Discriminating between similar alarm calls of contrasting function

Abstract

In a pioneering study of signal design, Marler (Marler 1955 Nature 176, 6–8. (doi:10.1038/176006a0); Marler 1957 Behaviour 11, 13–37. (doi:10.1163/156853956X00066)) argued that the contrasting acoustic design of hawk (seet) and mobbing alarm calls of European passerines reflected their contrasting function. Hawk alarms were high-frequency tones, warning conspecifics to flee but making localization difficult for predators, while mobbing calls were broadband and harsh, allowing easy localization and approach. Contrasting signal features are also consistent with signal detection theory. Discriminating these calls quickly is critical for survival, because hawk alarms require immediate escape. These signals should therefore be selected to be easy to discriminate, reducing the trade-off between immediate fleeing to hawk alarms and unnecessary fleeing to mobbing alarms. Despite these expectations, hawk and mobbing alarm calls of superb fairy-wrens, Malurus cyaneus, are surprisingly similar, raising the question of discriminability without contextual cues. We synthesized these calls on computer, made intermediates and used playbacks to test whether calls can be discriminated acoustically, and if so by what features. We found that birds used multiple acoustic features when discriminating calls, allowing fast discrimination despite overlap in individual parameters. We speculate that the similarity of fairy-wren alarm calls could enhance detectability of both signals, while multiple subtle acoustic differences reduce a trade-off with discriminability.

This article is part of the theme issue ‘Signal detection theory in recognition systems: from evolving models to experimental tests'.

1. Introduction

Many species of birds and mammals give alarm calls to warn conspecifics of danger, and these calls often convey specific information about the threat [1–4]. Individuals often vary the number of notes or acoustic frequency to convey information on the degree of danger [5,6], or give different alarm call types to communicate about different types of threat (‘functionally referential’ alarm calls; [7–9]). Such specific communication is beneficial because it allows listeners to take appropriate action. For example, vervet monkeys, Chlorocebus pygerythrus, look up and flee to cover after hearing alarm calls given to eagles, yet stand bipedally and look down after hearing alarm calls given to snakes [10]. Many bird species also give ‘aerial’ (hawk) alarm calls to warn of flying hawks and ‘mobbing’ alarm calls to perched or terrestrial predators [2,7,11,12]. These calls also prompt contrasting responses. Listeners flee after hearing aerial alarm calls, because flying hawks pose an immediate threat, yet often approach and harass a terrestrial or stationary predator after mobbing calls, because there is little current threat and callers benefit by gaining information and driving it away [13].

Aerial and mobbing alarm calls typically have strikingly different acoustic structures that in part reflect their functions. In a classic study of signal design, Marler [11] argued that their contrasting acoustic structures reflected their different functions. His sample of European passerines had aerial alarms that were constant, high-frequency tones of about 7 kHz, with a gradual start and end, making them difficult to localize. These ‘seet’ calls could therefore prompt conspecifics to flee to cover yet conceal the caller's location from the hawk. By contrast, mobbing calls were abrupt, lower-frequency and broadband calls that provide multiple location cues, and allow conspecifics to quickly localize the threat, allowing monitoring and harassment. Subsequent playback experiments confirmed that seet calls are harder for predators to locate than mobbing calls [2,14]. Furthermore, the high frequency of seet calls means that they attenuate quickly and are difficult for birds to hear [15,16], but particularly for hawks, which can have poorer hearing at higher frequencies than their smaller prey [17]. Among a broader range of species, there is great diversity in the specific features of aerial and mobbing alarm calls, with many aerial alarms being frequency modulated or of lower frequency, and many mobbing calls lacking abrupt notes and very broad frequency range [4,18]. Nonetheless, aerial alarms are typically more narrowband than mobbing calls given by the same species, which is likely to reduce localizability or audibility.

