Introduction

Vocal repertoires provide essential information to the study of how communication systems evolve (Maynard Smith and Harper 2003). For example, studies of nonhuman primate vocal communication have provided valuable contributions to the debate about the basis for the evolution of language in humans (Dunbar 2009). Nonhuman primate vocal repertoire size correlates with time spent grooming and with group size (McComb and Semple 2005), providing support for the theory that the complexity of human language has gradually evolved with the increase of social complexity (Dunbar 2009). However, comparative studies of repertoire size are often undermined by two factors. First, vocal repertoire data are derived from studies using different methods (McComb and Semple 2005). Second, identification of the signal categories has traditionally relied on human observers’ assessment of differences among vocalizations, and is thus subject to individual criteria. Although multivariate techniques have demonstrated that such categories may be appropriate (Fuller 2014; Gamba and Giacoma 2007; Maretti et al. 2010; Range and Fischer 2004), human assessment of vocalization types may reflect differences perceived by humans but not necessarily by the species (Fuller 2014; Green 1975; Hauser 1996).

New methodologies in the study of acoustic communication allow standardization across large datasets with limited assumptions (Clemins et al. 2006). These methods provide researchers with computer tools for exploring large databases without the disadvantages of subjective a priori classification, and are often referred to as “unsupervised” (Kogan and Margoliash 1997; Stathopoulos et al. 2014; Stowell and Plumbley 2014). Among the many methods (Garcia and Reyes Garcia 2003; Koolagudi et al. 2012), some used for automatic speech recognition, such as dynamic time warping, are increasingly used to investigate animal communication. Dynamic time warping has been useful for the classification of animal sounds in amphibians (Chen et al. 2012), birds (Anderson et al. 1996; Clemins and Johnson 2006; Ranjard and Ross 2008; Tao et al. 2008; Trawicki et al. 2005), marine mammals (Brown and Miller 2007), and primates (Riondato et al. 2013). These methods can be used to investigate the vocal repertoire across populations and species (Mercado and Handel 2012; Ranjard et al. 2010) and improve our ability to make inferences about the evolution of human language (Fedurek and Slocombe 2011). Although unsupervised classification cannot guarantee to classify calls in a way that is meaningful to animals, it does ensure quantitative objective classification (Pozzi et al. 2010).

Owing to their unique evolutionary history, lemurs are important subjects for comparative studies of vocal communication and may provide insights into the selective pressures that may have linked social and vocal complexity (Oda 2009). True lemurs (Eulemur spp.) are conspicuously vocal and their vocal repertoire comprises low-pitched and high-pitched sounds (Gamba and Giacoma 2005; Macedonia and Stanger 1994; Petter and Charles-Dominique 1979). The presence of various call variants and combinations has also been demonstrated qualitatively (Macedonia and Stanger 1994). Previous studies showed that vocal repertoire may differ between species in Eulemur fulvus (Paillette and Petter 1978), E. mongoz (Curtis 1997), E. macaco (Gosset et al. 2001), and E. coronatus (Gamba and Giacoma 2007).

The aim of this study was to investigate objectively the vocal repertoire across Eulemur species to understand whether different species show different repertoire size and vocalization types. We used an algorithm based on dynamic time warping to assess sound similarity (Ranjard et al. 2010). We then applied cluster analysis to identify groups of similar calls. To understand whether vocal repertoire size differs across Eulemur species we applied the same analytical process to datasets for different species, including the brown lemur (E. fulvus), the mongoose lemur (E. mongoz), the black lemur (E. macaco), and the crowned lemur (E. coronatus), whose repertoires were investigated in previous studies. We also analyzed three species that were not included in previous quantitative vocal repertoire studies: the red-bellied lemur (E. rubriventer), the rufous brown lemur (E. rufus), and the blue-eyed black lemur (E. flavifrons). Qualitative studies of Eulemur species have shown a degree of similarity in the acoustic structure of the calls but shed little light on the quantitative evaluation of similarities and differences, and suffered from subjective identification of the call types (Gamba and Giacoma 2005; Macedonia and Stanger 1994). No previous study has combined, to our knowledge, the study of lemurs’ vocal repertoire across different species using a quantitative unsupervised methodology.

