The alveolar trill is perceived as jagged/rough by speakers of different languages^a)

There is a large amount of research on sound symbolism, documenting how people often attribute meaning to speech sounds (Lockwood and Dingemanse, 2015). For example, experiments with speakers from different languages show that the high front vowel [i] is associated with the meaning of smallness, compared to low back vowels (Auracher, 2017; Hoshi , 2019; Knoeferle , 2017; Newman, 1933; Parise and Spence, 2012; Sapir, 1929; Tarte and Barritt, 1971). This pattern is hypothesized to stem from the fact that the high second formant frequency and large dispersion of the first and second formant frequencies of [i] correspond to the acoustics of small resonators (Fitch, 1994; Ohala, 1983; Winter , 2021). Importantly, this pattern has not only been attested in experiments, but is reflected in vocabulary structure across languages, with high front vowels occurring more frequently in words denoting smallness (Blasi , 2016; Fitch, 1994; Haynie , 2014; Huang , 1969; Johansson , 2020; Johnson, 1967; Levickij, 2013; Thorndike, 1945; Ultan, 1978; Winter and Perlman, 2021). This evidence from lexical and experimental studies is understood as a case of iconicity—a resemblance between the form of a signal (e.g., a word, gesture, or sign) and its meaning. A growing number of scholars argue that iconicity is a fundamental property of languages, spoken and signed (Dingemanse , 2015; Perniss , 2010).

In cases such as the association between vowels and size, iconicity is crossmodal, mapping between sound and qualities that are primarily related to different sensory modalities. Perhaps the most famous example of crossmodal iconicity is the bouba/kiki effect, where nonce words like bouba (or maluma) are matched to round shapes, as opposed to nonce words like kiki (or takete), which are matched to angular shapes (Köhler, 1929; Ramachandran and Hubbard, 2001). This association has been experimentally demonstrated across cultures with speakers of a large set of genealogically diverse spoken languages (Bremner , 2013; Ćwiek , 2022), and observational studies have found that roundness/angularity is statistically associated with bouba- and kiki-like speech sounds in the lexicon of English (Sidhu , 2021). Experimental evidence suggests that multiple analogies may underpin the perceived resemblance between bouba/kiki and round/angular shapes, including mediation through emotional arousal (Aryani , 2020) and through the similarity between the word bouba and the sounds produced by falling or bouncing round objects as opposed to angular ones (Fort and Schwartz, 2022).

The current study focuses on another case of crossmodal iconicity that may exert an influence on the phonological shape of words: the association of rhotic consonants with rough texture. An early study asking American English speakers to rate the qualities of speech sounds found that /r/ was judged as rougher than other phonemes (Greenberg and Jenkins, 1966). In line with this result, a cross-linguistic analysis of poetic texts found that /r/ was over-represented in poems with an aggressive rather than tender tone (Fónagy, 1961). It is worth noting that in those studies, it remains unclear from which specific realizations of the phoneme (as [r], [ɹ], [ʀ], or another speech sound) the conclusions are drawn. Recently the association between /r/ and roughness has been found to be widespread across spoken language vocabularies. Winter (2022) first showed that for a set of 100 English adjectives rated for roughness (e.g., rough, abrasive, prickly, smooth, coarse, cottony, silky, oily), the rhotic phoneme is statistically associated with descriptors of rough surfaces. This pattern was also found across 38 other Indo-European languages and replicated in Hungarian, a Uralic language. In a typological analysis of vocabulary data from lexical databases, trilled /r/ sounds, as indicated in the phonologically coded lexical databases, were found to be much more common in translational equivalents of “rough” rather than “smooth” for a diverse sample of 332 spoken languages from 84 phyla (see also Levickij, 2013). Considering that perceptual studies of surface touch suggest that the spatial frequency of grating patterns is a primary determinant of textural roughness (Hollins and Bensmaïa, 2007; Lederman, 1974, 1983), and that spatial frequency is perceptually associated with auditory amplitude modulations (Guzman-Martinez , 2012; Orchard-Mills , 2013; Sherman , 2013), Winter (2022) suggested that the intermittent tongue movements of trills and the resulting repetitive amplitude modulations [see Fig. 1(a)] might provide the iconic motivation behind this pattern.

However, a recent study by Anselme (2022) calls into question whether the statistical association with rough meanings in vocabulary data is specific to the trilled /r/ phoneme or whether it might be associated with rhotic consonants more broadly. Anselme (2023) provide evidence that the precise phonetic realization of rhotics has not always been accurately documented in the databases used by Winter (2022): Although /r/ technically symbolizes an alveolar trill according to the International Phonetic Alphabet (IPA), it is often used to represent a generic “r-like” sound, not reliably distinguishing whether it is typically realized as a trill. Anselme (2022) recoded a substantial portion of Winter (2022) typological data, and in their re-analysis found that “r-like” sounds in general, not just trills, are associated with roughness across spoken language lexicons. Thus, it is not entirely clear whether the patterns found in the analysis of Winter et al. of lexical data is rooted in an iconic association between roughness and /r/ realized specifically as an alveolar trill or whether it is driven by rhotics more generally, regardless of how they are phonetically realized.

