Social context information and acoustic convergence in elephant vocal choruses

Social species often use vocal communication to coordinate collective behaviors and mediate complex social interactions. Most studies of non-human vocal communication analyze calls as discrete units within vocal exchanges, and comparatively few have investigated the acoustic structure of choruses where many individuals vocalize simultaneously (but see¹). For animals that live in social groups, “choruses”, i.e., multi-caller, overlapping vocalizations, are common and serve diverse social functions. For example, chimpanzees (Pan troglodytes) use chorused calling to regulate group dynamics during fission-fusion events², sperm whales (Physeter macrocephalus) communicate in overlapping codas which are thought to reinforce social bonds between individuals³, groups of male Atlantic spotted dolphins (Stenella frontalis) produce overlapping whistles when forming coalitions against competitors⁴, and communally breeding cuckoos (Crotophaga major) perform group choruses that can support group formation and may play a role in collective decision making⁵. Like single calls, choruses may also contain acoustic signatures of identity. For example, green woodhoopoe (Phoeniculus purpureus) rallies (choruses of multiple group members calling simultaneously) are distinct at the group level and can be used by woodhoopoes to discriminate between neighbor and stranger groups⁶. Acorn woodpeckers (Melanerpes formicivorus) can determine whether two chorusing individuals belong to the same group as each other, likely by recognizing each individual caller within the overlapping chorus⁷. Given the prevalence of chorusing across taxa and its use in complex social interactions, the dynamics and patterns of overlapping multi-caller exchanges likely encode important social information in many species.

Vocal convergence, in which individuals modify their vocalizations to become more acoustically similar during interactions, has been observed in several species and shown to serve diverse functions across taxa. Rapid vocal convergence over the course of a single interaction has been documented in a variety of other animals, including songbirds^8,9, parrots^10,11, and dolphins¹². In songbirds, song-type matching or frequency matching between territorial neighbors is often a signal of aggressive intent^9,13. By contrast, bottlenose dolphins (Tursiops truncatus) sometimes copy a conspecific’s signature whistle during an interaction to address one another in an affiliative manner¹⁴. In orange-fronted conures (Eupsittula canicularis), vocal convergence or divergence during contact calling may help to coordinate fission-fusion events between flocks^10,15. Vocal convergence has also been documented in nonhuman primates^16,17, and has been extensively studied in humans, where it occurs at multiple linguistic levels (e.g., syntax^18,19, word choice²⁰, speech rate²¹, and numerous acoustic/phonetic features^22,23,24. In both humans and other primates, vocal convergence is hypothesized to signal social closeness²⁵.

African savannah elephants (Loxodonta africana) offer an opportunity to study overlapping calls and acoustic convergence given their large social groups and extensive use of flexible vocal communication to mediate complex social interactions. Female and juvenile savannah elephants live in fission-fusion societies with intricate social structures containing multiple organizational levels. Typically, related mothers and their dependent offspring form a tightly knit “family” group led by the oldest female, or “matriarch”, multiple related family groups form a more loosely knit “bond” group, and multiple unrelated bond groups form an even more loosely affiliated “clan”, with multiple clans coexisting in the same population^26,27. Although elephants most commonly associate with other members of their family units, they can recognize the voices of elephants in other families and bond groups in their broader clan²⁸ and in some cases use vocal communication to coordinate movements with clan members from different families²⁹. Elephants produce a variety of vocal signal types, the most common of which are very low frequency rumbles, which contain fundamental frequencies below the level of human hearing³⁰. Some rumbles such as Contact and Estrous rumbles can contain very high sound pressure levels³⁰, which have been empirically shown to be audibly detected by elephants up to 2 km away and may be detectable up to 10 km away for the loudest calls under certain atmospheric conditions^31,32. Playback experiments showed that elephants can recognize the voices of clan members at distances of up to 1.5–2 km³³.

Approximately 60% of elephant rumbles occur within choruses of two or more overlapping voices¹. Longer choruses are especially prevalent in high arousal social interactions such as greeting ceremonies and mating pandemoniums and may strengthen social bonds and group cohesion¹. Although elephants frequently conduct prolonged vocal exchanges that include both single rumbles and chorused rumbles, relatively little work has investigated the dynamics of chorusing and how these signals may differ structurally from single rumbles. Given that these extended multi-caller vocal exchanges are thought to be important for social affiliation and coordination, acoustic convergence may occur during these interactions, though this phenomenon has not been previously studied in elephants. Case studies of a few captive elephants have shown that they are capable of heterospecific vocal mimicry, which requires an unusually well-developed capacity for vocal production learning^34,35. Wild elephants, too, have been documented to create novel sounds¹, which family members have been heard to imitate (The Elephant Ethogram Croaking). It is unknown how vocal production learning might manifest in normal vocal exchanges in wild elephants, but vocal convergence during social interactions is one possible consequence of this ability.

Elephant rumbles encode several distinct types of information. Individuals have unique vocal signatures, enabling identity recognition³³, and rumbles vary with the age^1,36,37, sex³⁸, and emotional/physiological state of the caller^39,40. Acoustic differences in rumbles have been shown to exist between populations and to a subtle extent between social groups within a population⁴¹. Rumbles produced in different behavioral contexts have distinct acoustic characteristics^1,42,43, and the structure of some rumbles has also been shown to vary depending on the individual being addressed; i.e., they appear to serve as a form of “names” in this species⁴⁴.