We tested whether or not our unsupervised analyses identified the same vocalization types as previously described. Human sound recognition mechanisms are robust against noise changes and integrate many factors, resulting in accurate low-level acoustic classification. Humans can differentiate calls as discrete types when an unsupervised program, and possibly other species, would recognize a single type (Hauser 1996; Lippmann 1997). We, therefore, predicted that unsupervised clustering would find fewer vocalization types than previous studies. We also predicted that more variable vocalization types mask variation at a lower level, as in a clustering analysis of Guinea baboon calls (Papio papio: Maciej et al. 2013). Alternatively, cluster analysis may highlight variants of vocal types showing a particular contextual occurrence and other types that overlap with the a priori classification.

Methods

Subjects, Study Sites, Equipment, Data Collection, and Analysis

The recordings analyzed for the purpose of this study were part of a large collection of lemur sounds at the Department of Life Sciences and Systems Biology, University of Torino. The recordings originate from various recording campaigns focused on lemur vocal behavior that took place between 1999 and 2013. They were recorded in the wild and in captivity. The number of recording campaigns (hereafter corpora) and the number of calls within a corpus vary with species. We considered only calls emitted by adults. Detailed information about the corpora, sampling, data collection, and associated references is given in the Electronic Supplementary Material (ESM) Appendix S1.

Clustering Analyses

To identify independent groupings and to visualize emerging vocal types (Nowicki and Nelson 1990), we clustered vocalizations of each species on the basis of their degree of dissimilarity, as measured by the pairwise comparison using dynamic time warping (Ranjard et al. 2010). Detailed information about the calculation of dissimilarity indices is given in ESM Appendix S1. We used the affinity propagation tool (Frey and Dueck 2007) of the apcluster package in R (Bodenhofer et al. 2011; Hornik 2013). We labeled clusters with the “representative” vocalization (the “exemplar”), which was automatically chosen during the affinity propagation clustering process (see ESM Appendix S2). The cluster analysis used a squared negative Euclidean distance to measure dissimilarity and identify clusters. This clustering algorithm is based on similarities between pairs of data points. Affinity propagation clustering simultaneously considers all the data points as potential cluster centers (exemplars) and then chooses the final centers through an iterative process, after which the corresponding clusters also emerge. Although we did not define the number of clusters or the number of exemplars (Bodenhofer et al. 2011), the preference (p) with which a data point is chosen as a cluster center influences the number of clusters in the final solution. Because affinity propagation clustering does not automatically converge to an optimal clustering solution, we used two external validation procedures. The first validation was based on the q-scanning process (where q corresponds to the sample quantile of p, modified from Wang et al. 2007; see also Bodenhofer et al. 2011). We evaluated the clusters obtained using different preferences using the Adjusted Rand Index (Hubert and Arabie 1985) to assess the stability of successive cluster solutions (Hennig 2007). The second cluster validation procedure was based on the Silhouette Index, which reflects the compactness and separation of clusters in the final solution (Maciej et al. 2013). When ranked and averaged between species both procedures indicated the median of all the similarities between data points to be the optimal value for the preference. We kept all the analysis settings the same across all datasets. We used the calls used as exemplars in the final clustering solution to label the respective clusters.

A Posteriori Evaluation

We evaluated the agreement between the clustering analyses and the a priori classification using the Adjusted Rand Index (Hubert and Arabie 1985; Table I).

Table I Distribution of the vocalizations indicated a priori and as they emerged from the cluster analysis
Full size table

The terminology we use in the description of the polar dendrograms refers to Drout and Smith (2013). Each branch of the polar dendrogram is termed a “branch” or a “clade” while the terminal portion of each clade is called a “leaf.” Two-leaved clades are called “bifolious,” but the number of leaves in a clade is not limited. Although the horizontal orientation of dendrograms is irrelevant, its vertical arrangement is meaningful. The vertical position of the branch points indicates how similar or different they are from each other. Branches departing from the same branch point are most similar and belong to the same “level.” In the polar dendrograms, levels are numbered from the center (root) to the outer ring.

We also ran a stepwise discriminant function analysis (sDFA, IBM SPSS Statistics 21; Lehner 1996) using the acoustic parameters measured (ESM Appendix S3; see Gamba and Giacoma 2007 for details) using Praat (University of Amsterdam; Boersma and Weenink 2014). We used the sDFA to identify the weight of the different parameters contributing to the clustering process, although the acoustic analysis does not necessarily simulate feature extraction during the dynamic time warping. We ran the sDFA with the cluster information as the grouping variable to estimate how the acoustic parameters contributed to the classification of calls using leave-one-out cross-validation.

