Playback Experiments

Our playback experiments simulated secondary contact between geographically isolated populations and thus provide insight into the degree to which signal divergence in allopatry affects species recognition. For the 72 taxon pairs that are the main focus of the present study, we conducted a mean (± SD) of 12.1 ± 2.8 playback experiments population⁻¹ (range: 5–21; see  Supplemental Material Table S1). Sample sizes of the territories tested are similar for taxon pairs for which we have conducted playback experiments but for which we have not measured acoustic traits (Tables 1–3). Playback experiments followed a standard methodology (McGregor et al. 1992; see also Pegan et al. 2015). Briefly, each experiment measured the behavioral response of a territorial bird to 2 treatments: (1) song from the local population (sympatric treatment) and (2) song from the allopatric population (allopatric treatment). We alternated treatment order between territories; a previous analysis showed that treatment sequence does not influence behavioral response to allopatric song in this dataset (B. G. Freeman et al. personal observation). We used large banks of natural vocalizations in experiments, archived at xeno-canto.org and the Macaulay Library (Cornell Lab of Ornithology), to maximize independence of replicates (n = 6.6 ± 1.7 recordings treatment⁻¹), and we used a single recording in each treatment. We used recordings made within the geographic distribution of the population in question; sympatric recordings came from the vicinity of the region where we conducted playback experiments, and allopatric recordings came from the most geographically proximate portion of the allopatic populations' distribution. For example, for comparisons of Andean taxa found on either side of the Marañon Gap, we used recordings made near this biogeographic barrier (i.e. in southeast Ecuador or northern Peru) and did not use recordings made far from the barrier (even if they are classified as the same subspecies). We used good-quality recordings in playback experiments (most recordings were graded “A” on xeno-canto, or rated ≥3 stars on Macaulay Library) and did not normalize amplitudes.

TABLE 1.

Behavioral song discrimination in 9 allopatric taxon pairs that have recently been split, in part on the basis of divergent vocalizations. Song discrimination is the proportion of territories of the first-listed species that ignored song from the second species (sample size of territories tested is given in parentheses). These cases provide a yardstick for how much song discrimination is “enough” to merit classifying allopatric populations as distinct biological species when using our methodology. We include in this list 2 taxon pairs (Zimmerius and Cyanocompsa) for which proposals to define populations as distinct biological species are currently under consideration by the South American Classification Committee of the American Ornithological Society.

TABLE 2.

Twenty-one taxon pairs currently classified as subspecies that show strong behavioral discrimination against allopatric song. We suggest that these populations merit recognition as distinct biological species. Song discrimination is the proportion of territories of the first-listed subspecies that ignored song from the second, allopatric subspecies (sample size of territories tested is given in parentheses). We conducted reciprocal playback experiments for 5 pairs; in each of these cases, song discrimination was high in both directions (population A typically ignored song from population B, and population B typically ignored song from population A).

TABLE 3.

Pairs of taxa currently classified as subspecies that show strong behavioral discrimination against allopatric song and that likely meet in a contact zone. Song discrimination is the proportion of territories of the first-listed subspecies that ignored song from the second, allopatric subspecies (sample size of territories tested is given in parentheses). We conducted reciprocal playback experiments for 3 taxon pairs (but not near the contact zone); in each of these cases, song discrimination was high in both directions (population A typically ignored song from population B, and population B typically ignored song from population A). Field studies documenting genetic, vocal, and morphological variation in the presumed contact zones offer an opportunity to test our assumption that song discrimination in the allopatric portion of a population's range is associated with mate choice in the contact zone.

