Disentangling canid howls across multiple species and subspecies: Structure in a complex communication channel Authors: Arik Kershenbaum, Holly Root-Gutteridge, Bilal Habib, Janice Koler-Matznick, Brian Mitchell, Vicente Palacios, & Sara Waller NOTICE: this is the author’s version of a work that was accepted for publication in Behavioural Processes. Changes resulting from the publishing process, such as peer review, editing, corrections, structural formatting, and other quality control mechanisms may not be reflected in this document. Changes may have been made to this work since it was submitted for publication. A definitive version was subsequently published in Behavioural Processes, [VOL# 124, (March 2016)] DOI# 10.1016/j.beproc.2016.01.006 Kershenbaum, Arik , Holly Root-Gutteridge, Bilal Habib, Janice Koler-Matznick, Brian Mitchell, Vicente Palacios, and Sara Waller. "Disentangling canid howls across multiple species and subspecies: Structure in a complex communication channel." Behavioural Processes 124 (March 2016): 149-157. DOI: 10.1016/j.beproc.2016.01.006. Made available through Montana State University’s ScholarWorks scholarworks.montana.edu Elsevier Editorial System(tm) for Animal Behaviour Manuscript Draft Manuscript Number: Title: Disentangling canid howls across multiple species and subspecies: structure in a complex communication channel Article Type: UK Research paper Corresponding Author: Dr. Arik Kershenbaum, Corresponding Author's Institution: University of Cambridge First Author: Arik Kershenbaum Order of Authors: Arik Kershenbaum; Holly Root-Gutteridge; Bilal Habib; Janice Koler-Matznick; Brian Mitchell; Vicente Palacios; Sara Waller Abstract: Disentangling canid howls across multiple species and subspecies: structure in a complex communication channel Wolves, coyotes, and other canids are members of a diverse genus of top predators of considerable conservation and management interest. Canids show strongly cooperative behaviour, mediated in part by vocal communication. As such, they form an important model for the study of the evolution of human sociality and language. Many species are also the subject of conservation and management challenges, such as the critically endangered red wolf Canis rufus, and the grey wolf C. lupus where it comes into conflict with humans. Canid howls are part of a complex long- range communication channel, used both for territorial defence and group cohesion. Previous studies have shown that howls can encode individual and group identity. However, no comprehensive study has investigated the nature of variation in canid howls across the wide range of species. We analysed a database of over 2,000 howls recorded from 13 different canid species and subspecies. We applied a quantitative similarity measure to compare the modulation pattern in howls from different populations, and then applied an unsupervised clustering algorithm to group the howls into natural units of distinct howl types. We found that different species and subspecies showed markedly different use of howl types, indicating that howl modulation is not arbitrary, but can be used to distinguish one population from another. We give an example of the conservation importance of these findings by comparing the howls of the critically endangered red wolves to those of sympatric coyotes C. latrans, with whom red wolves may hybridise, potentially compromising reintroduced red wolf populations. We believe that quantitative cross-species comparisons such as these can provide important understanding of the nature and use of communication in socially cooperative species, as well as support conservation and management of wolf populations. Suggested Reviewers: Eloise Deaux Macquarie University eloise.deaux@gmail.com "I completed a Bachelor of Science, majoring in Brain, Behaviour and Evolution (Macquarie University, Sydney). Following, I completed an Honours Degree, researching dingo acoustic communication. I am currently doing my PhD Degree, focusing on canid (dingoes and dogs) social behaviours. I am particularly interested in questions relating to the evolution of acoustic and visual signals mediating social interactions. My research also has a strong focus in applications to deal with animal conservation and welfare issues." Simon Townsend University of Zurich simon.townsend@ieu.uzh.ch Research interests Animal vocal communication; acoustics; language evolution; syntax; phonology; social cognition; comparative psychology; animal behaviour; evolutionary anthropology; hormones; statistics Robert Lachlan Queen Mary University of London r.f.lachlan@qmul.ac.uk Research interests:The evolution of cultural communication systemsIn humans, and a disparate few groups of animals, communication systems are culturally transmitted. There is growing evidence that some of these systems, such as human speech and bird song, show deep homologies in their underlying genetics, development and neurobiology. I am interested in exploring why cultural transmission of communication signals evolves, focusing particularly on bird song as a model system. This is an inherently multidisciplinary endeavour, and has involved theoretical evolutionary analyses of the gene-culture coevolution between genes underlying development and cultural song traits; field behavioural experiments investigating why song is learned precisely; bioacoustic analysis of how constraints in song learning evolve; and evolutionary genomic analyses of the genes underlying song learning. Ultimately, I aim to tease apart the different evolutionary processes that have led to the incredibly precise cultural transmission of song that we see in many bird species. Simon Gadbois Dalhousie University sgadbois@dal.ca ANIMAL BEHAVIOUR ~ APPLIED OLFACTORY PROCESSING ~ MAMMALS (CANIDS) AND REPTILES (COLUBRIDS AND EMYDIDS) ~ SNIFFER DOGS FOR WILDLIFE CONSERVATION (SPECIES-AT-RISK) AND BIOMEDICAL APPLICATIONS Opposed Reviewers: Dear editors, We are pleased to submit our manuscript "Disentangling canid howls across multiple species and subspecies: structure in a complex communication channel" for consideration in Animal Behaviour. In our research, we performed a unique quantitative comparison of canid howling behaviour across multiple species and subspecies, using a database of recordings from a large number of sources. Rather than relying on subjective assessment of vocal sounds, we used analytical techniques to quantify a distance metric between pairs of howls, and unsupervised clustering to group howls into types. We showed that different species and subspecies of Canis make use of howl types in very different ways, indicating that howl use is non-arbitrary, and could play an important role in population-level processes, and in subsequent conservation efforts. As an example, we also examined the behaviour of three closely related species, the critically endangered red wolf C. rufus, coyote C. latrans, and eastern timber wolf C. lupus lycaon, and showed that their vocal behaviour may indicate the level of hybridisation between these species. We believe that such quantitative studies across a wide range of species and taxa can play a vital role in understanding the variation in behaviour, the evolution of distinct behaviours, and the conservation and management techniques that could help preserve biodiversity and reduce human- animal conflict. We look forward to your response. Yours, Arik Kershenbaum Cover letter TITLE 1 Disentangling canid howls across multiple species and subspecies: structure in a complex 2 communication channel 3 4 AUTHORS 5 Arik Kershenbaum1, Holly Root-Gutteridge2, Bilal Habib3, Jan Koler-Matznick4, Brian Mitchell5, 6 Vicente Palacios6, Sara Waller7 