Although much is known about some of the information encoded in single rumbles, rumble choruses are less studied and understood (but see¹). This gap exists partially because of the challenges of analyzing overlapping signals (the cocktail party problem), as well as the intensive data collection effort required to sample choruses from known individuals. Consequently, little is known about the degree to which chorused rumbles encode information about behavioral context and individual identity, as has been shown for single rumbles, as well as the structure and composition of multi-caller choruses.

Here, we study a wild population of African elephants and analyze how acoustic characteristics of single (non-overlapping) and chorused (overlapping) rumbles correspond to their varying social functions across individuals and families. We address three primary questions about the relation between single rumbles and choruses: (A1) Can behavioral context be predicted from acoustic characteristics of both single rumbles and choruses? (A2) Do rumbles from choruses encode context in the same way as single rumbles? (A3) Do rumbles become more similar to each other (i.e. converge) over the course of a multi-caller exchange? We also investigate, as secondary questions, whether the context-specific acoustic characteristics of single and chorused rumbles are the same across (B1) individuals from the same family and (B2) across individuals from different families within a population.

Context prediction from single and chorused rumbles (A1)

We successfully predicted context from single and chorused rumbles significantly better than chance (all p < 0.05; Fig. 1, blue bars for within-family training and testing). Weighted average precision and recall (respectively) for the four models were: EB single: 0.47, 0.44 (predicted distribution baseline (PDB) = 0.19, accuracy vs. PDB: p < 0.001); non-EB single: 0.45, 0.46 (PDB = 0.25, accuracy vs. PDB: p < 0.001); EB chorus: 0.37, 0.36 (PDB = 0.21, accuracy vs. PDB: p < 0.001) non-EB chorus: 0.31, 0.33 (PDB = 0.25, accuracy vs. PDB: p = 0.016) (see Table S1 for full classification metrics and Fig. S1 for feature importances). Despite achieving overall accuracy above chance, the classifiers frequently confused certain context type pairs. Among single rumbles, Coo, Let’s-Go, and Little-Greetings were most frequently confused, while Contact rumbles were the least likely to be confused of any context type. For chorused rumbles, Contact and Greeting rumbles were most often confused, and Greetings were also frequently confused with Little-Greetings and Cadenced rumbles. Confusion matrices for single and chorused rumbles from the EB family are shown in Fig. 2 (see Fig. S2 for results from all tests).

Relationship of single rumbles and choruses (A2)

When training on single rumbles and testing on chorused rumbles, our classifiers predicted the context of choruses significantly better than chance (all p < 0.01; Fig. 1: third blue bars; EB: weighted average precision = 0.31, recall = 0.25, PDB = 0.13, accuracy vs. PDB: p < 0.001; non-EB: weighted average precision = 0.40, recall = 0.32, PDB = 0.23, accuracy vs. PDB: p = 0.004). This suggests that single and chorused rumbles made within the same context share acoustic characteristics.

Vocal convergence within multi-caller vocal exchanges (A3)

We found evidence for acoustic convergence in three of the six context types. Cadenced (convergence rate (mean ± SE) = 0.364 ± 0.137; order-shuffled p < 0.001, exchange-shuffled p < 0.065), Contact (convergence score = 0.525 ± 0.190; p-values: 0.005, 0.005), and Greeting rumbles (convergence score = 0.367 ± 0.184; p-values: 0.010, 0.020) all showed statistically significant convergence, while Coo (− 0.121 ± 0.198), Let’s-Go (− 0.272 ± 0.106), and Little-Greeting (0.192 ± 0.169) did not (all p > 0.05 ; Fig. 3; Fig. S3).

Context prediction across individuals and families (B1, B2)

Within-family generalization (B1, EB family only): When testing on EB individuals that were held out during training, accuracy dropped substantially compared to within-family tests that allowed rumbles from the same individual to be included in both training and testing data, but remained significantly above chance (Fig. 1a: yellow bar vs. blue bar; weighted average precision = 0.35, recall = 0.33, PDB = 0.19, accuracy vs. PDB: p < 0.001). Recall was above 0 for each context type, suggesting that context-specific acoustic features can generalize across individuals despite individual variation.

Across-family generalization (B2): Cross-family training and testing led to additional drops in accuracy in some cases relative to within-family performance (Fig. 1: red bars), but still remained significantly above chance for all tests (accuracy range 0.19–0.32, PDB range 0.14–0.22, all p < 0.05). Performance was somewhat lower when classifiers were trained on single rumbles from one family and tested on chorused rumbles from another (accuracy 0.19–0.21) compared to same-call-type cross-family tests (single-to-single and chorus-to-chorus; accuracy 0.26–0.32). Among all six cross-family tests, weighted average precision ranged from 0.23 to 0.48 and weighted average recall ranged from 0.19 to 0.32 (see Table S1 for complete results).

