Vocal production learning, the ability to modify vocalizations in response to sounds made by others, was a critical prerequisite for the evolution of human speech but is rare among mammals. Elephants have exhibited this ability in captivity, yet its function in wild elephants remains unknown. Female African savannah elephants (Loxodonta africana) live in large societies with nested tiers of association in which vocal signatures of group identity could facilitate recognition of distant social affiliates. Vocal production learning allows the formation of such group signatures in many species and can also cause vocal differentiation between populations. However, the existence of vocal signatures of social group or population in elephants was unexplored. We recorded multiple social groups of wild elephants in two Kenyan populations (Samburu and Amboseli) and used random forest models to determine if calls could be assigned to individual callers, family groups, bond groups (collections of family groups), or populations based on acoustic structure. Calls were assigned by a random forest model to individual callers and populations with better-than-chance accuracy, demonstrating population-level divergence in vocalization structure. While random forest models failed to accurately assign calls to family or bond group, calls from the same family or bond group were significantly more similar (higher proximity scores) than calls from different groups, suggesting the existence of group signatures as well. We discuss possible drivers of this differentiation and argue that vocal learning is the most likely explanation for population- and group-level variation in elephants. The existence of group signatures suggests recognition of large numbers of individuals as a possible adaptive function for vocal production learning in elephants.

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Subjects were wild, free-ranging African elephants living in Amboseli National Park (n=43) and Samburu & Buffalo Springs National Reserves (n=158), Kenya. There were 149 females between 0–60 years of age (mean ± SD = 22.9 ± 13.3 years) and 52 males between 0–9 years of age (mean ± SD = 3.6 ± 2.7 years). Recordings from males older than 10 years were excluded, as we had few recordings from adult males other than acoustically distinctive musth rumbles, and we were primarily interested in the vocal behavior of female-calf groups. This study was strictly observational, and the elephants were not handled as part of the study. The study was approved by Colorado State University’s Institutional Animal Care and Use Committee under protocol #19-9229A.

METHOD DETAILS

Data collection

We recorded rumbles from wild female-calf elephant groups in two Kenyan populations with long-term individual-based monitoring projects: Amboseli National Park (“Amboseli”) and Samburu and Buffalo Springs National Reserves (“Samburu”). These two populations are 390 km apart with no current gene flow between them due to intervening urban development ²⁹. The field recording methods for this dataset are published as a preprint ⁵⁵. We recorded the identity of the caller and the behavioral context of each call, when known. Behavioral context was originally scored using slightly different ethograms in Amboseli ¹⁴ and Samburu ⁵⁵. To facilitate comparison between these two datasets, we concatenated behavioral context into 12 categories shared across both populations. In Amboseli, we recorded our confidence in the caller ID and behavioral context on a subjective ordinal scale (“certain”, “fairly confident”, “educated guess”, or “no idea”). In Samburu, we recorded the confidence as the inverse of the number of possible callers or behavioral contexts (1 if known for certain, 0.5 if one of two possibilities, etc.).

Acoustic analysis

We processed and measured the recordings following a previously published protocol ⁵⁵. In brief, we calculated the time series of energy values in 26 overlapping spectral bands between 0-500 Hz warped to the Mel scale ⁵⁶, as well as the time series of delta and delta-delta coefficients for each of these bands (representing the rate and acceleration of change over time in each spectral band, respectively). We also calculated the Hilbert amplitude envelope for each call, resulting in a total of 79 equal-length time series, or “acoustic contours”, for each call. We then performed a robust principal components analysis on the acoustic contours of all calls stacked end to end, decomposing the data into a sparse matrix, which described outliers in the data, and a low-rank matrix, which described the remaining observations ⁵⁷. We saved the means of each of the first 74 low-rank principal components for each call (describing 99.9% of the variation), 16 metrics describing the modulation of the low-rank principal components, and 4 metrics describing the sparse matrix as features for further analysis.

