Cross-linguistic variation in the detection of conversational click sounds

1 Introduction

Click consonants occur in the phonemic inventory of only a small number of languages, with some estimates placing this number at 0.5% of all languages (Brenzinger and Shah, 2023). These complex speech sounds are characterized by:

at least three major areas of complexity that are not found in most other consonants: (1) the double place of articulation features; (2) the overlap of the two constrictions for the length of the segments; and (3) the non-pulmonic airstream mechanism. (Miller, 2011, p. 416)

The release of the constrictions mentioned in (2) by Miller above allows for the flow of an ingressive airstream making a characteristic “popping sound” (Miller, 2011, p. 416) or “suction noise” (Sands, 2020, p. 2) which constitute clicks.

While these click consonants have received sustained and systematic study for many decades (Sands, 2020), similar sounds, sometimes called “non-phonemic clicks,” “non-consonantal clicks” or “paralinguistic clicks” (Gil, 2013), have until recently not received as much attention. Gil (2013) defines the “paralinguistic” nature of such clicks as follows:

Phonetically, they [para-linguistic clicks] involve sounds lying outside ordinary phonemic inventories. Grammatically, they are not integrated into morphological and syntactic structures. And semantically, they convey a very restricted range of meanings, some of which are associated with the expression of emotions. (Gil, 2013)

Over the past few decades, however, there has been an increased amount of attention to non-phonemic click sounds, with investigations in a variety of languages, including English, Spanish, French, German, Hebrew, Russian, Swedish, and Mandarin Chinese (Benor, 2004; Wright, 2005, 2007, 2011a,b; Ward, 2006; Reber, 2012; Gold et al., 2013; Ogden, 2013, 2020; González Temer, 2014; Trouvain, 2014, 2015; Trouvain and Malisz, 2016; Pinto and Vigil, 2018, 2019, 2020; Paschen, 2019; Pillion, 2019; Li, 2020; Pratchett, 2021; Ben-Moshe and Maschler, 2024; Schulte, 2020; Zellers, 2022). The articulation of these click sounds, generally classified as ingressive stops or affricates, has been reported as alveolar, post-alveolar or dental (Ward, 2006; Wright, 2011a; Gold et al., 2013; Ogden, 2013). Although production occurs primarily with the tip of the tongue, there is also velar contact (Wright, 2005, 2011a; Ogden, 2013). While these might be considered the most prototypical non-phonemic click sounds, the literature has documented two other less common places of articulation. Bilabial clicks (Wright, 2005, 2011a,b; Pinto and Vigil, 2018), which tend to be acoustically weaker, are similar to the action one does when blowing a kiss. Lastly, some clicks have a lateral release (Wright, 2005, 2011a; González Temer, 2014; Ogden, 2013), a sound that resembles the noise produced in some cultures when training or encouraging a horse.

Gil (2013) influential definition of paralinguistic clicks emphasized their affective display (e.g. as seen in the dental click used by English speakers to express disapproval, often represented as “tsk-tsk”) and logical meanings (e.g., as seen in Mauritania Fula where a single dental click is used for “yes” and two for “no”). More recent work has placed emphasis on the pragmatic and discursive functions of clicks in the context in which they are used (Ogden, 2013, 2020; Pinto and Vigil, 2018, 2019, 2020; Wright, 2005, 2007, 2011a,b; Ward, 2006). The different click functions that have been reported include sequence management, search, stance, and even a marker of authority and expertise, as observed by Schulte (2020).

For the most part, the literature on phonemic clicks and that on paralinguistic clicks have not come together (Brenzinger and Shah, 2023). A notable exception is a growing body of work proposing that paralinguistic clicks may have been introduced in communities speaking languages without phonemic clicks via contact with speakers of languages that do have clicks in their phonemic inventories (Grenoble et al., 2015). In the current study, we ask whether the perception of paralinguistic clicks is influenced by the languages we speak, offering what we believe is the first comparison of monolingual speakers of English and speakers of languages that include phonemic clicks. Specifically, we focus on the question of whether listeners can detect paralinguistic clicks. Little is known about the influence of native language on listeners' capacity to detect speech sounds more generally. In fact, to the best of our knowledge there is only some evidence for a native-language advantage in the detection of speech sounds in noise (Liu et al., 2015; Huo et al., 2016).