Each treatment consisted of placing a wireless speaker (UE Roll or JBL Charge 2+) within a territory, broadcasting song at natural amplitudes (∼80 dB at 1 m from the speaker) for 2 min, and observing behavioral responses during both the 2 min of playback and a subsequent 5 min observation period. We recorded multiple behavioral responses to playback for each treatment. Here, we focus only on closest approach to the speaker (in meters), a reliable indicator of behavioral response to playback (Martin and Martin 2001, Jankowski et al. 2010, Freeman and Montgomery 2016, Freeman et al. 2016). At the beginning of each treatment, territory owners were within hearing distance and out of sight (i.e. >15 m distant from the speaker) or, more uncommonly, visible >15 m distant from the speaker. If a bird was still responding to playback at the conclusion of the first treatment—for example, if the territorial bird(s) remained within 15 m of the speaker or continued to vocalize at an elevated rate—then we waited until 2 min after it stopped responding (i.e. moved >15 m away but stayed within hearing distance, or ceased vocalizing at an elevated rate) before initiating the second treatment.

In our experimental design, the sympatric treatment served as a positive control. That is, we expected territorial birds to respond aggressively to sympatric song playback and included only experiments in which this was indeed the case (i.e. birds approached to within 15 m of the speaker—typically to within 5 m—in response to sympatric playback). Our aim was to use behavioral response to playback experiments as a proxy for premating reproductive isolation based on song. Thus, we define “song discrimination” here as instances in which the territory owner(s) ignored allopatric song, defined as a failure to approach within 15 m of the speaker in response to the allopatric treatment (i.e. we distinguished between “response” and “failure to respond,” but not between “weak” and “strong” responses). We calculated song discrimination for each taxon pair as the percentage of territories that failed to approach the speaker in response to allopatric song. For example, a song discrimination score of 0.8 indicates that 80% of territorial birds (e.g., 8 of 10) ignored allopatric song while simultaneously actively defending a territory (as described above, all territories responded to sympatric song by approaching the speaker).

In our experiments, we played songs of populations A and B (where populations A and B comprise a taxon pair) to territorial birds of population A. Most taxon pairs were tested in only one direction. That is, in the majority of cases we asked whether population A discriminated against song from population B but not the reverse. To date, we have conducted reciprocal playback experiments for 23 taxon pairs (13 oscines and 10 suboscines; see  Supplemental Material Table S2) in which we measured both discrimination of population A to song from population B and also discrimination of population B to song from population A, and in which we conducted at least 4 experiments on each population (n = 11.5 ± 3.8 playback experiments population⁻¹; range: 4–23; see  Supplemental Material Table S2). Song discrimination in these reciprocal cases was highly correlated (r = 0.88, t = 8.4, df = 21, P < 0.0001; Figure 1), and we therefore assume that unidirectional data accurately describe song discrimination within taxon pairs in our database.

FIGURE 1.

Behavioral discrimination against allopatric song was symmetric within taxon pairs. Song discrimination within a taxon pair was highly correlated (r = 0.87, t = 7.95, df = 21, P < 0.0001) for the 23 taxon pairs (13 oscines and 10 suboscines; see  Supplemental Material Table S2) for which we measured both how population A discriminated against song from population B and the reverse. The dotted line shows the 1:1 line.

Birdsong functions in both mate choice and territorial defense. Our playback experiments measure territorial defense but are likely a conservative proxy for inferring the strength of a behavioral barrier to reproduction. This is because selection on females choosing mates is stronger than selection on birds (typically considered to be males) engaging in territorial defense. As a result, the response function to song is typically broader for birds engaging in territorial defense (again, typically considered to be males) and narrower for females choosing mates (Searcy and Brenowitz 1988, Seddon and Tobias 2010, Danner et al. 2011, Curé et al. 2012). That is, if a territorial bird ignores a song in a territorial context, it is likely that a female would also discriminate against that song in a mate choice context. We highlight that territorial defense is not the exclusive purview of males, and that female territorial defense is common in tropical birds (Odom et al. 2014). Indeed, we observed multiple individuals responding to playback in around half the cases when we observed aggressive responses to playback treatments; for sexually dichromatic taxa (e.g., antbirds) for which we could identify sex, multiple individuals typically consisted of a male and a female, presumably a mated pair; and we assume that female territorial defense was likewise common in the sexually monomorphic taxa we studied (B. G. Freeman et al. personal observation). In sum, we assume that taxon pairs for which a majority of territory-defending individuals (either males alone or mated pairs) fail to approach the speaker have evolved a substantial degree of premating reproductive isolation based on song.