Results

Vocal Repertoire

The cluster analysis showed variation in both the number of clusters and the distribution of calls across clusters with species (Table I; see ESM Appendix S5). Vocalizations of Eulemur fulvus were grouped into 11 clusters (Fig. 1; Table I). sDFA showed an overall correct classification of 84.2 % (cross-validated) when we used the clusters as the grouping variable. Signal duration (on the first discriminant function) and the first formant (F1, on the second discriminant function) had the highest loads in the model (Table II).

Fig. 1
figure 1

Polar dendrogram (center) showing how vocalizations of Eulemur fulvus cluster together (see ESM Appendix S4 for a detailed description of cluster topology). For each cluster, we show a spectrogram (the horizontal axis represents time; the vertical axis represents frequency) of the exemplar chosen during the affinity propagation process. All spectrograms are generated in Praat with the following parameters: window length: 0.025 s, time range as shown (0.25–2.50 s); frequency range: 0–10,500 Hz; dynamic range: 35–45 dB. The bar indicates 1 s duration. Exceptions are indicated as follows: * for 1.25 s, ** for 1.50 s, *** for 2.50 s. Values in parentheses indicate the percentage of the exemplar’s vocalization type in a cluster. Additional information is given in ESM Appendixes S4 to S6.

Full size image
Table II Stepwise discriminant analysis results for the seven Eulemur species
Full size table

Vocalizations of Eulemur rufus grouped into 10 clusters (Fig. 2; Table I). sDFA showed an overall correct classification of 94.7 % (cross-validated) when we used the clusters as the grouping variable. Signal duration (on the first discriminant function) and minimum fundamental frequency (MinF0, on the second discriminant function) had the highest loads in the model (Table II).

Fig. 2
figure 2

Polar dendrogram (center) showing how vocalizations of Eulemur rufus cluster together (see ESM Appendix S4). For each cluster, we show a spectrogram of the exemplar chosen during the affinity propagation process. All spectrograms are generated in Praat with the following parameters: window length: 0.025 s, time range as shown (0.25–2.00 s); frequency range: 0–10,500 Hz; dynamic range: 35–45 dB. The bar indicates 1 s duration. Exceptions are indicated as follows: * for 1.25 s, ** for 1.75 s, *** for 2.00 s. Values in parentheses indicate the percentage of the exemplar’s vocalization type in a cluster. Additional information is given in ESM Appendixes S4 to S6.

Full size image

Vocalizations of Eulemur rubriventer grouped into 14 clusters (Fig. 3; Table I). sDFA showed a correct classification of 73.5 % (cross-validated) when we used the clusters as the grouping variable. Signal duration (on the first discriminant function) and the second formant (F2, on the second discriminant function) had the highest loads in the model (Table II).

Fig. 3
figure 3

Polar dendrogram (center) showing how vocalizations of Eulemur rubriventer cluster together (see ESM Appendix S4). For each cluster, we show a spectrogram of the exemplar chosen during the affinity propagation process. All spectrograms are generated in Praat with the following parameters: window length: 0.025 s, time range as shown (0.25–0.75 s); frequency range: 0–10,500 Hz; dynamic range: 35–45 dB. The bar indicates 1 s duration. Values in parentheses indicate the percentage of the exemplar’s vocalization type in a cluster. Additional information is given in ESM Appendixes S4 to S6.

Full size image

Vocalizations of Eulemur mongoz grouped into nine clusters (Fig. 4; Table I). sDFA showed a correct classification of 69.2 % (cross-validated) when we used the clusters as the grouping variable. Signal duration and the third formant (F3) showed the highest loading values on the first and the second discriminant functions respectively (Table II).

Fig. 4
figure 4

Polar dendrogram (center) showing how vocalizations of Eulemur mongoz cluster together (see ESM Appendix S4). For each cluster, we show a spectrogram of the exemplar chosen during the affinity propagation process. All spectrograms are generated in Praat with the following parameters: window length: 0.025 s, time range as shown (0.25–1.25 s); frequency range: 0–10,500 Hz; dynamic range: 35–45 dB. The bar indicates 1 s duration. Exceptions are indicated as * for 1.25 s. Values in parentheses indicate the percentage of the exemplar’s vocalization type in a cluster. Additional information is given in ESM Appendixes S4 to S6.

