Network analysis reveals underlying syntactic features in a vocally learnt mammalian display, humpback whale song

Abstract

Vocal communication systems have a set of rules that govern the arrangement of acoustic signals, broadly defined as ‘syntax’. However, there is a limited understanding of potentially shared or analogous rules across vocal displays in different taxa. Recent work on songbirds has investigated syntax using network-based modelling. This technique quantifies features such as connectivity (adjacent signals in a sequence) and recurring patterns. Here, we apply network-based modelling to the complex, hierarchically structured songs of humpback whales (Megaptera novaeangliae) from east Australia. Given the song's annual evolving pattern and the cultural conformity of males within a population, network modelling captured the patterns of multiple song types over 13 consecutive years. Song arrangements in each year displayed clear ‘small-world’ network structure, characterized by clusters of highly connected sounds. Transitions between these connected sounds further suggested a combination of both structural stability and variability. Small-world network structure within humpback songs may facilitate the characteristic and persistent vocal learning observed. Similar small-world structures and transition patterns are found in several birdsong displays, indicating common syntactic patterns among vocal learning in multiple taxa. Understanding the syntactic rules governing vocal displays in multiple, independently evolving lineages may indicate what rules or structural features are important to the evolution of complex communication, including human language.

1. Background

While syntax in birdsong has been extensively studied, comparatively little is known about the rules governing non-human mammalian vocalizations which likely share a closer evolutionary lineage to our own [1]. Humpback whale (Megaptera novaeangliae) song provides a model to understand the evolution of complex vocal communication in mammals. Their song repertoire is large and versatile, analogous to particularly complicated birdsong displays. However, unlike birds, humpback whales produce songs in a stereotyped, nested multi-level hierarchy [2,3]. Individual sounds or ‘units’ are arranged in a stereotyped pattern called a ‘phrase’. Phrases then repeat multiple times to form a ‘theme’. A complete sequence of four to seven themes sung in a particular order comprises a ‘song cycle’, which is typically between 7 and 30 minutes long [3]. All male humpback whales in a population usually conform to one song pattern at any given time, though there are often within-year variants of that pattern. However, songs undergo incremental, progressive ‘evolutionary’ changes to their pattern each year, which singers adopt through social learning [4,5]. Songs in the South Pacific populations additionally undergo periodic cultural ‘revolutions’ in which the entire song pattern is replaced within a single breeding season [5,6]. Better understanding of the structure of humpback whale song and learning mechanisms involved may provide insight into how complex grammar and vocal learning evolved in a mammalian species [1,7].

Recent studies on syntax have begun to examine structural components such as connectivity (i.e. which sounds appear adjacent in a sequence) or transitions between connected sounds [8–13]. For example, if the sequence ‘AB’ occurs frequently, then A and B are highly connected and A transitions to B. For a longer sequence example ‘ABCDEF’, A and C are better connected than A and F because there is a short distance between A and C (one step), but a larger distance between A and F (four steps). Network modelling has emerged as a tool for investigating these structural components. ‘Small-world’ networks have certain elements which are more connected than others and any two elements can be linked within a few steps [14]. These small-world networks are common in some complex birdsong repertoires [8–13]. Additional common features are transition ‘motifs’ and unit usage. Transition motifs describe the frequency with which certain element arrangements (e.g. reoccurring patterns) occur [15] and can be either ‘deterministic’ or ‘non-deterministic’ [8]. Deterministic motifs occur when a particular sound type is only followed by a few specific other sound types, while sound types in non-deterministic arrangements may be followed by a wide variety of other sound types [8,12]. Unit usage quantifies the frequency with which each element occurs within the display. Zipf's law stipulates that word usage in human language has an inverse linear relationship with the rank of each word's use, meaning there are a few common words while most words are relatively rare [16,17]. Animal repertoires often have a convex departure from this linearity [11,18–20], known as a Zipf–Mandelbrot curve. The greater the departure, the more redundancy within a repertoire and less potential information is implied [21]. Unit usage therefore provides a baseline metric for estimating a system's potential capacity for complex communication [21,22].

