Emotions mediate nonlinear phenomena production in the vocalizations of two ape species

Abstract

Nonlinear phenomena (NLP) are widely observed in mammal vocalizations. One prominent, albeit rarely empirically tested, theory suggests that NLP serve to communicate individual emotional states. Here, we test this ‘emotional hypothesis’ by assessing NLP production in the vocalizations of chimpanzees and bonobos across various social contexts. These two species are relevant to test this hypothesis since bonobos are more socially opportunistic than chimpanzees. We found that both species produced, albeit at different frequencies, the same five distinct NLP types. Contextual valence influenced NLP production in both species with negative valence being associated with more frequent NLP production than positive and neutral valence. In contrast, using aggression severity and caller role as proxies for arousal, we found that in bonobos, but not in chimpanzees, vocalizations uttered during contact aggression or from victims and females contained more NLP. In contrast, the type of NLP produced was neither influenced by valence nor arousal in either species. Our study supports the emotional hypothesis regarding the occurrence of NLP production in mammals, particularly in opportunistics such as bonobos. This reinforces the hypothesis of an adaptative role of NLP in animal communication and prompts further investigations into their communicative functions.

This article is part of the theme issue ‘Nonlinear phenomena in vertebrate vocalizations: mechanisms and communicative functions’.

1. Introduction

Animals express their emotions through a variety of communicative channels [1–7]. Acoustic communication is one of the most widely used communicative channels in the animal kingdom, as it enables efficient, instantaneous and long-distance communication. The link between an individual’s emotional state and the acoustic parameters of the vocalizations produced has been extensively studied in mammals because vocalizations are thought to be reliable indicators of their internal status [5,6,8]. Beyond the information conveyed by the call type itself (e.g. [9–11]) or, in some species, by call sequences [12–15], additional subtle acoustic features can also encode the emotional state of the caller. Receivers can potentially rely on these features to assess the emotional state of the caller, which may affect the meaning of the message [16–18]. Complex neural pathways have been demonstrated whose activation induces fine modifications of the vocal tract in response to the emotions experienced [6,19]. For example, in response to an emotional state, certain muscles contract while others are inhibited, resulting in changes in the acoustic properties of the caller’s vocalization [20,21]. Emotional state is commonly characterized by its arousal (from low to high) and valence (positive, neutral or negative) [22]. Arousal refers to the intensity of the emotion, ranging from low levels, such as calmness or boredom, to high levels, like excitement or anxiety. It reflects the physiological and psychological activation that accompanies an emotion. On the other hand, valence describes the emotional value or quality, which can be positive (e.g. happiness or joy), neutral (e.g. indifference) or negative (e.g. sadness or anger). While arousal captures how energized or subdued an emotion feels, valence indicates whether the emotional experience is pleasant or unpleasant. Together, these two dimensions form a comprehensive framework for analysing emotions, allowing us to categorize feelings on a broad spectrum [22].

Pitch is an excellent marker of the arousal of an emotion. The stronger the arousal, the higher the pitch produced [23–25]. The vocal expression of emotional valence has been less intensively studied than acoustic markers of arousal but the duration of the vocalization was shown to be one of the most common markers of emotional valence. In particular, longer vocalizations reflect negative emotional valence [6,26–28]. Another emerging and promising marker of emotion is the presence of nonlinear phenomena (NLP) [29–33]. NLP may carry information about emotional arousal and valence. NLP result from perturbations in the typical rhythmic vibration of the vocal folds that cause deviations from regular, tonal voice production [34–37]. Different types of NLP are described: deterministic chaos, sidebands, vibrato, frequency jump and subharmonics (their definitions are given in §2). Their modes of production are still not totally understood. NLP give the voice a perceptual quality of harshness, roughness or instability [38].

For a long time, it was thought that NLP were simple by-products of vocalizations, caused for example by anatomical disorders [39]. NLP were thus thought to bring no additional layer of information to receivers and thus were mostly ignored in the analyses of vocal production. Interest in the study of NLP has been growing over the last decade and recent findings suggest that the mechanisms underlying their production are multiple and complex [35,40–42]. For example, in the non-verbal communication of humans (laughter, roar, moan, cry, etc.), nonlinearities in the vocalizations appear to reflect the emotional arousal and valence of the caller [38]. In baby cries, NLP are a reliable indicator of the level of pain an infant is experiencing [43]. The production of NLP can also be driven by some socioecological factors (body size, aggression). Indeed, the presence of NLP in vocalizations can lead to manipulation of the audience by exaggerating harshness in order to increase the perceived aversiveness of vocalizations, and exaggerate the body size [44]. Finally, NLP production is, at least partially, under the control of hormones, for example increasing in women during the menopausal period [45] or in neutered male dogs experimenting with decreasing levels of testosterone [29].

There is no longer any doubt that NLP are widespread in mammal vocalizations and carry several layers of information about the caller: identity [46,47], physical condition [48] or emotional arousal [38,49,50]. NLP can be perceived as a way to catch the attention of the audience [41,42] or to exaggerate, the size or the motivational state of the caller [35]. However, to our knowledge, NLP in non-human primate (hereafter primates) communication have been largely ignored [48,51]. In chimpanzees, NLP are routinely present in the pant hoots [51] and may signal the physical condition of the caller [48]. Yet, to our knowledge, the prevalence of NLP across the entire vocal repertoire of chimpanzees and other primate species, and how NLP production is linked to emotion, have never been investigated.

To illustrate, chimpanzees modify their screams according to their social role during agonistic interactions and can exaggerate the severity of aggression experienced as a function of the audience composition [52]. Although the prominence of NLP in chimpanzee screams was not described and included in the acoustic analyses of this study, the acoustic data provided in this study suggest the presence of NLP in their screams. This raises questions pertaining to the functional value of NLP.