Finally, we note that we did not expect our song discrimination scores to equal either 0 or 1 even if populations' species recognition capacities were lacking or complete, respectively. This is because measurement error in our field experiments can only bias our estimate of song discrimination up from 0 (if “true” song discrimination equals 0) or down from 1 (if “true” song discrimination equals 1). In particular, we note that we did not include a negative control in our playback experiments. Thus, we are unable to define the background rate at which territorial individuals approach the speaker during the allopatric treatment simply by happenstance, because they are curious about a novel sound. As a consequence, we expect song discrimination scores to be <1, even if song were, in fact, a complete behavioral barrier to reproduction. For example, we previously quantified the relationship between song discrimination and genetic distance using Michaelis-Menten models, finding that song discrimination asymptotes were ∼0.78 when taxon pairs have been isolated for ∼5 million yr (B. G. Freeman et al. personal observation). This suggests that over evolutionary time, for the genetic distances in our dataset, song discrimination peaks not at 1 but at ∼0.78. As a consequence, we interpret taxon pairs with song discrimination scores around this value to represent cases where song discrimination is essentially complete.

Acoustic Trait Analysis

We analyzed 1,087 songs from 72 taxon pairs (8.0 ± 1.8 songs population⁻¹; all >5 songs population⁻¹) using the software Raven Pro 1.5 (Bioacoustics Research Program 2014; see  Supplemental Material Table S3). We measured acoustic traits for the same recordings used in playback experiments, supplementing these with additional recordings downloaded from xeno-canto ( http://xeno-canto.org) and the Macaulay Library of Natural Sounds ( http://macaulaylibrary.org) to boost sample sizes. For a representative song from each recording, we measured 7 song variables (total note count, mean note rate, mean note length, peak frequency, low frequency, mean note frequency range, and total song frequency range), following Mason et al. (2014). Within Raven, we used a Hann spectrogram window with 512 samples, a time grid with an overlap of 50% and a hop size of 256 samples, and a frequency grid with discrete Fourier transform set at 512 and grid spacing of 86.1 Hz.

We analyzed patterns of variation in acoustic traits for each taxon pair by log-transforming total note count and running a principal component analysis (PCA), using the correlation matrix based on the centered and scaled dataset wherein all variables were set to mean = 0 and SD = 1. This method produces distinct PCAs for each of the 72 taxon pairs. For these 72 comparisons, PC1 explained, on average, 48.2% of variation in multidimensional acoustic space (range: 30.8–73.8%). To compare taxon pairs' divergence in acoustic space using a common currency, we quantified standardized acoustic divergence between populations within a taxon pair as the distance between population means along PC1, measured in units of pooled standard deviations.

Statistical Analysis

All statistical analyses were carried out in R (R Development Core Team 2014). We first used the “cor.test” function to quantify the correlation between standardized acoustic divergence and behavioral song discrimination. Because standardized acoustic divergence was nonlinearly related to song discrimination, we used nonlinear regressions to investigate this relationship. Specifically, we used Michaelis-Menten curves, fit using the “nls” function, to investigate whether this nonlinear relationship differed depending on taxonomic rank (intraspecific vs. interspecific) or clade identity (oscine vs. suboscine). Our null model fit the formula y = ax/(b + x) to the full dataset, where a is the asymptote and b is a measure of the rate of increase. Our alternative model fit the formula y = ax/(b + cδ + x), where δ is an indicator variable of taxonomic status or clade identity and c is the difference in rate between the 2 groups. We then used the “anova” function to compare the relative fit of null and alternate models. Finally, we used t-tests to test whether species and subspecies differed in their song discrimination for both oscines and suboscines.