Full size image

Vocalizations of Eulemur coronatus grouped into 13 clusters (Fig. 5; Table I). sDFA showed a correct classification of 83.4 % (cross-validated) when we used the clusters as the grouping variable. Signal duration (on the first discriminant function) and the first formant (F1, on the second discriminant function) had the highest loads in the model (Table II).

Fig. 5
figure 5

Polar dendrogram (center) showing how vocalizations of Eulemur coronatus cluster together (see ESM Appendix S4). For each cluster, we show a spectrogram of the exemplar chosen during the affinity propagation process. All spectrograms are generated in Praat with the following parameters: window length: 0.025 s, time range as shown (0.25–1.00 s); frequency range: 0–10,500 Hz; dynamic range: 35–45 dB. The bar indicates 1 s duration. Values in parentheses indicate the percentage of the exemplar’s vocalization type in a cluster. Additional information is given in ESM Appendixes S4 to S6.

Full size image

Vocalizations of Eulemur flavifrons grouped into 10 clusters (Fig. 6; Table I). sDFA showed a correct classification of 71.4 % (cross-validated) when we used the clusters as the grouping variable. Signal duration and the first formant had the highest loads on the first two discriminant functions (Table II).

Fig. 6
figure 6

Polar dendrogram (center) showing how vocalizations of Eulemur flavifrons cluster together (see ESM Appendix S4). For each cluster, we show a spectrogram of the exemplar chosen during the affinity propagation process. All spectrograms are generated in Praat with the following parameters: window length: 0.025 s, time range as shown (0.25–2.50 s); frequency range: 0–10,500 Hz; dynamic range: 35–45 dB. The bar indicates 1 s duration. Exceptions are indicated as follows: * for 1.25 s, ** for 1.75 s, *** for 2.00 s. Values in parentheses indicate the percentage of the exemplar’s vocalization type in a cluster. Additional information is given in ESM S4 to S6.

Full size image

Vocalizations of Eulemur macaco grouped into 10 clusters (Fig. 7; Table I). sDFA showed a correct classification of 82.0 % when we used the clusters as the grouping variable. Duration and F1 showed strongest correlation with the first two discriminant functions, respectively (Table II).

Fig. 7
figure 7

Polar dendrogram (center) showing how vocalizations of Eulemur macaco cluster together (see ESM Appendix S4). For each cluster, we show a spectrogram of the exemplar chosen during the affinity propagation process. All spectrograms are generated in Praat with the following parameters: window length: 0.025 s, time range as shown (0.25–1.00 s); frequency range: 0–10,500 Hz; dynamic range: 35–45 dB. The bar indicates 1 s duration. Values in parentheses indicate the percentage of the exemplar’s vocalization type in a cluster. Additional information is given in ESM Appendixes S4 to S6.

Full size image

External Cluster Evaluation

The agreement between the a priori classification and the grouping identified by the clustering analysis was relatively low across the species, ranging from 0.18 to 0.32 (Table I).

Discussion

Our approach succeeded in categorizing vocalizations emitted by seven species using dissimilarity indices. Dissimilarity indices have the advantage of being synthetic and convenient but lack the detail of acoustic analysis (Maciej et al. 2013; Riondato et al. 2013). The discriminant model based on measures of temporal and frequency parameters demonstrated that true lemur calls can be assigned to independently derived clusters identified on the basis of dissimilarity indices with a high rate of correct classification. Furthermore, the accuracy achieved is in the range of that found when the combination of pitch and filter features is classified a priori (Gamba 2006; Gamba and Giacoma 2005).

Diversity of the Vocal Repertoire

True lemurs differ remarkably in their social organization and ecology (Mittermeier et al. 2008; Tattersall and Sussman 1998). Thus we predicted differences in their vocal communication signals, in line with previous studies (Macedonia and Stanger 1994; McComb and Semple 2005). Our results support this prediction: we found that different species show different repertoire size and vocalization types. The audio-visual identification of vocal categories varied from a minimum of 7 vocalization types in Eulemur coronatus to 14 types in E. fulvus, E. rubriventer, and E. mongoz. The overall range obtained by the unsupervised analysis was similar, ranging from 9 to 14 clusters. Thus, audio-visual identification and unsupervised classification of vocalization types gave comparable estimates.