Most studies examine the broad context of production of each call type (e.g. food call, rest call, alarm call, travel call) in order to derive its meaning and function, largely ignoring acoustic variation within the same call type in the same context but with different social situations (e.g. severity of aggression) or acoustic similarity across call types and contexts with similar features (e.g. highly arousing contexts) [53]. Examining how the acoustic features of a vocalization—especially the presence or absence of NLP—are influenced by the overall valence or arousal of the situation, rather than the specific context, may reveal an additional layer of information in the vocalization, beyond any direct reference to the situation itself [52].

Here, we offer the first comprehensive description of the production of NLP in the vocalizations of our two closest living relatives, the chimpanzees (Pan troglodytes) and the bonobos (Pan paniscus). In particular, we focus on distinct social contexts with diverse emotional valence and arousal to determine whether nonlinearities in their vocal production are associated with the valence and the degree of arousal of their emotional state. Chimpanzees and bonobos are two closely related species that share a similar social structure (large multi-male multi-female groups with a high degree of fission–fusion dynamics) but exhibit striking differences in their social dynamics [54,55]. The social strategies used by males and females and by aggressors and victims differ between the two species, which may affect the emotional state of individuals in each species differently. This in turn may lead to differences in the presence of NLP in their respective vocalizations. Male chimpanzees, in particular, form long-lasting social bonds with specific other individuals [56]. As in other primates, bond formation operates through grooming interactions [57] and grooming partners frequently form coalitions with each other [56,58]. Male chimpanzees engage more frequently in coalitions than females [58–60]. Furthermore, triadic knowledge of others’ bond strength is used by chimpanzees in their behavioural decisions in agonistic [61] and affiliative contexts [62]. Social interactions in chimpanzees are therefore highly audience dependent, social decisions are taken strategically and social interactions are thus relatively predictable. In contrast, in bonobos, there is no clear link between grooming interaction and coalition formation. Coalitions are often formed opportunistically [63], and unlike in chimpanzees, bonobo females are the ones forming the most coalitions both with strongly and weakly bonded partners, and most often target males [60,63]. Social decisions in bonobos might thus be more opportunistic and less predictable. Chimpanzees and bonobos also differ in terms of intersexual dominance relationships with male chimpanzees outranking all females [64] whereas female bonobos are higher ranking than most males in the community with variation in the degree of female dominance across communities [65,66]. Differences in intersexual dominance relationships co-vary with sexual dimorphism in body size and vocal tract anatomy. Whereas in chimpanzees males have a forearm length 14.5% longer than females on average, the sexual dimorphism is much lower in bonobos (0.3% difference between males and females) and some adult females are larger and heavier than some adult males [67]. Furthermore, whereas the male and female bonobos have similar maximum fundamental frequency (max F0), the max F0 of chimpanzee males is higher than that of females [68].

Given the general lack of knowledge about NLP production in the vocalizations of both species, the first aim of our study was to provide an in-depth description of the occurrence and types of NLP present in the vocalizations of bonobos and chimpanzees. Beyond this overall description, we aim to test the ‘emotional hypothesis’. This hypothesis postulates that the presence of NLP in vocalizations is a reliable indicator of the emotional valence and/or arousal of the caller.

If this emotional hypothesis is true, we expect more NLP in vocalizations in contexts with negative valence than in contexts with positive and neutral valence in both chimpanzees and bonobos (prediction P1). Moreover, in the negative valence context and especially in the agonistic context, we expect more NLP in vocalizations produced during social interactions with high arousal levels. Since agonistic interactions are in general less predictable in bonobos than in chimpanzees, which may increase the emotional arousal of the situation, we predict that the vocalizations uttered in agonistic contexts contain more NLP in bonobos than in chimpanzees (prediction P2).

Given the differences between bonobos and chimpanzees in terms of conflict predictability, coalition formation patterns and inter-sexual dominance relationships (see above), we expect that the combined effect of the sex of the caller and its role in the aggression on the production of NLP varies between the two species. More specifically we predict that

• —Between species, the role in the conflict (i.e. victim versus aggressor) will have a stronger effect on NLP production in bonobos than in chimpanzees since victims in bonobos might be more surprised given the larger level of unpredictability in their conflicts (prediction P3).
• —Within species, the role in the conflict should play a stronger role in female than in male chimpanzees and in male than in female bonobos (prediction P4). In fact, female chimpanzees are often targets of male displays or redirected aggressions that are less predictable than directed conflicts (between males). In contrast, bonobo males are most often the targets of female coalitionary aggressions.

2. Methods

(a). Ethical approval

All methods employed in this study were purely observational and non-invasive.

For bonobos, data were collected in zoological parks, all certified by the European Associate of Zoos and Aquaria (EAZA). All data collection protocols were performed in accordance with the relevant guidelines and regulations and were approved by the Institutional Animal Ethical Committee of the University of Lyon/Saint-Etienne, under authorization no. 42-218-0901--38 SV 09 (Lab ENES).

For chimpanzee, the ‘Ethikrat’ of the Max Planck Society, the Ministére de l'Enseignement supérieur et de la Recherche scientifique, the Ministére des Eaux et Forêts of Côte d'Ivoire and the Office Ivoirien des Parcs et Réserves all gave ethical approval for this study. Researchers at the Taï Chimpanzee Project follow strict hygiene rules to prevent zoonotic transmission of pathogens.

(b). Sites and subjects

S.K. recorded bonobo vocalizations in three zoos (Apenheul (Netherlands), Planckendael (Belgium) and La Vallée des Singes (France)). All study subjects live in multi-male and multi-female groups and comprise both immature and adult individuals (see electronic supplementary material, table S1 for details of each population recorded) that reproduce the group structure observed in the wild [69]. In the three zoos, bonobos had access to indoor and outdoor enclosures, and daily routines and diets were similar in each zoo. For this study, we used only vocalizations recorded from adult individuals: four at Apenheul Zoo (three females and one male), eight at Planckendael Zoo (four females and four males) and 11 at la Vallée des Singes (seven females and four males).