7 1Department of Zoology, University of Cambridge, UK 8 2Department of Biology, Syracuse University, USA 9 3 Department of Animal Ecology and Conservation Biology, Wildlife Institute of India, Dehradun, 10 India 11 4 The New Guinea Singing Dog Conservation Society, Central Point, OR, USA 12 5 The Rubenstein School of Environment and Natural Resources, University of Vermont, USA 13 6Instituto Cavanilles de Biodiversidad y Biología Evolutiva. University of Valencia, Spain 14 7Montana State University, Bozeman, MT, USA 15 16 CORRESPONDING AUTHOR: Arik Kershenbaum, arik.kershenbaum@gmail.com 17 18 WORD COUNT: 5376 19 20 21 Title Document ABSTRACT 1 2 Wolves, coyotes, and other canids are members of a diverse genus of top predators of considerable 3 conservation and management interest. Canids show strongly cooperative behaviour, mediated in part 4 by vocal communication. As such, they form an important model for the study of the evolution of 5 human sociality and language. Many species are also the subject of conservation and management 6 challenges, such as the critically endangered red wolf Canis rufus, and the grey wolf C. lupus where it 7 comes into conflict with humans. Canid howls are part of a complex long-range communication 8 channel, used both for territorial defence and group cohesion. Previous studies have shown that howls 9 can encode individual and group identity. However, no comprehensive study has investigated the 10 nature of variation in canid howls across the wide range of species. We analysed a database of over 11 2,000 howls recorded from 13 different canid species and subspecies. We applied a quantitative 12 similarity measure to compare the modulation pattern in howls from different populations, and then 13 applied an unsupervised clustering algorithm to group the howls into natural units of distinct howl 14 types. We found that different species and subspecies showed markedly different use of howl types, 15 indicating that howl modulation is not arbitrary, but can be used to distinguish one population from 16 another. We give an example of the conservation importance of these findings by comparing the 17 howls of the critically endangered red wolves to those of sympatric coyotes C. latrans, with whom red 18 wolves may hybridise, potentially compromising reintroduced red wolf populations. We believe that 19 quantitative cross-species comparisons such as these can provide important understanding of the 20 nature and use of communication in socially cooperative species, as well as support conservation and 21 management of wolf populations. 22 Keywords: Bioacoustics, Coyote, Dog, Howling, Jackal, Social communication, Wolf 23 *Manuscript Click here to view linked References INTRODUCTION 24 25 The genus Canis comprises several species and subspecies that share many ecological and 26 behavioural similarities (Bekoff et al. 1981). Most are apex predators, and although some hunt in 27 packs and others alone, all species are strongly social, living in groups ranging in size from a handful 28 of close family members, e.g. coyotes Canis latrans (Bekoff 1977), to large groups of 20 or more 29 animals, e.g. Ethiopian wolves C. simensis (Sillero-Zubiri & Gottelli 1994). For humans, one of the 30 most familiar canid behaviours is the howl, a long-range communication channel thought to play a 31 role both in territorial advertising and in group cohesion (Theberge & Falls 1967; Harrington & Mech 32 1979; Harrington 1987). Howling is most familiar in grey wolves C. lupus (Harrington et al. 2003), 33 but all species in the genus produce howl-like vocalisations in addition to other, shorter range 34 communication, such as barks, yips, and growls (Cohen & Fox 1976). These diverse short-range 35 vocalisations are thought to mediate much of canid social behaviour, such as maintaining dominance 36 relationships, but it has been speculated that howling too plays a role in inter- and intra-group 37 dynamics (Harrington & Mech 1979; Jaeger et al. 1996; Gese 2001). Support for this hypothesis 38 includes recent studies showing that wolves recognise the howl vocalisations of familiar individuals, 39 and show changes in affective behaviour in response to the howls of individuals that have been 40 removed from the group (Mazzini et al. 2013; Palacios et al. 2015). 41 Canids of all species pose a number of management and conservation challenges. As apex predators, 42 canids have a major influence on prey populations, and changes in canid numbers can result in trophic 43 cascades (Elmhagen & Rushton 2007; Beschta & Ripple 2009). Some species, such as the Ethiopian 44 wolf (Sillero-Zubiri & Gottelli 1994) and the red wolf C. rufus (Paradiso & Nowak 1972) are 45 critically endangered, whereas the grey wolf C. lupus is frequently in conflict with human populations 46 due to livestock depredation (Sillero-Zubiri & Laurenson 2001), and golden jackals C. aureus and 47 domestic dogs C. familiaris are considered to be significant reservoirs of rabies (Seimenis 2008; 48 Davlin & VonVille 2012). Management of these issues requires an in-depth understanding of the 49 behavioural ecology of these species and subspecies, which would appear to be incomplete without an 50 understanding of the role of long-range vocal communication. In addition, phylogenetic relationships 51 in the genus as a whole are unclear (Bardeleben et al. 2005; Koepfli et al. 2015), with most 52 component species being capable of producing fertile hybrids, and there is considerable lack of 53 agreement over the status of several grey wolf subspecies and populations (Chambers et al. 2012). As 54 a result, the possible role of vocal behaviour as an isolating factor (or otherwise) between populations 55 is important for the conservation of genetic diversity in subspecies that, while genetically compatible, 56 maintain considerable phenotypic adaptation to their local habitats (Chambers et al. 2012). 57 Partly because of the lack of agreement on the taxonomic status of many canid species and 58 subspecies, and partly for reasons of simplicity, in this paper we will use the term "species" as a 59 shorthand for "species and subspecies". 60 Early studies of canid howling behaviour emphasised qualitative descriptions of howl types 61 (McCarley 1975; Cohen & Fox 1976; Tembrock 1976; Lehner 1978) and overall acoustic 62 characteristics, such as mean fundamental frequency and frequency range, as well as modulation 63 shape measures (Theberge & Falls 1967; Tooze et al. 1990). Multiple variables describing changes in 64 the frequency and amplitude of the howl over time can be used for individual discrimination, among 65 which important discriminative variables are the mean, maximum, and coefficient of variation of the 66 fundamental frequency, and the amplitudes of the various harmonics (Root-Gutteridge et al. 2014a; 67 Root-Gutteridge et al. 2014b). However, there are reasons to consider that information exists in the 68 precise frequency modulation of wolf howls, as well as in simpler acoustic characteristics. Firstly, 69 howls are predominantly narrow-band vocalisations, meaning that most of the acoustic energy is 70 concentrated at a small range of frequencies at any one time. Further, this well-defined frequency 71 varies throughout the course of the howl (Figure 1). This "frequency modulation" is known to be used 72 to encode information in other species with similar vocalisations; particularly bottlenose dolphins 73 Tursiops truncatus (Janik & Slater 1998; Quick & Janik 2012), in which individual identity can be 74 reliably extracted from the frequency modulation patterns of whistles (Kershenbaum et al. 2013). In 75 addition, considerations of signal transmission indicate that long-range communication in an 76 absorptive environment (e.g. forest) would tend to favour narrow-band frequency modulation over 77 other encoding modalities (Henry & Lucas 2010). Therefore, we consider it appropriate to analyse the 78 frequency modulation of canid howls in a similar way to that of dolphin whistles, to test for 79 information content and characteristic differences between species and populations. 