Taxonomic Implications

Our results are relevant to the taxonomic classification of allopatric populations under the BSC. We focus on our song discrimination data, because our finding that song discrimination and standardized acoustic divergence are not linearly related suggests that song discrimination is a more direct proxy for whether song constitutes a behavioral barrier to reproduction. Surprisingly, for suboscines, song discrimination was significantly greater in intraspecific taxon pairs (mean song discrimination = 0.70, n = 18) than in interspecific taxon pairs (mean song discrimination = 0.47, n = 11; t = −2.1, df = 23.4, P = 0.047). In oscines, song discrimination was unrelated to taxonomic rank (t = 0.91, df = 29.2, P = 0.37). These mismatches between song discrimination and taxonomic rank, particularly for suboscines, indicate that current taxonomy does not reflect the role of song divergence as a barrier to reproduction between allopatric populations of Neotropical passerines. To partially address this mismatch, we suggest that geographically isolated taxon pairs that are currently classified as conspecific, but that show strong song discrimination to allopatric song, merit classification as distinct biological species. We apply this logic to evaluate species limits for all taxon pairs for which we have conducted playback experiments (i.e. including taxon pairs for which we have not measured divergence in acoustic traits).

As described above, we do not expect our methodology to produce song discrimination scores of 1 (100% of territories ignore allopatric song), even if song in fact constituted a complete behavioral barrier to reproduction between 2 isolated populations. To provide a yardstick of how much song discrimination in our dataset is “enough” to merit classifying allopatric populations as distinct biological species, we consider song discrimination scores for the 9 interspecific taxon pairs in our dataset that have been split into distinct species within the past 2 decades in part on the basis of their divergent vocalizations (Table 1). Song discrimination in these cases averaged 0.59 (6 of 10 territorial birds ignored allopatric song). We use this average as a rough benchmark and suggest that the 21 taxon pairs currently classified as conspecific that show song discrimination greater than ∼0.6 merit classification as distinct biological species (Table 2).

In Table 3, we highlight 8 additional taxon pairs in which populations showed strong behavioral discrimination (>0.6) toward allopatric song, but in which it is likely that the 2 populations meet in a contact zone. Our playback experiments in these cases suggest that song constitutes a behavioral barrier to reproduction. But because our experiment took place far from the presumed contact zone, further work in the putative contact zones is needed to determine whether this is correct. We note that these cases thus represent tests of our assumption that high song discrimination between allopatric populations indicates that populations would also show high song discrimination and fail to interbreed when given the opportunity to actually interact (Hudson and Price 2014).

Acoustic Trait Analyses

Our conclusion that acoustic divergence is weakly related to behavioral discrimination when acoustic divergence is low (Figure 2), and thus that playback experiments provide better inference of reproductive isolation under these conditions, relies on our methodology for measuring and analyzing acoustic traits. We quantified 7 acoustic traits, measured from an average of 8 recordings per population, and used multivariate statistics to measure acoustic divergence between populations. By contrast, studies of acoustic variation among closely related populations often consider many dozens of traits, some of which are customized to the specific vocalization of the taxa under consideration, measure hundreds of recordings, and also conduct univariate comparisons for each individual acoustic trait (e.g., Ng et al. 2016). Because the present study measured acoustic divergence across a diverse set of Neotropical passerines that differ dramatically in song, we focused on a small number of acoustic traits that are likely to be broadly relevant across taxa (and also likely to have low collinearity). Our sample sizes of recordings analyzed per population per taxon pair were not large enough to permit meaningful univariate analyses; we note that conclusions from multivariate acoustic analyses are typically similar to those from univariate analyses (e.g., Ng et al. 2016).

It is possible that a different methodology for measuring acoustic divergence could recover a tighter correspondence between acoustic divergence and behavioral discrimination, but we highlight that populations may be statistically different in their acoustic traits but fail to discriminate against allopatric song in playback experiments (e.g., Nelson 1998, Soha et al. 2016). If our findings are valid, we suggest that acoustic trait analyses are best employed (1) as an initial exploration of song divergence (i.e. a stopgap measure pending experimental data), especially when acoustic divergence is weak or moderate, and (2) in concert with playback experiments; for example, the shortcomings associated with conducting playback experiments in a small number of geographic locations can be overcome by comparing vocal characters from across a broader geographic spread. Future research can assess the validity of our results by focusing on single clades and by combining playback data with acoustic analyses (with acoustic traits tailored to the clade in question) to quantify the correlation between these 2 methods for measuring song divergence.