In a previous study on paralinguistic click detection by Vigil and Pinto (2020), native English speakers were presented with pairs of short video clips, where one clip included a click and the other had been edited to remove the click. The clicks were all so-called search clicks, with the speaker searching for how to start or continue her sentence. Participants were asked to indicate whether they noticed any difference in the speaker's oral expression between the two videos (yes/no), and if so, they had to report the correct timestamp based on the timer in the video. The participants were successful in less than 20% of the cases, which the authors argued is below chance level. This result is surprising given that paralinguistic clicks are acoustically salient: they stand out because they typically occur during pauses and co-occur with low-salient fillers such as um (Wright, 2005; Trouvain and Malisz, 2016; Moreno, 2019). In addition, detecting phonemic clicks is very easy, as anyone who has ever listened to a language with such clicks can testify.0F¹

The aim of the present study is to investigate whether speakers of click-consonant using languages (henceforth, CU languages, as in Brenzinger and Shah, 2023) show more sensitivity to the presence of paralinguistic clicks. Unfortunately, not much is known about the presence of specific paralinguistic clicks in CU languages. That is, while these languages use clicks not only as consonants, but also as discourse markers and in interjections (Brenzinger and Shah, 2023), we are only aware of one reported case, in siSwati, of a paralinguistic click that is different from the language's phonemic clicks (Rycroft, 1982). However, according to Brenzinger and Shah (2023) it is highly unlikely that this is an isolated case. It is also unknown whether the loudness of paralinguistic clicks in CU languages is comparable to that of phonemic clicks or to that of paralinguistic clicks in English and other non-CU languages.

In the present study, we administered the click detection task from Vigil and Pinto (2020) using the same stimuli and procedure. One modification is that we tested participants online, allowing us to contrast monolingual English speakers with speakers of languages containing phonemic clicks.

2 Experiment 1: monolingual English speakers

All of the experimental design and analyses were pre-registered online at https://osf.io/uyv4d.

2.1 Methods and materials

2.1.1 Participants

A total of 118 monolingual English speakers from the United States were recruited on Prolific.

2.1.2 Stimuli

We used the same stimuli as Vigil and Pinto (2020) — eight video clips of 9-10 second duration, taken from a scientific presentation given by a native speaker of American English. Each video in the test condition contained one paralinguistic search click. In Vigil and Pinto (2020), a second version of each video was created by silencing the click, creating four pairs of target stimuli, one with a click and one without. In addition to the target stimuli, there were four pairs of distractor videos that did not include any clicks. The target pairs were divided into two test conditions, “opening,” when the click occurred at the beginning of a sentence, and “mid,” when the click appeared in the middle of a sentence. Within each pair, the two video clips were played with an inter-stimulus interval (ISI) of 5 s.

2.1.3 Procedure

The procedure was identical to that of Vigil and Pinto (2020), except for the fact that participants were tested individually and online rather than in groups in a classroom.

Participants were randomly assigned to either the opening or the mid condition and to one of three pseudo-randomized orders of presentation of the eight pairs, such that not more than two target or distractor pairs occurred in a row. For half of the target pairs, the click appeared in the first video, and for the other half, it appeared in the second video.

After giving consent, participants were shown one pair of practice videos to familiarize themselves with the stimuli to be used; these videos (from the same speaker) were identical and contained no click. Participants at this stage were also prompted to adjust the volume of their device. They then moved on to the test phase, during which the eight pairs of videos were presented. The time elapsed was shown prominently on the bottom of each video while it played (see Figure 1).