Our results support the prediction that average group size influences vocal repertoire size in part. Both audio-visual identification and unsupervised classification of vocalization types provide a repertoire size estimate of 14 calls for Eulemur rubriventer, an estimate that is surprisingly larger than those observed for other species except E. coronatus, which have group sizes of 8.4 (Kappeler and Heymann 1996), whereas E. rubriventer has a mean group size of just 3 (Overdorff 1996) or 3.2 (Kappeler and Heymann 1996). E. mongoz have a similar average group size of 3.0–3.5 (Kappeler and Heymann 1996; Nadhurou et al. 2015) and show a repertoire size of 9 calls. Several authors have suggested a relationship between a species’ social organization and its communication, proposing that an egalitarian social structure or stable social groups may favor diversity in communication signals (Mitani 1996). E. rubriventer is the only species we studied to have a stable, pair-bonded group structure (Tecot 2008). The other species live in one-male, multifemale groups or multimale, multifemale groups (Fuentes 2002). The social organization in E. mongoz varies between populations, and includes both pair bonding and one-male, multifemale groups (Fuentes 2002). The larger distribution of E. rubriventer may also influence the diversity of vocal communication, as may the fact that we included only captive E. rubriventer in the analysis. However, vocal repertoire appears to be consistent across captive, wild-caught individuals (Colombo, unpubl. data), suggesting that other factors may have a stronger effect than the distribution range size. The strong relationships between repertoire size and stable social organization have been proposed for facial expressions (Preuschoft and van Hooff 1995) and the rate of vocal emissions (Mitani 1996), and further studies are needed to clarify whether pair-bonding also “places a selective premium” (Mitani 1996, p. 246) on vocal repertoire size. In support of this proposal, pair-bonding is considered a key factor favoring the convergent evolution of complex singing displays (Geissmann 2000; Torti et al. 2013) in the “singing primates” (Indri indri, Tarsius spp., Presbytis spp., and Hylobates spp.: Haimoff 1986; Indri indri: Bonadonna et al. 2014).

We predicted that the unsupervised procedure would recognize a lower number of vocalization types. This was true for Eulemur fulvus (11 in the unsupervised analysis vs. 14 in the audio-visual a priori assessment), E. mongoz (9 vs. 14), E. rufus (10 vs. 12) and E. macaco (10 vs. 11). The repertoire estimate derived from a previous study of E. macaco (N = 13; Gosset et al. 2001) exceeds both that observed during the reassessment process (N = 10) and the result of the cluster analysis (N = 10). Although the calls in our sample may be incomplete, we suspect that this discrepancy arose due to the different criteria used to assess vocalization types in these studies.

Our prediction that the unsupervised procedure would recognize a lower number of vocalization types was not supported in two cases: Eulemur coronatus (13 unsupervised vs. 7 audio-visual vocal types) and E. mongoz (14 vs. 9). In both cases, the unsupervised procedure recognized more than one type of alarm call. Previous studies of these species estimated a vocal repertoire size of 15 vocalizations for E. mongoz (9 validated using sDFA; Nadhurou et al. 2015) and 10 vocalizations for E. coronatus (all validated using DFA; Gamba and Giacoma 2007). It is clear that different methods led to different estimates, but interesting that, in principle, dynamic time warping allows the identification of vocalization types using a smaller number of calls than sDFA. Whether these differences in vocal repertoire size reflect different arousal states or contexts is an interesting direction for future research.

Cluster vs. A Priori Classification

Agreement between the clustering process and the a priori criteria was low, with values of the Adjusted Rand Index ranging between 0.18 (in Eulemur rubriventer) and 0.32 (in E. coronatus and E. macaco and E. rufus). This supports the prediction that unsupervised clustering of the vocalizations would not find the vocalization types identified in previous studies. However, despite the differences with the a priori classification, the clusters obtained using dynamic time warping–generated dissimilarity indices revealed a remarkable potential for grouping calls on the basis of acoustic measurements of different parameters. Among the parameters, duration showed the heaviest loadings on the first discriminant function. Thus, the mismatching between the a priori classification and cluster analysis is in line with the suggestion that humans tend to recognize as discrete vocal types sounds that may be grouped into a single type when perceived by other species or classified by quantitative analyses (Hauser 1996).