Implications for Species Limits

Avian taxonomy typically follows the BSC, which defines species as reproductively isolated populations. Applying the BSC to geographically isolated populations that do not have the opportunity to interbreed is challenging. In practice, ornithologists use a comparative framework to assess whether isolated populations are likely to interbreed (or not) should they come into sympatry (Isler et al. 1998, Tobias et al. 2010b). Here, we focus on divergence in song—a single, behavioral barrier. Although a variety of ecological, morphological, behavioral, and genetic factors can all constitute barriers to reproduction, including divergence in calls and other vocalizations (reviewed in Price 2008), song divergence is an important barrier to reproduction in birds, perhaps particularly so in suboscines (Edwards et al. 2005, Tobias et al. 2012).

We use song discrimination scores for taxon pairs that have recently been split, in part on the basis of their divergent vocalizations, as a yardstick for when to split taxa on the basis of our song discrimination data (Table 1). The average song discrimination scores for these taxa are ∼0.6, which means that 6 of 10 territorial birds ignored allopatric song. We found 21 intraspecific taxon pairs that cleared this threshold, and we suggest that these populations merit status as distinct biological species (Table 2). In many cases, these taxon pairs have been previously considered to represent distinct species on the basis of obvious vocal differentiation. For example, Striped Woodhaunter (Automolus subulatus) populations east and west of the Andes differ in voice and have thus been considered to represent distinct species by some authorities (e.g., Ridgely and Greenfield 2001). It will thus surprise few ornithologists familiar with Neotropical birds to learn that territorial woodhaunters from west of the Andes typically ignore playback of Amazonian song. However, we also document strong song discrimination between many populations that are known to differ in song but that have not been previously considered to represent distinct species—for example, Andean Solitaire (Myadestes ralloides) populations north and south of the Marañon Gap in Peru. Finally, we document high song discrimination in 8 cases of allopatric populations that likely meet in a contact zone (Table 3). Such examples offer an opportunity to test our assumption that song discrimination in the allopatric portion of the range is indeed associated with mate choice in the contact zone.

We suggest elevating the taxa listed in Table 2 to species rank. One complicating factor to carrying out this suggestion in practice is that there may be more than 2 taxa involved in many of these species. That is, we argue that allopatric populations A and B (for which we conducted playback experiments) should be considered distinct biological species, but where does allopatric population C (for which we did not conduct experiments) fit in? In most cases, major geographic barriers can provide a reasonable basis for placing additional populations. For example, the spine of the Andes divides Striped Woodhaunter populations in two (east and west), while the Marañon Gap divides Andean Solitaire populations in two (north and south). In other cases, geographic barriers are not complete—for example, distributions extend across the northern reaches of the Andes, as in the Yellow-olive Flycatcher (Tolmomyias sulphurescens)—and the question of affinities of additional allopatric populations is more difficult to address. It may be that such cases require a more comprehensive analysis, for example by conducting playback experiments between each of dozens of populations. Nevertheless, our data suggest that multiple biological species lurk within these complicated species complexes; if published information on song variation can be used to group additional allopatric populations not considered in our playback experiments, then the data we present here may be sufficient to redefine species limits in these complexes.

Finally, we refrain from interpreting low song discrimination values as indicating an absence of premating reproductive isolation that might constitute evidence for “lumping” populations currently considered distinct species into a single taxon. We take this cautious approach for 2 reasons. First, 2 allopatric populations may be nearly identical in song, with low song discrimination scores, but divergent in other signaling traits (e.g., plumage or call notes) that generate substantial premating reproductive isolation. Second, we argue that if a bird ignores a song in a territorial context, it is likely that a female would also discriminate against that song in a mate choice context. However, whether the reverse is true is less clear. Indeed, there are cases where males respond to a song in a territorial context but females discriminate against this same song in a mate choice context (note that, in this example, females may cue to male calls more then male songs; Seddon and Tobias 2010). In sum, we interpret a lack of response to allopatric song as indicating a behavioral barrier to reproduction, but we do not interpret a response to allopatric song as indicating the absence of a behavioral barrier to reproduction.