T.B. recorded the vocalizations of wild chimpanzees in the Taï National Park (Ivory Coast) [70] in three communities: South, East and North. In the South community, she recorded vocalizations of 19 adult chimpanzees, 14 females and five males. In the East community, she recorded vocalizations of 15 adults: eight females and seven males. In the North group, she recorded vocalizations of 10 adults: eight females and two males (see electronic supplementary material, table S2).

(c). Vocal recordings

(i). Bonobos

S.K. recorded all bonobo vocalizations over a 1 year period using ad libitum protocol [71] without focusing on specific events or subjects. In Planckendael Zoo, bonobo recordings were performed from 20 March 2013 to 10 May 2013 and from 20 February to 4 March 2014, from 10.00 to 18.00, for a total of 190 hours of recording time. In Apenheul Zoo, vocalizations were recorded daily between 9.00 and 17.00 from 14 May 2013 to 6 July 2013, and on 12 March 2014, totalling 75 hours of recording time. At La Vallée des Singes, bonobo vocalizations were recorded between 9.00 and 17.00 from 28 October 2013 to 25 November 2013, amounting to 115 hours of recording time. In the three zoos, vocalizations were recorded both indoors and outdoors following the location of the subjects [72].

In the three zoos, vocalizations were recorded using a Zoom H4 Digital Multi-track Recorder (sample rate: 44.1 kHz, 16 bits per sample). One channel was connected to a Sennheiser MKH70-1 ultra-directional microphone to record bonobo vocalizations and the other channel was connected to an AKG MPA II micro-tie audio recorder allowing the observer to describe the behaviour associated with the vocalizations in real time.

(ii). Chimpanzee

T.B. recorded vocalizations on wild chimpanzees over nine months across two field periods: 1 January to 23 May 2019 and 5 December 2019 to 19 March 2020 (see details in [73]). Each day, one community was followed and T.B. recorded all vocalizations produced by a focal subject for six consecutive hours (two subjects per day, 6.30−1230 and 12.30–18.30). In addition, during focal follow, T.B. recorded vocalizations ad libitum from other individuals present at the party when she could clearly identify the caller and the context (see below). In parallel to vocalization recordings, T.B. recorded for each vocalization the behaviour of the caller, the general activity and other social interactions in the group occurring around the vocalization produced using a smartphone and Cybertracker software. She used a Tascam DR-40X digital recorder (sample rate: 48 kHz and 24 bits per sample) to record all the chimpanzee vocalizations. This recorder was connected to a Sennheiser ME67 ultra-directional microphone.

(iii). Characteristics of the sample size

Since age can influence the production of NLP in mammals [29,47,74–78], we discarded all individuals under 10 years of age [72,73] (keeping only subadult and adult) in order to analyse the stable vocal patterns of NLP for both species and to assess the effect of the production context and caller sex. In total, vocalizations from 22 adult bonobos and 29 adult chimpanzees were analysed.

(d). Contexts of vocalizations

In this article, we defined vocalization as a single unit (or single call) that could be emitted alone or may belong to a sequence of two or more vocalizations.

The context of vocalizing was determined by the behaviour or context in which the signaller was engaged at the moment of vocal emission (see above). For both species, we retained five general common contexts (from [53] for bonobos, and from [73] for chimpanzees): agonistic, food anticipation (in bonobos: vocalizations given directly prior to or at the start of scheduled feedings; in chimpanzees: vocalizations given when they arrive at food source), foraging, affiliation (in both species: give or receive affiliative gestures including hug, arm reach, kiss, touch but not groom) and grooming. We then classified these different contexts into three categories of valence: positive valence (food anticipation), neutral valence (foraging and feeding peacefully (with no agonistic behaviours)) and negative valence (agonistic interaction in which the caller was directly involved). In positive valence, we did not take groom and affiliation as chimpanzees did not produce vocalizations during these interactions in our recordings. These broader categories make the data comparable between the two species, despite the fact that we recorded bonobos in captivity and chimpanzees in the wild.

The results from our first analysis (see §3) showed that the largest production of NLP occurred in agonistic contexts. In order to address the ‘arousal’ hypothesis, we focused on the agonistic contexts to examine the link between the NLP production and the severity of the interaction (with or without contact (pushing, slapping, kicking, grabbing or biting) and/or the role of the caller (victim or aggressor)). The aggression severity and the role played during an agonistic interaction can reflect the level of arousal of the caller [79].

(e). Nonlinear phenomena analysis

(i). Types of nonlinear phenomena

We considered five types of nonlinear phenomena [34,35,39]:

1. Deterministic chaos: asynchronous non-periodic regime vibrations of the vocal folds, resulting in a large frequency bandwidth and a harsh voice.

2. Sidebands: variation of the fundamental frequency at a fast modulation rate (between 30 and 100 Hz), resulting in horizontal lines parallel to and equally spaced above and below the fundamental frequency on a spectrogram. Sidebands are perceived as making the voice sound harsher.

3. Vibrato: variation of the fundamental frequency at a slower modulation rate than the sidebands (between 5 and 15 Hz), resulting in a sinusoidal shape of the fundamental frequency on the spectrogram. Vibrato is perceived as a trembling of the voice.

4. Subharmonics: vibration of one vocal fold at a periodic fractional value of the second vocal fold.

5. Frequency jump: abrupt change of the vibration of the vocal folds leading to a sudden jump of frequency.

The mechanisms of production and the communicative functions of NLP have attracted increasing interest in recent years, and their definitions are still under discussion (this Special issue). Deterministic chaos, sidebands, subharmonic and frequency jump are classically classified as NLP, but the case of vibrato is debated. Although less common, some studies in humans (singers) consider vibrato alone to be an NLP type [80]. In our study, we made the reflective choice to consider vibrato as an NLP type in order to provide a comprehensive analysis of vocal production in chimpanzees and bonobos. Indeed, a pilot study of their spectrograms indicated that these two species differ in the occurrence of slow frequency modulation (vibrato) production. Here, we provide a clear and non-overlapping range of frequency modulation rates to discriminate vibrato from sidebands (see above).