Repertoire complexity (i.e. size and variety of sounds) has been quantified in humpback whale song [23], but quantitative studies are limited for structural complexity and syntactic rules [23–26]. In the current study, recordings of the east Australian humpback whale song were assessed over 13 consecutive years. We used directed network modelling, which considers unit order, to measure three specific structural features of songs: (i) the overall network structure, (ii) the distribution of unit usage, and (iii) transition patterns. It was hypothesized that, given the structure of humpback whale song and its similarities to complex birdsong, it will contain both small-world network structure and a Zipf–Mandelbrot distribution as these features are common among similar complex and learned vocal displays. Each of the three structural features was first assessed separately per year. Then, each feature was assessed across all years combined to quantify consistent patterns irrespective of song arrangement [8]. Identifying shared syntactic rules in animal vocal learning displays will help determine common vocal learning strategies or evolutionary pathways across multiple taxa [27,28].

2. Methods

(a). Data collection

Recordings were made from 2002 to 2014 at Peregian Beach (26°30′ S, 153°05′ E) and at Point Lookout on North Stradbroke Island (27°43′ S, 153°53′ E), both on the coast of southeast Queensland, Australia. Recordings were made using a variety of methods over this time including moored hydrophone arrays, boat-based recordings, and passive acoustic monitoring. Methods and sample sizes are summarized in electronic supplementary material, table S1.

(b). Song transcriptions and theme classifications

Spectrograms were generated using Raven Pro 1.5 so that recordings could be inspected visually and aurally for quality. Songs of sufficiently high quality (at least 10 decibels (dB) above background noise with no other overlapping singers) were transcribed at the unit level into numerical sequences using an acoustic dictionary of 149 distinct units classified using a self-organizing map in MATLAB (see Allen et al. [29] for method details). Thirty-six complete song cycles from a minimum of six different singers were selected randomly from the available high-quality recordings and transcribed from each year from 2002 to 2014. The exceptions were 2006 (n = 12 song cycles from 2 singers) and 2007 (n = 4 song cycles from 1 singer) due to insufficient high-quality recordings. As recordings were made along the migratory pathway, songs recorded on separate days were assumed to be from different individuals. In total, 412 song cycles were transcribed from 95 singers. Song cycles had an average length of 172 total units (range: 101–228) and an average duration of 7.2 min (range: 4.3–11.1).

Qualitatively assigned themes were verified quantitatively using a weighted Levenshtein similarity index [24,30], with weightings (β = 1) based on acoustic similarity of units [29,31]. Hierarchical cluster analysis then grouped identified themes onto major dendrogram branches based on the similarity of their unit sequence patterns, confirming that theme assignments were robust. All dendrograms had a cophenetic correlation coefficient greater than 0.8, indicating a good representation of the data [32].

(c). Small-world network analysis

One network model for each year was generated based on the 36 song cycles from that year (n = 12 in 2006, n = 4 in 2007), accounting for the consistent population-wide conformity to one general song pattern in each separate year [3,5]. A separate model was generated based on all years combined (n = 412 for 2002–2014). Network analysis was performed using the PajaroLoco software package [33] in Mathematica v. 10.4 [34]. Models were directed networks to account for sequential unit order. Each network model quantified connectivity among units (i.e. units which were adjacent in a sequence) in the respective dataset and compared it to a random Erdös–Renyi network which permuted the same number of units and connections as the observed data [14,35]. This measure, termed the small-world coefficient (S), is calculated following Humphries & Gurney [35] as

$$S=\frac{C/C_{\mathrm{rand}}}{L/L_{\mathrm{rand}}}$$

where C is the clustering coefficient for the study dataset, L is the characteristic path length for the study dataset and C_rand and L_rand are values for C and L calculated for the randomly permuted Erdös–Renyi network. ‘Small-world’ networks are characterized by (i) a small-world coefficient (S) greater than 1 [35] and (ii) clusters of units above a certain degree of modularity, referred to here as ‘network communities’ and defined by having a higher number of connections among each other than with units outside of the cluster [11,14,33,35].

(d). Transition motifs

Transition patterns or ‘motifs’ were based on the sequential arrangements of units. For each year's dataset, the number of different unit types that immediately preceded another unit was counted (P_i) and averaged across all unit types $(\overline{P}=(\mathrm{\Sigma }{P}_i)/N).$ The number of different unit types that immediately followed another unit was also counted (F_i) and averaged across all unit types $(\overline{F}=(\mathrm{\Sigma }{F}_i)/N)$ ([8,12]). These averages were used to calculate transition patterns, representing the frequency with which those patterns occurred across all song variants in a given year. For example, if one male sung ‘AB’ and another sung ‘BC’, the motif calculations would reflect how often A immediately preceded B and how often C immediately followed B across the dataset. Motifs fell into four categories as defined by Sasahara et al. [8] (figure 1):

• 1.Bottleneck—for a given unit type (X), a greater than average number of unit types precedes and a less than average number of unit types follows ($P_x>\overline{P}$ and $F_x<\overline{F}$).
• 2.Hourglass—for a given unit type (X), a greater than average number of unit types both precede and follow ($P_x>\overline{P}$ and $F_x>\overline{F}$).
• 3.Branch—for a given unit type (X), a less than average number of unit types precedes and a greater than average number of unit types follows ($P_x<\overline{P}$ and $F_x>\overline{F}$).
• 4.One-way—for a given unit type (X), a less than average number of unit types both precedes and follows ($P_x<\overline{P}$ and $F_x<\overline{F}$).