QUANTIFICATION AND STATISTICAL ANALYSIS

All statistical analyses were performed in R v. 4.2.2 ⁵⁸ and 0.05 was used as the significance threshold for all tests. For all random forest models, we only included calls with a high degree of confidence in the identity of the caller (“certain” or “fairly confident” in Amboseli, “1” in Samburu).

To test the hypothesis that Amboseli and Samburu elephants exhibit population-level acoustic differences, we ran a random forest (500 trees, 6 variables/node, 60% of observations/tree, minimum node size = 1, no maximum tree depth) to predict population as a function of the acoustic features. This analysis included 1258 calls from 158 individuals in Samburu and 597 calls from 43 individuals in Amboseli. We randomly selected 20% of the callers with at least five calls each from each population and allocated all calls from these callers to the test set, with the remaining calls allocated to the training set. We calculated the proportion of observations in the test set that were classified correctly (classification accuracy) and ran a one-tailed exact binomial test comparing the classification accuracy to the proportion that would have been classified accurately if the model always guessed the most common group in the training set (majority classifier). We repeated this process 1000 times and calculated the median P-value across all runs. The mean ± SD proportion of calls allocated to the test set was 0.18 ± 0.043.

To test the hypothesis that elephants exhibit vocal signatures of group identity at the bond group or family group level, we ran two additional random forest models (same hyperparameters) to predict bond group and family group, respectively, as a function of the acoustic features using only data from Samburu. We defined social groups based on previous work and continuous monitoring ^12,59. We excluded Amboseli from the bond and family group models because 85% of our Amboseli recordings came from one family group. To ensure that the models could only use group-specific features and not individual-specific features to predict bond or family group, we randomly selected 20% of the callers with at least five calls each from each group and allocated all calls from these callers to the test set with the remaining callers allocated to the training set ⁶⁰. We only included groups that had at least two individuals with at least five calls each, so that each group would always have at least five calls in both the training and test sets. This resulted in 1170 calls from six bond groups for the bond group model and 1112 calls from seven family groups for the family group model. We ran 1000 iterations, calculating the classification accuracy and p-value for each run as before.  The mean ± SD proportion of calls allocated to the test set was 0.18 ± 0.054 for the bond group model and 0.16 ± 0.055 for the family group model.

To test the hypothesis that elephant rumbles are individually specific, we ran a fourth random forest (same hyperparameters) to predict individual ID as a function of the acoustic features. As calls produced by the same caller on the same date might exhibit similar features due to temporary circumstances such as caller’s internal state, behavioral context, and ambient conditions, we randomly selected one date for each caller and held out all calls from these caller-dates as the test set ⁶⁰. We used callers from both populations for this analysis but only included callers that produced at least three calls on at least two different dates each (505 calls from 22 callers in Samburu, 324 calls from 11 callers in Amboseli). We calculated the classification accuracy and p-value for each of 1000 iterations as before. The mean ± SD proportion of calls allocated to the test set was 0.24 ± 0.026.

To determine if there were acoustic differences among populations, bond groups, or family groups that could not be explained by behavioral context or caller age, we calculated random forest proximity scores between each possible pair of calls. The random forest proximity score for a given pair of calls was the proportion of trees for which both calls were classified in the same terminal node, adjusted for the size of the node, and represented a metric of call similarity in terms of the acoustic features most relevant to predicting the response variable ⁶¹. We calculated proximity scores from three different random forests (population, bond group, and family group), using the same hyperparameters and subsets of the data as before except that we increased the number of trees to 8000 and did not hold out any observations as a test set. For each set of proximity scores, we ran a generalized linear model with a gamma error distribution modeling proximity score as a function of whether the two calls in a given pair came from the same “class” (population, bond group, or family group, depending on the model) and the absolute value of the age difference between the two callers. As proximity scores could be 0, we added 0.0001 to all proximity scores so all the values would be positive. To control for behavioral context and individual ID, we only included pairs of calls with the same behavioral context and different callers. We also only included calls for which we had high confidence in the behavioral context (“certain” or “fairly confident” in Amboseli, “1” in Samburu). This resulted in a sample size of 217,389 call pairs for the population model, 101,344 call pairs for the bond group model, and 93,935 call pairs for the family group model.