Signal detection theory provides an important insight into the challenges facing individuals during communication, including about danger. Animals face the daunting task of both detecting and discriminating signals, often degraded or attenuated during transmission, against a background of environmental noise and different signals from their own and other species [19,20]. These challenges mean that there will at times be overlap in sensory input between any specific signal type and other stimuli, resulting in uncertainty over the presence or type of signal. Signal detection theory suggests that, in consequence, animals adjust their threshold for response to perceptual input to optimize the trade-off between responding when a signal does occur (correct detections), and falsely responding to other stimuli that bear some resemblance (false alarms) [19,21]. Individuals cannot simultaneously maximize correct detections and minimize false alarms because, while having a lower threshold means response to a greater proportion of true signals (correct detections), it has the cost of also increasing responses to other stimuli (false alarms). Similarly, lowering the threshold reduces the risk of not responding to a true signal (missed detections) but at the cost of reducing appropriate lack of response when the signal does not occur (correct rejections). The optimal threshold will depend on the payoffs for these behavioural decisions [21]. In the case of aerial alarm calls, there should be selection for a low threshold for response, resulting in few missed detections but many false alarms. This is because the cost of not responding when there is an alarm, which could be death, is greater than the energetic cost of fleeing unnecessarily. In this situation, animals are likely to be responsive to any stimuli that resemble aerial alarm calls, which Wiley [19,20] calls ‘adaptive gullibility’.

Signal detection theory suggests three ways in which signallers can enhance communication by reducing the receivers' trade-off between correct detections and false alarms [19]. First, a signaller can enhance detectability by increasing the acoustic contrast between a signal and its background and enhance discriminability between signals by making them acoustically distinct. Both detectability and discriminability depend on signal contrast, under control of the caller, as well as the sensory attributes of the receiver. More distinct signals will lead to fewer errors, with both more correct detections and fewer false alarms. This is similar to Darwin's principle of antithesis, in which signals with opposite meaning have contrasting features [22], and provides another explanation for the differences between aerial and mobbing alarm calls [23]. Second, signals are more detectable and discriminable when there is repetition. Third, using a smaller number of signal types increases detectability. Uncertainty about what signals to listen for reduces the ability to detect any one of them, so that it is more difficult to detect (and discriminate) signals from a repertoire when it is uncertain what signal is to be used. This is because animals can only attend to a tiny fraction of sensory input at any one time, so that attention is limited and focus on one aspect of the world reduces attention to another [24]. Given the importance of predation, senders should be selected to produce alarm calls that are acoustically distinct from background sounds and other calls, repeat them and use a small but stereotyped repertoire. These are indeed all common attributes of alarm calls [3].

Here we examine how superb fairy-wrens, Malurus cyaneus, discriminate between conspecific aerial and mobbing alarm calls. These call types provoke contrasting anti-predator behaviours in birds, and in most species are acoustically very distinct, consistent with their function and with principles of signal detection theory. In fairy-wrens, like other species, aerial alarm calls are given in response to flying predatory birds and provoke immediate flight to cover and silence, while mobbing calls are given to perched or terrestrial predators and prompt vigilance, approach and calling [25–29]. However, in contrast with expectation, the two call types are surprisingly similar acoustically (figure 1). Both have peak frequencies of over 7 kHz [25,27], which is similar to or above the frequency of seet alarm calls, both are rapidly frequency modulated and both lack the harsh, broadband structure common among mobbing calls. These similarities are surprising, given the need to flee immediately only to aerial alarm calls. Mobbing call notes appear to differ in two ways from aerial call notes: a lower-frequency ‘hook’ to start and then a slightly declining overall frequency, a ‘slope’. Aerial calls lack these attributes. We analysed acoustic differences between call types and used playback experiments of computer-synthesized aerial, mobbing and intermediate calls to test if birds can discriminate these calls without contextual cues and, if so, by what acoustic features.

Figure 1.

(a) Aerial and (b) mobbing alarm call elements of superb fairy-wrens, each showing waveform (above) and spectrogram (below). Spectrograms were created in Raven Pro 1.5 and used a Blackman window function with a 5.33 ms size, 0.208 ms hop size and 188 Hz frequency grid. (Online version in colour.)