Both duration and formants contributed to the identification of clusters in almost all the species considered. Formants are known to be crucial for the identification of vocalization types (Gamba 2014; Gamba and Giacoma 2007; Giacoma et al. 2011) and have the potential to provide listeners with individual and species-specific cues (Gamba et al. 2012a).

Snorts, clicks, and hoots were not selected as cluster representatives and were often grouped with different vocalization types to form fairly dishomogeneous clusters. This result is consistent across the species and is in line with previous data that suggest that low-pitched calls may be part of a graded system more than discrete emissions (Gamba and Giacoma 2007). Identifiable vocalization types are common, but calls with intermediate acoustic structure may also occur and may be either “oversplit” by human listeners or not recognized as discrete by the unsupervised methodology we adopted. Low-pitched calls of Eulemur (grunts, clicks, grunted hoots, hoots, snorts, and possibly long grunts) are usually classified as contact calls (Gamba and Giacoma 2005, 2007; Gamba et al. 2012a, b; Pflüger and Fichtel 2012; Rendall et al. 2000). These low-pitched signals, especially grunts, are the most frequently emitted call type in Eulemur (Gamba and Giacoma 2005; Gamba et al. 2012a; Pflüger and Fichtel 2012). However, whether acoustic variation in low-pitched signals plays a role in encoding information other than emitter position is still unclear (Pflüger and Fichtel 2012).

The context of call emission is a powerful indicator of their social function and may provide crucial information to the investigation of acoustic structure (Gros-Louis et al. 2008; Rendall et al. 1999). Future studies are necessary to explore the contextual variation of the vocalization types, how the occurrence of vocal signals relates to their acoustic structure, and how this information can be integrated into unsupervised analyses.

Although there was low agreement between cluster analysis and a priori classification, distinct types of grunts and/or grunted hoots emerge in all species. In addition, grunts emitted by Eulemur coronatus are identified as three different types. Long grunts, which are reported to denote contexts of disturbance and potential territorial predation, or are emitted during locomotion (Gamba and Giacoma 2005, 2007; Pflüger and Fichtel 2012), occur in E. mongoz and E. fulvus. Associations between low-pitched calls and tonal calls emerged as distinct clusters (grunt-tonal calls, long grunt-tonal calls) in all species except E. rufus, and have been reported for many species (Macedonia and Stanger 1994).

Our findings support the prediction that variation in particular vocal types may mask variation at a lower level, in agreement with a study of Guinea baboon calls (Maciej et al. 2013). In baboon calls, variation in screams was stronger than for other vocalization types. In five of six Eulemur species, we found that screams represented more than one (usually homogeneous) cluster (E. flavifrons did not emit screams in the same situation in which other species emitted them). In E. fulvus and E. rufus, we identified three clusters of territorial calls, while alarm calls formed three clusters in E. coronatus and five clusters in E. flavifrons. The fact that cluster analysis identified more than one cluster of alarm calls, screams, and territorial calls indicates variability that has not been reported in previous studies (Gamba and Giacoma 2007; Macedonia and Stanger 1994). These results represent an operationally useful indication for future studies, which may link vocal variation with factors such as level of arousal, social interactions, or audience composition (Clay and Zuberbühler 2012; Fichtel and Hammerschmidt 2002; Slocombe and Zuberbühler 2007; Stoeger et al. 2011).

In conclusion, dynamic time warping appears to be a promising method for deepening our knowledge of how lemurs encode information in their vocal signals, and allows the objective identification of vocalization types. We envisage the use of unsupervised classification in different circumstances, including field studies. For example, various researchers report that the classification of calls to be used in playback experiments is particularly challenging. Acoustic analysis may reveal that recorded calls may in fact be different signals (Rendall et al. 1999). Researchers can face the problem of classifying calls in different groups when in the field. In these situations, the unsupervised classification of a small number of calls can be very helpful to provide the investigator with an interpretable quantitative analysis, which may result in improved experimental design and aid in the evaluation of the results (Seiler et al. 2013).