However, in order to provide a comparable approach to the current majority of studies in mammals, and to allow possible future straightforward comparisons, we reran all our analyses removing vibrato as an NLP type. The results of this additional analysis are provided in electronic supplementary material, section S6.

(ii). Annotation of nonlinear phenomena

We performed a manual acoustic quantitative description of nonlinear phenomena on all recorded vocalizations for which the identity of the caller and the associated production context were known. For each vocalization, F.F. noted the presence and the type of NLP present using Praat software [81]. For each NLP type, F.F. noted its presence or absence (0 for the absence and 1 for the presence), its duration for sidebands, deterministic chaos, vibrato and subharmonic (excluding frequency jump) and the time between the start of the call and the start of the NLP. When one vocalization contained more than one NLP type, we labelled it as ‘multiple NLP’, and we removed these from our analysis (n = 22 vocalizations). Interobserver reliability was performed by a second person (A.G.) on 198 vocalizations from both species, with 97.4% agreement in NLP coding between F.F. and A.G.

Thus, we quantified the proportion of vocalizations with NLP, with which type of NLP, and the proportion of each type of NLP in a given vocalization (dividing the duration of the NLP by the total duration of the vocalization).

(f). Statistical analysis

Within each species, the intergroup variation in the NLP production was tested. To do this, we compared the proportion of NLP production in the three zoos for the bonobos together and the three communities for the chimpanzees together. We did not find any differences when we took community/zoo identity into account (electronic supplementary material, table S3).

To test our overall hypothesis that emotional arousal influences NLP production and the type of NLP produced we ran four generalized linear mixed models (GLMMs).

In the first model (model 1), we assessed the effect of the valence on the presence of NLP in the vocalizations using a GLMM with a Bernoulli error distribution. Each vocalization constituted a data point and we modelled whether this vocalization had NLP (1) or not (0) as the response variable. In model 1, we used the two-way interaction between the level of valence (positive, neutral and negative) and the species as predictors. This interaction was included to test the hypothesis that changes in valence have a stronger effect in bonobos than in chimpanzees because of the larger unpredictability of social situations and in particular of aggressions in bonobos as compared with chimpanzees. In model 1, we also controlled for the duration of the call to account for the possibility that NLP are more likely to occur in longer vocalizations, and because the duration of vocalizations appears to be linked with the valence. In order to account for repeated vocal sampling of the same individual and the presence of different vocalizations within the same vocal sequence we also added the identity of the caller and the identity of the sequence as random factors.

A second model (model 2) was run to test the hypothesis that, when NLP are being produced, some NLP type are more likely to occur than others in some valence level. In model 2, we assessed the effect of contextual valence on the type of NLP produced using a GLMM with a categorical family. This model was run on a subset of data including only vocalizations that had NLP. Here we focused only on vocalizations where NLP was present, as we did not want to compare each type of NLP with the absence of NLP, but each type of NLP with the other types of NLP. Each vocalization with NLP constituted a data point and we modelled the type of NLP produced (chaos, sidebands, vibrato, subharmonics, frequency jump) as the response variable. We used the two-way interaction between the level of valence (positive, neutral and negative) and the species as predictors. We also controlled for the duration of the call. We added the identity of the caller and the identity of the sequence as random factors. All of this is for the same reason mentioned above.

Because model 1 revealed that NLP were mostly produced in agonistic contexts (i.e. negative valence contexts, see §3), we then focused on these contexts to examine whether the arousal level influenced the production of NLP within consistent valence.

In the third model (model 3), we tested the hypothesis that the conflict severity and the role played by individuals in conflict should impact the arousal of the caller and therefore the likelihood of producing NLP using a GLMM with a Bernoulli error distribution. We ran this model on a subset of data comprising only vocalization uttered during agonistic interaction involving the caller. Each vocalization constituted a data point and we modelled whether this vocalization had NLP (1 or 0) as response variable. In this model, we used first a two-way interaction between the role in the aggression (victim and aggressor) and the species and a three-way interaction between the role in the aggression, the sex of the caller and the species. In addition, we tested the two-way interaction between the severity of the aggression (i.e. with or without contact) and the species and the three-way interaction between the severity of the aggression, the sex of the caller and the species. This interaction was included to test the hypothesis that changes in arousal have a stronger effect in bonobos than in chimpanzees because of the greater unpredictability of their aggression. Furthermore, since the role of each sex differs in each species, emotional arousal is expected to covary with sex and species. As for models 1 and 2, we controlled for the duration of the vocalization. As previously, we also added the identity of the caller and the identity of the vocal sequence as random factors.

In the last model (model 4), we assessed the effect of contextual arousal on the type of NLP produced using a GLMM with a categorical family. This model was run on a subset of data including only vocalizations that had NLP. Each vocalization with NLP constituted a data point and we modelled the type of NLP produced (chaos, sidebands, vibrato, subharmonics, frequency jump) as the response variable. We used the three-way interaction between the role of aggression, the sex and the species as predictors. In addition, we added the interaction between the severity of the aggression (i.e. with or without contact), the sex of the caller and the species. We also controlled for the duration of the call. We added the identity of the caller and the identity of the sequence as random factors. This model was used to test the hypothesis that the NLP types may be related to emotional arousal.