2. Material and Methods

(a). Study site and species

We studied fairy-wrens resident in and adjacent to the 40 ha Australian National Botanic Gardens, Canberra, from September to December 2017. These birds are small (ca 10 g), cooperatively breeding passerines, in the family Maluridae, that forage primarily on the ground [30]. The sexes differ in appearance after gaining adult colours [30]. Their aerial alarm calls signal more urgent danger by repeating elements; birds flee to cover after multi-element calls and remain longer in cover after more elements [26,27]. Mobbing alarm calls are given to terrestrial or stationary threats, including snakes, foxes and perched predatory birds [29,31], and prompt calling, vigilance and approach [25]. As part of another project, territories were mapped and all birds were marked with coloured leg bands [32]. Resident predators of fairy-wrens included predatory birds such as collared sparrowhawks, Accipiter cirrhocephalus, pied currawongs, Strepera graculina, grey butcher birds, Cracticus torquatus and southern boobook owls, Ninox novaeseelandiae [12,33], as well as foxes, Vulpes vulpes, and brown snakes, Pseudonaja textilis (R. Turner 2020, personal communication).

(b). Recording and acoustic analyses

Aerial and mobbing alarm calls were each recorded from 16 different fairy-wren territories using a Marantz PMD661 audio recorder sampling wave files at 48 kHz and 24 bits. Aerial calls were prompted with a gliding model collared sparrowhawk, while mobbing calls were prompted with taxidermic models of perched boobook owls (see electronic supplementary material for details).

We measured 12 acoustic features of alarm call elements in Raven Pro 1.5 [34]. Six features measured frequency: (i) peak frequency (Hz), at which amplitude is greatest; (ii) low frequency (Hz), above which 95% of the energy occurs; (iii) high frequency (Hz), below which 95% of the energy occurs; (iv) bandwidth (Hz), the difference between high and low frequency; (v) slope (Hz), the change in carrier frequency, excluding the introductory hook of mobbing calls; and (vi) frequency irregularity (Hz), the standard deviation of absolute differences in frequency between successive inflection points. There were four temporal features: (vii) rate of frequency modulation (Hz); (viii) duration (ms), measured from the waveform; (ix) centre time (ms), the time at which half the energy has occurred; and (x) time irregularity (ms), the standard deviation in intervals between successive inflection points. We also measured: (xi) aggregate entropy (bits), which ranges from 0 (pure tone) to 1 (white noise); and (xii) power uniformity (dB), the average power after the element is normalized to a constant value, which is greater if the element has a uniform compared with fluctuating power. As well as measuring whole elements, we also measured all acoustic parameters after selecting only the sloping section of mobbing elements after the introductory hook. Electronic supplementary material, figure S1 shows measurement methods.

(c). Call synthesis

We used the graphical synthesizer in Avisoft-SASLab Pro [35] to create synthetic versions of natural elements and intermediate variants (figure 2). Call synthesis allows control of acoustic parameters, permitting tests of the role of specific acoustic features in discrimination and recognition in a variety of species [36,37], including fairy-wrens [38]. We retained natural variability by creating a synthetic version of each of the aerial and mobbing elements with the highest recording quality from each territory. To do so, each element was synthesized based on measurements of the time, frequency and amplitude of each inflection point. We then created six intermediate variants by modifying the three features that appeared to distinguish mobbing and aerial calls. (i) Hook: the introductory ‘hook’ component of each mobbing element was removed and then added to a randomly selected (matched) aerial element, creating ‘hook’ and ‘non-hook’ variants of each element. (ii) Slope: the non-hook component of mobbing elements was ‘flattened’, by setting all the low inflection points to the peak frequency of this component, which always occurred at a low inflection point. High inflection points were then adjusted to retain the original differences between sequential low and high inflection points. (iii) Peak frequency: all inflection points in the mobbing element were raised to make the peak frequency of each mobbing element the same as its randomly selected aerial element. Finally, we created a synthetic version of crimson rosella, Platycercus elegans, piping contact calls as a control for playback of synthesized sound of the same amplitude. Piping calls are unrelated to danger and ignored by fairy-wrens [38]. Acoustic analyses showed that synthetic aerial and mobbing calls were close representations of natural calls, even to the extent of capturing variation among individual elements (electronic supplementary material, figure S2 and tables S1, S2).

Figure 2.