80 81 82 Figure 1. Example spectrogram of multiple wolves howling. 83 84 A few studies have examined frequency modulation in canid howls, e.g. in coyotes (Hallberg 2007) 85 and Iberian wolves (Palacios et al. 2007), by defining stereotyped modulation patterns such as, 86 "rising", "step down", and "warble to flat". However, these arbitrary categories may not be perceived 87 as distinct units by the focal animal (Kershenbaum et al. 2014), and are potentially subject to selective 88 bias by researchers focusing on "interesting" spectral patterns. Therefore, a thorough analysis of 89 frequency modulation must include (a) a quantitative measure of howl similarity (Deecke & Janik 90 2006), and (b) an objective method for grouping howls into distinct howl types. The latter requirement 91 is particularly acute, as a quantitative comparison between the vocal behaviours of different 92 populations is problematic if both repertoires include vocalisations of different qualitative types. For 93 example, comparing the howls of one population to the barks of another would be an unproductive 94 effort. Therefore, an alternative paradigm is required that takes into account the partitioning of a vocal 95 repertoire into distinct types. We propose that, where multiple distinct vocalisation types are used 96 with overlapping repertoires between populations, the only meaningful way to compare behaviour is 97 to compare the vocalisation type histograms, rather than compare the individual vocalisations. This 98 approach has also been carried out in previous studies of birdsong syntax (Jin & Kozhevnikov 2011). 99 In essence, we interpret the howl type usage histograms as a "fingerprint" of vocal behaviour. 100 In this work, we define and implement a howl similarity metric, as well as an automated clustering 101 technique, and analyse a large database of over 6,000 howls from 21 different species of canids. We 102 classify these howls into distinct types, and compare the relative use of this global repertoire by 103 different populations, thereby testing for objective differences that distinguish between different 104 species. Although we do not explicitly test for contextual reference in canid howling, our results show 105 the diversity of use of different howl types, raising the possibility that specific types may be more 106 common in some behavioural contexts than others. 107 108 109 METHODS 110 111 We collected a database of canid howling recordings from a wide range of sources. Altogether, we 112 analysed 6,009 howls from 21 distinct species, from 207 sources. Recordings were made both of 113 captive and wild animals. The number of sources for each species varied from one (dingo C. lupus 114 dingo, Tibetan wolf C. l. chanco, and others) to 23 (eastern timber wolf C. l. lycaon). However, we 115 excluded all species with only a single source to avoid confounding individual distinctiveness with 116 species distinctiveness, providing a dataset with 13 distinct species from 131 sources. Of these, 2,005 117 howls were considered to be of sufficient quality for further analysis (no overlapping howls, sufficient 118 signal strength). A breakdown of the recordings is given in Table 1. For each howl, we traced the 119 frequency modulation using a combination of manual and automatic extraction tools (Kershenbaum & 120 Roch 2013; Root-Gutteridge et al. 2014b). Each analysis was reviewed by both AK and HRG for 121 validation. 122 123 Table 1. Number of howls, and number of recording sources (packs) for each of the species in the 124 database. 125 Canid species Common name Reference Number of howls Number of sources C. aureus Golden jackal 28 3 C. latrans Coyote 187 4 C. rufus Red wolf (Chambers et al. 79 4 2012) C. lycaon or C. l. lycaon Eastern wolf (Chambers et al. 2012) 510 20 C. lupus Grey wolf C. l. occidentalis Mackenzie Valley wolf (Chambers et al. 2012) 127 8 C. l. baileyi Mexican wolf (Chambers et al. 2012) 31 2 C. l. arctos Arctic wolf (Chambers et al. 2012) 26 7 C. l. lupus European wolf (Nowak 1995) 65 13 C. l. signatus Iberian wolf (Vilà et al. 1999) 25 3 C. l pallipes Indian wolf (Nowak 1995) 175 7 C. l. lupaster North African wolf (Rueness et al. 2011) 33 5 C. familiaris or C. l. familiaris Domestic dog (as companion animal) 375 53 C. hallstromi or C. l. hallstromi New Guinea singing dog (Koler-Matznick et al. 2003) 344 2 126 127 Once the frequency modulation of the howls had been recorded, we compared every howl pairwise to 128 generate a 2,005 x 2,005 matrix of howl similarity/dissimilarity. We used dynamic time warping 129 (DTW) (Kruskal 1983) to deliver a quantitative metric of this distance (or dissimilarity) between 130 every pair of howls. Dynamic time warping has been widely used for comparing frequency data such 131 as these, particularly in the analysis of dolphin vocalisations (Buck & Tyack 1993; Deecke & Janik 132 2006; Sayigh et al. 2007). The DTW algorithm allows the time points of each sampled frequency 133 measurement to vary freely, until an optimum match between the two curves is achieved. The amount 134 of time-axis distortion necessary to achieve this match is then taken as a measurement of the 135 quantitative difference between the curves. 136 Using the dissimilarity matrix obtained by DTW, we applied the k-means unsupervised clustering 137 algorithm to group the howls into natural clusters based on their similarity. We chose the appropriate 138 number of clusters using a combination of cluster purity, measured as the mean cluster silhouette 139 value (Rousseeuw 1987), and stability using a bootstrap technique; repeatedly applying the clustering 140 to a random subset of 90% of the howls, and measuring similarity of the clustering results using 141 normalised mutual information (Zhong & Ghosh 2005). 142 We then examined the usage of each of the different howl types by the 13 different species. As 143 multiple recordings were obtained from the same individual, or from individuals within the same 144 pack, potential issues of pseudoreplication arise if howls are directly compared to each other; howls 145 from the same individual are likely to be more similar to each other than howls from separate 146 individuals or from different packs. Therefore, rather than analysing howl difference distributions 147 directly, we examined only differences in the use of different howl types, by calculating the 148 proportion of howls that belong to each howl type, for each species. This provides a "fingerprint" of 149 howl type usage, which can then be compared between species. We calculated the sum of squared 150 differences between the howl type histograms of different sources (packs) within each of the 13 151 species, and used an exact test (Fisher 1925) to estimate the significance of the similarity within a 152 species, and between species pairs. We randomised the howl type distributions 105 times within each 153 species to generate a null distribution of sum of squared differences, and calculated the proportion of 154 randomised differences that were less than the measured inter-species difference. We also identified 155 the most common howl type in each species and examined various exemplar howls of this type. To 156 test the ability of the howl type usage fingerprint to identify canid species, we measured the similarity 157 between each source (pack) and mean histograms of each species (with the target source excluded), 158 and recorded which species was most similar to the target source. 