Taken together, evidence from lexical databases suggests that the trilled phoneme /r/ is associated with roughness (Winter , 2022). However, the overuse of the IPA symbol /r/ to represent in writing various r-like sounds without, in many cases, specifying unambiguously which particular speech sounds it stands for prevents us from concluding that it is specifically the trill that bears this semantic association (Anselme , 2022). The current study addresses this ambiguity by directly testing the connection between specifically the alveolar trill [r] realized in a controlled and explicit manner and roughness across a diverse language sample in order to explore the cross-linguistic potential of the hypothesized association. We follow up on the lexical pattern found by Winter (2022) with a perception experiment to assess whether the alveolar trill [r] is perceived as rough by speakers of 28 different languages. We tested [r] against the lateral [l], another liquid with an alveolar place of articulation, but with no strong repetitive amplitude modulations [see Fig. 1(b)]. By presenting acoustic stimuli to speakers of a diverse set of languages, we were also able to assess the extent to which speakers of languages with different phonetic realizations of rhotic consonants differ in their crossmodal associations of the alveolar trill [r]. Given that phonemes primarily serve a contrastive function to distinguish words within languages, it is possible that speakers of a language that uses an alveolar trill [r] as the primary phonetic variant may treat this phoneme as relatively more “arbitrary,” and less imbued with meaning. Therefore, we wanted to assess whether having the alveolar trill [r] as the primary allophone of the /r/ phoneme in one's language could potentially diminish the strength of its iconic association. Similarly, we were able to explore whether distinguishing /r/ and /l/ sounds phonemically in one's grammar also plays a role in modulating the perceived crossmodal iconicity of the alveolar trill [r].

In our experiment, we used the shapes shown in Figs. 1(c) and 1(d) as visual stimuli, asking participants to imagine what it feels like to touch these surfaces. Thus, the connection specifically between touch and sound, which we were seeking to investigate, is indirect with these visually presented stimuli, despite highlighting haptic touch to our participants via the instructions. However, the use of visual stimuli rather than felt surfaces was necessary to conduct the experiment online (see below), which prevented the use of textural stimuli. In selecting these visual shapes as representations of textures, the contrast between a jagged shape and a flat shape was motivated by studies suggesting that the frequency of spatial grating predicts roughness (Hollins and Bensmaïa, 2007; Lederman, 1974, 1983). Furthermore, surface texture is what is known as a “common sensible,” a percept that can be perceived through multiple different modalities (Marks, 1978). Roughness, in particular, can also be perceived via vision (Lederman and Abbott, 1981) and audition (Lederman, 1979) and has similar psychometric functions in these modalities. Nevertheless, our experiment is somewhat ambiguous with respect to vision and touch, testing a stimulus that can be perceived either as “jagged,” relating to the construct of “shape,” or as “rough,” relating to the construct of “texture.”

The experiment reported here was conducted as a part of a larger study (Ćwiek , 2021; Ćwiek , 2022), which included both an online web experiment and an on-site field experiment. Our overarching goal for conducting the same experiment online and in the field was to maximize linguistic and cultural diversity, with the goal of targeting non-WEIRD (non-Western educated industrialized rich democratic) communities (Blasi , 2022; Henrich , 2010). Participation in the web experiment required literacy as well as access to and experience with the Internet. The experiment conducted on site did not require participants to be literate, and thus it allowed us to target speakers with limited formal education as well as limited access to the Internet and globalized culture.

As the web and field experiments differ only slightly (see Sec. II C) and are also analyzed in the same statistical model, for ease of presentation (see Sec. II E), we treat them as two separate samples from the same study. All of the data and code for the experiments are available in an Open Science Framework repository at https://osf.io/mjcnq/.

We used opportunity sampling for both the web experiment and the field experiment. All participants indicated their informed consent and completed the study on a voluntary basis.

For the web experiment, we distributed the survey online via social media or via directly contacting native speakers and asking them to share the link to the experiment with their friends and family. Our initial convenience sample for the web experiment included data from 975 participants. We excluded participants who indicated they did not speak the language of the survey (n = 9), who failed to provide both responses (n = 38), or who selected a response without playing back the sound (n = 22). Additionally, we did not obtain enough Tamil and Malagasy data (one and two speakers, respectively) for them to be included in our analysis. In total, we excluded 72 participants (7.4%) from the analysis, leading to a final sample with data from 903 participants representing 25 languages from nine language families, as detailed in Table I. A total of 781 participants (86.5%) spoke a second language, and 122 participants (13.5%) self-reported to be monolingual. Of the participants who were not native English speakers, 727 (84.1%) spoke English as a second language. In terms of gender composition, our sample included 681 female speakers (75.4%) and 222 male speakers (28.3%). Participants ranged from 18 to 84 years of age (mean 32.9 years, median 29 years).