Synthetic alarm call elements, arranged roughly from a normal aerial element, through intermediates, to a normal mobbing element. (a) Aerial; (b) aerial with hook added; (c) mobbing with hook removed, slope flattened and raised to aerial peak frequency; (d) mobbing with hook removed, slope flattened; (e) mobbing with hook removed and raised to aerial peak frequency; (f) mobbing with hook removed; (g) mobbing with slope removed; (h) mobbing. Spectrogram settings are the same as figure 1, but the scale of the figure is different.

(d). Playback experiments

We used a sequence of four experiments to test if birds could discriminate alarm call types, and if so by what acoustic features. To reduce disruption and the risk of habituation, each experiment had a small number of treatments. (i) Experiment 1 tested the adequacy of synthetic calls and whether individuals could discriminate aerial from mobbing calls without any contextual information. The five treatments were natural and synthetic aerial alarms; natural and synthetic mobbing alarms; and synthetic rosella control. The following experiments used only synthetic calls, to test what acoustic features allowed discrimination. (ii) Experiment 2 tested whether the presence of a hook or frequency slope allowed discrimination. The four treatments were aerial call, mobbing call, mobbing call with hook removed and mobbing call with flattened frequency slope. (iii) Experiment 3 further tested the importance of the hook, using two treatments: aerial call and aerial call with hook added. (4) Experiment 4 tested whether a combination of features allowed discrimination. There were four treatments: aerial call; mobbing call with hook removed and flattened slope; mobbing call with hook removed and raised to the same peak frequency as the matched aerial call; and mobbing call with hook removed, flattened slope and raised to the same peak frequency as the matched aerial call. This last treatment removed all mobbing-element features suspected to allow discrimination between aerial and mobbing calls. All alarm calls consisted of two elements given at the same natural interval and natural amplitude (electronic supplementary material, methods).

Each experiment had a matched design, with focal individuals within 16 territories receiving a unique set of all playbacks. All focal individuals were adults, with an equal number of each sex. Only males in full blue plumage were included. Focal individuals on a specific territory could differ between experiments, and they always differed between territories; 22 territories were used over the four experiments, with 16 in any one experiment. Within Experiments 3 and 4, the focal individual was always the same for each treatment, but within Experiments 1 and 2, a bird of the same sex and plumage was occasionally substituted for the focal bird (5/16 and 4/16 territories, respectively). In total, there were 41 focal birds used, including the nine ‘substitutes’. Playbacks were carried out when the observer, with playback equipment, was about 10 m from the focal bird, and the immediate response was scored as: (i) no response; (ii) scanning; and (iii) fearful response, including startling, freezing, or fleeing. Birds usually respond fearfully to aerial alarm calls [4], and fairy-wrens in particular respond fearfully to aerial but not mobbing calls [27,31], so our playbacks provided an assay of whether birds can discriminate rapidly between aerial and mobbing calls. Further details are given in the electronic supplementary material.

(e). Statistical analyses

We used linear mixed models, with restricted maximum likelihood (REML) and with territory as a random effect, to compare aerial and mobbing calls, with and without the hook included in the mobbing call. The analysis was carried out in R 3.4.3 [39] using the nlme package [40]. We used principal components analysis, in R stats, to visualize acoustic differences between aerial and mobbing calls. Acoustic features included were peak frequency, low frequency, high frequency, slope, frequency irregularity, frequency modulation rate, duration and power uniformity; other measures were highly intercorrelated and therefore excluded. Analyses of experiments were based on their matched design and used Cochran's Q tests (Experiments 1, 2 and 4) and exact, two-tailed McNemar tests (Experiment 3, and paired contrasts in the other experiments), which are appropriate for analysis of matched, dichotomous data. Birds almost always responded fearfully or with scanning to playback of all alarm call variants, so for analysis the responses were dichotomized into fearful responses versus other responses (no response or scanning). Analyses used exact2 × 2 and RVAideMemoire packages in R 3.4.3 [41,42]. Mixed models were not used because two of the experiments had a uniform response to one or more treatments, and such models are unreliable when estimated values approach 0 or 1 [43,44]. We also used a generalized linear mixed model analysis (lme4 package in R [45]) of all playback experiments combined, because each gave an incomplete picture yet they had overlapping sets of playback treatments, with some variability in response to all treatments. This analysis also increased sample size for any given acoustic feature and so reduced the problem of type II error, given the smaller sample in any one experiment. The analysis therefore provided the most powerful test of the effects of acoustic features on response to playback. Focal individual and experiment number were random terms, and analysis examined the effect of four acoustic features on response: (i) hook presence; (ii) frequency slope; (iii) peak frequency; and (iv) ‘base’ call type, which identified whether the synthetic variant was based on an aerial or mobbing call. Base call type was included because any acoustic features that were not manipulated, such as frequency modulation rate or frequency irregularity, would remain in the base call and therefore affect response if used in discrimination.