Chorused rumbles encode behavioral context

Our results show that there are systematic differences in rumble structure, at least on average, between the six context types. Context type prediction was significantly above chance for all tests (with one exception; see below), reproducing previous findings that rumble acoustic structure varies with context type^1,43 and extending these findings to multi-caller choruses. This study demonstrates that chorused rumbles encode context type using acoustic features similar to single rumbles, allowing for classification of choruses by context type and showing that overlapping multi-caller signals preserve context-specific information even when individuals vocalize simultaneously. This finding is significant because rumbles produced under natural conditions often exhibit substantial temporal overlap, and most previous studies have ignored these rumbles due to an inability to separate them and measure them individually. Our results suggest that future studies may be able to achieve greater sample sizes by including overlapping rumbles and/or multicaller exchanges. Moreover, our finding that multi-caller rumble exchanges exhibit context-specific acoustic convergence suggests that multi-caller signals may contain an additional layer of more fine-grained information within a context (see vocal convergence section below) that is not accessible from individual rumbles analyzed independently. Recent work with sperm whales and orcas (Orcinus orca) has also successfully predicted broad behavioral contexts from sequences of calls produced by multiple callers^45,46, albeit using different analysis techniques. Including choruses in analyses attempting to link call acoustic structure to behavioral context is likely to be especially valuable for studying acoustic communication in systems where it is often difficult to identify individual callers, such as cetaceans and elephants.

Individual and family-level variation

Prediction accuracy remained above chance even when testing across family groups (B2), demonstrating that context-specific acoustic features generalize partially across individuals and family groups. However, accuracy decreased considerably when testing on held-out individuals (B1, EB only), or across families (B2), suggesting that, in addition to context type, rumbles also encode information about caller and perhaps family identity. The existence of individual differences in rumbles produced by different elephant callers is well established^33,47,48. Additionally, female African savannah elephants have previously been shown to exhibit acoustic differences in rumble structure between different populations and to a lesser extent between different social groups within the same population⁴¹. While the decrease in accuracy for cross-family tests (Fig. 1, red bars) relative to within-family tests (Fig. 1, blue bars) could reflect the presence of family-level acoustic differences, these results could also be explained by individual-level variation alone, which may nonetheless allow elephants to differentiate among family groups²⁸. The degree to which family-level and/or population-level vocal differences persist in this species, as well as the forces shaping them (e.g. learned vocal convergence with a family/population, genetic or environmental effects, environmental adaptation, or some combination of these factors) remains an open question.

As an exploratory analysis, we performed the held-out caller analysis on the non-EB dataset, which includes individuals from multiple families. Prediction accuracy was not significantly above chance (p = 0.41), consistent with our finding that both individual- and family-level variation limit generalization. The non-EB dataset also had high imbalance among individuals and context types, such that caller identity and context type were highly confounded, unlike the EB dataset, which contained only 16 individuals and was relatively balanced across context types and individuals. This likely prevented the model from learning generalizable context-specific acoustic features; rather, the model may have learned individual-specific features that happened to correlate with specific contexts during training.

Context-specific patterns in classification

Although our models were able to predict context type above chance, more rumbles were classified incorrectly than correctly. This supports previous findings that the acoustic structure of rumbles is highly graded with substantial overlap between contexts^1,43,49,50. Moreover, behavioral contexts themselves often gradually transition from one to another, and it is sometimes difficult for observers in the field to objectively identify the boundaries between consecutive behavioral contexts¹. Additionally, it seems likely that context type labels we applied to rumbles are imperfectly correlated with the contextual information content of the rumbles as perceived by the elephants. If different rumbles within the same broad context type are functionally distinct, this could also explain some of the misclassifications.

Despite these limitations to our prediction models, the variation in accuracy among context types reveals interesting patterns. For single rumbles, contact calls frequently had the highest precision and recall rates among all context types, including in the experiments in which we trained on EB and tested on non-EB rumbles. This aligns with the qualitative observation by¹ that contact rumbles exhibit a distinctive structure. Single Coo, Let’s-Go, and Little-Greeting rumbles were frequently confused with one another (range 15–30% misclassification among these classes; Fig. S2). Accuracy was low for single Greeting rumbles, which were often misclassified as Coo or Little-Greeting; this may be explained by the relatively small number of single Greeting rumbles in our dataset (37 samples across all family groups).

When analyzing chorused rumbles, prediction accuracy was equivalently high for both Contact and Greeting rumbles; the increased accuracy in identifying greetings is likely owing to the presence of roars and trumpets in greeting ceremonies in addition to overlapping Greeting rumbles¹. Despite the relatively high accuracy for Contact and Greeting choruses compared to other context types, they were still commonly confused with each other (33% of Contact incorrectly classified as Greeting, 25% of Greeting incorrectly classified as Contact), and Cadenced and Coo choruses were also frequently misclassified as Greetings (23% in both cases; Fig. S2). Some of these confusions may be attributable to common behavioral sequences in this species. For example, Contact rumble exchanges typically occur when the caller and receiver are separated by several hundred meters, but may continue as the callers move closer and closer to one another. Eventually, the contact rumbles transition into Greeting rumbles as the callers meet, which may then transition into Cadenced rumbles as the group attempts to coordinate subsequent activity¹. In addition to gradation in signals that may occur during transitions between behavioral contexts, it is possible that some rumbles produced during transitions were mislabeled due to this ambiguity, which could account for some misclassifications. Interestingly, prediction accuracy was low for chorused Let’s-Go rumbles (only 4% correctly classified). This likely reflects the small sample size (23 samples across all family groups), as well as potential labeling ambiguity that could arise from Let’s-Go rumbles often transitioning into Cadenced rumbles¹.