For the field experiment, opportunity sampling involved collaborating with linguists who were going on field visits during the period of the study. The field experiment was conducted at 6 sites, with a total of 133 participants who were speakers of six different languages from four families, including Palikúr, Brazilian Portuguese, Daakie, Tashlhiyt, German, and English (see Table II). Four of these language groups (Palikúr, Brazilian Portuguese, Daakie) were targeted as non-WEIRD communities, with limited formal education and access to the internet and globalized culture. Palikúr data were collected at the banks of Oyapock River near St. Georges de l'Oyapock in French Guayana (at the border with Brazil). Brazilian Portuguese data were collected with a quilombo community from the Cametá region in Brazil. Both Palikúr and Brazilian Portuguese speakers live in Amazonia and are rural communities of farmers/hunters who sell their goods on the market. Daakie data were collected with a farming/hunting community living in Port Vato on Ambrym, Vanuatu. All three communities do not have regular access to electricity, and the use of mobile phones is highly limited because of a lack of resources and connection service. Access to education is limited in these communities. For comparison, English, German, and Tashlhiyt speakers were recruited so that we could disentangle the effects of task (web experiment versus field experiment) from characteristics of the participant sample. English and Tashlhiyt data were collected among university students in Birmingham, UK, and Agadir, Morocco, respectively. German data were collected among residents of a holiday resort in Lubmin, Germany. The specific settings and participant samples for the field experiment differed across the six language groups, reflecting the various on-site conditions. The Daakie speakers took part in the study in a small concrete building belonging to the Presbyterian Church, seated at a table on a bench, with efforts made to minimize distractions from bystanders. Brazilian Portuguese participants performed the task in their homes; Palikúr speakers performed the task in a communal building, where they were interviewed one-on-one in a separate room. English and Tashlhiyt participants performed the task in a quiet room in a university; German speakers performed the task in a quiet bungalow in a holiday resort.

Sample size in the field experiment varied based on the availability and willingness of participants on each site. We excluded six participants (4.5%) who failed to provide responses for each sound stimulus, leaving us with a sample of 127 participants. Of these, 75 speakers (59.1%) spoke a second language, and 52 speakers (40.9%) self-reported to be monolingual. Specifically for the three target languages—Palikúr, Brazilian Portuguese, Daakie—the number of second language speakers was 21 (63.6%), in contrast to 12 monolingual speakers (36.4%). Only 1 participant from the target languages (3.0%) self-reported to know English, as opposed to 32 who did not (97.0%). The final sample included 91 female participants (71.7%) and 36 male participants (28.3%). Ages ranged from 18 to 75 years (mean 28.6 years, median 20.0 years).

The acoustic stimuli included a recording of the alveolar trill [r] and a recording of the lateral alveolar approximant [l] [see Figs. 1(a) and 1(b)]. These sounds were produced by a native Polish speaker with training in phonetics (author A.Ć.). The sounds were produced in isolation, without any carrier phrase or vocalic context. The rough and smooth textures were represented with two line drawings: one jagged/rough, one flat/smooth [Figs. 1(c) and 1(d)]. Participants were instructed to imagine moving their finger along the lines to emphasize the touch dimension.

In addition to the current experiment, the complete study included a main task involving guessing the meaning of novel iconic vocalizations (Ćwiek , 2021) and an additional task involving bouba/kiki (Ćwiek , 2022), with the current study always coming last. Thus, these other tasks were both related to different kinds of vocal iconicity and sound symbolism. Importantly, however, participants were not provided with any feedback on their guessing in either of the previous experiments. For the entire set of studies, we collaborated with native speakers who translated the consent forms and instructions (Ćwiek , 2021, Ćwiek , 2022). The web experiment was hosted on the Percy platform (Draxler, 2011) and accessed by participants via their personal computer, smartphone, or tablet over the Internet. The field experiment was conducted orally in the native language of the participants, including the consent process and all instructions. The consent and the instructions were read to the participants, and they also had the opportunity to read these themselves. All participants provided signed consent. For English, German, and Tashlhiyt speakers, this procedure and the experiment were conducted by linguists who were also native speakers. In the case of Daakie and Brazilian Portuguese speakers, this was done by linguists who knew the respective languages. For Palikúr speakers, the field linguist conducted the experiment in Brazilian Portuguese with an on-site interpreter translating into Palikúr.

The task was identical across all languages, but differed slightly between the web and field experiments. In the web experiment, participants were presented with images of the two lines next to each other on a screen. They then listened to each auditory stimulus separately, making a response after hearing each sound (sequential rather than paired matching). The order of presentation of the auditory stimuli, as well as the images (left versus right), was randomized. For the field experiment, participants were simultaneously presented with both lines and were played both sounds via laptop speakers of the respective experimenter before making their response, enabling paired matching after listening to both sounds. The lines were printed out on white paper in A5 format and presented on a table in front of the participant. In contrast to the web experiment, the presentation order of the line drawings (left versus right) and the sounds was not recorded and not controlled for.

In the web experiment, participants could click to replay each sound, and in the field experiment, they could ask the experimenter to play a sound again. After completing the full study, participants were asked for background information on their sex, age, native language(s), and other known languages—via written questions in the web experiment and via oral questions in the field experiment. The web experiment additionally asked for the participants' country of residence and the place where they entered primary school. Additionally, we inquired about the environment in which they completed the survey, the input device and audio output device they used, and their hearing ability.

