Integration.

In all languages investigated, the sound of the OIR interjection shows some degree of calibration to local linguistic systems. Vowel targets are language-specific and appear to be drawn to existing phonemic targets, e.g. /e/ for Spanish and /a/ for Cha'palaa. Intonation melodies appear to be linked to the interrogative prosodic system, which may differ from language to language. The occurrence and quality of consonant onsets is related to the consonant inventory of the language. The interjection is also part of a larger paradigm of expressions for the other-initiation of repair, including, in English, other items like what? and pardon? [5], [15]. Huh? is thus an item integrated in several linguistic subsystems, from segmental and prosodic phonology to conversational structure.

Conventionalisation.

Huh? exhibits linguistic conventions that speakers need to learn in order to use the form properly. A learner of Spanish has to know that repair is initiated with the mid front unrounded vowel “e↗”, a learner of Cha'palaa has to know that the form is more like “a” with falling intonation, and a learner of Dutch has to know that a glottal fricative at onset is common: “h”. Its acquisition follows a normal trajectory, at least in American English-speaking children [37]. Second language learners' reports confirm that the precise form of this interjection has to be learnt, and that intuitions are not necessarily a reliable guide in this process [38].

Perhaps there is a continuum from non-linguistic vocalisations like sneezing and crying to prototypical conventional lexical items like bless you and pain [39]. Our evidence suggests that huh? is more on the word side of that continuum. Based on the fact that huh? is integrated in multiple linguistic subsystems and conventionalised in language-specific ways we conclude that huh? a lexical word.

Narrow bandwidth of variation.

We have already shown that the uniformity of the interjections is in striking contrast to the question words that languages can recruit for the same function (Table 1). Another way to appreciate the small range of cross-linguistic variation exhibited by this form is to consider it in context of the possibility space for words in spoken languages. Across languages, words can consist of one or more syllables, but the OIR interjection was never longer than one syllable in the languages we have studied, even in those like Murrinh-Patha, for which phonological words are generally longer than one syllable. Across languages, syllables can have rich internal structure, but the only structure attested in the OIR interjection is (C) V, i.e. a vowel V with an optional onset consonant C, even in languages like Dutch, where CVC syllables are common.

Similarity in vowels and consonants.

Strong constraints on variation are also seen in the vowels and consonants employed. Vowel space can be depicted as a two-dimensional plane formed by height and backness. On average, languages have around 6 vowel phonemes [40], which tend to be maximally spread across this space to increase perceptual distinctiveness [41]. Given this fact, it is striking that the vowels of OIR interjection tokens are only found in the low front central corner of vowel space (Figure 2), and that on a third dimension of lip rounding OIR interjections are only found on the ‘unrounded’ side. Consonants are articulated at different locations throughout the vocal tract (lips, teeth, alveolar ridge, palate, velum, uvula, pharynx, epiglottis) and with different manners of articulation, from plosives, nasals and trills to taps, fricatives, and glides – a multidimensional possibility space in which the International Phonetic Alphabet records at least 64 simple phonemic consonants (and three times as many complex variants) attested in the world's languages [4]. Out of this enormous range of possibilities, only two basic sounds, the glottal consonants [ʔ] and [h], are found in the OIR interjection across languages.

Such limited variation and striking similarity across languages is wholly unexpected on the basis of the principle of the arbitrariness of the sign. Does this mean that huh? is a universal word? We propose a qualified yes. Qualified, because huh? is clearly not phonetically the same word across languages – if Cha'palaa tokens were cross-spliced into Spanish dialog, Spanish speakers would likely be confused. What appears to be universal is the function of this interjection along with a set of constraints determining its form. Other-initiated repair sequences have been found in all languages investigated so far, and no language appears to lack an interjection for this function. Thus huh? is universal in the sense that a short, questioning interjection like it with the function of initiating repair is likely to be attested in all natural spoken languages.

Innateness.

On one account, huh? may be similar across languages because it is a natural symptom with a biological basis, like laughs and screams – a “grunt” [8], [10]. Such qualifications, common in the wider literature on interjections, place huh? in a position close to instinctive cries [7], [42]. This would be one explanation for its similarity: it is innate, therefore all humans share it, therefore it assumes roughly the same form in all languages. This view is as hard to support as it is to discount, but we note four doubts.

No known phylogenetic precursor.