159 Finally, we examined more closely the similarity in the howling behaviour of two sympatric species, 160 the red wolf and coyote. Red wolves and coyotes hybridise in the wild, which poses a threat to 161 reintroduction programs for the critically endangered red wolf (Hinton et al. 2013; Gese et al. 2015). 162 We tested for significant differences between the howls of these two species, to determine whether 163 howling behaviour may potentially provide a form of behavioural isolation, or alternatively encourage 164 admixing and introgression. We also compared these two species to a subspecies of grey wolf, the 165 eastern timber wolf C. l. lycaon, whose taxonomic status is unclear, but is considered to be very 166 closely related to C. rufus, if not conspecific (Wilson et al. 2000; Koblmuller et al. 2009; Chambers et 167 al. 2012). We reclustered the DTW data, using only howls from the red wolf, coyote, and eastern 168 wolf. We then repeated the sum of square difference analysis, comparing the red wolf-coyote-eastern 169 difference to a null distribution generated by randomising the order of the histogram of howl types, as 170 well as comparing the histogram fingerprints between sources, as with the full data set. 171 172 173 RESULTS 174 175 Applying multidimensional scaling (Cox & Cox 2000) to the full 2,005 x 2,005 matrix of howl 176 distances found 37 significant dimensions, which were then passed to the k-means clustering 177 algorithm. Analysis of silhouette values in k-means led to 21 distinct clusters (howl types). Figure 2 178 shows the howl distance matrix reduced to two dimensions (for visualisation), with cluster assignment 179 indicated. The clustering appeared robust; 99.3% of all howls were classified with posterior 180 probability > 0.5. Bootstrapping and re-clustering with 80% of the data produced a normalised mutual 181 information in comparison to the full data set of 0.760 ± 0.033, i.e. 76% of the cluster assignment 182 information was retained even when applying the algorithm to a reduced data set. 183 184 185 Figure 2. Multidimensional scaling of the 2,005 x 2,005 howl distance matrix into two dimensions. 186 Each point is a howl, and points closer together are more similar than those further apart. Colours 187 indicate k-means clustering assignment. The size of each point is for ease of visualisation only. 188 189 Within-species comparisons show that for the eastern timber wolf, the domestic dog, the coyote, the 190 red wolf, the North African wolf C.l. lupaster, and the Arctic wolf C. l. arctos, howl type usage was 191 more similar among sources of that species than would be expected by chance (Table 2). This 192 indicates that in these species, the different sites from which recordings were taken showed a species-193 specific pattern of howl type usage. 194 195 Table 2. Exact test of similarity of howl type use within each species. The p-value represents the 196 proportion of randomised trials where the mean difference between sources within a particular species 197 was less than the actual mean difference within the species. Starred values are significant at 5%. 198 Species p Number of sources Golden jackal 0.718 3 Coyote 0.019 * 4 Red wolf 0.007 * 4 Eastern Timber wolf 0.014 * 20 Mackenzie Valley wolf 0.955 8 Mexican wolf 0.891 2 Arctic wolf 0.006 * 7 European wolf 0.237 13 Iberian wolf 0.935 3 Indian wolf 0.144 7 North African wolf <0.001 * 5 Domestic dog 0.003 * 53 New Guinea singing dog 0.899 2 199 200 The use of each howl type, adjusted for overall howl use frequency, for each of the species show 201 species-specific fingerprints (Figure 3). The red wolf and coyote share howl type 3 as the most 202 common; the European C. l. lupus and Iberian C. l. signatus wolves share type 5; and the Mackenzie 203 Valley C. l. occidentalis, Indian C. l. pallipes, and Mexican C. l. baileyi wolves share type 7. Each 204 other species has a distinct call type that is most commonly used, relative to its overall usage in the 205 sample database. Apart from these distinctive howl types, the different species have different 206 repertoire diversities, with for instance the North African wolf making use of many fewer howl types 207 than the golden jackal, despite being represented by a similar overall number of sources and howls 208 (Figure 3). One qualitative trend noticeable from the exemplar howls (chosen as those nearest to the 209 cluster centroid) is that the smaller species (red wolf, coyote, New Guinea singing dog, domestic dog, 210 golden jackal) favoured howls that ended with a sharp drop in frequency, whereas larger species 211 (arctic wolf, eastern timber wolf, European wolf, Mackenzie Valley wolf) used howls with much less 212 frequency modulation, particularly at the end of the howl (Figure 4), although this may be an artefact 213 of the lower fundamental frequency used by larger species. 214 215 216 Figure 3. Howl use histograms for each of the 13 species, showing the relative use of each of the 21 217 howl types, adjusted for overall howl type frequency. Red bars show the most commonly distinctive 218 howl type for each species, with the index number of that type appearing above each histogram. N 219 indicates the number of howls, and S indicates the number of sources. 220 221 222 Figure 4. Three examples of the howls of the particular howl types identified as characteristic of each 223 species, and represented in Figure 2 by the red bars. Note that the howls within a type are similar in a 224 dynamic time warping sense, although they may vary somewhat in length. 225 226 227 The confusion matrix for the identification of species by source, and the results of the species 228 identification assessment (Table 3) shows that the coyote, Arctic wolf, and North African wolf all 229 were well identified by howl usage fingerprint comparison, with identification of the red wolf and 230 Mackenzie Valley wolf also higher than expected. The New Guinea singing dog C. l. hallstromi, 231 domestic dog, golden jackal, and North African wolf appeared to form a cluster of similar howl usage 232 types, and the coyote and red wolf seem to form a separate cluster, with heavy use of type 15 howls 233 (which only seem to be used by 3 other species, and at very low frequency). 234 235 Table 3a. Classification success by comparing howl type usage histograms as fingerprints. The % 236 correct column indicates how many recording sources (animal packs) were correctly identified as their 237 particular species when compared to all other sources in the database. The Best guess column 238 indicates which species were most frequently identified as the most similar species to the target 239 source. 240 Species % correct Best guess Golden jackal 33.3 Domestic dog Coyote 50 Coyote Red wolf 25 Red wolf, Coyote, Domestic dog, Arctic Eastern Timber 5 Arctic Mackenzie Valley 25 Mackenzie Valley, Indian Mexican 0 Red wolf Arctic 57.1 Arctic European 0 Mackenzie Valley Iberian 0 Eastern Timber , European , Mackenzie Valley Indian 0 Mackenzie Valley North African 40 North African Domestic dog 13.2 North African New Guinea Singing Dog 0 Domestic dog, North African 241 Table 3b. Confusion matrix showing the number of sources identified as each species type. 242 Predicted species RW COY NGSD DD GJ ARC ETIM EUR IBER MV MEX NAFR IND A ct ua l s pe ci es RW 1 1 1 1 COY 1 2 1 NGSD 1 1 DD 3 3 1 7 4 3 8 3 2 2 9 8 GJ 2 1 ARC 1 1 4 1 ETIM 3 5 1 1 3 4 3 EUR 2 2 2 2 3 1 1 IBER 1 1 1 MV 1 1 1 1 2 2 MEX 1 1 NAFR 1 1 1 2 IND 1 1 1 4 243 244 In the reduced analysis of just red wolf, coyote, and eastern timber wolf, there were a total of 776 245 howls, 510 eastern timber wolf, 187 coyote, and 79 red wolf. Applying multidimensional scaling to 246 the full 776 x 776 matrix led to 42 significant dimensions, and 11 k-means clusters. All howls (100%) 247 were classified with posterior probability > 0.5, and bootstrapping followed by reclustering led to a 248 normalised mutual information of 0.706 ± 0.059. With these data (Table 4), the red wolf and coyote 249 also showed significant similarity between the different packs of the same species (p=0.006 and 250 p=0.009 respectively), whereas the eastern timber wolf was only marginally significant (p=0.052). 