Vocal convergence in multi-caller exchanges

Vocal convergence occurs across diverse timescales and social contexts in animals. Convergence over the course of days or weeks is frequently related to social bonding, as during pair-bonding in budgerigars (Melopsittacus undulatus) and common marmosets (Callithrix jacchus)^51,52 and formation of social bonds in cooperative vampire bats (Desmodus rotundus)⁵³. By contrast, rapid acoustic convergence over the course of a single vocal exchange, such as what we observed in this study, has a variety of functions in animals, ranging from social bonding in humans²⁵ to territorial aggression in songbirds (Contopus virens)⁹ to logistical coordination in parrots Eupsittula canicularis)¹⁵. If the primary function of rapid vocal convergence in elephants is social bonding, we would expect it to occur in calls related to bonding, such as Coo, Greeting, and Little-Greeting rumbles. Conversely, if the primary function of rapid vocal convergence in elephants is coordinating movement, we would expect it to occur primarily in calls related to logistical coordination, such as Contact, Let’s-Go, and Cadenced rumbles.

We found evidence for vocal convergence in two logistical contexts (Cadenced and Contact rumbles) and one social bonding context (Greeting rumbles), but no evidence for convergence in one logistical context (Let’s-Go rumbles) or two social bonding contexts (Coo and Little-Greeting rumbles). The absence of convergence in Coo and Let’s-Go should be interpreted cautiously given small sample sizes. Additionally, these contexts differ functionally from those showing convergence; Coo and Little-Greeting exchanges are typically brief (often a single rumble or just two overlapping rumbles), while Let’s-Go often involves the same individual repeatedly calling or multiple individuals attempting to influence the group to move in different directions¹. In contrast, Cadenced and Contact rumbles often occur in extended exchanges among multiple callers coordinating a shared plan of action, suggesting that the primary function of rapid vocal convergence in elephant rumble exchanges may be to facilitate logistical coordination. This interpretation aligns with findings in parrots, where rapid vocal convergence between members of different flocks is thought to signal a desire to join flocks, while rapid vocal divergence is thought to signal the opposite^10,15.

The capacity for vocal production learning that presumably underlies the rapid vocal convergence observed in our study could also be the cause of the vocal differences between populations and social groups observed by⁴¹, via long-term vocal convergence among members of the same population/social group. Indeed, most known examples of vocal convergence occur on a much longer timescale than a single vocal interaction. In many species, acoustic convergence over a period of weeks or more leads to vocal group signatures or dialects, in which members of the same social group or individuals within a particular geographical area have calls that are more similar to one another than to the calls of other conspecifics^54,55,56,57. This connection between vocal convergence and dialect acquisition has been suggested in humans. For example, convergence is stronger when the two parties have different dialects or accents⁵⁸, especially for second language speakers⁵⁹. Interestingly, vocal convergence in humans is context-dependent⁵⁹, paralleling what we observed in elephants. Future work to help explain the relationship between short term vocal convergence and long-term dialect acquisition will require closer examination of levels of convergence across different contexts and measuring whether groups with higher levels of short-term convergence also show greater vocal similarity over longer timescales.

Implications, limitations, future work

Our findings suggest that both single and chorused rumbles can simultaneously encode multiple types of information and that individuals coordinate vocal behavior by using context-specific rumbles in choruses and by exhibiting context-dependent vocal convergence in multi-caller exchanges. Limitations to this work included dataset imbalance and limited samples from multiple family groups, particularly in the non-EB dataset, and the absence of playback experiments to further support observed patterns of vocal convergence. We recommend that future studies address these shortcomings with balanced data collection across multiple families and playback experiments to evaluate responses to convergent vs. non-convergent exchanges.

An additional limitation was the determination of context type and caller ID in rumble choruses. This coding relied on real time judgment by experienced observers (> 10 years) present at the time of data collection. Data were recorded by a single designated observer per session (J.P., or a research assistant in 2009), accompanied by a second experienced observer (N.N. or P.G.), who was consulted in cases of uncertainty. Because judgments were not independently recorded by both observers for comparison, formal inter-rater reliability could not be assessed. Although our dataset included only samples in which context type was deemed ‘certain’, meaning that behavioral context was unambiguous, we acknowledge that this could add noise to our data, and view this as an inherent feature of studying animal behavior. Additionally, a sensitivity analysis restricted to only exchanges with ‘certain’ caller IDs supported the robustness of our results (Table S2). To address this limitation, we recommend that future work pair video recording with independent, retrospective coding by additional observers to formally assess inter-rater reliability.However, even when using video, context determination often requires knowledge of events that happened many minutes or even hours before the event in question, as well as knowledge of which family members are and are not present when the call is made, which is hard to assess from a video that has a limited field of view. The use of stenomasks, which let observers dictate detailed real-time notes on a separate audio channel without diverting their attention away from subjects, who may surround the vehicle or may require binoculars to see, may help address this challenge.

