Long-distance vocalizations of spotted hyenas contain individual, but not group, signatures

Abstract

In animal societies, identity signals are common, mediate interactions within groups, and allow individuals to discriminate group-mates from out-group competitors. However, individual recognition becomes increasingly challenging as group size increases and as signals must be transmitted over greater distances. Group vocal signatures may evolve when successful in-group/out-group distinctions are at the crux of fitness-relevant decisions, but group signatures alone are insufficient when differentiated within-group relationships are important for decision-making. Spotted hyenas are social carnivores that live in stable clans of less than 125 individuals composed of multiple unrelated matrilines. Clan members cooperate to defend resources and communal territories from neighbouring clans and other mega carnivores; this collective defence is mediated by long-range (up to 5 km range) recruitment vocalizations, called whoops. Here, we use machine learning to determine that spotted hyena whoops contain individual but not group signatures, and that fundamental frequency features which propagate well are critical for individual discrimination. For effective clan-level cooperation, hyenas face the cognitive challenge of remembering and recognizing individual voices at long range. We show that serial redundancy in whoop bouts increases individual classification accuracy and thus extended call bouts used by hyenas probably evolved to overcome the challenges of communicating individual identity at long distance.

Optimising source identification from marmoset vocalisations with hierarchical machine learning classifiers

“…We showed that at large sample sizes, mean precisions and recalls across datasets increased by more than eight percentage points with this hierarchical approach (table 3). We found the accuracies of the hierarchical classifier on the largest balanced datasets to remain satisfactory and higher than most recent studies classifying animal vocalisations using MLAs [62,76,77], even though the number of samples per class of data was lower than those studies. The same was reflected in the performance of the hierarchical classifier on the originally collected calls (supplementary figure 2).…”

“…We showed that at large sample sizes, mean precisions and recalls across datasets increased by more than eight percentage points with this hierarchical approach (table 3). We found the accuracies of the hierarchical classifier on the largest balanced datasets to remain satisfactory and higher than most recent studies classifying animal vocalisations using MLAs [62,76,77],…”

Improving the workflow to crack Small, Unbalanced, Noisy, but Genuine (SUNG) datasets in bioacoustics: The case of bonobo calls

“…A thorough discussion is beyond the scope of this paper, but a more traditional representation based on performing a cepstral analysis at the middle of the call and adding delta and delta-delta coefficients performs worse (See S1 Text for more details). The better performance achieved with the bioacoustic set is consistent with the fact that the bioacoustic features are the cornerstone on which each call type is primarily defined by expert bioacousticians and primatologists (e.g., [87,89,90] for recent perspectives). Finally, the fact that the performance reached with the DCT set is almost as good as with the bioacoustic set is very encouraging: it indicates that a small number of acoustic descriptors succeed in capturing most of the relevant information present in the signal.…”

The value of individual identification in studies of free-living hyenas and aardwolves

“…Acoustic variation among different individuals (signal strength) and consistency of individuals’ acoustic signatures over time (signal stability) are necessary for acoustic identification of individuals (AIID; Linhart et al 2022 ). If the acoustic variation quantified by Lehmann et al ( 2022 ) proves to be consistent over time, as suggested by East and Hofer ( 1991a ), then AIID could become a powerful tool for studying spotted hyenas.…”

“…The most vocal hyaenids, spotted hyenas, may be individually distinguishable not only through visual identification, but possibly through acoustic identification as well (Lehmann et al 2022 ). Spotted hyenas’ loudest vocalizations, called “whoops,” are emitted in bouts and can be heard from up to five kilometres away (Kruuk 1972 ; East and Hofer 1991a ).…”