Figure 1.

Diagrams representing the four types of transition motifs, adapted from Sasahara et al. [8]: (a) bottleneck, (b) hourglass, (c) branch, and (d) one-way.

‘Deterministic’ motifs (bottleneck and one-way) have fewer than average units following any particular unit type $(F_x<\overline{F})$. ‘Non-deterministic’ motifs (hourglass and branching) have greater than average units following any particular unit type $(P_x<\overline{P})$ [8]. Each of the four transition motifs were counted per song cycle (including all phrase repetitions of all themes) and averaged for each year's dataset separately. All years were then combined and similarly analysed as a single dataset. The proportions of deterministic and non-deterministic transitions were compared to one another within each year and among all years to determine whether their relative proportions were consistent across years.

(e). Unit usage

Units were ranked based on how frequently they occurred in each year's dataset. Rankings and frequency of occurrence were plotted logarithmically, generating a distribution of unit usage for each year [18–20,36,37] to determine the ‘openness’ of the repertoire. A repertoire with a slope of −1 (a ‘Zipf curve’) indicates that a small number of sounds are used frequently and many sounds are used infrequently [10,16,20,37]. A ‘closed’ system has a slope of less than −1 indicating that a few specific units dominate use, whereas an ‘open’ system has a slope between 0 and −1, indicating that many units are used with similar frequency. The same calculation was repeated using all years combined based on the 412 song cycles. Unit usage parameters were calculated using the PajaroLoco software package [33] in Mathematica v. 10.4 [34].

3. Results

(a). Small-world networks

Directed network models for each year's song had clear small-world network structure (i.e. small-world coefficient (S) > 1), with an average S value of 3.0 (range: 1.2–5.1 per year) (figures 2 and 3 and table 1). Therefore, unit types within song sequences clustered into highly connected groups with short distances between unit types. Song arrangements for each year contained an average of 4.7 (range: 3–7, table 1) network communities (i.e. clusters of highly connected unit types) and a short average path length (mean = 2.1, range: 1.8–2.4, table 1), meaning that any pair of unit types only needed a few steps to connect them. A directed network model for all years combined had a small-world coefficient of 2.6, average path length of 1.9, and 12 network communities (table 1). Network communities for each year contained an average of 10 unit types (range: 1–38). However, song themes contained an average of three unit types (range: 1–10), suggesting that network communities did not correspond with song themes (stereotyped patterns of units within a song type). Additionally, there was no difference between small-world coefficients and whether songs changed by evolution or revolution (heteroscedastic t-test, p = 0.28).

Figure 2.

Example of a directed network representation of units for the song sequences for 2002 (S = 3.3, N = 36) with an average amount of small-world structure (average S = 3.0). Units served as the vertices and the transitions between units served as the directed edges (or connections) between vertices. Arrows indicate transition direction between units, and line thickness indicates the frequency of the transitions. There are high amounts of clustering between certain groups of units (network communities) circled in different colours. Units within network communities have more transitions between each other than with units outside their own community. Only a few transitions connect units between separate network communities. Network representations for each year (2002–2014) and for all years combined can be found in electronic supplementary material, figures S1–S14. (Online version in colour.)

Figure 3.

Small-world coefficient (S) values. S values are shown for each year (average S = 3.01) based on all of the song cycles for that year (N = 12 in 2006, N = 4, in 2007, N = 36 per year 2002–2005 and 2008–2014; light grey bars), as well as the coefficient for all of the song cycles (412 for 2002–2014) in all years combined (dark grey bar). The black line marks the threshold for small-world topography (S = 1.0).

Table 1.