For each pair, participants were asked whether they perceived a difference between the two videos. It should be noted, however, that the removal of the click only occurred in the audio, though participants were purposely not informed about this. If informants reported having perceived a difference, they were asked to report at what timestamp the difference occurred, and to qualitatively describe the difference. As Vigil and Pinto (2020) pointed out, these written responses can be difficult to interpret, and as such, they add a somewhat subjective element into the data analysis. Following these authors, we therefore did not incorporate these last qualitative responses into the main, pre-registered analyses. We nonetheless kept them for comparability of the design, and we propose a post-hoc analysis and an extensive presentation of the responses in the Supplementary Material. At the end of the experiment, participants filled in a simple demographic questionnaire.

2.2 Results

Following Vigil and Pinto (2020), we coded participant responses in target conditions as timely detections of a difference if participants (i) correctly identified that there was a difference between the videos and (ii) the reported timing of the difference was within one second before or after the occurrence of the click. These results are reported in Figure 2, and they are the core of our pre-registered analyses. Qualitatively similar results were obtained for other metrics. First, we counted perceived differences, which include any report of a difference between two videos, independently of the reported time of the difference. This measure also applies to the control conditions in which there was no difference. Second, we counted click mentions: here we include any report of a difference between two videos that additionally mentions a click (after manual inspection, we included reports that contained a word starting with “click”, “lip”, “mouth”, “smack” or “tongue”). See the Supplementary Material for details.

In their analysis, Vigil and Pinto (2020) performed separate one-sample proportion tests for correct detection frequency counts of Opening clicks and Mid clicks. Chance was estimated as a rate of correct responses of 1/2. The present analysis differed in two ways. First, we used a more stringent chance value of 1/6. This is considering that there is a 1 in 2 chance of participants saying there is a difference between the video pair, and also that for 9-10 second videos there is approximately 1 in 3 chances of entering the correct timestamp within +/- 1 second. Second, we used bootstrapping to estimate a confidence interval for the accuracy scores (including for those of Vigil and Pinto, 2020), which provides a principled way to evaluate the plausibility of the null hypothesis, and also allows for an easy reinterpretation of the results if a different chance level were to be chosen. All bootstrapping simulations were based on 1,000 resamples with replacement, of which we report confidence intervals at 95%.

Figure 2 shows the full results of this experiment, as well as those of Vigil and Pinto (2020) for reference. For the current experiment, the closest replication of Vigil and Pinto's experiment, one-sample tests of proportions on timely detections reveal that accuracy on target pairs was not better than the set chance level of 17% in either the Opening or the Mid condition (both z < −2.7, p >0.9 in both cases). A two-sample proportion test showed no significant difference between the two conditions either (z = −0.07, p >0.9). We note that Vigil and Pinto (2020) found a “trend” here, with a p-value of.068. Finally, comparing the present results to those of Vigil and Pinto (2020) in two-sample proportion tests, we observed lower accuracy in the Opening condition (z = 2.9, p = 0.004) but not in the Mid condition (z = 0.99, p = 0.32).

For completeness, Figure 3 reports d′ values for perceived differences, with the opening and mid conditions shown both separately and pooled. The d′ values show the same qualitative pattern as the raw response proportions. We computed d′ at the aggregate level rather than per participant, because each participant contributed only four trials per condition, making per-participant d′ estimates unstable. The inferential tests reported above are therefore the primary, and almost identical, analyses.

2.3 Discussion

These findings largely replicate the results of Vigil and Pinto (2020), again showing that the ability of monolingual English speakers to detect clicks is low. The accuracy of our participants is even lower than what was reported in the original study. This suggests that our online testing method may negatively affect detection in comparison to the previous in-classroom version. We also note that the false detection rate in the original study was very high, with more reported differences in distractors than in target pairs, which was not the case here.

Our results confirm that the chance levels previously used may have been overestimated. Vigil and Pinto (2020) already reported that participants were reliably below the then chosen chance level of 50% (mean timely detections on target items: 17.4%), and our participants are even lower (mean timely detections on target items: 11.0%), suggesting that the true chance level is probably even further down than the 17% we proposed above.