3. Results

(a). Acoustic analysis of natural calls

Principal components analysis showed an overall acoustic separation of aerial and mobbing elements, yet there was overlap in almost all individual acoustic features. PC1 accounted for 57% of the variation and separated the call types, with aerial elements having higher values (figure 3). Higher scores on PC1 were associated particularly with higher frequencies, greater power uniformity, lower frequency irregularity and a flatter frequency slope (electronic supplementary material, table S3). Aerial and mobbing alarm calls also differed in the mean value of almost all of the 12 acoustic features measured, but overlapped in all acoustic properties except that mobbing elements always had a lower low frequency and a greater frequency irregularity (electronic supplementary material, table S4). Furthermore, there was an overlap in all 12 features when comparing aerial calls with the component of mobbing calls after the hook (electronic supplementary material, table S4).

Figure 3.

Principal components plot of the acoustic features of aerial and mobbing alarm call elements. PC1 explained 57% of the variation and higher scores meant higher frequencies, greater power uniformity, lower frequency irregularity and a flatter frequency slope (electronic supplementary material, table S3 shows PC1 and PC2 loadings). N = 32 aerial and 48 mobbing elements (electronic supplementary material). (Online version in colour.)

(b). Playback experiments

Experiment 1 showed that birds responded similarly to synthetic and natural alarm calls, and with appropriate responses to aerial and mobbing calls (figure 4a). Fairy-wrens responded fearfully to natural and synthetic aerial playbacks, whereas they usually responded with scanning but not fearfully to natural and synthetic mobbing call playbacks (Cochran Q, fearful versus other = 40.2, d.f. = 3, p < 0.001; birds never responded to rosella control calls, so they were excluded from this analysis). The lack of any response to rosella control calls contrasted with all other playbacks (no response versus any response: McNemar exact p < 0.001 for all pairwise comparisons with control calls). Pairwise comparisons revealed no difference in response to natural and synthetic versions of aerial (McNemar exact p = 1.0) or natural and synthetic mobbing playbacks (p = 0.25).

Figure 4.

Immediate response to playback of natural and synthetic alarm calls, in (a) Experiment 1, (b) Experiment 2, (c) Experiment 3, and (d) Experiment 4. The response was scored as none, scan or fearful (flee, freeze or startle); n = 16 territories in each experiment, with focal individuals receiving a full set of matched treatments. Experiment 1 also included a control playback of synthetic rosella piping calls, to which no birds responded. All calls are synthetic unless noted; +hook = hook added; –hook = hook removed; flat = frequency slope flattened; a.p.f. = raised to aerial peak frequency.

In Experiment 2, neither the hook nor frequency slope of mobbing alarm calls explained how these calls were discriminated from aerial alarms (figure 4b). Birds responded more fearfully to synthetic aerial alarm calls than to synthetic mobbing calls and their variants (Cochran Q, all four treatments = 29.6, d.f. = 3, p < 0.001). We detected no significant change in response to mobbing calls if the element's hook was removed (McNemar exact p = 0.125), or the frequency slope was flattened (p = 0.625). Therefore, neither of the two features that most obviously distinguish call ‘shape’ significantly explained the ability of birds to discriminate call types.

Experiment 3 confirmed in a different context that the hook component of mobbing calls did not allow full discrimination of alarm call types (figure 4c). We detected no significant change in the number of individuals that responded fearfully when a hook was added to an aerial call (McNemar exact p = 0.219). Although we detected no significant change within either Experiment 2 or Experiment 3, we observed a smaller number of fearful responses to calls with a hook component. Focusing on territories where there was a change in fearfulness between playbacks, 6/7 focal individuals responded more fearfully when the hook was removed from a mobbing call in Experiment 2, while 5/6 responded less fearfully when a hook was added to an aerial call in Experiment 3 (figure 4).