251 Comparison of the howl type fingerprints (Table 5) showed that the coyote was well identified from 252 most sources (3 out of 4 sources correctly identified), whereas the red wolf and eastern timber wolf 253 were often misclassified one as the other, with the red wolf identified as eastern timber wolf in 2 out 254 of 4 sources, and eastern timber wolf as red wolf in 6 out of 20 sources. Red wolves and coyotes share 255 their most common howl type – type 3 – which is rarely used by timber wolves. Red wolves will often 256 use howl type 6, which coyotes and timber wolves rarely use, and may be intermediate in 257 characteristics between coyote (type 3) and timber wolf (type 11) howls, by being lower in frequency 258 and flatter (Figure 5). 259 260 Table 4. Exact test of similarity of howl type use, reclustered using only data from the three species, 261 red wolf, coyote, and eastern timber wolf. The p-value represents the proportion of randomised trials 262 where the mean difference between sources within a particular species was less than the actual mean 263 difference within the species. Starred values are significant at 5%. 264 Species p Red wolf 0.0060 * Coyote 0.0090 * Eastern Timber wolf 0.0520 265 266 Table 5. Confusion matrix showing the number of sources of red wolf, coyote, and eastern timber 267 wolf identified as each of the three species, from comparing howl type histogram similarity. 268 Predicted Red wolf Coyote Eastern Timber wolf A ct ua l Red wolf 1 1 2 Coyote 1 3 0 Eastern Timber wolf 6 2 12 269 270 271 Figure 5. Examples of coyote howls of type 3 (left), red wolf howls of type 6 (middle) and eastern 272 timber wolf howls of type 11 (right). Type 6 howls are rarely used by coyotes and eastern timber 273 wolves, but commonly used by red wolves, and may represent an intermediate form. 274 275 276 DISCUSSION 277 278 In this study we analysed over 2,000 howls from 13 different species and subspecies belonging to the 279 genus Canis around the world. Using dynamic time warping as a quantitative measure of howl 280 dissimilarity, we applied an objective unsupervised clustering algorithm to group the howls into 281 distinct howl types. The k-means algorithm produced 21 clusters that were stable to bootstrapping, 282 and that probably represent genuine howl type categories, which we define without the need for 283 subjective description of howl characteristics. 284 Each population recorded made different use of these 21 howl types, with many species/subspecies 285 having a particular howl type that was characteristic of that species/subspecies. Within six of the 286 species - the eastern timber wolf, the domestic dog, the coyote, the red wolf, the North African wolf, 287 and the Arctic wolf - a statistically significant similarity existed in their howl type usage. Further, we 288 found that individual populations of five of the species - coyote, red wolf, Arctic wolf, North African 289 wolf, and Mackenzie Valley wolf - could be identified using the howl type histograms of the 290 remaining populations in the data set. 291 In general, we conclude that canid howling is not an arbitrary signal, but possesses species-specific 292 information, which may reflect adaptive and/or neutral processes of isolation. 293 We also performed a more detailed analysis of the howls of three American canids - the red wolf, the 294 coyote, and the eastern timber wolf - because of the conservation importance of hybridisation between 295 the critically endangered red wolf and the coyote, as well as continuing disagreement over the 296 phylogenetic relationship between the red wolf and both the coyote and eastern timber wolf. We 297 found that coyote and red wolf howl type usage differs significantly, which could be a useful tool for 298 managing red wolf conservation in the face of competition from sympatric coyotes. Red wolf howling 299 was similar to that of eastern timber wolves, further complicating the challenge of red wolf 300 introduction both at the southern end of its range (coyotes) and at the northern end (eastern). In 301 contrast, red wolves and coyotes share their most common howl type, whereas red wolves will often 302 use howl type 6, which coyotes and timber wolves almost never use. The intermediate nature of howl 303 type 6 may provide potential evidence of hybridization between these sepecies. 304 We note in passing that the smaller species - the red wolf, domestic dog, New Guinea singing dog, 305 golden jackal - show a greater diversity of howl types than the larger species, and are similar to each 306 other in their howl type usage. We lack sufficient data to examine this further; however, this 307 phenomenon could be due to peculiarity of the habitat or niche of these smaller species, or could be 308 due to a different emphasis on long and short range communication between larger and smaller 309 species, or a different emphasis on the social role of howling. 310 Given the diverse and non-arbitrary nature of howl differences, it is natural to ask whether variations 311 in howl structure reflect referential or context-specific information. Early studies of wolf 312 communication pointed out that different vocalisation types (e.g. howl vs. growl, yelp, etc.) were 313 associated with different behavioural contexts (Cohen & Fox 1976; Tembrock 1976), but stopped 314 short of suggesting that particular features within howls themselves represented certain arousal states 315 or environmental contexts (Theberge & Falls 1967; Lehner 1978). More recent studies have begun to 316 address this question in dingoes (Déaux & Clarke 2013), as well as dogs (Faragó et al. 2014), and 317 there is some evidence that vocal communication may be used in Canis to coordinate hunting activity 318 (Muntz & Patterson 2004). Experimentally, it has been shown that howl modulation patterns convey 319 individual identity, and that animals attend to this information (Palacios et al. 2015). Thus, individual 320 identity in howl structure is more than just an epiphenomenon, and may be of relevance to 321 conservation and management programs (Llaneza et al. 2005; Terry et al. 2005; Brennan et al. 2013; 322 Hansen et al. 2015). Depredation of livestock by coyotes (Knowlton et al. 1999) and wolves (Sillero-323 Zubiri & Laurenson 2001), in particular, is a cause for concern, but attempts to use vocalisation 324 playbacks as active deterrents have largely been unsuccessful (Gable 2010). 