We ran all models within a Bayesian framework and fitted the models using the function ‘brm’ from the brms package v. 2.20.4 [82] in R v. 4.3.1, using default priors for the models about the presence of NLP as the response variable and normal (0,2) for the models about the type of NLP. For the duration, we reported the estimate (denoted Est) and the 95% credible interval (denoted 95% CI) of the slope (percentage of increase in the presence of NLP per unit of time). We also reported the percentage of posterior distribution (denoted P) in support of the hypothesis that the slopes are strictly positive, which indicates the probability that the model parameters (slopes) are greater than zero. For all the other models, we reported the odds ratio (denoted OR), which represents the ratio of the odds of the event occurring between two groups, based on the posterior distribution of the model parameters and the 95% CI for this ratio. Lastly, we reported the percentage of the posterior distribution in support of the hypothesis that the odds ratio is either strictly greater than or strictly less than one (without including the null effect). In each model, we scaled the variable ‘call duration’ to a mean of 0 and a s.d. of 1.

3. Results

(a). Description of the occurrences of nonlinear phenomena in bonobos and chimpanzees

In both chimpanzees and bonobos, NLP were present in about a quarter of their vocalizations (1331 out of 5347 vocalizations from 1681 sequences, i.e. 24.9% in bonobos; and 394 out of 1604 vocalizations from 188 sequences, i.e. 24.6% in chimpanzees). Although this proportion was similar in both species, it is worth noting that this figure did not take into account the possible differences in the proportion of contexts in which the vocalizations were recorded for each species.

Interestingly, in both species, we found that the presence of NLP was linked with the duration of the vocalization, with a 76% increase in the proportion of vocalizations with NLP when the duration of the vocalization increases by one second (model 1, Est = 75.7%, 95% CI: [73.1%, 78.2%], P = 100%).

Besides, the relative mean duration of each NLP type in vocalizations (see figure 1) was correlated with the type of NLP (table 1). As a consequence, we did not study the relative duration of NLP in our analyses but rather the presence/absence of each NLP type (excluding frequency jump, which is an instantaneous event (without duration, see figure 1E,H).

Figure 1.

Spectrograms and schematic representations (drawn by hand) of the five types of NLP described in bonobos: (A) deterministic chaos, (B) sidebands, (C) vibrato, (D) subharmonics, (E) frequency jump; and in chimpanzee: (F) chaos, (G) sidebands and (H) frequency jump, subharmonics and vibrato. See definitions in §2. The recording of each vocalization presented in figure 1 can be found using the following QRcode.

Table 1.

Relative mean duration of each NLP type in vocalizations and its standard error in percentage for the bonobo and the chimpanzee. For instance, for bonobos, when there was chaos in vocalization, it represented 87.2% on average of the total duration of vocalization. When there was more than one NLP type, the relative duration was accumulated and given in the NLP category ‘multiple’. Here, 100% represents a vocalization without any tonal phase and 0% represents a vocalization without NLP.

NLP type	bonobo	chimpanzee
chaos	87.2% [85.1%; 89.4%]	92.0% [89.1%; 94.8%]
sidebands	52.8% [50.5%; 55.1%]	82.7% [76.9%; 88.6%]
vibrato	44.7% [31.8%; 57.5%]	61.8% [57.3%; 66.3%]
subharmonics	32.3% [22.2%; 42.4%]	35.3% [10.0%; 73.9%]
frequency jump	NA	NA

(b). Effect of the valence on the nonlinear phenomena production in bonobos and chimpanzees

Both chimpanzees and bonobos produced NLP, at least sometimes, in context with positive, neutral and negative valence.

However, model 1 showed that, in bonobos, NLP were 15 times more likely to be present in agonistic contexts (i.e. negative valence) than in contexts with neutral valence (model 1, agonistic versus foraging: OR = 15.4, 95% CI: [6.0, 45.2], P = 100%) and 14 times more than in contexts with positive valence (agonistic versus food anticipation: OR = 14.5, 95% CI: [6.8, 33.8], P = 100%). We could not detect a meaningful difference between positive and neutral valence (food anticipation versus foraging: OR = 0.94, 95% CI: [0.4, 2.3], P = 55.5 %; figure 2). Similar results were found in chimpanzees. NLP were more than 1000 times more often produced in agonistic contexts than in contexts with neutral valence (model 1, agonistic versus foraging: OR = 1176.1, 95% CI: [70.3, 41831.7], P = 100%) and more than 500 times more often produced in agonistic contexts than in contexts with positive valence (model 1, agonistic versus food anticipation: OR = 524.8, 95% CI: [35.8, 13935.3], P = 100%). As in bonobos, we could not detect a meaningful difference between positive and neutral valence (food anticipation versus foraging: OR = 0.4, 95% CI: [0.0, 4.2], P = 76%) (see electronic supplementary material, section S2, figure S1 and table S4 for the summary of model 1).

Figure 2.

Proportion of vocalizations with nonlinear phenomena of bonobos (in purple) and chimpanzees (in orange), in three levels of valence: positive (food anticipation context), neutral (foraging context) and negative (agonistic context).

Comparing the species, we found that chimpanzees tend to produce three times more NLP in negative contexts than bonobos (chimpanzee versus bonobo: OR = 3.3, 95% CI: [0.9, 13.0], P = 95.9%), and bonobos produced 11 times more NLP in positive contexts and 23 times more NLP in neutral contexts than chimpanzees (bonobo versus chimpanzee in positive valence: OR = 11.0, 95% CI: [0.8, 238.8], P = 96.6 %; bonobo versus chimpanzee in neutral valence: OR = 23, 95% CI: [1.4, 714.1], P = 98.7 %; figure 2). (See electronic supplementary material, section S2 for the summary of model 1.)

We reran the same analysis removing vibrato as an NLP type, which did not impact our results (see electronic supplementary material, section S6 and figure S7).