Lastly, extraction of our acoustic contour features required resampling each rumble’s spectrogram to a common time-axis length, which does not preserve absolute call duration. Notably, we found that adding duration as an additional feature produced only small, inconsistent changes in context type classification of single rumbles (Table S3), which suggests that context-relevant information is sufficiently captured by other features included in our analysis. However, we were not able to conduct an analogous test for chorused rumbles, because no single duration measure is directly comparable across choruses with varying numbers of rumbles and overlap. We therefore cannot rule out the possibility that this time-axis normalization removes context information from choruses, and recommend that future studies explore approaches for incorporating duration information when analyzing choruses. Despite these limitations, our findings offer novel insight into how chorusing and multi-caller exchanges are composed and how vocal coordination may facilitate group coordination in this species.

Study population

We collected audio recordings of wild elephant family groups in Amboseli National Park, Kenya, for periods in 1986–1990, 1998–2009, 2020 and 2022. The population has been continuously monitored by the Amboseli Elephant Research Project (AERP) since 1972, and all individuals can be individually identified by external ear morphology²⁶. The population grew from ~ 600 individuals (~ 28 families) in the 1990s to ~ 2000 (60 families) by 2020 (AERP long-term records).

Acoustic data collection

We recorded elephants from a vehicle during daylight hours using all-occurrence sampling for vocalizations⁶⁰. Recording equipment included a Nagra IV-SJ reel-to-reel tape recorder with a Sennheiser MHK 110 microphone; an HHB PDR 1000 DAT or factory modified Nagra Ares BB digital recorder with an Earthworks QTC1 microphone (4 to 40 kHz range) or a Sound Devices MixPre3 with an Earthworks QTC1 or QTC40 microphone (3–40 kHz ± 1 dB frequency response). Audio was recorded at a 48 kHz sampling rate with 16-bit resolution and stored as either 48 kHz–2 kHz wav files.

To collect field recordings, we parked near a sub-group of elephants with good visibility (5–20 m to the nearest elephant). We repositioned the vehicle when the elephants moved greater than ~ 25 m away. Although many elephant calls persist over long distances, the best quality calls and field data are typically from individuals within 25 m; accordingly, we varied the nearest elephant to ensure high-quality recordings from a range of individuals. In 1999–2000, we focused recording sessions on the ‘EB’ family of 28 individuals to increase repeated recordings from known individuals. Recordings from other families (N non-EB families = 39) are primarily from either before or after this time frame.

Field data included the general area and specific location, date, tape and track number, track start and end times, channel settings, elephant group size and type (all male, females and calves, mixed) and the identified families and individuals present. When a call was heard, the elapsed time was noted as well as, when known, the call type and context type (see below), the caller, and distance to the source. Any contextual or other comments about the situation or the behavior of the calling animal were noted. The calling elephant was identified using behavioral and contextual cues, such as an open mouth, raised head, flapping ears, or engaged in behavior known to be associated with calling, such as touching an infant, separation from the main body of the group, or being the only individual of the right age class in the immediate vicinity (adult and young calf calls are audibly distinct).

The term ‘call type’ refers to broad structurally differentiated categories of sounds (e.g. rumble, roar, trumpet, combination call;^1,61,62, and the term ‘context type’ refers to subtypes of a single broad ‘call type’ based on a combination of the behavioral context and the sound quality, assessed with the aid of long-term experience (Table 1;^1,30,63. Caller ID and context type were recorded with level of confidence (certain, fairly confident, educated guess, no idea). For this study we only used records with the call type of ‘rumble’ and context type labelled as ‘certain’. Caller ID criteria varied by analysis (see below). Context type was assigned in real time by an experienced observer during data collection, based on direct observation of the calling elephant’s behavior and social surroundings. The ‘certain’ confidence level was reserved for cases in which the behavioral context was unambiguous to the observer at the time of recording. Despite selecting only calls for which context type was labelled ‘certain’ there is, nonetheless, a great deal of variation in the precise social context surrounding the production of each call and the age and internal state of the callers. As elephant rumbles vary with behavioral context, age and the emotional state of the caller^1,36,39 this contextual heterogeneity of the recordings added substantial noise to the data.