REFERENCES

1. Lattenkamp, E.Z., and Vernes, S.C. (2018). Vocal learning: a language-relevant trait in need of a broad cross-species approach. Curr. Opin. Behav. Sci. 21, 209–215. 10.1016/j.cobeha.2018.04.007.
2. Janik, V.M., and Knörnschild, M. (2021). Vocal production learning in mammals revisited. Philos. Trans. R. Soc. B Biol. Sci. 376. 10.1098/rstb.2020.0244.
3. Carouso-Peck, S., Goldstein, M.H., and Fitch, W.T. (2021). The many functions of vocal learning. Philos. Trans. R. Soc. B Biol. Sci. 376. 10.1098/rstb.2020.0235.
4. Colombelli-Négrel, D., Hauber, M.E., Robertson, J., Sulloway, F.J., Hoi, H., Griggio, M., and Kleindorfer, S. (2012). Embryonic learning of vocal passwords in superb fairy-wrens reveals intruder cuckoo nestlings. Curr. Biol. 22, 2155–2160. 10.1016/j.cub.2012.09.025.
5. Hersh, T.A., Gero, S., Rendell, L., Cantor, M., Weilgart, L., Amano, M., Dawson, S.M., Slooten, E., Johnson, C.M., Kerr, I., et al. (2022). Evidence from sperm whale clans of symbolic marking in non-human cultures. Proc. Natl. Acad. Sci. U. S. A. 119, 1–8. 10.1073/pnas.2201692119.
6. Boughman, J.W., and Wilkinson, G.S. (1998). Greater spear-nosed bats discriminate group mates by vocalizations. Anim. Behav. 55, 1717–1732. 10.1006/anbe.1997.0721.
7. Wright, T.F., and Dahlin, C.R. (2017). Vocal dialects in parrots: Patterns and processes of cultural evolution. Emu 118, 50–66. 10.1080/01584197.2017.1379356.
8. Campbell, P., Pasch, B., Pino, J.L., Crino, O.L., Phillips, M., and Phelps, S.M. (2010). Geographic variation in the songs of neotropical singing mice: Testing the relative importance of drift and local adaptation. Evolution (N. Y). 64, 1955–1972. 10.1111/j.1558-5646.2010.00962.x.
9. Ey, E., and Fischer, J. (2009). The “acoustic adaptation hypothesis”- a review of the evidence from birds, anurans, and mammals. Bioacoustics 19, 21–48. 10.1080/09524622.2009.9753613.".
10. Stoeger, A.S., Mietchen, D., Oh, S., de Silva, S., Herbst, C.T., Kwon, S., and Fitch, W.T. (2012). An Asian elephant imitates human speech. Curr. Biol. 22, 2144–2148. 10.1016/j.cub.2012.09.022.
11. Poole, J.H., Tyack, P.L., Stoeger-Horwath, A.S., and Watwood, S. (2005). Elephants are capable of vocal learning. Nature 434, 455–456. 10.1029/2001GL014051.
12. Wittemyer, G., Douglas-Hamilton, I., and Getz, W.M. (2005). The socioecology of elephants: analysis of the processes creating multitiered social structures. Anim. Behav. 69, 1357–1371. 10.1016/j.anbehav.2004.08.018.
13. Archie, E.A., Moss, C.J., and Alberts, S.C. (2011). Friends and relations: kinship and the nature of female elephant social relationships. In The Amboseli elephants: a long-term perspective on a long-lived mammal, C. J. Moss, H. Croze, and P. C. Lee, eds. (University of Chicago Press), pp. 238–245. 10.7208/chicago/9780226542263.001.0001.
14. Poole, J.H. (2011). Behavioral contexts of elephant acoustic communication. In The Amboseli elephants: a long-term perspective on a long-lived mammal, C. J. Moss, H. Croze, and P. C. Lee, eds. (University of Chicago Press), pp. 125–159.
15. Soltis, J. (2010). Vocal communication in African elephants (Loxodonta africana). Zoo Biol. 29, 192–209. 10.1002/zoo.20251.