Additionally, the type of NLP produced changed according to the contextual valence (model 2). However, we only found strong evidence for an effect of valence on the type of NLP produced in bonobos (interaction valence × species). In bonobos, the production of chaos and frequency jumps was meaningfully influenced by the emotional valence. Chaos was produced six times more often in negative contexts compared with positive contexts (agonistic versus food anticipation: OR = 5.9, 95% CI: [1.229, 32.0], P = 98.5%) while there was no consistent difference between positive and neutral contexts (food anticipation versus foraging: OR = 0.2, 95% CI: [0.0, 2.0], P = 91.7%) and between negative and neutral contexts (agonistic versus foraging: OR = 1.1, 95% CI: [0.2, 9.0], P = 55.1%). Frequency jumps were produced 1.6 times less in negative contexts compared with positive contexts (agonistic versus food anticipation: OR = 0.0, 95% CI: [0, 0.5], P = 99.5%) while there was no difference between positive and neutral contexts (food anticipation versus foraging: OR = 4.3, 95% CI: [0.2, 100.6], P = 81.7%), and between negative and neutral contexts (agonistic versus foraging: OR = 0.2, 95% CI: [0.0, 3.3], P = 87.7%). Moreover, there was no difference between negative and positive contexts, between positive and neutral contexts and between negative and neutral contexts in the production of the other NLP types (all 95% CI overlapped 1 and all P were inferior to 84.5%; figure 3A; see electronic supplementary material, section S3 for summary details of model 2).

Figure 3.

Comparison of the proportion of the different types of NLP (chaos in red, frequency jumps in yellow, sidebands in green, subharmonics in light blue, vibrato in dark blue) present in vocalizations on different levels of valence (negative, neutral, positive) in bonobos (A) and chimpanzees (B).

In chimpanzees, the valence did not have a consistent effect on the type of NLP produced. There were no differences between negative and positive contexts, between positive and neutral contexts and between negative and neutral contexts (all 95% CI overlapped 1 and all P were inferior to 94.5%, figure 3B; see electronic supplementary material, section S3 for summary details of model 2).

We reran the same analysis removing vibrato as an NLP type, which did not impact our results (see electronic supplementary material, section S6 and figure S9).

(c). Effect of arousal in agonistic contexts on the nonlinear phenomena production in bonobos and chimpanzees

Focusing on agonistic contexts, we investigated whether the severity of interactions and the role played by the caller during interactions influenced the NLP production (considering the presence or absence of NLP). We used both parameters as proxies for the emotional arousal of the caller. We analysed 986 bonobo vocalizations from 187 sequences and 302 chimpanzee vocalizations from 39 sequences. We recorded 395 vocalizations during interactions with physical contact (366 in bonobos and 29 in chimpanzees) and 893 during non-contact interactions (620 in bonobos and 273 in chimpanzees). Victims produced 761 vocalizations (605 in bonobos and 156 in chimpanzees) and aggressors 527 vocalizations (381 in bonobos and 146 in chimpanzees).

The impact of a three-way interaction (role × sex × species) on NLP production was not supported by our statistical analysis (see electronic supplementary material, section S4 for more details about model 3).

The effect of the severity of the interaction on the NLP production differed between the two species. NLP proportion was six times higher in agonistic interactions with contact than without contact in bonobos (contact versus no contact: OR = 6.3, 95% CI: [2.1, 22.1], P = 99.9%) but this difference was not supported in chimpanzees (contact versus no contact: OR = 0.4, 95% CI: [0.0, 2.9], P = 80.1%). This difference between species was supported by 99.5% of the posterior distribution (figure 4A, see electronic supplementary material, section S4 for more details about model 3).

Figure 4.

Effect of contact (A,B,C), protagonist role (D,E,F) and sex (female in black and male in grey) during conflicts on the proportion of vocalizations with NLP in bonobos (in purple, A and D) and in chimpanzees (in orange, A and D). Comparison of the proportion of vocalizations containing NLP of protagonists during agonistic interactions with or without contact considering the species (A) and the sex (B for bonobos and C for chimpanzees) and depending on their role (aggressor and victim) considering the species (D) and the sex ((E) for bonobos and (F) in chimpanzees).

The effect of the severity of the agonistic interaction on the NLP proportion differed between sexes. NLP proportion was eight times higher during agonistic interactions with contact than without contact in female bonobos (contact versus no contact: OR = 8.8, 95% CI: [2.1, 40.5], P = 99.9%). No such effect was found in female chimpanzees (agonistic interaction with contact versus without contact: OR = 0.4, 95% CI: [0.0, 2.9], P = 79.7%), with a difference in the tendency between species supported by 99.5% of the posterior distribution (figure 4B,C). We found no robust evidence that the severity of interaction affected the likelihood of vocalizations to contain NLP in males of either species (in bonobo: contact versus no contact: OR = 3.0, 95% CI: [0.4, 20.6], P = 87.5%, and in chimpanzee: contact versus no contact: OR = 1.2, 95% CI: [0.0, 1005.1], P = 52.7%). There was no consistent difference between species (differences supported by only 58.9% of the posterior distribution; figure 4B for bonobos and figure 4C for chimpanzees; see electronic supplementary material, section S4 for more details about model 3).

The effect of the role of the caller on the NLP proportion differed between species. The proportion of NLP was four times higher in agonistic interactions when the caller was a victim in bonobos (victim versus aggressor: OR = 4.2, 95% CI: [1.3, 14.4], P = 99.3%) but this difference was not supported in chimpanzee (victim versus aggressor: OR = 0.6, 95% CI: [0.1, 3.2], P = 71.8%), the differences between species were supported by 95.5 % of the posterior distribution (figure 4D; see electronic supplementary material, section S4 for more details about model 3).