To investigate the effect of language background on participants' judgments, we coded what rhotic variant characterized each of the languages spoken by our participants. The coding procedure was based on Anselme (2023) and used resources from large databases with phonemic and phonetic information on the languages spoken by our participants, especially PHOIBLE (Moran , 2014) and Glottolog (Hammarström , 2020). However, language-specific sources were consulted for each language separately, including recordings such as those available in the DoReCo corpus (Seifart , 2022). All of the information on the procedure, the sources, and the individual sources consulted can be found in the OSF repository at https://osf.io/mjcnq/.

We coded rhotic variants separately for each speaker's first and second languages. There were three dimensions of coding, each binary-coded as occurring (1) or not (0). First, we coded the languages for whether they have a phonemic contrast between /r/ and /l/. Second, we coded whether each language uses an alveolar trill [r] as the main r-sound. Third, we coded whether [r] can feature as an allophone in each language.

When coding foreign languages reported by the participants, we marked a variable as present for this participant if any of the languages they spoke had the variable we were looking for. For example, if a participant reported speaking Polish and German as foreign languages, they would be marked as “1” for “[r] as the main r-sound,” as they spoke at least one language in which this was the case. The results of the coding for each language can be found in Tables I and II for the web and the field experiment, respectively.

In the web experiment, 143 participants (16%) lacked an r/l contrast in their first language, with only one of the participants also not using the r/l contrast in any second language. A total of 372 participants (41.2%) spoke a first language that uses the alveolar trill [r] as the primary r-sound; 531 participants (41.2%) spoke a first language for which the trill was not the primary variant. A total of 295 participants (32.7%) spoke at least one second language that uses the alveolar trill [r] as the primary r-sound, as opposed to 486 participants (53.8%) with a second language or languages in which the alveolar trill was the primary variant [122 participants (13.5%) did not speak any second language].

For the field experiment, only eight participants (6% of the sample) spoke a first language that lacks an r/l contrast; all of these eight participants also knew a foreign language that distinguishes phonemically between /r/ and /l/, which suggests that the entire sample knew at least one language that feature an r/l contrast. A total of 32 participants (25.2%) spoke at least one language natively in which the alveolar trill [r] was the primary r-sound, as opposed to 95 participants (74.8%) who spoke native languages where this was not the case. A total of 46 participants (36.2%) reported speaking at least one second language in which the alveolar trill [r] was the primary r-sound, 28 participants (22.0%) spoke a second language without an alveolar trill as the primary variant, and 53 participants (41.7%) reported speaking no second language.

As can be seen across both Tables I and II, participants from almost all languages, except for Palikúr, spoke at least one language in which trilled [r] could feature as an allophone. Here, our definition of allophones is intentionally broad, encompassing any variant of a phoneme that may appear in specific contexts or as free variation. This includes cases where [r] might be considered a non-standard or a less common variant. For example, while the r-sound of standard German is not an alveolar trill, it does feature in certain dialects and is traditionally also associated with singing and theatre performances (Siebs, 1930). Similarly, although [r] is not the main allophone in French, it is retained by some speakers and in certain regional varieties. Likewise, while Japanese is not typically known for having trilled [r] as a primary or standard allophone, this sound can occur in certain forms of speech, such as “gangster speech” (Sreetharan, 2004, p. 97). Also, in American English, which does not have trilled [r] as part of its standard phonemic inventory, one can find instances thereof in comedic displays or advertisements, where it is used for expressive purposes (Winter , 2022, p. 5). To establish that trilled [r] can occur as an allophone, we collected data systematically through published literature, and, where necessary, through online sources or direct recordings, without any a priori assumptions about what to expect from each language. When a language is coded as “[r] as allophone” but not “[r] as the main sound,” this implies that the trilled [r] is less frequent in those languages, as it appears in specific contexts rather than being a primary feature of the language's phonological system. However, its presence as an allophone indicates that it is still embedded within the language's phonology, albeit in a more limited and context-dependent manner.

Throughout all analyses, we use r (R Core Team, 2019) together with the tidyverse package (Wickham , 2019) for data processing and visualization. All statistical models are a version of multilevel Bayesian logistic regression implemented in brms (Bürkner, 2017). In both the web experiment and the field experiment, each participant contributed two data points. We collapsed both data points into a single data point per participant, a variable we call “match,” and the main dependent variable of our logistic regression models. For this variable, we only coded cases as match (1) when they were complete matches: i.e., a participant matched the jagged line to [r] and they matched the flat line to [l]. Complete mismatches (matching [l] to the jagged line and [r] to the flat line) as well as partial matches (e.g., matching [r] to both the jagged line and the flat line) were both coded as mismatch (0) (cf. Ćwiek , 2022). If we assume that both responses are independent, chance for the match variable would be at 25%. However, it is likely that the second response is influenced by the first one, in which case chance would exceed 25% and would be 50% if the second response was entirely locked to the first. Especially because for the field experiment, both sound files were presented first, we took a conservative approach by assuming complete dependence between the responses and chose 50% as our chance level baseline to measure matching performance.