In Experiment 4, birds still responded more fearfully to aerial calls than to mobbing calls from which we removed the obvious distinguishing features (figure 4d; Cochran Q = 14.3, d.f. = 3, p = 0.002). Remarkably, birds could often discriminate mobbing and aerial calls when all three main features of mobbing calls were modified to match aerial calls. Birds still responded less fearfully to mobbing calls after removing the hook, flattening the frequency slope and increasing the peak frequency (McNemar exact p = 0.008).

The combined analysis of all experiments revealed that the presence of the hook and the acoustic base both contributed to discrimination of alarm call types (figure 5). The ‘base’ type identified whether the synthesized call was based on a mobbing or aerial element. The generalized linear mixed model found that the presence of a hook (estimate ± s.e.: −2.17 ± 0.52, z = −4.20, p < 0.001) and having a mobbing base (−3.83 ± 0.78, z = −4.93, p < 0.001) decreased the probability of responding fearfully, but that neither having a frequency slope (−0.46 ± 0.50, z = −0.93, p = 0.36) nor having an aerial peak frequency (−0.49 ± 0.58, z = −0.84, p = 0.40) affected response.

Figure 5.

Predicted probability of responding fearfully to playback of alarm calls depending on their acoustic attributes. Probabilities are based on estimated marginal means of a generalized linear mixed model that included all four playback experiments, using the emmeans package in R. Predictions are separated according to whether the call had ‘aerial features’ (aerial base, no hook, no slope, high peak frequency) or ‘mobbing features’ (mobbing base, hook, declining slope, lower peak frequency). (Online version in colour.)

4. Discussion

Fairy-wrens discriminated immediately between aerial and mobbing alarm calls, in playback experiments that excluded contextual cues and consisted of only two elements. The two alarm call types were acoustically separable overall, yet overlapped in most individual features. Natural and synthetic alarm calls were acoustically similar and birds responded to them in the same way. Birds responded fearfully to aerial calls but usually only with scanning to mobbing calls. Individual playback experiments did not identify a key acoustic attribute that allowed discrimination of call types, suggesting that birds used a combination of features. Analysis of experiments combined showed that birds discriminated between the calls using both the fine acoustic structure (‘base’ properties) and the presence of the introductory ‘hook’ of mobbing call elements. Overall, birds used multiple features to rapidly discriminate aerial from mobbing calls, as needed given the context of immediate danger, and despite overlapping acoustic features. The similarity of these call types might arise to increase detection, yet it does not appear to compromise discrimination.

Fairy-wren aerial and mobbing calls differed statistically overall and in most individual acoustic features, although the difference between call types was less extreme than commonly described for these call types [11,46,47]. The mean differences between these calls were consistent with expectation, with aerial calls being of higher frequency, more narrowband and more tone-like than mobbing calls. These differences should make the calls less easy to localize or to hear at a distance than mobbing calls, and so reduce the risk of eavesdropping by predatory birds. Nonetheless, both call types are short, with rapid frequency modulation, and peak frequencies at or above those described for ‘seet’ aerial calls in other passerines [46,47]. Furthermore, most acoustic features overlapped in values between call types, with all overlapping if excluding the mobbing call's introductory hook. Therefore, fairy-wrens face a potentially difficult perceptual task of discriminating these calls quickly, particularly in natural situations with variable background noise, attenuation and degradation.

Fairy-wrens rapidly discriminated and responded appropriately to aerial and mobbing calls, even with only two-element calls that lasted on average less than 0.5 s, and with no contextual cues from predators or conspecifics. The fearful response to aerial calls, yet scanning to mobbing calls, is appropriate to the threat and consistent with previous observations and experiments that have separately studied responses to these two call types [25–27]. The combination of acoustically different calls given to discrete types of threat, and appropriate responses according to call type, confirms that fairy-wrens have functionally referential alarm calls. They therefore join a small number of bird species for which this has been shown experimentally [7,8,12].