16. Soltis, J., Leong, K., and Savage, A. (2005). African elephant vocal communication II: Rumble variation reflects the individual identity and emotional state of callers. Anim. Behav. 70, 589–599. 10.1016/j.anbehav.2004.11.016.
17. Clemins, P.J., Johnson, M.T., Leong, K.M., and Savage, A. (2005). Automatic classification and speaker identification of African elephant (Loxodonta africana) vocalizations. J. Acoust. Soc. Am. 117, 956.
18. de Silva, S. (2010). Acoustic communication in the Asian elephant, Elephas maximus maximus. Behaviour 147, 825–852. 10.1163/000579510X495762.
19. Stoeger, A.S., and Baotic, A. (2016). Information content and acoustic structure of male African elephant social rumbles. Nat. Publ. Gr., 1–8. 10.1038/srep27585.
20. Wierucka, K., Henley, M.D., and Mumby, H.S. (2021). Acoustic cues to individuality in wild male adult African savannah elephants (Loxodonta africana). PeerJ 9, 1–14. 10.7717/peerj.10736.
21. McComb, K., Moss, C., Sayialel, S., and Baker, L. (2000). Unusually extensive networks of vocal recognition in African elephants. Anim. Behav. 59, 1103–1109.
22. Shannon, G., Slotow, R., Durant, S.M., Sayialel, K.N., Poole, J., Moss, C., and McComb, K. (2013). Effects of social disruption in elephants persist decades after culling. Front. Zool. 10, 62. 10.1186/1742-9994-10-62.
23. Rhodes, J.S., Cutler, A., and Moon, K.R. (2022). Geometry- and accuracy-preserving random forest proximities. arXiv, 1–18. arXiv:2201.12682.
24. Stoeger, A.S., and Manger, P. (2014). Vocal learning in elephants: neural bases and adaptive context. Curr. Opin. Neurobiol. 28, 101–107. 10.1016/j.conb.2014.07.001.
25. Podos, J., and Warren, P.S. (2007). The evolution of geographic variation in birdsong. Adv. Study Behav. 37, 403–458. 10.1016/S0065-3454(07)37009-5.
26. Knörnschild, M., Nagy, M., Metz, M., Mayer, F., and Von Helversen, O. (2012). Learned vocal group signatures in the polygynous bat Saccopteryx bilineata. Anim. Behav. 84, 761–769. 10.1016/j.anbehav.2012.06.029.
27. Hile, A., Plummer, T., and Striedter, G. (2000). Male vocal imitation produces call convergence during pair bonding in budgerigars, Melopsittacus undulatus. Anim. Behav. 59, 1209–1218. 10.1006/anbe.1999.1438.
28. Keen, S.C., Meliza, C.D., and Rubenstein, D.R. (2013). Flight calls signal group and individual identity but not kinship in a cooperatively breeding bird. Behav. Ecol. 24, 1279–1285. 10.1093/beheco/art062.
29. Chiyo, P.I., Grieneisen, L.E., Wittemyer, G., Moss, C.J., Lee, P.C., Douglas-Hamilton, I., and Archie, E.A. (2014). The influence of social structure, habitat, and host traits on the transmission of Escherichia coli in wild elephants. PLoS One 9. 10.1371/journal.pone.0093408.
30. MacDougall-Shackleton, E.A., Derryberry, E.P., and Hahn, T.P. (2002). Nonlocal male mountain white-crowned sparrows have lower paternity and higher parasite loads than males singing local dialect. Behav. Ecol. 13, 682–689. 10.1093/beheco/13.5.682.
31. Ey, E., Pfefferle, D., and Fischer, J. (2007). Do age- and sex-related variations reliably reflect body size in non-human primate vocalizations? A review. Primates 48, 253–267. 10.1007/s10329-006-0033-y.