The first model we report is a logistic regression model that includes two fixed effects: an intercept, and a fixed effect for “experiment,” which is a treatment-coded indicator variable representing the difference between the web experiment (0 = reference level) and the field experiment (1). This model includes random intercepts for language, family, and AUTOTYP area, defined by Nichols (2013). The AUTOTYP areas are geographic regions grouping languages based on shared linguistic features and historical interactions, rather than genetic relationships. We then assess the impact of the language-level predictors in line with our rhotic coding, as described in Sec. II D, with one model testing the fixed effects “has trilled [r] in L1” and “has trilled [r] in L2,” and another model testing the fixed effects “has r/l contrast in L1” and “has r/l contrast in L2.” These variables were treatment-coded, with not having [r] or not having an r/l contrast as the reference level (= 0). We fitted separate models for these two types of predictors because data for the r/l contrast variable were heavily unbalanced, with very few languages not making this contrast. We did not fit a model for the “[r] as allophone” variable because, as Tables I and II show, there is not enough variation between languages to test the impact of this factor.

All models included the same random intercepts as described above. Random slopes for the rhotic predictors were impossible to implement as there was generally no variation for these predictors within language family or AUTOTYP area (cf. Tables I and II). The only random slope that was possible to implement due to having enough variation within random effects levels was “has trilled [r] in L2” for language family, which we added to the model testing for these fixed effects. As “presentation order” was only controlled for in the web experiment, we tested this variable in a separate model fitted to data from the web experiment only (with by-language, by-family, and by-AUTOTYP area random slopes for order). As this predictor was roughly balanced (467 participants in the web experiment heard [r] first, 436 heard [l] first; 51.7% versus 48.2%), we sum-coded this predictor ( –1 = [r] first, +1 = [l] first) to aid the interpretation of the intercept, which then represents the grand average matching probability.

We used Student t distributed priors for the intercept (degrees of freedom = 3, scale = 2.5) and random effect standard deviations. We also used Student t distributed priors (degrees of freedom = 5, scale = 2.5) for all fixed effects slopes. We used LKJ(2) priors for all random effect correlation terms. Prior predictive simulations showed that these priors accommodate our data well. We additionally verified that our fitted models adequately captured plausible data-generating processes via posterior predictive simulations. All models were estimated using Markov chain Monte Carlo simulation with four chains at 10 000 iterations (4000 warm-up samples excluded, thin = 2 to reduce disk space for fitted models), which resulted in 12 000 posterior samples used for inference.

We list descriptive percentages for “match” in Tables I and II. The estimates of individual languages stemming from the statistical model seen in Fig. 2 differ from the descriptive values due to shrinkage: In multilevel models, information from the group level results is used to inform individual random effects estimates, which are drawn towards the mean.

On average, matching probability was very high, with the descriptive mean lying at 88.5% across both the web and the field experiments. Average matching was high for speakers from all languages in the sample, with the highest being 100% for Estonian and Finnish speakers and the lowest being 70% for Albanian and Mandarin Chinese speakers.

The multilevel logistic regression coefficient estimates the average matching for web experiment as 88.2%, with a 95% credible interval (CrI) of [81.7%, 92.9%]. For the field experiment, the posterior mean is 97.5%, 95% CrI [92.9%, 99.2%]. The credible intervals for both experiments are far above the chance threshold, with the posterior probability of exceeding chance being p(>50%) = 1.0 for both samples. This indicates that given this data, model, and priors, we can be very certain that the cross-linguistic average in both samples exceeds chance. The slope of the fixed effect of “experiment” was positive, indicating higher average matching for the field experiment than the web experiment (logit estimate = +1.64, standard error [SE] = 0.68, 95% CrI [0.57, 2.79]). The posterior probability of this coefficient having the same sign was p(β > 0) = 0.99, indicating high certainty that matching was higher in the field experiment than the web experiment. Figure 2 shows the posterior estimates and 95% CrIs for all languages sorted by average matching, with the box highlighting that results from the field languages all have the highest averages.

With respect to the rhotic predictors, descriptive statistics indicate that the participants whose native language has the alveolar trill [r] as the primary rhotic variant have a slightly lower proportion of matches (86.6%) than those without (89.8%). This difference, although small in terms of effect size, is indicated to be quite certain given this model, data, and priors: The posterior probability of this coefficient having the same sign was high: p(β < 0) = 0.99 (logit coefficient −0.93, SE = 0.4, 95% CrI [−1.6, −0.3]). This effect can be seen in Fig. 2, where languages in which the alveolar trill [r] is the main allophone (colored orange) appear relatively more towards the left of the plot, compared to the other languages (colored blue). There was no such effect for speaking a second language with an alveolar trill [r] (logit estimate = +0.5, SE = 0.93, 95% CrI [−0.8, +2.2]), with the posterior probability of being positive at p(β > 0) = 0.71, indicating that this specific result is bound up with considerable uncertainty. Similarly, results from the model including the predictors for whether r/l are phonologically distinguished in the language were inconclusive {coefficient for r/l in L1 –0.29, SE = 0.79, 95% CrI [−1.0, +1.6], p(β > 0) = 0.66; r/l in L2 –1.33, SE = 2.5, 95% CrI [−5.8, +2.0], p(β > 0) = 0.69}. This is also apparent when looking at the plot in Fig. 2, where circles indicate languages without an r/l contrast, which appear amongst those languages with the highest average matching (field experiment Palikúr), as well as amongst those languages with the lowest average matching (web experiment Mandarin Chinese) and everything in between.