Data were logged in a FileMaker Database. To the basic field notes we added caller absolute age (years) and age class divided into nine age categories (infant (< 6 months), calf (0–4.9 years), juvenile (5–9.9 years), and adult (> 10 years; subdivided into six age classes). As the database was populated, additional information was included where relevant. For example, due to the very low frequency nature of elephant rumbles field notes might have mentioned one caller while the spectrogram clearly showed two, in which case we inferred the likely identity of the second caller using the behavioral and acoustic cues described above, e.g., via calling-associated behavior. For example, if a female rumbled when approached by another and two overlapping rumbles were observed in the spectrogram, we assumed the second caller was the approaching female; when two females were jointly touching an infant and one of them rumbled, we assumed the second rumble was the approaching female; when a female gave a contact rumble and two overlapping calls were seen ion the spectrogram, one fainter than the other, we assumed that the fainter (more distant) call was produced by a family/bond group member of the caller was separated from the main body of the group. Within contexts, rumble fundamental frequency decreases with age and was, therefore, also used to narrow down caller identity between candidate callers. We required ‘certain’ caller IDs for single rumbles but allowed some uncertain caller IDs for chorused rumbles (where identifying all callers is inherently more difficult), and for convergence analyses (see below) we also included inferred caller identifications. Based on visual inspection of spectrograms, we noted whether the file contained single or multiple calls, with further categorization as ‘a series of single calls’, ‘two overlapping calls’, ‘multiple overlapping calls’ or ‘single and overlapping calls’ for files with multiple calls. Spectrograms containing multiple calls were categorized as containing separate calling events, rather than biphonation, based on known differences in vocal production mechanisms and the temporal structure of overlapping calls (see Supplementary Materials). Finally, we recorded the number of rumbles in each saved file (range: 1–24) and, where possible, created a unique reference number for each rumble.

For this study, we analyzed six rumble context types: three are given in a logistical context (Contact, Let’s-Go, and Cadenced) and three occur in an affiliative context (Coo, Little-Greeting and Greeting) (Table 1; sample spectrograms shown in Fig. 4). These were among the most commonly heard adult calls among elephant family groups and provided the largest sample sizes of known contexts and known callers. We limited our analyses to rumbles produced by adult females in order to avoid confounding variation between context types with variation among sexes and age groups, which has been previously documented in elephants^36,38.

Audio preprocessing

All acoustic preprocessing and subsequent analyses were conducted in Python 3.11 using librosa⁶⁴, scipy⁶⁵, and scikit-learn⁶⁶. We first downsampled all audio files to 2 kHz and manually split recordings into separate .wav files at points where silences occurred between rumbles. We categorized audio files as containing either a single rumble or a sequence of overlapping calls produced by multiple individuals, hereafter referred to as “single rumbles” and “chorused rumbles”, respectively. We removed any recordings that contained significant background noise or signals from other species. Calves (of both sexes) were excluded because their calls have different characteristics than those of adults, such as higher fundamental frequency and shorter duration, and because they participate unevenly across the six context types used in our study. Adult males were excluded because they rumble much less frequently than adult females overall and only in some of the contexts we looked at (e.g. Let’s-Go and Cadenced), and because the adult male rumbles in our dataset were mostly recorded during musth, a reproductively active state associated with structurally distinct rumbles. This resulted in a final dataset of N = 902 rumble recordings (EB family: 319 single and 285 chorused rumbles; non-EB families: 150 single and 148 chorused rumbles). Bouts of overlapping rumbles ranged in length from 2 rumbles to 18 rumbles, with 90% of choruses having 8 or fewer overlapping rumbles. Elephant vocal exchanges can last for several minutes and may include periods of silence between rumbles; therefore, the samples in this analysis likely represent only fragments of complete exchanges. To facilitate analysis with standard audio-processing tools, which are optimized for typical audio sample rates and frequency ranges, we pitch shifted audio files by a factor of 8 (e.g. a 10 Hz tone was shifted to 80 Hz, with the new sample rate of the signal being 16 kHz). Lastly, we zero-centered the audio files and applied z-score normalization to ensure that relative differences in amplitude among audio files did not impact our analyses.

Acoustic feature extraction

We extracted two complementary sets of acoustic features from each pitch-shifted wav file, which we then used for context type prediction: (1) mean log power in frequency bands of 10 Hz width, ranging from 10 to 300 Hz (30 features; note that these are represented as 80 Hz bands between 80 and 2400 Hz in the pitch-shifted audio files) and (2) a simplified version of the acoustic contour features used in⁴⁴. These feature sets are representative of the variation found in elephant rumbles, capturing spectral energy distribution and spectro-temporal dynamics, both previously found to carry biologically meaningful information. For example, both spectral and temporal characteristics of rumbles have been shown to reflect the emotional state of callers and vary with social context^{39, 43, 67}.

To calculate acoustic feature set 1 (power band features), we computed spectrograms (n_fft = 800 samples, hop size = 128 samples, window length = 512 samples) and averaged log-power spectral density within evenly spaced frequency bands (30 bands, 80 Hz bandwidth each in pitch shifted audio, corresponding to 10 Hz bandwidth in unshifted audio). We selected this bandwidth as the fundamental frequency of elephant rumbles is between approximately 12–19 Hz¹, and thus we expect that every frequency band would capture roughly one harmonic. We set the upper bound of this range at 300 Hz because it comfortably encompasses the fundamental frequency and several harmonics of rumbles. Because the lowest power band is below elephants’ vocal range and therefore contained only noise, it was zeroed out and excluded from the remaining analyses, yielding 29 power band features. Measuring energy in uniformly spaced frequency bands also makes minimal assumptions about the underlying structure of the signal (as opposed to handcrafted features that presuppose a particular shape of contour).