32. Parker, J.M., and Wittemyer, G. (2022). Orphaning stunts growth in wild African elephants. Conserv. Physiol. 10, 1–10. 10.1093/conphys/coac053.
33. Okello, J.B.A., Masembe, C., Rasmussen, H.B., Wittemyer, G., Omondi, P., Kahindi, O., Muwanika, V.B., Arctander, P., Douglas-Hamilton, I., Nyakaana, S., et al. (2008). Population genetic structure of savannah elephants in Kenya: Conservation and management implications. J. Hered. 99, 443–452. 10.1093/jhered/esn028.
34. Hedwig, D., Poole, J., and Granli, P. (2021). Does social complexity drive vocal complexity? Insights from the two african elephant species. Animals 11, 1–21. 10.3390/ani11113071.
35. Roca, A.L., Georgiadis, N., Pecon-Slattery, J., and O’Brien, S.J. (2001). Genetic evidence for two species of elephant in africa. Science (80-. ). 293, 1473–1477.
36. Pardo, M.A., Poole, J.H., Stoeger, A.S., Wrege, P.H., O’Connell-Rodwell, C.E., Padmalal, U.K., and De Silva, S. (2019). Differences in combinatorial calls among the 3 elephant species cannot be explained by phylogeny. Behav. Ecol. 30, 809–820. 10.1093/beheco/arz018.
37. Kershenbaum, A., Ilany, A., Blaustein, L., and Geffen, E. (2012). Syntactic structure and geographical dialects in the songs of male rock hyraxes. Proc. R. Soc. B Biol. Sci. 279, 2974–2981. 10.1098/rspb.2012.0322.
38. Reyes-Arias, J.D., Brady, B., Ramos, E.A., Henaut, Y., Castelblanco-Martínez, D.N., Maust-Mohl, M., Searle, L., Pérez-Lachaud, G., Guzmán, H.M., Poveda, H., et al. (2023). Vocalizations of wild West Indian manatee vary across subspecies and geographic location. Sci. Rep. 13, 11028. 10.1038/s41598-023-37882-8.
39. Bader-El-Den, M., Teitei, E., and Perry, T. (2019). Biased random forest for dealing with the class imbalance problem. IEEE Trans. Neural Networks Learn. Syst. 30, 2163–2172. 10.1109/TNNLS.2018.2878400.
40. Bates, L.A., Sayialel, K.N., Njiraini, N.W., Poole, J.H., Moss, C.J., and Byrne, R.W. (2008). African elephants have expectations about the locations of out-of-sight family members. Biol. Lett. 4, 34–36. 10.1098/rsbl.2007.0529.
41. Carlson, N. V., Kelly, E.M.K., and Couzin, I. (2020). Individual vocal recognition across taxa: A review of the literature and a look into the future: Individual vocal recognition. Philos. Trans. R. Soc. B Biol. Sci. 375. 10.1098/rstb.2019.0479.
42. Yurk, H., Barrett-Lennard, L.G., Ford, J.K.B., and Matkin, C.O. (2002). Cultural transmission within maternal lineages: vocal clans in resident killer whales in southern Alaska. Anim. Behav. 63, 1103–1119. 10.1006/anbe.2002.3012.
43. Rendell, L.E., and Whitehead, H. (2003). Vocal clans in sperm whales (Physeter macrocephalus). Proc. R. Soc. B Biol. Sci. 270, 225–231. 10.1098/rspb.2002.2239.
44. Whitehead, H., Dillon, M., Dufault, S., Weilgart, L., and Wright, J. (1998). Non-geographically based population structure of south pacific sperm whales: Dialects, fluke-markings and genetics. J. Anim. Ecol. 67, 253–262. 10.1046/j.1365-2656.1998.00187.x.