325 Our results have shown clear differences in howl structure between populations. Whether populations 326 in geographical proximity represent separate species, subspecies, or otherwise, it seems clear that 327 distinct ecotypes exist. The presence of discrete differences in vocal behaviour suggests that 328 consideration should be given to conservation of populations such as C. rufus and C. lupus lycaon, 329 even if genetic isolation does not exist. Recent studies have shown multiple examples of dialects not 330 just in birdsong (Kroodsma 2004), but also in multiple mammalian taxa including rodents 331 (Slobodchikoff & Coast 1980; Gannon & Lawlor 1989), primates (de la Torre & Snowdon 2009; 332 Thinh et al. 2011; Meyer et al. 2012), and hyraxes (Kershenbaum et al. 2012). Our study adds to 333 recent work showing dialectic differences between the howls of wolves in Europe and North America 334 (Palacios et al. 2007), and fits into an increasingly important trend of understanding the proximal 335 causes and ultimate significance of dialectic variation (Lameira et al. 2010). 336 In the case of the critically endangered red wolf, hybridisation with coyotes represents the largest 337 threat to reintroduced populations (Hinton et al. 2013; Gese et al. 2015). Although howling behaviour 338 has long been identified in Canis as a mechanism for separating competing populations (Harrington & 339 Mech 1979; Jaeger et al. 1996; Gese 2001), and vocal behaviour as a mechanism for genetic isolation 340 in other mammalian taxa (Braune et al. 2008), to our knowledge no studies have addressed the 341 question whether vocal differences can act to reduce interspecific hybridisation in Canis, or may in 342 fact be the result of past hybridisation. Coyotes fail to respond to stimuli of wolf howling (Petroelje et 343 al. 2013); detailed analysis of C. rufus recordings have uncovered non-howl vocalisations that have 344 not been reported in C. latrans (Schneider & Anderson 2011); and the behavioural responses of 345 individual wolves vary according to the familiarity of playback howls (Mazzini et al. 2013; Palacios 346 et al. 2015). All these findings raise the possibility that vocal differences between C. rufus and C. 347 latrans may have conservation significance. Our work adds to this body of evidence, and should 348 encourage further investigation of the possibility of behavioural isolation between these populations. 349 Our study made use of data sources of widely varying size and quality - something inevitable when 350 integrating recordings from around the world and from species of greatly varying abundance. We 351 have endeavoured to minimise the statistical artefacts arising from this imbalance, and have been 352 careful to use the recording source (essentially, a single pack) as the unit of comparison. Some 353 pseudo-replication may remain, as we cannot ensure that the proportion of howls in each type is 354 constant for a species. However, in most cases there are insufficient howls from specific individuals to 355 look at how the pattern varies by individual within species. Despite these statistical limitations, we 356 believe that such broad comparative studies have great value in understanding behaviour across a 357 wider taxonomic basis than just the species, and we hope that this utility compensates somewhat for 358 the patchy nature of the data sources. 359 Automatic clustering using unsupervised algorithms is potentially problematic, as the presence of 360 computer-identified clusters does not guarantee that these elements have cognitive significance for the 361 animals involved. Indeed, we have no mechanistic indication that canids perceive and compare howls 362 in a way similar to our dynamic time warping. To date, what we know is that wolves detect changes in 363 the fundamental frequency of howls outside their natural range of variability, and changes in the 364 frequency modulation pattern of howls (Palacios et al. 2015). However we feel confident that DTW 365 provides a useful comparative tool, because consideration of acoustic propagation would indicate that 366 frequency modulation of howls is likely an important signal channel in long-range communication. 367 Also, we took care to evaluate our clustering results using multiple metrics, and assessing their 368 stability in the face of bootstrapping, to maximise confidence that the howl type partitions did, in fact, 369 represent a division of howls into realistic howl types. 370 This study has involved only correlative analyses, but we believe that this kind of quantitative 371 categorisation of vocalisation types is necessary before carrying out manipulative and playback 372 experiments. Being armed with an objective set of howl types, or a methodology for arriving at such a 373 definition, allows researchers to test the cognitive significance of different howl compositions, and 374 look for potential behavioural correlates, such as territorial advertising and group cohesion. Any 375 experimental work with critically endangered species such as the red wolf can be problematic, but we 376 hope that with a firmer understanding of the vocal behaviour of these animals, it will be possible to 377 design experiments that will benefit the conservation and management of this and other species. 378 Howling is a social communication process that is likely of major importance in the overall behaviour 379 of all canid species. A deeper understanding of their social behaviour is not possible without a 380 framework within which to understand their vocal behaviour. In particular, quantitative and objective 381 assessment of howling is highly preferable to subjective interpretation by humans, who lack the 382 auditory and cognitive instruments of the focal animals. We believe that further experiments using 383 this method may reveal functionally referential elements to the canid howl repertoire, which would be 384 a highly significant finding for two reasons. Firstly, canid conservation and management can benefit 385 from acoustic methods for surveying and assessing population size and health/genetic purity (Llaneza 386 et al. 2005; Brennan et al. 2013), which can be difficult using traditional methods, particularly when 387 snow is absent (Blanco & Cortés 2011). Active acoustic deterrence has also been suggested as a tool 388 in the control of animal movements for mitigating wolf conflict with farmers (Gable 2010), but such 389 techniques cannot be successfully implemented without understanding the message being transmitted. 390 Secondly, the role of vocal communication in mediating social behaviour in canids may contribute to 391 understanding the evolution of human language (Seyfarth & Cheney 2014). To our knowledge, no 392 animal species other than humans possess any form of true language, not even any form of "proto-393 language". Therefore, it has been problematic to explain the evolution of human language as a 394 continual progression from "non-language" to "language", through increasing adaptive advantage at 395 each step (Tomasello 2008). The presence of complex referential communication in species that must 396 cooperate to survive was probably a crucial step in the evolution of language (Jackendoff 1999). 397 Understanding the communication systems of extant social species is essential to understanding the 398 potential evolutionary trajectories to more complex communication that have occurred in the past, 399 eventually leading to human linguistic abilities. Although taxonomically distant from early hominids, 400 wolves and other canids show remarkable parallels with humans in, for example, social behaviour, 401 intelligence, and vocal communication (Miklósi 2014). Further understanding of canid social 402 communication could serve as a model system for the evolution of more complex vocal 403 communication and language. 404 405 406 407 References 408 Bardeleben, C., Moore, R. L. & Wayne, R. K. 2005. A molecular phylogeny of the Canidae based 409 on six nuclear loci. Molecular Phylogenetics and Evolution, 37, 815-831. 410 Bekoff, M. 1977. Canis latrans. Mammalian species, 79, 1-9. 