Whereas laughter and pain cries (and by extension the conventionalized interjections associated with them) have demonstrable phylogenetic precursors in other mammals [43]–[46], there is, to the best of our knowledge, no evidence for an animal precursor of huh?. Nor is it obvious what the function and biological survival value of this precursor would be in primates which lack the kind of shared intentionality that underlies human cooperative communication [47], [48].

Not an involuntary response.

Grunts and other non-linguistic vocalisations such as sneezes and pain cries are typically direct, involuntary responses to stimuli [49]. In contrast, the OIR interjection is selected for a specific purpose at a specific juncture in conversation from a larger system of alternative formats for initiating repair [15], [50]. A greater degree of agency over utterance and selection is characteristic of linguistic rather than instinctive expressions.

Acquired like a normal word.

Whereas non-linguistic vocalisations like sneezes, cries and smiles are present at birth or develop soon after [51], [52], the acquisition of huh? follows a trajectory that is very similar to that of other linguistic items. In American English-speaking children, it is employed and responded to somewhat erratically at 2.5 years but perfectly at 5 years [37]. Related to this, the variability of laughs and screams appears to be much greater than what we find for huh?, and is not as strongly regimented by language [44], [53]–[55].

Parsimony.

In terms of evolutionary history, language is a recent arrival that shows clear signs of being a bio-cultural hybrid: a complex adaptive system in continuous cultural evolution within a landscape of cognitive, cultural, and communicative factors [56]–[59]. Although some of our linguistic abilities are no doubt underpinned by genetic infrastructure, positing innateness for specific linguistic items would hardly be realistic given the timescale involved. Strong cultural universals do not necessarily imply strong innate biases [60] and strong innate biases are unlikely to evolve in cultural systems [61]. If there is a mechanism that can explain cross-linguistic similarity on a more proximate timescale, without resorting to genetic encoding, this is preferred on scientific principles of parsimony.

Convergence.

A more plausible mechanism for the cross-linguistic similarity of huh? is convergent cultural evolution. This proposal sees huh? not as an arbitrary grunt but as a product of cultural evolution in the adaptive context of its interactional environment. The basic principle is well-known from biology: similar environmental constraints have led to the independent evolution of similar body plans in sharks and dolphins, and in the placental mammals of North-America and the marsupials of Australia. Likewise, we propose that the similarity of huh? in unrelated and distantly related languages is due to the fact that it is found in a strongly similar environment in all these languages. What is this environment like?

Conversations are built out of sequences of communicative moves between speakers [62], [63]. These moves –or ‘turns at talk’– are allocated in systematic ways and bear special sequential relations to each other [64], [65]. For instance, a question sets up an expectation that the addressee will provide a fitted response –in this case an answer– in the next move. Speakers inspect moves for their fittedness and aim to minimize gaps and overlaps between them. Speaker change most often takes only between 100–300 milliseconds, and deviations from the timing target can be treated as problematic [66], [67]. In order for this tight timing to work, planning a next turn often has to start well before the end of the preceding turn [68], [69]. Trouble in hearing or understanding is a regular feature of conversation [5], [47]. In the case of such trouble, planning and producing a fitted and timely response will be harder (indeed at times impossible), but the pressure to produce one will be just as strong. Given these pressures of turn-taking and formulation in conversation, a signal that indicates trouble should be minimal and easy to deploy. At the same time, given the communicative importance of indicating trouble (which if not solved might derail the conversation), such a signal should also clearly indicate a knowledge deficit and push for a response. These requirements are met rather precisely in the combination of minimal effort and questioning prosody that characterises the OIR interjection across languages.

Minimal effort.

Many of the formal aspects of the OIR interjection minimize articulatory effort. The codaless monosyllable is the least marked syllable type across languages [68], [70]. The glottal onset, where present, is simply some constriction at the narrowest place in the vocal tract, and the unrounded low front central vowel is close to the neutral state of the articulators – both requiring minimal encoding, planning, and articulation [34]. Additionally, for Spanish phonetic corpus studies show that the vowel target of the interjection is the most frequently attested vowel [71], making retrieval, planning, and production easier [72]. These features render the OIR interjection well-fitted to the interactional environment of other-initiated repair. For the person initiating repair, the OIR interjection is quickly deployable from intention to articulation [68], and therefore easy to produce even under conditions of cognitive duress. For the addressee, the minimal form is a word that is unlike most content words and therefore –by Darwin's principle of antithesis [45]– a good signal that the other has no contentful response on offer.

Questioning prosody.