Network features for each year and all years combined. Features include song type (whether a song resulted from incremental evolutionary (E) or rapid revolutionary (R) changes identified in Allen et al. [23]), the small-world coefficient (S), average path length (L), clustering coefficient (C), number of network communities (NC), unit repertoire size (Rep), slope of unit usage (Slope), percentage of deterministic motifs (DM), and percentage of non-deterministic motifs (NDM).

year	type	S	L	C	Rep	NC	Slope	DM (%)	NDM (%)
2002	E	3.3	2.4	0.5	39	5	−2.2	70	30
2003	R	1.7	2.1	0.4	28	5	−2.2	57	43
2004	E	4.2	2.2	0.6	57	7	−1.9	63	37
2005	E	3.8	2.4	0.6	56	5	−2.1	55	45
2006	R	2.2	2.1	0.5	40	5	−1.6	50	50
2007	R	1.2	2.3	0.3	21	5	−1.1	62	38
2008	E	2.4	2.1	0.4	55	4	−2.1	58	42
2009	R	2.6	1.8	0.7	44	5	−2.1	57	43
2010	E	3.1	2.2	0.6	65	5	−1.9	57	43
2011	R	3.6	2.2	0.6	67	3	−1.9	61	39
2012	E	3.2	2.1	0.6	73	4	−1.8	53	47
2013	R	5.1	2.0	0.7	48	4	−2.2	71	29
2014	R	2.5	2.0	0.5	41	4	−2.1	54	46
all years	n.a.	2.6	1.9	0.6	142	12	−1.5	56	44

(b). Transition motifs

Deterministic motifs (i.e. few unit types following any particular unit) were more common than non-deterministic motifs (i.e. many unit types following any particular unit) in every year (average of deterministic motifs = 59%, range: 50–71%), except for 2006 which had equal proportions of transitions. When all years were combined into a single analysis, the percentages of deterministic and non-deterministic motifs were 56% and 44%, respectively. Therefore, songs have more motifs which indicate stability than motifs indicating variability.

One-way motifs (where few unit types precede and follow a unit) were the most common in each year (range: 43–63%, electronic supplementary material, table S2). Hourglass motifs, which have many units both preceding and following a unit, were the second most common motif (range: 20–41%). Branch motifs (few units precede a unit, but many units follow) and bottleneck motifs (many units precede a unit, but only a few units follow) were far less common in each year (branch range: 6–20%, bottleneck range: 4–19%). These trends suggest that song sequences primarily contain sequences with either restricted, stable patterns (one-way motifs) or variable patterns (hourglass motifs). This is reinforced by how rarely transitions occurred between stable and variable patterns (bottleneck and branch motifs).

(c). Unit usage

The repertoire size for each year ranged from 21 to 73 (average = 45) unit types per year and a total of 142 unit types over the 13-year study period. Unit types were shared by an average of four separate years (range: 1–11) and an average of 50% of unit types in a given year were shared between at least two song themes (range: 29–73%). The Zipf–Mandelbrot curve was present for each separate year, as well as for all years combined, with a plot of log rank of usage versus log frequency of occurrence displaying a clear convex departure from linearity (figure 4). All linear regressions had negative slopes with an average of −1.9 (range: −1.1 to −2.2), indicating a more ‘closed’ system in which song repertoires were composed of a few (approx. 10% of units) common units, while most units were rarely used. There was no difference between slopes of songs changing via evolutions versus revolutions (heteroscedastic t-test, p = 0.51). Commonly used units near the top of the curve were mainly hourglass units, while most rarely used units near the bottom of the curve were one-way units (figure 4).

Figure 4.

Unit usage distribution for all years combined. The logarithmic distribution of rank of unit use (x-axis) is modelled as a function of the frequency of occurrence (y-axis) for all units used across every song type in every year (N = 412 song cycles). The dashed line represents a linear regression line of best fit. Each data point is labelled according to which transition motif (hourglass, branching, bottleneck, one-way) that unit was categorized as during analysis. Distributions for each year can be found in the electronic supplementary material, figures S1–S14. (Online version in colour.)

4. Discussion

Humpback whale song contained several features indicating a degree of structural complexity and syntax. All analysed song patterns (2002–2014) displayed a small-world network structure. Small-world structure was also present when data were combined into one network model. Changes to small-world structure were not driven by changes to song arrangement since there were no significant structural differences between songs that underwent evolutionary (incremental) versus revolutionary (rapid) changes. Therefore, while song patterns changed to varying degrees each year [5], some degree of small-world structure was always present. This suggests that small-world structure is a consistent feature of general song structure and not of any one particular arrangement.