Taking another perspective, it is tempting to interpret these results on the contrary as suggestive evidence of an–albeit weak– capacity of our participants to detect clicks. First, the confidence interval of the timely detection does not intercept with zero, which it should if participants were not detecting any clicks (and reporting this faithfully rather than responding randomly). Furthermore, participants perceived more differences in target than in distractor pairs. Finally, they correctly described having perceived the absence vs. presence of a click or click-like sound only in response to target pairs (see Supplementary Material for details about the qualitative reports).

Overall, we conjecture that the capacity of English speakers to detect paralinguistic clicks appears to be weak, but not null.

3 Experiment 2: CU language speakers

Both the experimental design and analyses were pre-registered online at https://osf.io/9zp2b.

In this follow-up experiment, we use the same stimuli and procedure to examine whether speakers of click-consonant using languages (CU languages) have a higher capacity at detecting paralinguistic clicks than native English speakers.

3.1 Methods and materials

3.1.1 Participants

As Prolific does not include any CU language in its list for participant filtering, participants were initially selected as being born and residing in South Africa, and as English-speaking multilinguals. Given the languages spoken by the natives of the country, this provided a high chance of encountering speakers of CU languages. By asking in a posttest questionnaire what language(s) participants spoke well, we were able to only include in our analysis participants who listed at least one CU language. Given the sociolinguistic situation in South Africa,1F² it is likely that all of them were in fact native speakers of at least one CU language and few of them native speakers of English.

Overall, 124 people participated, 54 of whom spoke at least one CU language and were hence included in our analysis. Of these, 46 reported Zulu as one of the languages they spoke, 16 Xhosa, 14 Sotho, 9 Swati and 3 Ndebele, with 27 speaking two or more of these languages.2F³

3.1.2 Stimuli and procedure

We used the exact same stimuli and procedure as those in Experiment 1.

3.2 Results

We carried out the same analyses as for Experiment 1. The results are shown in Figure 2.

One-sample tests of proportions showed that accuracy on the target pairs was not better than the set chance level of 17% in the Mid condition (z = −0.89, p = 0.81), and only marginally so in the Opening condition (z = 1.33, p =0.092). A two-sample proportion test showed no significant difference between the two conditions either (z = −1.54, p = 0.12). Finally, we compared performance in this experiment to that in Experiment 1. A two sample test of proportions revealed higher accuracy in the present experiment in the Opening condition (z = −2.7, p =0.006), but not in the Mid condition (z = −0.67, p =0.50).

Next, we carried out non pre-registered explanatory analyses. First, perceived differences (second panel in Figure 2) are more numerous in target than in distractor pairs, with the confidence interval of the Opening condition not including these controls. Second, click mentions are absent from the distractor pairs altogether, but to a certain extent present for the target pairs. The same results in the form of d' from detection theory can be found in Figure 3, they yield the same qualitative conclusions.

3.3 Discussion

Similar to the native English speakers in Experiment 1, speakers of CU languages showed an overall low ability to detect paralinguistic clicks. However, performance on sentence-initial clicks (Opening condition) was numerically better than that on sentence-medial ones (Mid condition), and also significantly better than that of native English speakers in the same condition. For sentence-medial clicks, performance was not different from that of our native English participants. Overall, these results suggest that speakers of a CU language as well as English have some ability to detect clicks, and that their ability is greater than that of monolingual English speakers.

4 Conclusion

In two online experiments, we examined listeners' awareness of paralinguistic clicks. In Experiment 1, we replicated the previous finding of Vigil and Pinto (2020) that native English speakers have a low ability to detect the presence of such clicks in two otherwise identical video clips. In Experiment 2, we used the same stimuli to test fluent speakers of a click-consonant using language who are proficient in English. These participants did show better performance than the English speakers from Experiment 1, but only on sentence-initial clicks. Even in this condition, they did not perform better than our preset chance level of 17%.