The effect of the role of the caller on the NLP proportion differed between sexes. The NLP proportion was four times higher when the caller was a victim in female bonobos (victim versus aggressor: OR = 4.5, 95% CI: [1.2, 18.3], P = 98.9%) but this difference was not found in female chimpanzees (victim versus aggressor: OR = 0.4, 95% CI: [0.0, 2.3], P = 84.1%), with a difference between species supported by 98.7 % of the posterior distribution (figure 4D). We found no consistent effect on the NLP proportion between victim and aggressor in males of both species (in bonobos, victim versus aggressor: OR = 5.5, 95% CI: [0.7, 50.3], P = 94.8 %; in chimpanzees, victim versus aggressor: OR = 8.7, 95% CI: [0.1, 1264.5], P = 81.2%), with a difference between species supported by 68.7 % of the posterior distribution (figure 4E for bonobos and figure 4F for chimpanzees, see electronic supplementary material, section S4 for more details about model 3).

We reran the same analysis removing vibrato as an NLP type, which did not impact our results (see electronic supplementary material, section S6 and figure S11).

Our last model, model 4, supported the hypothesis that arousal had no consistent effect on the type of NLP produced in vocalizations in both species. First, the interactions with or without contact did not change the types of NLP produced (all 95% CI overlapped 1 in both species and all P were inferior to 61.4%; figure 5A,B; see electronic supplementary material, section S5 for more details about model 4). Second, the role of the caller did not influence the type of NLP produced (all 95% CI overlapped 1 in both species and all P were inferior to 87.5%; figure 5C,D; see electronic supplementary material, section S5 for more details about model 4).

Figure 5.

Comparison of the proportion of the different types of NLP (chaos in red, frequency jump in yellow, sidebands in green, subharmonics in light blue, vibrato in dark blue) present in vocalizations, comparing with and without contact in bonobos (A), chimpanzees (B), and victims and aggressors in bonobos (C) and chimpanzees (D).

We reran the same analysis removing vibrato as an NLP type, which did not impact our results (see electronic supplementary material, section S6 and figures S13, S15).

4. Discussion

In this study, we provide one of the most extensive investigations of the factors influencing the presence of NLP in non-human primate vocalizations. Focusing on our two closest living relatives, we found that both chimpanzees and bonobos frequently produced NLP in their vocalizations (about 25% of vocalizations contained NLP). The production of NLP was much more likely in longer than in shorter vocalizations in both species. While both species produced all five types of NLP—deterministic chaos, sidebands, vibrato, subharmonics and frequency jumps—they were used in different proportions in the two species. Vibrato was the most common NLP produced by chimpanzees, while bonobos mostly used sidebands and frequency jumps.

We do not yet know whether these differences reveal functional information or are just random noise production that does not relate to predictable information transfer. Bonobos have shorter vocal folds than chimpanzees and produce vocalizations with higher fundamental frequencies [68]. How this might impact NLP production is however unclear. As far as we know, the link between anatomical structure and NLP production remains to be investigated. In addition to emotional effects, Reide et al. [39] suggest a different hypothesis promoting NLP production— that for long-distance chimpanzee vocalizations, extreme air pressure required to produce high amplitude vocalizations that travel 1 km might destabilize the vocal cords resulting in NLP. Unfortunately, the data available for this study do not allow us to know the distance to which each vocalization travels and so to measure the amplitude of vocalizations, as they were recorded under different conditions and at different distances from the observer. However, we attenuated these limitations using call duration as a proxy for amplitude and controlled for call duration in all our analyses [83].

Beyond the first general description of the NLP types in non-human primates, we found support for the emotional hypothesis of NLP production in both species, since this production was influenced both by the contextual valence and the arousal of the caller. First, the emotional valence of the context surrounding vocal production influenced the likelihood of producing NLP. More NLP were produced in contexts with negative valence than in contexts with neutral or positive valence for both species. In bonobos, the type of NLP produced varied with the contextual valence: in negative valence contexts, they produced more deterministic chaos and fewer frequency jumps than in other valences, whereas no such effect of valence was found in chimpanzees.

Our findings indicating that NLP are more prevalent in contexts with negative valence are aligned with psychoacoustic studies on human perception, which show that vocalizations with NLP are classified as aversive vocalizations regardless of the valence of context [50]. However, it can be challenging to disentangle valence and arousal in certain contexts. For instance, it is debatable whether the arousal levels in agonistic interactions (negative valence) and food anticipation (positive valence) are comparable. More vocalizations in contexts with positive valence and higher degrees of arousal, such as play, would be necessary. Lastly and interestingly, we observed that bonobos produced more NLP in neutral and positive contexts than chimpanzees. It may be the case that the emotional commitment in social interactions is greater in bonobos than in chimpanzees due to their highly social and playful nature [84,85].

Secondly, focusing on negative valence interactions, we observed that the level of arousal of the caller in bonobos (assessed by the severity of interactions) affected the production of NLP. Bonobos produced more NLP when the conflict led to physical contact than when it remained limited to threats and displays. Furthermore, when bonobos were the target of an agonistic interaction, they produced more NLP than the aggressors. Conversely, our findings indicated that neither the interaction severity nor the role played in the interaction had an impact on the production of NLP in chimpanzees. It is noteworthy that the NLP type produced did not appear to be associated with emotional arousal in either bonobos or chimpanzees.

Our findings in bonobos are in line with previous studies on mammals showing a link between the production of NLP and the emotional arousal of the caller [75,86,87]. Victims of aggression may be more emotionally aroused than aggressors because they may be surprised to be targeted, and have less control over it. Similarly, physical aggression in conflict can result in injury and thus lead to higher levels of arousal associated with the risk involved. Interestingly, however, we only found an effect of conflict severity and caller role in bonobos and not in chimpanzees. This species difference in NLP production may stem from the species difference in the nature and configuration of agonistic interactions. Chimpanzees form long-term social bonds with a selective set of individuals in their community and take into account dyadic bond strength in their decisions to engage in dyadic or polyadic aggressions [61]. In contrast, in bonobos coalition formation or dyadic conflicts are not predicted by the bonding index such as the grooming index, which renders the conflict setting more unpredictable, even more so for the victims [88,89]. Furthermore, agonistic interactions appear to be less common in bonobos than in chimpanzees [90,91]. Our observations in bonobos are along these lines, with the victims appearing more vocally stressed than the aggressors [92]. The overall idea is that emotion affects vocal production more in bonobos, which is reflected both in the role and severity of aggression on NLP production and in the type of NLP produced. Bonobo vocalizations would thus contain more emotional information than chimpanzee vocalizations.