Acoustic feature set 2 (acoustic contour features) aimed to quantify the distribution of power across frequencies and how these changed over time in each audio file. These methods were adapted from⁴⁴ (see Supplementary Methods for details). We first calculated a mel spectrogram for each audio file (fmin = 80 Hz, fmax = 2000 Hz, Hamming window) using the librosa melspectrogram function and converted it to a dB scale. Our classification models required fixed-size inputs, but rumble recordings in our dataset varied in duration. To create fixed-size feature vectors for classification, we resampled spectrograms to a common time-axis length, which was set as the median duration m for all N rumbles within the subset each audio file belonged to (i.e., EB single rumbles, EB chorused rumbles, non-EB single rumbles, non-EB chorused rumbles). For this resampling, we used the scipy resample function to linearly interpolate new samples along the time axis. We then stacked the N spectrograms into a tensor of shape (N,13,m), and applied PCA (10 components) along the frequency axis after robust scaling using scikit-learn, reducing the shape to (N,10,m). Using a common time-axis of length m was necessary because this PCA step required all rumbles to have identical input dimensions (see Supplementary Materials for sensitivity testing of this transformation, Table S3). For each principal component’s time series, we computed the median, skewness, minimum, and interquartile range, giving 40 features total for each file. We then applied an FFT over time and extracted indices and values of the 4 largest peaks (80 additional features, yielding a total of 120 acoustic contour features).

Multi-caller exchange dataset

The acoustic convergence analysis (see below) required a separate dataset containing temporal sequences of rumbles of the same context type. For this, we grouped together subsequent audio files containing rumbles of the same context type that were separated by no more than 180 s, using the metadata to ensure that rumbles were recorded on the same tape and track, and the elapsed time field as reference for the time between recordings. We selected the 180 s threshold based on the distribution of inter-rumble intervals calculated from metadata files, which showed that 80.7% of gaps between successive rumbles were below this value. We used only rumbles from EB elephants to reduce noise that might be introduced by family-level differences in convergence patterns. This selection process allowed both non-overlapping rumbles and overlapping rumbles from different recordings to be included in the same grouping, hereafter referred to as an exchange. To independently measure acoustic characteristics of each rumble within an exchange, we used manually annotated rumble onset and offset times to extract only the non-overlapping portions of each rumble. We excluded exchanges with fewer than three rumbles that had non-overlapping portions. To ensure that this analysis focused on multi-caller interactions rather than single caller sequences, we restricted the dataset to exchanges containing rumbles produced by two or more individuals with no single caller contributing more than 70% of all rumbles. This therefore depended on caller ID for each rumble, including chorused rumbles for which some identities were uncertain (see Acoustic Data Collection section above). To ensure that caller ID uncertainty did not influence our results, we applied these criteria under the most conservative, worst-case interpretation of ambiguous IDs. We also separately re-ran our analyses on a further-restricted subset containing only exchanges with no caller-ID ambiguity whatsoever. Results were consistent between datasets (see Supplementary Materials, Table S2). We also truncated 18 exchanges containing more than 10 rumbles to include only the final 10 rumbles in order to maintain a dataset of exchanges of comparable length. This resulted in a dataset of 92 exchanges (mean ± SE per context: 15.33 ± 2.91; mean ± SE rumbles per exchange: 5.64 ± 0.28; mean ± SE callers per exchange: 3.09 ± 0.43 (range 1–8)). Lastly, we extracted log power band features (acoustic feature set 1) from the non-overlapping portion of each rumble in this dataset as described above.

Statistical analyses

We addressed five research questions (Table 2) using logistic regression models and a bespoke acoustic convergence model developed for this study. For context prediction (A1, A2, B1, B2), we treated each single rumble or chorus as a discrete acoustic unit and classified them based on extracted features. For the acoustic convergence analysis (A3), we examined temporal sequences of rumbles within multi-caller exchanges, measuring how individual rumbles change acoustically over the course of an exchange. Because samples from the EB family made up approximately half of our dataset, we analyzed rumbles from EB elephants separately from rumbles produced by non-EB elephants. All logistic regression models predicted context from acoustic features, but with varied train and test sets. All regression models were trained using balanced class weights to account for unbalanced sample sizes among contexts. We trained each context regression model with 20-fold cross validation unless otherwise specified.

(A1) Context prediction of single and chorused rumbles.

To test whether the acoustic structure of rumbles encodes information about behavioral context, we trained separate classification models to predict the context of single and chorused rumbles, resulting in four different models: EB single rumbles, EB chorused rumbles, non-EB single rumbles, and non-EB chorused rumbles.

(A2) Relationship between single rumbles and choruses.

To determine whether chorused rumbles share acoustic characteristics with single rumbles, which may give insight into how choruses are composed, we trained classifiers to predict the context of choruses, using single rumbles as training data and testing on chorused rumbles. We created separate models for EB and non-EB rumbles.

(A3) Convergence within multi-caller exchanges.

To evaluate whether rumbles become more similar over the course of a multi-caller exchange, we calculated a “convergence rate” for each exchange in the multi-caller exchange dataset. Using the feature vectors extracted from every rumble, we computed the Euclidean distance between successive rumble pairs in the 29-dimensional feature space, log-transformed these distances, and fit a linear regression to the sequence. The convergence rate was defined as the negative slope of log-transformed successive distances multiplied by the number of successive pairs, producing a measure of total log-distance change over the exchange. For example, if there are five rumbles in an exchange, and the distances between successive pairs are 2, 3, 2, and 1 (log-distances: 0.69, 1.11, 0.69, 0), then the regression slope for this exchange is -0.25, so the convergence rate is 0.25 × 4 = 1.0. Positive rates indicate that rumbles become more acoustically similar over time (i.e., convergence), and negative convergence rates indicate divergence. See Supplementary Materials for further details.