45. Van Cise, A.M., Mahaffy, S.D., Baird, R.W., Mooney, T.A., and Barlow, J. (2018). Song of my people: dialect differences among sympatric social groups of short-finned pilot whales in Hawai’i. Behav. Ecol. Sociobiol. 72. 10.1007/s00265-018-2596-1.
46. Ford, J.K.B. (1991). Vocal traditions among resident killer whales (Orcinus orca) in coastal waters of British Columbia. Can. J. Zool. 69, 1454–1483. 10.1139/z91-206.
47. Henry, L., Barbu, S., Lemasson, A., and Hausberger, M. (2015). Dialects in animals: evidence, development and potential functions. Anim. Behav. Cogn. 2, 132–155. 10.12966/abc.05.03.2015.
48. Briefer, E.F., and McElligott, A.G. (2012). Social effects on vocal ontogeny in an ungulate, the goat, Capra hircus. Anim. Behav. 83, 991–1000.
49. Volodin, I.A., Volodina, E. V, Lapshina, E.N., Efremova, K.O., and Soldatova, N. V (2014). Vocal group signatures in the goitred gazelle Gazella subgutturosa. Anim. Cogn. 17, 349–357. 10.1007/s10071-013-0666-3.
50. Zürcher, Y., Willems, E.P., and Burkart, J.M. (2019). Are dialects socially learned in marmoset monkeys? Evidence from translocation experiments. PLoS One 14, e0222486. 10.1371/journal.pone.0222486.
51. Mitani, J.C., Hunley, K.L., and Murdoch, M.E. (1999). Geographic variation in the calls of wild chimpanzees: A reassessment. Am. J. Primatol. 47, 133–151. 10.1002/(SICI)1098-2345(1999)47:2<133::AID-AJP4>3.0.CO;2-I.
52. Tanaka, T., Sugiura, H., and Masataka, N. (2006). Cross-sectional and longitudinal studies of the development of group differences in acoustic features of coo calls in two groups of Japanese macaques. Ethology 112, 7–21. 10.1111/j.1439-0310.2006.01103.x.
53. Stoeger, A.S., Zeppelzauer, M., and Baotic, A. (2014). Age-group estimation in free-ranging African elephants based on acoustic cues of low-frequency rumbles. Bioacoustics 23, 231–246.
54. Lameira, A.R., Delgado, R.A., and Wich, S.A. (2010). Review of geographic variation in terrestrial mammalian acoustic signals: Human speech variation in a omparative perspective. J. Evol. Psychol. 8, 309–332. 10.1556/JEP.8.2010.4.2.
55. Pardo, M.A., Fristrup, K., Lolchuragi, D.S., Poole, J., Granli, P., Moss, C., Douglas-Hamilton, I., and Wittemyer, G. (2023). African elephants address one another with individually specific calls. bioRxiv. 10.1101/2023.08.25.554872.
56. Stevens, S.S., Volkmann, J., and Newman, E.B. (1937). A scale for the measurement of the psychological magnitude pitch. J. Acoust. Soc. Am. 8, 185–190. 10.1121/1.1915893.
57. Sykulsi, M. (2015). rpca: RobustPCA: decompose a matrix into low-rank and sparse components.
58. R Core Team (2022). R: A language and environment for statistical computing.
59. Goldenberg, S.Z., and Wittemyer, G. (2017). Orphaned female elephant social bonds reflect lack of access to mature adults. Sci. Rep. 7, 1–7. 10.1038/s41598-017-14712-2.
60. Lehmann, K.D.S., Jensen, F.H., Gersick, A.S., Strandburg-Peshkin, A., and Holekamp, K.E. (2022). Long-distance vocalizations of spotted hyenas contain individual, but not group, signatures. Proceeding R. Soc. B 289, 20220548.
61. Rhodes, J.S., Cutler, A., and Moon, K.R. (2023). Geometry- and Accuracy-Preserving Random Forest Proximities. IEEE Trans. Pattern Anal. Mach. Intell. 10.1109/TPAMI.2023.3263774.