Contrary to our expectation, we found an effect of sex on NLP production in bonobos but not in chimpanzees. Bonobo females produced more NLP than males, with a large variability, regardless of their role in the agonistic interactions and regardless of the severity of the aggression, whereas we could not detect a consistent effect of caller sex on the presence of NLP in chimpanzees. The sex differences in bonobos might rather be explained by differences in the social role of each sex. Indeed, female bonobos support each other in conflict much more than female chimpanzees and any bonobo female can be, in principle, the supporter of another one [93]. In bonobos, females form large-scale coalitions, whereas males have fewer coalitionary partners [88,94]. On the contrary, chimpanzees present fewer female–female coalitions and it is rather males who form the coalitions with each other [60].

This study is a pioneering investigation into the NLP of vocalizations in great apes. Interestingly, our study reveals that the vibrato, which is usually ignored as an NLP type, is more present in chimpanzees than bonobos. The production mechanisms and the communicative functions of such slow frequency modulation remain an open question. However, the fact that we compare captive bonobos with wild chimpanzees may be a limitation of our conclusions. We cannot totally rule out that the species differences highlighted could result from the difference in how the environment impacts valence rather than true species differences. Indeed, in fission–fusion societies, such as those of both bonobos and chimpanzees, individuals involved in a conflict may temporarily distance themselves from the group, and group fission may occur. Despite the best efforts to reproduce this fission–fusion dynamic in captivity in order to manage the conflictual situations, the conditions of captivity inevitably alter this dynamic and may, for example, favour the occurrence of reconciliatory behaviours in zoos. However, the likelihood of recording mild aggression or quieter positive or neutral interactions was similar between the two environments. Indeed, the high level of habituation level of the wild chimpanzee communities to humans allows us to approach them at close range, similar to the situation in captivity. We were, therefore, confident that we would have similar chances of detecting an arousal effect between wild and captive data. Furthermore, most of the behavioural patterns and species differences, such as high rates of aggression and strategic decision-making are found in both captive and wild environments [95].

5. Conclusion

In conclusion, our study brings new light to the socio-emotional factors triggering the production of NLP in our two closest living relatives. We show that the production of NLP in vocalizations reflects both emotional valence (in both species) and arousal (at least in bonobos). However, while valence and arousal influence the occurrence of NLP production, this is not the case for the types of NLP produced. Whether great apes can control their NLP production in their vocalizations and intentionally exaggerate or reduce their production according to the audience, or whether the production of NLP and NLP types is simply a by-product of the emotion themselves remains to be determined. Regardless of the process, however, the presence of NLP in vocalizations may provide an additional informative layer pertaining to the emotional state of the caller, in addition to the semantic content or function of the vocalization types themselves. This extra layer may be more important in species where social interactions are opportunistic and the identity of allies to recruit is more unpredictable, such as in bonobos, than in the more strategic species, like chimpanzees. Future studies should determine whether the presence of NLP in vocalizations influences receiver reaction and in particular their likelihood to provide support in conflicts. This knowledge is essential to fully understand the potential adaptive value of NLP production in non-human primates.

This study is the first to quantify NLP in bonobos and chimpanzees during conflict and opens up a wide range of new questions, the most important of which is whether the presence of NLP in vocalizations affects the behavioural responses of receivers.

We provide an overview of the production of NLP in the bonobo and chimpanzee vocalizations. Arousal level and valence might be linked to NLP production and could be used as a proxy for the assessment of the emotional state of the individuals.

Ethics

All methods employed in this study were purely observational and non-invasive. For bonobos, data were collected in zoological parks, all certified by the European Associate of Zoos and Aquaria (EAZA). All data collection protocols were performed in accordance with the relevant guidelines and regulations and were approved by the Institutional Animal Ethical Committee of the University of Lyon/Saint-Etienne, under authorization no. 42-218-0901-38 SV 09 (Lab ENES). For chimpanzee, ‘Ethikrat’ of the Max Planck Society, the Ministére de l'Enseignement supérieur et de la Recherche scientifique, the Ministére des Eaux et Forêts of Côte d'Ivoire and the Office Ivoirien des Parcs et Réserves all gave ethical approval for this study. Researchers at the Taï Chimpanzee Project follow strict hygiene rules to prevent zoonotic transmission of pathogens.

Data accessibility

All datasets as well as R codes for analysing data can be downloaded from [96].

Supplementary material is available online [97].

Declaration of AI use

We have not used AI-assisted technologies in creating this article.

Authors’ contributions

F.F.: conceptualization, data curation, formal analysis, methodology, visualization, writing—original draft; L.P.: formal analysis, writing—review and editing; C.G.-B.: writing—original draft; S.K.: data curation, investigation, writing—review and editing; T.B.: data curation, investigation, writing—review and editing; R.M.W.: funding acquisition, resources, supervision, writing—review and editing; C.C.: funding acquisition, supervision, writing—review and editing; F.L.: conceptualization, funding acquisition, methodology, project administration, supervision, writing—original draft.

All authors gave final approval for publication and agreed to be held accountable for the work performed therein.

Conflict of interest declaration

We declare we have no competing interests.

Funding

Funding was provided by the Ecole Normale Supérieure de Lyon (PhD grant to F.F.), the university of Saint-Etienne (research funding (F.L. and L.P.), the Institut universitaire de France (F.L.) and the CRNS (C.G.B.; R.W.; C.C.). Chimpanzee data collection was funded by Max Planck Society Evolution of Brain Connectivity Funds (M.IF.EVAN 8103).