(B1 and B2) Assessing generalizability across individuals within and across families.

To investigate whether the acoustic characteristics of rumbles that correspond to different contexts are stable across different individuals within a family (research question B1), we trained classifiers to predict the context of single rumbles using a held-out caller cross-validation approach. For each iteration, we trained a classifier on single rumbles from all individuals in the EB family except one, then generated test predictions for all rumbles from the held-out caller, repeating this step until we made predictions for all samples. This ensures the classifier never trained on the individual it was testing, and thus tested whether context type-specific acoustic features were consistent across callers. We applied this approach to only the EB data as we had sufficient samples from multiple individuals within this family to perform held-out caller cross-validation; the non-EB data, drawn from several families with few samples each, did not allow for robust testing for cross-individual, within-family generalization. Additionally, we did not conduct this test on chorused rumbles because each chorus contains multiple vocalizing individuals, and therefore creating folds that hold out callers would require excluding every chorus containing any held-out individual, thereby removing a large proportion of the already limited chorus dataset and resulting in many folds with too few (or no) training examples to fit a classifier.

We then more explicitly tested whether context type-specific acoustic features are stable across families (research question B2), by training classifiers to predict context, using rumbles from one family category as training data and testing on rumbles from a different family category (i.e. training on EB rumbles and testing on non-EB rumbles, or vice versa). We performed this cross-family-category analysis for both single and chorused rumbles.

Analysis of statistical significance

To determine whether classification accuracy was significantly greater than chance, we compared against a predicted distribution baseline that randomly assigns context labels in the same proportion as the classifier’s predictions. For example, if the full classifier predicts 20% of examples to be “Cadenced”, the baseline also assigns “Cadenced” to 20% of examples, but randomly. This baseline controls for whether the classifier has simply learned which classes are more frequent, as opposed to learning which acoustic features correspond to which classes, which is particularly important for our dataset given its class imbalance. For each random baseline we ran 10,000 random permutations and counted the proportion whose accuracy was greater than or equal to that of the full classifier. This proportion is equivalent to a p-value under the null hypothesis that the classifier performs no better than chance.

We also measured the acoustic feature importances for each classification model, both globally, (i.e., which features are most useful for improving overall accuracy), and within each context (i.e., which features are most associated with that context). We computed the global feature importances using the permutation importance method⁶⁸. For the context-specific feature importances we used the classifier weight from the feature to the context, as in⁶⁹. As the classifier was trained and tested with 20-fold cross validation, we averaged these importance scores across the 20 folds.

To test whether the convergence rate of a rumble sequence was significantly above zero, we used two different permutation tests with 200 iterations each to determine whether convergence in each context was higher than expected by chance. First, we conducted an order-shuffled test, in which we randomized the rumble order within each exchange, creating randomly shuffled versions of each exchange in our dataset. We calculated the convergence rate for each shuffled exchange and then computed the mean convergence for each context type. We repeated this 200 times, creating a permuted convergence rate distribution for each context type. The p-value was calculated as the proportion of the permuted mean convergence rates that were equal to or greater than the observed mean for that context type. This test controlled for the possibility that rumbles occurring later in exchanges were inherently more similar to one another across all exchanges, which could create an apparent convergence pattern without actual within-exchange convergence.

Second, we conducted an exchange-shuffled test, testing whether the rumbles simply converged globally to the same general features at the end of every exchange, versus converging locally and becoming specifically more similar to the preceding and following rumbles in the same exchange. For this we maintained rumble positions within an exchange but randomly shuffled rumbles among all exchanges in the dataset. For example, such a shuffled exchange could consist of the first rumble in exchange A, followed by the second rumble in exchange B, followed by the third rumble in exchange C. We selected only from the other exchanges with the same number of rumbles as the original. We created 200 exchange-shuffled versions of the dataset, and again we calculated the convergence rates across all the exchanges in our dataset, keeping the original context type labels for each exchange, and then compared this to the observed mean convergence rates in each context type category. As above, the p-value was the fraction of exchange-shuffled datasets whose convergence score was greater than or equal to the true mean convergence rate. This test controlled for whether rumbles simply converged to a common “end-state” across all exchanges rather than converging specifically within an exchange.

Caller identity was used differently across our analyses. For A1 and A2, context type is classified directly from acoustic features, and context type labels were assigned in the field independently of caller identity, meaning that these analyses do not depend on caller-identification accuracy. B1 depends directly on caller identity, because it is used to create held-out-caller cross-validation folds; inaccurate caller IDs would cause leakage between training and test sets. For this reason, B1 was restricted to single rumbles from the EB family with “certain” caller IDs (see Table 2). B2 (across-family generalization) trains and tests on different families, so no individual ever appears in both sets regardless of caller ID accuracy. Only the acoustic convergence analysis (A3) relied on inferred caller identifications, as described above in Multi-caller exchange dataset.