Experiments with modified synthetic calls showed that fairy-wrens used both the mobbing ‘hook’ and subtle ‘base’ features to discriminate between aerial and mobbing calls. Analyses of Experiments 2 and 3 separately did not reveal a significant effect of the hook, although we observed a smaller number of individuals responding fearfully to calls with a hook than the same calls without the hook (figure 4b,c). The two other features tested in Experiments 2 and 4—frequency slope and peak frequency—had no effect on responses. The more powerful, combined analysis of all four experiments revealed, however, that both the hook component and the base type contributed to discrimination. The ‘base’ type was defined by whether the synthetic call was based on an aerial or mobbing element and contributed to discrimination even when frequency slope had been flattened and peak frequency matched (Experiment 4). There must therefore be fine-scale features of aerial and mobbing calls that contribute to discrimination, in addition to the introductory hook.

We suggest that the hook component of the mobbing element helps reduce false alarms to mobbing calls, given the subtle base differences between call types. Adding an additional component to the mobbing rather than aerial call makes sense because the default response should be to flee if uncertain about the call type (adaptive gullibility), so that missing the extra component would not jeopardize safety. Consistent with this possibility, a simple acoustic prefix allows conspecific song recognition in white-crowned and golden-crowned sparrows, Zonotrichia leucophrys and Zonotrichia atricapilla [48,49]. The hook is less likely to be an ‘alerting’ component, to draw attention to the subtle differences in base features between mobbing and aerial calls, because that would be expected to be included in both call types. The most likely base features allowing discrimination are those whose standard deviations do not overlap between call types. These are frequency bandwidth, frequency irregularity, aggregate entropy and power uniformity, in which mobbing elements on average have greater and more irregular frequency, have more variable amplitude and are less tonal. All would be amenable to experimental assessment. Many other vertebrates use multiple features to recognize or discriminate between acoustic signals (e.g. [50–54]), which––like repetition––may reduce errors in communication [55]. Related advantages of using multiple features are that some attributes are more vulnerable to attenuation and degradation during transmission [16], or more vulnerable to acoustic masking [56,57]. In fairy-wrens, the subtle base features, including rapid temporal and frequency changes, will be vulnerable to degradation [16], increasing the value of the hook component.

If it is critical to rapidly discriminate aerial from mobbing alarm calls, why do fairy-wrens have calls that share so many features, including high frequency and rapid frequency modulation? Fairy-wrens are not constrained to produce such calls, as other calls in their repertoire are of lower frequency, harmonically broadband or lacking rapid frequency modulation (e.g. [28,58]). We consider two possibilities. First, call similarity could increase detectability, following the principle that listening for a diversity of stimuli reduces the probability of detecting any one [24]. Birds listening for high-frequency, rapidly modulated calls would be ‘primed’ to detect both types of alarm calls. Second, the high frequency of mobbing calls may make them less locatable or detectable at a distance than more-typical, lower-frequency, harsh mobbing calls [2]. Individuals could then be less vulnerable to the predator being mobbed, particularly to a perched raptor, or to secondary predators attracted by the commotion of mobbing [13,59].

Overall, our work using synthetic alarm calls reveals that fairy-wrens rapidly discriminate alarm call types despite acoustic similarities, which raises questions about the evolution of call diversity. We suggest that signal detection theory must be considered when assessing the design of alarm signals. Contrasting alarm call design could depend on discriminability, and not simply the independent function of alarm calls and risk of eavesdropping by predators. In addition, we suggest that receiver cognition could constrain call design because greater diversity may reduce detectability.

Ethics

The work was carried out under permit A2015/67 from the Australian National University Ethics Committee.

Data accessibility

Key data are presented in figures and tables in the main text and electronic supplementary material. All raw data have been lodged in the Dryad repository: https://dx.doi.org/10.5061/dryad.f4qrfj6sb [60].

Authors' contributions

The authors were both involved in designing the project and writing the manuscript. R.M. conceived the study, and N.T. carried out the fieldwork and did the statistical analyses.

Competing interests

We declare we have no competing interests.

Funding

This study was supported by the Research School of Biology, Australian National University, and Australian Research Council grant no. DP150102632.