 4500 — lists 4,500 entries. That sounds enormous until you put it beside a vocabulary of 20,000 to 40,000 word families for advanced reading. Sound symbolism is a well-lit corner of the lexicon, not the building. Treat it as an accelerator for a specific, high-value subset: manner, texture, motion, sound, intensity — the words that make speech vivid and that dictionaries translate worst.

Translation flattens them

Most ideophones do not map onto a single English word. Kirakira is not simply "sparkle"; it is a light, repeated, scattered glitter. Gira gira, with the voiced consonant, is a hard, glaring shine — and also the feeling of being on edge. If your card says "sparkle" on the back, you have learned a word that will be wrong in half the sentences you meet it in. The correct card is a scene, not a gloss.

Where the Gold Is

  • Japanese mimetics. Three classes, heavy reduplication, and a voicing alternation that functions as an intensity dial: sara sara (smooth, dry) against zara zara (rough, gritty); kira kira (sparkling) against gira gira (glaring, grating). Learn the pairs and the dial comes with them.
  • Korean ideophones. A vowel-brightness dial for size and weight (반짝 / 번쩍, 졸졸 / 줄줄) and a consonant-force dial for intensity: plain 달가닥, tense 딸가닥, aspirated 탈가닥 — three gradations of the same clatter.
  • English phonaesthemes. gl-, sn-, sl-, str-, -ump. Not a rule you can apply mechanically, but a family you can card as a set, which is far cheaper than carding the members separately.
  • Diminutives and augmentatives. Spanish -ito, Italian -ino, German -chen, Portuguese -inho — all built on high front vowels, all consistent with the cross-linguistic size mapping. Same for the English nursery pattern: teeny, little, bitty.

Building It Into a Deck

  1. Card the pair, not the word. 반짝 alone teaches you one word. 반짝 against 번쩍 teaches you the dial, and the dial generalises to words you have not met. This is the single highest-return change you can make to an ideophone deck.
  2. Put the sound on the front. Audio first, meaning second. The whole advantage is auditory and articulatory; a text-only card reduces an iconic word to an arbitrary one.
  3. Test perception before production. Early cards should ask "which is bigger, banjjak or beonjjeok?" That trains the mapping. Only later should the card ask you to produce the word for a scene.
  4. Verify before you trust the sound. Treat iconicity as a hypothesis, not a translation. Guess, then check. Log every word whose sound misled you; that log is the most valuable deck you will ever build, because it is the list of errors you would otherwise repeat with confidence.
  5. Write scenes, not glosses. Back of the card: one sentence in the target language plus a plain-language gloss of the effect. 雨がざあざあ降っている — rain coming down hard and steady — not "rain sound".
  6. Keep a voicing-contrast subdeck. Voiced against voiceless minimal pairs are the highest-density information in Japanese and Korean mimetics, and they are exactly the pairs that a one-word gloss destroys.
  7. Interleave the iconic words with the arbitrary ones. A session of nothing but mimetics teaches you "today's answer is vivid". Mixed sessions force you to retrieve the actual meaning, which is the same desirable-difficulty logic that governs spacing.
  8. Card phonaesthemes as one family card. Front: gl-. Back: the six or seven members plus the shared semantic thread. One card, one pattern, seven words.
  9. Say it out loud at least once per session. The mapping runs through the articulators. Silent review keeps the meaning and loses the cue.

What Does Not Work

  • Inventing sound-alike stories for arbitrary words. The keyword method is a real technique, but it manufactures a private iconicity. Mixing manufactured hooks with genuine ones destroys your ability to tell which is which, and the manufactured ones do not transfer to new words.
  • Assuming the direction is universal. Saji's voicing result shows a shared feature with opposite mappings. Guess in the target language's direction, or do not guess.
  • Lists of a hundred ideophones with one-word translations. This is the standard appendix in a textbook, and it produces words you recognise on the page and cannot use in a sentence.
  • Skipping the audio. A text-only mimetic card is an arbitrary card with extra steps.

The Honest Limits

Three qualifications keep this honest. First, the effect sizes are modest. Lockwood's advantage is real and replicated, but it is a boost to first encoding, not a substitute for retrieval practice; nobody has shown that sound symbolism reduces the number of reviews a word needs over a year. Second, the corpus evidence (r² below 0.002) says arbitrariness dominates the adult vocabulary, so a learner who chases iconicity will spend most of their time on the wrong words. Third, almost all of the learning experiments use ideophones and mimetics — the most iconic class in any language. The results may not extend to the abstract vocabulary that carries most of the meaning in real texts.

What survives is a narrow, useful claim: for the subset of vocabulary that depicts sensory experience, form carries information about meaning, learners are sensitive to it across cultures, and pairing form with meaning improves learning. That is enough to change how you build a deck. It is not enough to change how you study everything.

The Bottom Line

Language is mostly arbitrary and occasionally not. The non-arbitrary part is concentrated in exactly the words that dictionaries handle worst: manner, texture, motion, intensity, sound, and the feeling of things. Those words are also the difference between speech that is correct and speech that is alive.

So the practical advice is short. Learn the pairs, not the words, because the pair contains the rule. Put the sound in the card, because the cue lives in the ear and the mouth. Guess before you check, then check, because a confident wrong guess is the most expensive thing in a deck. And keep the arbitrary words in a separate pile, because no amount of kiki will help you with nevertheless.

FluentCards is built for the part that remains: audio on every card so the sound is present at review, phrase-level answers that force you to produce rather than recognise, and FSRS scheduling that brings the arbitrary majority back often enough to stay. Sound symbolism gives you a head start on the vivid words. Retrieval does the rest.