As discussed in Sec. II, the effect of order ([r] played first versus [l] played first) was controlled only for the web experiment. When looking at the first trial only, the jagged line was chosen 94.2% of the time when [r] was played first, and the flat line was chosen 83.7% of the time when [l] was played first. This indicates that both [r] and [l] alone are matched correctly, but [r] more consistently so. The model with an effect of order fitted to the subset of the data only from the web experiment indicates that the difference between [r]-first and [l]-first trials is relatively certain (logit estimate for [l] first −1.0, SE = 0.6, 95% CrI [−2.1, −0.1]), with a high posterior probability of being of the same sign, p(β > 0) = 0.96.

Typological analyses of spoken vocabularies have found a statistical bias towards the occurrence of /r/ in words that refer to rough qualities of texture (Winter , 2022). The source of this bias has been hypothesized to be an iconic correspondence between roughness and the acoustic/articulatory properties of the alveolar trill [r] in particular (although see Anselme , 2022), but experimental evidence for this connection was lacking. Here, to investigate the basis for the correspondence between r-sounds and rough meanings, we conducted an experiment to test whether speakers of different languages associate an alveolar trill [r] with a jagged/rough line, and in contrast, an alveolar lateral approximant [l] with a flat/smooth line, in a task that emphasizes touch as much as possible by asking participants to imagine moving their finger across each line. In two experiments, one online and the other on-site, participants—including speakers of 28 different languages from ten language families—listened to recordings of an [r] and an [l] and were asked to match each sound to an image of either a jagged line or a straight line.

We found a strong effect overall: Participants matched [r] with the jagged line/rough surface and [l] with the smooth line/smooth surface an estimated 88% of trials for the online experiment and 98% of trials for the field experiment, well above the conservative baseline level of 50%. It is noteworthy that this matching probability is about 15% higher than what was observed for the bouba/kiki effect in a study using the same sample of speakers and a highly comparable experimental design that also involved sequential matching (Ćwiek , 2022). Moreover, in stark contrast to our current experiment, the bouba/kiki effect was found to have exceptions among language groups, with some groups not showing the effect. In the present data, all of the language groups in our sample showed the effect, i.e., the pattern is exceptionless, with each group showing a matching probability that is well above chance. These results indicate that the [r]/[l] crossmodal correspondence is extremely strong and one of the most cross-culturally robust cases of sound symbolism documented to date.

There are several other results worth highlighting. First, past research on the bouba/kiki effect has shown that tasks involving paired matching greatly amplify the effect (see discussion in Nielsen and Rendall, 2011, 2012). We believe this is the most likely explanation for why, in the present study, the field experiment showed overall higher matching than the web experiment, by about 10%. In the web experiment, each response followed each auditory stimulus (see Sec. II C). In the field experiment, people gave their two responses only after hearing both sounds, thus facilitating paired matching.

Another notable result was the order effect observed in the web experiment, such that on first trials, [r] was matched to the jagged line more consistently than [l] was matched to the flat line, by about 10%. The fact that both these percentages were well above 50% for first trials indicates that both [r] and [l] independently carried strong iconic associations with their respective lines/textures, even as the effect was somewhat stronger for [r]. This highlights the advantage of sequential matching, which allows teasing apart the relative contribution of each stimulus, in contrast to paired matching, for which it is unknown how much each stimulus contributes to the overall picture. Notably, previous studies using a sequential matching design found a similar pattern with the bouba/kiki effect, where the bouba stimulus is more consistently associated with the round shape, than kiki with the angular shape (Ćwiek , 2022; Fort , 2018; Margiotoudi , 2019; Yang , 2019).

Importantly, the effect we observed here is clearly present for speakers of languages with differing r-sounds and phoneme inventories with respect to these sounds: Matching exceeds chance regardless of whether speakers spoke a first or second language in which [r] was the primary realization or not and regardless of whether they spoke a language that phonologically distinguished between /r/ and /l/ sounds. Even though matching was high regardless of the phonological and phonetic characteristics of rhotics in speakers' first and second languages, we found a small but reliable effect where matching was reduced for speakers of languages in which trilled [r] is the primary variant. One possible explanation for this result is that when this sound is used as a contrastive phoneme within a language, and therefore regularly serves the phonemic function of distinguishing arbitrary words, its iconic associations may be reduced. This suggests that the extent to which a sound triggers iconic associations is malleable and modulated by the degree to which a sound is embedded within the phonological grammar of a language.