Unexpectedly, each year's network communities (clusters of highly connected units) did not coincide with that year's identified song themes (the repeating and stereotyped patterns of units). Instead, on average, about half of all unit types in a given year occurred in at least two song themes rather than each song theme containing a unique set of units. Similarly, unit types occurred across multiple years, even in instances where song patterns were unrelated due to a cultural revolution event. Thus, even unrelated song types did not seem to have entirely unique repertoires. Potentially shared units between both song themes and song types from different years could facilitate song learning because even when the song arrangement is novel, some unit types may be familiar.

Unit type arrangement within the network structure showed that common song features were versatile in some aspects, yet restricted in others. This dichotomy stemmed from the prevalence of both one-way and hourglass motifs. One-way motifs indicate that many unit types ‘direct’ song patterns towards a specific sequence, limiting variability. These units were typically found on the network's periphery because they connected with few unit types and are likely to only appear within a single year's song variants. Conversely, hourglass motifs demonstrate the common use of some units as ‘hubs’ or points of high connection that facilitate sequence variability [9]. Their role as hubs placed hourglass units at the centre of the network communities within broader song networks, as they occurred in multiple themes, within-year song variants, or yearly song types. While repertoires contained more one-way unit types, hourglass unit types occurred more frequently within song sequences themselves. Therefore, hourglass units mostly comprise the top of the unit usage curve, while one-way units tended to make up the lower portion of the curve (figure 4). Based on visual inspection of the units classified, hourglass units were primarily low frequency (approx. 300–500 hertz (Hz)), flat, tonal calls such as ‘moans’, ‘groans’, or ‘cries’, while one-way units covered the spectrum of acoustic features. The dominant use of stable one-way and diverse hourglass transitions reflects the combination of stereotypy and variability [8] which is observed in multiple levels of humpback whale songs [3,24]. The presence of these patterns across song variants and unrelated song types further suggests fixed roles in the song structure. The common use of hourglass units may allow incorporation of novelty by individuals, while the range of one-way units support stability in that year's general arrangement which all singers must conform to.

Overall, song structure featured stability and stereotypy by having considerably more deterministic motifs (e.g. one-way and bottleneck) than non-deterministic motifs (e.g. hourglass and branching). Deterministic motifs encourage stereotypy by limiting which units follow them, thus restricting variation [8]. Unit usage (overall and per year) always followed a similar distribution to that of human languages [16]. In other words, each song type contained sequences with a few commonly used unit types, while most units were rare. However, all plots were convex rather than linear, displaying the ‘Zipf–Mandelbrot’ curve that is often observed in animal communication [20,22,36]. This indicates that each song's repertoire contained a high level of redundancy, likely due to its repetitive hierarchical structure. This is consistent with the relatively low estimates of mean unit-level entropy (approx. 1.0 bits) found in other studies across song sequences within both this population [23,26] and Hawaii [25], which indicates predictability in unit arrangements and further supports the presence of stereotypy and redundancy within humpback song. Disproportionately high usage of the few central hourglass units (e.g. ‘moans’, ‘groans’, or ‘barks’) encouraged small variations within the stereotyped song sequences, while the remainder of the units were used in these small variations. Although following Zipf's law is not sufficient evidence of language, it does illustrate that some animal vocal repertoires contain a surprising amount of complexity in their organization [21].

The same syntactic features we have identified in humpback whale song (e.g. small-world structure, deterministic motifs, and redundant unit usage) also occur in the songs of certain songbirds [8–13,20]. Small-world coefficient values in humpback whale songs are comparable to those calculated for birdsongs (S = 1.69–4.70 for seven species, summarized in table 2). Reflecting network structure, deterministic transition motifs also occurred in similar proportions across these species. One-way transitions are the most commonly used motif, followed by hourglass ‘hub’ elements central to small-world structure (table 2). Additionally, the network communities identified here are equivalent to the ‘small-world themes’ identified in some bird songs [11,12]. Such parallels indicate that the network structures of these displays cluster their song elements (units in humpback whales, phrases in birds) in similar ways and encourage stability in their song arrangements. Stability through small-world structures may therefore be taxon-general based on their presence in the vocal learning displays of multiple, taxonomically diverse species.

Table 2.

Network features in seven birdsong displays compared with humpback whale song (current study, averaged over all 13 study years). S, small-world coefficients are given, as well as each of the four transition motifs (one-way, hourglass, bottleneck, branch). All transitions are presented as percentages of the total number of transitions found in those displays.

species	S	deterministic		non-deterministic
		one-way (%)	bottlenecks (%)	hourglass (%)	branches (%)
Western tanagera	2.10	57	0	43	0
Cassin's vireoa	4.70	38	21	26	15
Black-headed grosbeaka	2.70	34	18	25	23
Redthroata	4.10	55	7	32	5
Sage thrashera	2.86	62	10	20	9
California thrashera	2.60	51	3	27	19
California thrasherb	1.69	50	7	32	11
Nightingalec	4.29	35	15	35	14
Humpback whale	3.00	51	8	30	11

^aTaylor & Cody [12].