If ‘minimal’ were the only design requirement, the most low-effort form possible would be enough. But to carry out the work of initiating repair, the OIR interjection also has to signal a knowledge deficit and indicate that a response is needed. We have seen that the intonation of the interjection appears to be calibrated to local systems of questioning prosody. In many languages this means that it has rising intonation – a contour that requires more effort than falling intonation [73], and (in English-speaking infants) has been shown to elicit greater attention [74]. In Cha'palaa and Icelandic, where the OIR interjection has falling intonation, it has a low central unrounded vowel – the vowel that is inherently most sonorous and acoustically salient due to the wide open oral cavity [75]. We propose that the questioning prosody and the acoustic salience of the interjection render it more adaptive for the function of OIR. As a question word devoid of semantic content, it expediently returns the floor to the original speaker and signals that there is trouble to be fixed.

In effect, huh? is an easy to produce, maximally underspecified question word – a tight fit of form and function found in language after language. We propose that this is the result of convergent cultural evolution: the interactional environment of other-initiated repair, present in every language investigated so far, provides a set of selective pressures that pull the interjection towards a similar form and that keep regular processes of language change from affecting the item. This process of convergent evolution explains the narrow bandwidth of the variation, but also the language-specific calibration of the items. To minimize articulatory effort, the OIR interjections of different languages will end up in the same low-effort area of the phonetic possibility space; yet to be recognised as questioning expressions, they will be calibrated to local phonological and prosodic subsystems.

We use ‘convergent evolution’ as a general term for the independent evolution of similarities in form and function. When ancestral forms are known, a distinction can be made between form/function convergence in species that are closely related (‘parallel evolution’) versus in species that are not closely related (‘convergent evolution’). However, this distinction is not always consistently made in biology and recently there have been proposals to use ‘convergent evolution’ as a general term [76]. We use the term in this general sense. Our proposal accounts for the present-day cross-linguistic similarity of huh?, but has to remain agnostic as to its ultimate origins – in the absence of historical language data it is impossible to tell whether the present-day forms go back to one ancestral form (a stabilising evolution scenario [77]) or whether they arose independently in different languages (an independent convergent evolution scenario [78]). In either case, the selective pressures are the same.

The convergent evolution proposal explains the forms documented so far, but also generates the prediction that in undescribed languages as well as newly emerging ones, we can expect to find a similar repair initiator that minimizes articulatory effort while making use of questioning prosody. Independently emerged sign languages of the deaf, though in a different expressive modality (visual-only instead of audio-visual), provide a good test case. Consistent with our proposal, in Argentinian Sign Language, repair can be initiated with a minimal sign that involves a raising of the eyebrows, the semiotic equivalent of questioning prosody [14],[79].

Conversational infrastructure and convergence of linguistic form.

Apart from its explanatory and predictive value, the convergent evolution proposal offers a more general mechanism. For most words in most languages, there is no necessary connection between form and function. This is why words can change over time, and why we expect even words with similar functions to have different forms in unrelated languages. Accordingly, cultural evolutionary models of language change have tended to depict languages as collections of words evolving in utterances [80], [81], with various social and cognitive biases influencing transmission [82], [83] and with frequency of use as a primary factor influencing rates of change and divergence [84], [85]. However, our study points to a factor that may constrain divergence or diachronic drift: the selective pressures of specific conversational environments, which may cause convergent cultural evolution.

The possibility should not be surprising. After all, words evolve in utterances in conversation, so conversational infrastructure is part of the evolutionary landscape for words. We are referring here to the sequential infrastructure that serves as the common vehicle for language use – an infrastructure that may well predate more complex forms of language and that seems largely independent of sometimes radical differences between individual languages [63], [66], [86], [87]. A clear effect of this conversational ecology on the cultural evolution of linguistic items has not, to our knowledge, been observed before.

Though we have focused on huh? as a case study, the mechanism we propose has wider relevance. In our corpora, we have noted other items that are strongly similar in form and function across unrelated languages: continuers like mm/m-hm [88], hesitation markers like uh/um [89], [90], and change of state tokens like oh/ah [91]. It would be neither plausible nor parsimonious to propose that all of these have precursors in distinct innate grunts. Instead, we observe that these interjections all serve important discourse regulatory functions, and we propose that the reason they are so similar across languages is that common communicative needs and conversational infrastructure conspire to create, for each of them, a set of similar selective pressures constraining their evolution. The ultimate fit to the tight constraints of their conversational environments, these words stay put and help us conduct conversation in optimal ways. The approach followed in this study can be systematically extended to the larger set of discourse regulatory expressions and beyond, to explore further effects of conversational ecologies on language structure.