This result may be reflected in historical situations in which languages come to acquire an alveolar trill as part of their standard phonemic inventory through contact with other languages. For example, Campbell (2004, p. 68) discusses a scenario where speakers of two Mayan languages, Chol and Tzotzil, had no trilled [r] sound before exposure to Spanish. After the sound was introduced into both of these languages via loan words, this new foreign sound, “which apparently seemed exotic to the speakers of these Mayan languages” (p. 68), came first to be employed specifically in onomatopoeias and expressive vocabulary and only later ventured into the general lexicon, where it featured in arbitrary contrasts. This historical case suggests that for speakers of languages that do not already have an alveolar trill within their native phoneme inventory, this sound initially carries high expressive potential before it becomes more embedded within the grammar. However, more work is needed to ascertain whether it is specifically the conventional use of the alveolar trill [r] that is driving the weakening of the effect in our study. One way of testing this hypothesis more directly would be to quantify the functional load carried by /r/ sounds in different languages, e.g., in terms of how many meanings are distinguished by /r/ in the lexicon (cf. Wedel , 2013a; Wedel , 2013b). Our current results would predict that the more meanings depend on /r/, i.e., the higher its functional load, the more its expressive associations should be diminished.

Our results also speak to a long-standing debate in sound symbolism research: whether the analogies underpinning the iconicity of speech sounds are primarily rooted in acoustic or articulatory factors (e.g., Diffloth, 1994; Margiotoudi , 2019; Sapir, 1929; Sidhu and Vigliocco, 2023; Thompson and Do, 2019; Vainio and Vainio, 2021). Are iconic correspondences based on the acoustics of speech sounds, or are they based on articulatory factors, including proprioception (how it feels to articulate the sounds with the vocal tract) and vision (the visible features of the mouth and face involved in articulating the sounds)? For the bouba/kiki effect, it has been found that resemblances based on acoustic factors are sufficient to carry the effect, as it also occurs with reversed speech that cannot be articulated (Passi and Arun, 2024), as well as in stimuli that are filtered to be non-speech objects to listeners (Silva and Bellini-Leite, 2020). Moreover, the effect is not modulated by seeing videos of speakers pronouncing the nonce words bouba and kiki, and if anything, it is weakened by viewing such articulations (Sidhu and Vigliocco, 2023).

The alevolar trill is interesting from this perspective, as it is a sound that is notoriously difficult to articulate, requiring precise articulatory and aerodynamic control (Solé, 2002). To achieve the distinctive mode of trilled tongue movement, speakers must “position the tongue and apply the correct amount of pressure against the alveolar ridge” to allow pressure to “overcome occlusion while maintaining ability for the tongue to recoil” (Olsen, 2016, p. 317). Evidence from first and second language acquisition shows that alveolar trills are acquired late (Ball , 2001; Jiménez, 1987; Kehoe, 2018; Carballo and Mendoza, 2000), and indeed, some native speakers never learn to articulate the sound (Solé, 2002, p. 656)—an outcome that is common enough to receive a label in some languages, such as Italian erre moscia “weak r,” used to refer to Italian speakers, including native speakers, who cannot master trills. From this perspective, it is interesting that speakers of Palikúr, the only language in which trilled [r] never occurs as an allophone, performed matching at ceiling (100%). Together with the evidence from languages in which the trill is not regularly used, such as Mandarin Chinese and Japanese, this shows that even when speakers cannot produce the trilled [r], they still perceive the sound to be more fitting for the jagged rather than the flat line. Indeed, models of the acquisition of non-native consonants, such as the perceptual assimilation model (PAM) (Best , 2001), suggest that non-native sounds that cannot be assimilated to any existing sounds in a language may be perceived as non-speech sounds, which essentially are mere acoustic objects without learned articulatory representations. Thus, similar to what has been observed for bouba/kiki (Passi and Arun, 2024; Silva and Bellini-Leite, 2020), this suggests that the acoustics of the alveolar trill [r] alone are sufficient to carry the effect. Future work performing acoustic manipulations of [r] sounds similar to those that have been conducted for bouba/kiki (Passi and Arun, 2024; Silva and Bellini-Leite, 2020) could be used to lend further support for this interpretation of our results.

In using only one /r/ and one /l/ stimulus each, our study was not explicitly set up to experimentally manipulate acoustic factors and test what specific cues drive the matching we observed. That being said, a very likely cognitive mechanism that explains our results is the fact that independent of speech, people associate spatial frequencies crossmodally with the frequency of amplitude modulation (Guzman-Martinez , 2012; Orchard-Mills , 2013; Sherman , 2013). Our stimuli differ exactly in these two characteristics, albeit in a categorical manner: One stimulus has spatial frequency, the other one does not. One sound involves repeated closure (and hence cyclical amplitude modulation), the other one does not. Given that prior research on perception has shown a correspondence between the same visual and auditory features that are also contrastive in our study, we think that this is the most likely mechanism. Interestingly, amplitude modulation also turns out to be an important cue for the bouba/kiki effect (Fort and Schwartz, 2022). It has to be borne in mind, however, that Anselme (2022) have recoded the phonetic characteristics of r-sounds in the lexical data of Winter et al. which suggested that all r-like sounds may be associated with roughness in texture vocabularies. This suggests that there may be other aspects to the perceived roughness of r-sounds, on top of the amplitude modulation that differed saliently in the current study.