There may be several reasons as to why the overall performance was low. First, we used audiovisual stimuli, since in audio-only stimuli clicks can be confused with other sounds such as mouse clicks (Trouvain and Malisz, 2016). Yet, perhaps click detection is easier in the absence of video. In this light, it is interesting to mention a study with audio-only stimuli that focused on the filler um (Christenfeld, 1995). In this study, um was quite well detected by native listeners when they were asked to attend to speech style rather than to its content. Second, we used only search clicks, presented in emotively-neutral speech. This raises the question of whether other types of paralinguistic clicks may be easier to detect. For instance, in emotive-laden speech, stance clicks can be expected to occur; their detection may therefore be easier. Finally, we only tested the detection of clicks in a language that is spoken by all participants, either natively or as an L2. We cannot exclude that performance would be better when presented with an unknown language. Indeed, clicks are acoustically salient, and-as mentioned in the introduction - phonemic clicks are very easy to detect by listeners of non-CU languages. More generally, listeners show awareness of salient sounds in languages they don't speak (Lev-Ari and Peperkamp, 2016), which should facilitate their detection in fluent speech.

Our finding that speakers of CU languages showed a slightly better ability than monolingual English speakers at detecting non-phonemic clicks reveals that the presence of phonemic clicks in people's native language(s) yields an advantage in the detection of non-linguistic clicks, at least when listening to a non-CU language that they speak too. The effect we find is small, but we note that it was detected despite an adversarial online test setting, yielding less sensitivity compared to Vigil and Pinto (2020) study in a classroom setting, and with 9-10 s long stimuli. As to the restriction of the effect to opening positions, it is worth noting that click phonemes tend to appear at the beginning of words in CU languages (Best, 2020). Perhaps this explains why CU speakers' detection of paralinguistic clicks is enhanced in opening positions.

More generally, one may ask what aspects of CU languages influence the perception of paralinguistic clicks, and further research is needed in this area. Are paralinguistic clicks easier to detect depending on their acoustic similarities to phonemic clicks that hearers have been exposed to? Does the position of consonant clicks influence where paralinguistic clicks will be best detected (see above)? Does the function of a paralinguistic click matter? CU languages are thought to represent a small proportion of the languages of the world (0.5% according to Brenzinger and Shah, 2023), yet they offer a unique opportunity for scientific inquiries. Clicks are a rare case of an alternation between meaningful units for paralinguistic or linguistic purposes across languages. We propose a lightweight method to engage native speakers of CU languages and answer these questions. While it may not be as scalable as other methods, we hope that our work can help evaluate the feasibility of addressing click-related questions with a diverse audience, and that such efforts can yield insights about the intricacies and interactions of human language and human perception that can hardly be reached otherwise.

Correction note

This article has been corrected with minor changes. These changes do not impact the scientific content of the article.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/flang.2026.1932401/full#supplementary-material

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Footnotes

Edited by: Juhani Järvikivi, University of Alberta, Canada

Reviewed by: Xiaoting Li, University of Alberta, Canada

Filip Nenadić, University of Potsdam, Germany

Footnotes

YouTube contains many videos with speech in a language with phonemic clicks; see e.g. https://www.youtube.com/shorts/qbEmB8jFPo0.

Most people in South Africa are multilingual. An excellent overview of the country's languages and the people that speak them, based on census data from 2022, can be found here: https://southafrica-info.com/arts-culture/the-languages-of-south-africa/.

In full detail, the included participants reported among the languages they spoke the following sets of CU languages: [Zulu] (20), [Xhosa, Zulu] (7), [Sotho, Zulu] (6), [Swati, Zulu] (4), [Sotho, Xhosa, Zulu] (4), [Sotho] (3), [Swati] (3), [Swati, Xhosa, Zulu] (2), [Ndebele, Zulu] (2), [Ndebele, Xhosa, Zulu] (1), [Sotho, Xhosa] (1), [Xhosa] (1).