^bCody et al. [11].

^cWeiss et al. [9].

Small-world structure is also found in human language [22,38], with a relatively small number of key words used often as important syntactic components. Evidence suggests that small-world structure results from the need for language to have ‘optimal navigation’, in which word arrangement can express an intended message efficiently using the smallest number of steps [39,40]. While humpback whale songs do not contain information in the same way as human language, consistent small-world structure is likely to increase song learning efficiency. This may explain how singers learn a song's pattern so quickly, as well as the song's ability to spread through an entire population within a single year or across large spatial scales as shown in the South Pacific [5].

The concept of song structure facilitating learning is reinforced by the common presence of repeating patterns across song types, shown as loops within the networks (figure 2; electronic supplementary material, figure S2). Doublets, triplets, and alternating repetitions were common one-way transitions. Longer bouts of multiple repetitions of the same unit type appeared to involve mainly short duration, low frequency, broadband calls such as ‘grunts’ or ‘croaks’. Several previous studies have found repetitions to be common through fine-scale qualitative analyses [41–43]. Both Guinee & Payne [42] and Payne [43] characterized these repetitions as ‘rhyme-like’, suggesting that they could be used as mnemonic devices to better remember song content. The prevalence of repetitions quantified here supports this suggestion that repetitions may aid in song learning by making songs easier to remember.

5. Conclusion

The current study quantified fine-scale structural complexity and syntactic patterns of humpback whale song. Although similar network modelling has been applied to birdsong [8–13], studies are lacking for mammalian songs or hierarchical displays. By filling these gaps using humpback whale song, the vocally learned displays of multiple species can be compared. The convergence of small-world structure occurs in complex birdsong arrangements [13]; its presence in humpback whale song indicates that this convergence occurs in complex songs across multiple vocal learning species. While these vocal displays may serve different functions within their respective species, they share a common need for individuals to learn them efficiently. If small-world structure does facilitate learning, then it is likely to be an important feature of vocally learned displays. The ability to learn syntactic rules for syllable arrangement is a fundamental component of vocal learning and development in humans [44,45], birds [46], and humpback whales [47]. Studies that quantify syntactic rules across distantly related species are therefore invaluable for understanding the origin and evolution of vocal learning and language.

Ethics

The following ethics approval numbers pertain to data collected for this study (2002–2004) the University of Queensland Office of Research and Postgraduate Studies (2005–2007), the University of Queensland Research and Research Training Division (2008–2011), the University of Queensland Research Management Office (2012–2014): 2002–2003: ZOO/ENT/250/02/USNR/DSTO; 2003–2004: ZOO/ENT/216/03/UNNR/DSTO; 2004–2005: ZOO/ENT/239/04/USNR/DSTO; 2005–2006: SVS/381/05/DSTO & US ONR; 2006–2007: SVS/870/06/DSTO and US ONR; 2007–2008: SVS/203/07/DSTO and US ONR; 2008–2009: SVS/299/08/ACAMMS; 2010–2012: SVS/230/10/(NF); 2012–2013: SVS/403/12/EPSML; 2013–2014: CURTIN/SVS/283/13. Certificates available upon request.

Data accessibility

Data for this manuscript are available on the Dryad Digital Repository: https://doi.org/10.5061/dryad.2bvq83bkv [48].

Authors' contributions

R.A.D. and M.J.N. provided extensive raw song data. E.C.G., M.J.N., and J.A.A. developed the research concept. J.A.A. transcribed all song recordings, developed the methodology, and analysed all data. R.A.D., E.C.G., M.J.N., and J.A.A. interpreted the analyses and wrote the manuscript.

Competing interests

We declare we have no competing interests.

Funding

J.A.A. was funded by an Australian Government Research Training Program Scholarship and the Australian American Association University of Queensland Fellowship. E.C.G. was funded by a Royal Society Newton International Fellowship and a Royal Society University Research Fellowship. HARC was funded by the US Office of Naval Research, the Australian Defence Science and Technology Organisation, and the Australian Marine Mammal Centre. BRAHSS was funded by the E&P Sound and Marine Life Joint Industry Programme and the US Bureau of Ocean Energy Management.