Finally, another point for future research relates to our use of visual images to represent rough and smooth textures, with instructions for participants to imagine the feeling as they move their finger along the lines. Our study aimed to shed experimental light on the source of the lexical patterns found by Winter (2022) related to words describing rough textures, and yet, our evidence is somewhat indirect, mediated through the use of visual images. In this respect, it is interesting to note the deep similarity between roughness and jaggedness, which can be seen as related multisensory properties that vary in spatial frequency. For comparison to the effect found here, the bouba/kiki effect also exhibits a strong tactile component, having also been obtained with felt rather than seen shapes (Ciaramitaro , 2021; Fryer , 2014; Graven and Desebrock, 2018; Sakamoto and Watanabe, 2018). Evidence obtained with Italian speakers shows that bouba/kiki-type words are not only matched to shapes, but also to surfaces differing in roughness (Etzi , 2016). Indeed, the crossmodal correspondence between spatial frequency and amplitude modulation also works between vibrotactile frequency and amplitude modulation (Guzman-Martinez , 2012), suggesting that the same feature—spatial frequency—matters for both modalities. Thus, even though our experimental stimuli are ambiguous with respect to vision/touch, the dimension of shape we investigate is conceptually similar to, and associated with, textural roughness. The fact that bouba/kiki effects work in both vision and touch, including with stimuli differing in roughness only, suggest that our results might also carry over to an experimental design that involved a genuine touch component, which was not possible in our web-based experiment. Future work can use textural stimuli to further hone in on the connection between speech sounds and touch alone.

To conclude, we found—in a large cross-linguistic experiment spanning a diverse sample of participants speaking 28 languages from 12 different language families, including participants from cultures with little access to technology and globalized culture—that trilled [r] was overwhelmingly associated with a jagged/rough line and, correspondingly, [l] was associated with a flat/smooth line. While the average effect was always found regardless of the phonetic and phonological characteristics of rhotics in participants' respective languages, it was somewhat weakened for speakers who use trilled [r] as the primary variant, suggesting the conventional use of this sound as a phoneme may diminish its iconic power. Nevertheless, the effect was extremely strong—even stronger than what has been observed for the widely studied bouba/kiki effect. In contrast to the bouba/kiki effect, which is not obtained for speakers from all languages (Ćwiek , 2022; Styles and Gawne, 2017), the r/l effect observed here was obtained without exception for all languages in our sample, suggesting it may be one of the most cross-linguistically robust cases of sound symbolism documented to date.

S.F. and A.Ć. were supported by the German Research Foundation Grant No. FU 791/6-1. B.W. was supported by the UKRI Future Leaders Fellowship MR/T040505/1. The authors wish to thank the co-authors of the main study: Christoph Draxler, Eva Liina Asu, Katri Hiovain, Sofia Koutalidis, Manfred Krifka, Pärtel Lippus, Gary Lupyan, Nathalie Schümchen, Ádám Szalontai, and Özlem Ünal-Logacev. We also thank Mohammad Ali Nazari, Samer Al Moubayed, Anna Ayrapetyan, Carla Bombi Ferrer, Nataliya Bryhadyr, Chiara Celata, Ioana Chitoran, Taehong Cho, Soledad Dominguez, Cornelia Ebert, Mattias Heldner, Mariam Heller, Louis Jesus, Enkeleida Kapia, Soung-U Kim, James Kirby, Jorge Lucero, Konstantina Margiotoudi, Mariam Matiashvili, Feresteh Modaressi, Scott Moisik, Oliver Niebuhr, Catherine Pelachaud, Zacharia Pourtskhvanidze, Pilar Prieto, Vikram Ramanarayanan, Oksana Rasskazova, Daniel Recasens, Amélie Rochet-Capellan, Mariam Rukhadze, Johanna Schelhaas, Vera Scholvin, Frank Seifart, Stavros Skopeteas, SOS Children's Village Armenia, Katarzyna Stoltmann, and Martti Vainio for being involved in the translation or distribution of the survey. We thank Timo Roettger for supplying the plot theme function. Finally, we thank Catherine Best for her advice on the perception of the alveolar trill. A.Ć., S.F., M.P., and B.W. conceived and designed the study. A.Ć., D.D., S.F., S.K., G.E.O., J.P., M.P., C.P., S.R., R.R., J.Z., and B.W. translated (or arranged translation of) and distributed the surveys. A.Ć. and R.A. collected the data on r-variants and performed the rhotic coding. A.Ć., D.D., and B.W. performed the statistical analyses. A.Ć. wrote the first draft. M.P. and B.W. revised the manuscript. A.Ć., D.D., S.F., S.K., M.P., and B.W. edited revisions of the manuscript.

The authors have no conflicts to disclose.

Ethics approval has been granted to the PSIMS project by the German Linguistics Society Ethics Commission under No. 2018-02-180912.

The data that support the findings of this study (all data, scripts, and models) are openly available in the OSF repository at https://osf.io/mjcnq/.