411 Bekoff, M., Diamond, J. & Mitton, J. B. 1981. Life-history patterns and sociality in canids: body 412 size, reproduction, and behavior. Oecologia, 50, 386-390. 413 Beschta, R. L. & Ripple, W. J. 2009. Large predators and trophic cascades in terrestrial ecosystems 414 of the western United States. Biological Conservation, 142, 2401-2414. 415 Blanco, J. C. & Cortés, Y. 2011. Surveying wolves without snow: a critical review of the methods 416 used in Spain. Hystrix, the Italian Journal of Mammalogy, 23, 35-48. 417 Braune, P., Schmidt, S. & Zimmermann, E. 2008. Acoustic divergence in the communication of 418 cryptic species of nocturnal primates (Microcebus ssp.). BMC Biology, 6, 19-7007-6-19. doi: 419 10.1186/1741-7007-6-19. 420 Brennan, A., Cross, P. C., Ausband, D. E., Barbknecht, A. & Creel, S. 2013. Testing automated 421 howling devices in a wintertime wolf survey. Wildlife Society Bulletin, 37, 389-393. 422 Buck, J. R. & Tyack, P. L. 1993. A quantitative measure of similarity for Tursiops truncatus 423 signature whistles. The Journal of the Acoustical Society of America, 94, 2497-2506. 424 Chambers, S. M., Fain, S. R., Fazio, B. & Amaral, M. 2012. An account of the taxonomy of North 425 American wolves from morphological and genetic analyses. North American Fauna, 77, 1-67. 426 Cohen, J. & Fox, M. 1976. Vocalizations in wild canids and possible effects of domestication. 427 Behavioural Processes, 1, 77-92. 428 Cox, T. F. & Cox, M. A. A. 2000. Multidimensional Scaling. Berlin: Springer. 429 Davlin, S. L. & VonVille, H. M. 2012. Canine rabies vaccination and domestic dog population 430 characteristics in the developing world: a systematic review. Vaccine, 30, 3492-3502. 431 de la Torre, S. & Snowdon, C. T. 2009. Dialects in pygmy marmosets? Population variation in call 432 structure. American Journal of Primatology, 71, 333-342. doi: 10.1002/ajp.20657. 433 Déaux, É C. & Clarke, J. A. 2013. Dingo (Canis lupus dingo) acoustic repertoire: form and contexts. 434 Behaviour, 150, 75-101. 435 Deecke, V. B. & Janik, V. M. 2006. Automated categorization of bioacoustic signals: avoiding 436 perceptual pitfalls. The Journal of the Acoustical Society of America, 119, 645-653. 437 Elmhagen, B. & Rushton, S. P. 2007. Trophic control of mesopredators in terrestrial ecosystems: 438 top-down or bottom-up? Ecology Letters, 10, 197-206. 439 Faragó, T., Townsend, S. & Range, F. 2014. The information content of wolf (and dog) social 440 communication. In: Biocommunication of Animals (Ed. by G. Witzany), pp. 41-62. Springer. 441 Fisher, R. A. 1925. Statistical Methods for Research Workers. Edinburgh: Genesis Publishing Pvt 442 Ltd. 443 Gable, E. 2010. Wolves: 'Howl boxes' could provide new tool to reduce conflicts. 444 http://www.eenews.net/stories/93834. 445 Gannon, W. L. & Lawlor, T. E. 1989. Variation of the chip vocalization of three species of 446 Townsend chipmunks (genus Eutamias). Journal of Mammalogy, 70, 740-753. 447 Gese, E. M. 2001. Territorial defense by coyotes (Canis latrans) in Yellowstone National Park, 448 Wyoming: who, how, where, when, and why. Canadian Journal of Zoology, 79, 980-987. 449 Gese, E. M., Knowlton, F. F., Adams, J. R., Beck, K., Fuller, T. K., Murray, D. L., Steury, T. D., 450 Stoskopf, M. K., Waddell, W. T. & Waits, L. P. 2015. Managing hybridization of a recovering 451 endangered species: The red wolf Canis rufus as a case study. Current Zoology, 61, 191-205. 452 Hallberg, K. I. 2007. Information in a long-distance vocal signal: Chorus howling in the coyote 453 (Canis latrans). PhD thesis, The Ohio State University. 454 Hansen, S. J., Frair, J. L., Underwood, H. B. & Gibbs, J. P. 2015. Pairing call-response surveys 455 and distance sampling for a mammalian carnivore. The Journal of Wildlife Management, 79, 662-671. 456 Harrington, F. H., Asa, C. S., Mech, L. & Boitani, L. 2003. Wolf communication. In: Wolves: 457 Behavior, Ecology, and Conservation (Ed. by L. D. Mech & L. Boitani), pp. 66-103. 458 Chicago:University of Chicago Press. 459 Harrington, F. H. 1987. Aggressive howling in wolves. Animal Behaviour, 35, 7-12. 460 Harrington, F. H. & Mech, L. D. 1979. Wolf howling and its role in territory maintenance. 461 Behaviour, 68, 207-249. 462 Henry, K. S. & Lucas, J. R. 2010. Habitat-related differences in the frequency selectivity of auditory 463 filters in songbirds. Functional Ecology, 24, 614-624. 464 Hinton, J. W., Chamberlain, M. J. & Rabon, D. R. 2013. Red Wolf (Canis rufus) Recovery: A 465 Review with Suggestions for Future Research. Animals, 3, 722-744. 466 Jackendoff, R. 1999. Possible stages in the evolution of the language capacity. Trends in Cognitive 467 Sciences, 3, 272-279. 468 Jaeger, M. M., Pandit, R. K. & Haque, E. 1996. Seasonal differences in territorial behavior by 469 golden jackals in Bangladesh: howling versus confrontation. Journal of Mammalogy, 77, 768-775. 470 Janik, V. M. & Slater, P. J. B. 1998. Context-specific use suggests that bottlenose dolphin signature 471 whistles are cohesion calls. Animal Behaviour, 56, 829-838. 472 Jin, D. Z. & Kozhevnikov, A. A. 2011. A compact statistical model of the song syntax in Bengalese 473 finch. PLoS Computational Biology, 7, e1001108. 474 Kershenbaum, A. & Roch, M. A. 2013. An image processing based paradigm for the extraction of 475 tonal sounds in cetacean communications. Journal of the Acoustical Society of America, 134, 4435-476 4445. 477 Kershenbaum, A., Sayigh, L. S. & Janik, V. M. 2013. The encoding of individual identity in 478 dolphin signature whistles: how much information is needed? PLoS One, 8, e77671. doi: 479 10.1371/journal.pone.0077671. 480 Kershenbaum, A., Ilany, A., Blaustein, L. & Geffen, E. 2012. Syntactic structure and geographical 481 dialects in the songs of male rock hyraxes. Proceedings of the Royal Society B: Biological Sciences, 482 279, 2974-2981. 483 Kershenbaum, A., Blumstein, D. T., Roch, M. A., Akçay, Ç, Backus, G., Bee, M. A., et al. 2014. 484 Acoustic sequences in non-human animals: a tutorial review and prospectus. Biological Reviews, 485 (online before print). 486 Knowlton, F. F., Gese, E. M. & Jaeger, M. M. 1999. Coyote depredation control: an interface 487 between biology and management. Journal of Range Management, 52, 398-412. 488 Koblmuller, S., Nord, M., Wayne, R. K. & Leonard, J. A. 2009. Origin and status of the Great 489 Lakes wolf. Molecular Ecology, 18, 2313-2326. 490 Koepfli, K., Pollinger, J., Godinho, R., Robinson, J., Lea, A., Hendricks, S., Schweizer, R. M., 491 Thalmann, O., Silva, P. & Fan, Z. 2015. Genome-wide Evidence Reveals that African and Eurasian 492 Golden Jackals Are Distinct Species. Current Biology, (online before print). 493 Koler-Matznick, J., Brisbin, I. L., Feinstein, M. & Bulmer, S. 2003. An updated description of the 494 New Guinea singing dog (Canis hallstromi, Troughton 1957). Journal of Zoology, 261, 109-118. 495 Kroodsma, D. 2004. The diversity and plasticity of birdsong. In: Nature’s Music: The Science of 496 Birdsong (Ed. by P. R. Marler & H. Slabbekoorn), San Diego, CA:Elsevier Academic Press. 497 Kruskal, J. B. 1983. An overview of sequence comparison: time warps, string edits, and 498 macromolecules. SIAM Review, 25, 201-237. 499 Lameira, A., Delgado, R. & Wich, S. 2010. Review of geographic variation in terrestrial mammalian 500 acoustic signals: Human speech variation in a comparative perspective. Journal of Evolutionary 501 Psychology, 8, 309-332. 502 Lehner, P. N. 1978. Coyote vocalizations: a lexicon and comparisons with other canids. Animal 503 Behaviour, 26, 712-722. 504 Llaneza, L., Ordiz, A., Palacios, V. & Uzal, A. 2005. Monitoring wolf populations using howling 505 points combined with sign survey transects. Wildlife Biology in Practice, 1, 108-117. 506 Mazzini, F., Townsend, S. W., Virányi, Z. & Range, F. 2013. Wolf howling is mediated by 507 relationship quality rather than underlying emotional stress. Current Biology, 23, 1677-1680. 508 McCarley, H. 1975. Long-distance vocalizations of coyotes (Canis latrans). Journal of Mammalogy, 509 56, 847-856. 510 Meyer, D., Hodges, J. K., Rinaldi, D., Wijaya, A., Roos, C. & Hammerschmidt, K. 2012. 511 Acoustic structure of male loud-calls support molecular phylogeny of Sumatran and Javanese leaf 512 monkeys (genus Presbytis). BMC Evolutionary Biology, 12, 16-2148-12-16. doi: 10.1186/1471-2148-513 12-16. 514 Miklósi, Á. 2014. Dog Behaviour, Evolution, and Cognition. Oxford: Oxford University Press. 515 Muntz, E. M. & Patterson, B. R. 2004. Evidence for the use of vocalization to coordinate the killing 516 of a white-tailed deer, Odocoileus virginianus, by coyotes, Canis latrans. The Canadian Field-517 Naturalist, 118, 278-280. 518 Nowak, R. 1995. Another look at wolf taxonomy. Ecology and conservation of wolves in a changing 519 world, 375, 357-389. 520 Palacios, V., Font, E., Márquez, R. & Carazo, P. 2015. Recognition of familiarity on the basis of 521 howls: a playback experiment in a captive group of wolves. Behaviour, 152, 593-614. 522 Palacios, V., Font, E. & Márquez, R. 2007. Iberian wolf howls: acoustic structure, individual 523 variation, and a comparison with North American populations. Journal of Mammalogy, 88, 606-613. 524 Paradiso, J. L. & Nowak, R. M. 1972. Canis rufus. Mammalian species, 22, 1-4. 525 Petroelje, T. R., Belant, J. L. & Beyer Jr, D. E. 2013. Factors affecting the elicitation of vocal 526 responses from coyotes Canis latrans. Wildlife Biology, 19, 41-47. 527 Quick, N. J. & Janik, V. M. 2012. Bottlenose dolphins exchange signature whistles when meeting at 528 sea. Proceedings of the Royal Society B: Biological Sciences, 279, 2539-2545. 529 Root-Gutteridge, H., Bencsik, M., Chebli, M., Gentle, L. K., Terrell-Nield, C., Bourit, A. & 530 Yarnell, R. W. 2014a. Identifying individual wild Eastern grey wolves (Canis lupus lycaon) using 531 fundamental frequency and amplitude of howls. Bioacoustics, 23, 55-66. 532 Root-Gutteridge, H., Bencsik, M., Chebli, M., Gentle, L. K., Terrell-Nield, C., Bourit, A. & 533 Yarnell, R. W. 2014b. Improving individual identification in captive Eastern grey wolves (Canis 534 lupus lycaon) using the time course of howl amplitudes. Bioacoustics, 23, 39-53. 535 Rousseeuw, P. J. 1987. Silhouettes: a graphical aid to the interpretation and validation of cluster 536 analysis. Journal of Computational and Applied Mathematics, 20, 53-65. 537 Rueness, E. K., Asmyhr, M. G., Sillero-Zubiri, C., Macdonald, D. W., Bekele, A., Atickem, A. & 538 Stenseth, N. C. 2011. The cryptic African wolf: Canis aureus lupaster is not a golden jackal and is 539 not endemic to Egypt. PloS one, 6, e16385. doi: 10.1371/journal.pone.0016385. 540 Sayigh, L. S., Esch, H. C., Wells, R. S. & Janik, V. M. 2007. Facts about signature whistles of 541 bottlenose dolphins, Tursiops truncatus. Animal Behaviour, 74, 1631-1642. 542 Schneider, J. N. & Anderson, R. E. 2011. Tonal vocalizations in the red wolf (Canis rufus): 543 potential functions of nonlinear sound production. The Journal of the Acoustical Society of America, 544 130, 2275-2284. 545 Seimenis, A. 2008. The rabies situation in the Middle East. Developments in Biologicals, 131, 43-53. 546 Seyfarth, R. M. & Cheney, D. L. 2014. The evolution of language from social cognition. Current 547 Opinion in Neurobiology, 28, 5-9. 548 Sillero-Zubiri, C. & Laurenson, M. 2001. Interactions between carnivores and local communities: 549 Conflict or co-existence? In: Carnivore Conservation (Ed. by J. Gittleman, K. Funk, D. Macdonald & 550 R. Wayne), pp. 282-312. Press Syndicate of the University of Cambridge. 551 Sillero-Zubiri, C. & Gottelli, D. 1994. Canis simensis. Mammalian Species, 485, 1-6. 552 Slobodchikoff, C. & Coast, R. 1980. Dialects in the alarm calls of prairie dogs. Behavioral Ecology 553 and Sociobiology, 7, 49-53. doi: 10.1007/BF00302518. 554 Tembrock, G. 1976. Canid vocalizations. Behavioural Processes, 1, 57-75. 555 Terry, A. M., Peake, T. M. & McGregor, P. K. 2005. The role of vocal individuality in 556 conservation. Frontiers in Zoology, 2, 10. 557 Theberge, J. B. & Falls, J. B. 1967. Howling as a means of communication in timber wolves. 558 American Zoologist, 7, 331-338. 559 Thinh, V. N., Hallam, C., Roos, C. & Hammerschmidt, K. 2011. Concordance between vocal and 560 genetic diversity in crested gibbons. BMC Evolutionary Biology, 11, 36. 561 Tomasello, M. 2008. Origins of Human Communication. Cambridge, MA: MIT press. 562 Tooze, Z., Harrington, F. & Fentress, J. 1990. Individually distinct vocalizations in timber wolves, 563 Canis lupus. Animal Behaviour, 40, 723-730. 564 Vilà, C., Amorim, I. R., Leonard, J. A., Posada, D., Castroviejo, J., Petrucci‐Fonseca, F., 565 Crandall, K. A., Ellegren, H. & Wayne, R. K. 1999. Mitochondrial DNA phylogeography and 566 population history of the grey wolf Canis lupus. Molecular Ecology, 8, 2089-2103. 567 Wilson, P. J., Grewal, S., Lawford, I. D., Heal, J. N., Granacki, A. G., Pennock, D., Theberge, J. 568 B., Theberge, M. T., Voigt, D. R. & Waddell, W. 2000. DNA profiles of the eastern Canadian wolf 569 and the red wolf provide evidence for a common evolutionary history independent of the gray wolf. 570 Canadian Journal of Zoology, 78, 2156-2166. 571 Zhong, S. & Ghosh, J. 2005. Generative model-based document clustering: a comparative study. 572 Knowledge and Information Systems, 8, 374-384. 573 574  Wolves and other canids pose major conservation and management challenges  Howling is an important behaviour for all canid species  We analysed over 2,000 howls from 13 species and subspecies of wolves, coyotes, etc  Quantitative methods showed different species use different patterns of howl types  Understanding howl type usage is important both for conservation and management. Highlights (for review) ACKNOWLEDGEMENTS AK thanks Sheela Hira and Knoxville Zoo, Laura Pearson and Asheville Zoo, and Chris Lasher and North Carolina Zoo for access to record their red wolves, and Arthur Povey for assistance coding the data. Recording work was approved by the Institutional Animal Care and Use Committee of the University of Tennessee. AK is supported by a Herchel Smith postdoctoral fellowship at the University of Cambridge. Part of this work was carried out while AK was a Postdoctoral Fellow at the National Institute for Mathematical and Biological Synthesis, an Institute sponsored by the National Science Foundation through NSF Award #DBI-1300426, with additional support from The University of Tennessee, Knoxville. BH is thankful to the State Forest Departments of Himachal Pradesh, J&K, and Maharashtra, and to various zoos in India for permitting us to record howls. HRG is grateful to all who helped with the project: the staff at Colchester Zoo; the Wildwood Trust, the Borror Laboratory of Bioacoustics; the British Library; Lupus Laetus; Polish Mammal Research Institute; Tigress Productions; the BBC Natural History Unit; Longleat Safari Park; Tierstimmen Archiv; Wild Sweden; Wolf Park; the Macaulay Sound Library and the UK Wolf Conservation Trust; and Mike Collins, Teresa Palmer, Monty Sloan, Karl-Heinz Frommolt, Yorgos Iliopoulos, Christine Anhalt, Louise Gentle, Richard Yarnell, Victoria Allison Hughes and Susan Parks. BRM thanks the USDA/APHIS/WS/National Wildlife Research Center for supporting his doctoral research and providing access to captive coyotes; recording work was approved by the NWRC IACUC. SW thanks Mariana Olsen for assistance with data collection, and Yellowstone National Park for permission to record. Acknowledgments