Playback Experiments Can Help Address the Latitudinal Gradient in Taxonomy

It is widely acknowledged that species are more narrowly defined in the temperate zone than in tropical latitudes (for birds, see Tobias et al. 2008, Weir 2009, Milá et al. 2012). In birds, this bias arises for both biological and nonbiological reasons. First, temperate zone species tend to attain secondary contact more quickly than tropical species (Martin et al. 2010, Weir and Price 2011), with the consequence that genetically divergent, allopatric populations are more common in the tropics than in the temperate zone (Martin and Tewksbury 2008). Second, temperate zone scientists carry out the bulk of avian systematic research, and many tropical taxa are poorly studied. This geographic bias in taxonomy has practical consequences for comparative studies and our understanding of avian diversity. Studies of diversification along latitudinal gradients are hampered by variation in taxonomy (e.g., Tobias et al. 2008), as are investigations of the evolutionary dynamics of particular clades. For example, analyses of the tempo and drivers of diversification in Scytalopus would presumably have reached different conclusions if they were conducted in 1990, when 13 species were recognized (Sibley and Monroe 1990), compared with at present, when 42 species are recognized (Remsen et al. 2017).

There is widespread sentiment that the latitudinal gradient in taxonomy ought not to exist, but little agreement on how to implement the widespread taxonomic changes necessary to address this bias. The status quo is that authors publish taxonomic studies that assess species limits within a complex, and taxonomic committees apply the BSC to pass (or reject) these studies' recommendations. This results in a slow but steady flattening of the latitudinal gradient in taxonomy; for example, the South American Classification Committee of the American Ornithologists' Union recognized 52 additional species of continental landbirds found in South America in the 5 yr period between 2012 and 2016 (Remsen et al. 2017), far more than the 2 additional species of continental landbirds found in the United States and Canada that the North American Classification Committee recognized during the same period (Chesser et al. 2017). However, others advocate greatly accelerating the push to rank tropical taxa at the species level (e.g., Gill 2014); for example, the Handbook of the Birds of the World recently applied species delimitation criteria (Tobias et al. 2010b) to elevate >1,000 populations of birds (primarily in the tropics) to species status (see  http://www.lynxeds.com/product/hbw-and-birdlife-international-illustrated-checklist-birds-world-0). Needless to say, this proposal to increase extant species-level diversity of birds by ∼10% in one fell swoop is controversial (Remsen 2015, Collar et al. 2016).

The approach we describe here—conducting playback experiments to measure species recognition between allopatric populations—offers a way forward to addressing the latitudinal gradient in taxonomy, for 4 reasons. First, it maintains the primacy of the BSC, which has many advantages—primarily, that a focus on reproductive isolation means that “species” is the only taxonomic rank that has a biological basis (e.g., Toews 2015). Second, it focuses on a trait directly relevant to reproductive isolation—song. Tropical birds often exhibit minimal plumage variation, and, because genetic divergence is only loosely correlated with reproductive barriers (Price and Bouvier 2002), focusing on song divergence is an especially profitable approach to assessment of species limits in allopatric tropical taxa. Third, it directly assesses whether song divergence is relevant to reproductive isolation. While the Handbook of the Birds of the World approach considers vocal differentiation as evidence for species status (Collar et al. 2016), not all vocal differences are created equal (see Figure 2); playback experiments “ask the birds themselves” if differences between populations matter to species recognition. Fourth, playback experiments are easier than ever to carry out. Field guides and online resources provide descriptions of geographic variation in song that suggest populations to target for playback studies; large repositories of high-quality, publicly available recordings provide recordings to use in playback experiments; and affordable and lightweight wireless speakers make conducting fieldwork relatively straightforward. We envision that citizen scientists and ornithologists, particularly those residing in tropical nations, will organize and carry out playback studies to measure species recognition between allopatric populations of tropical birds on an increasingly broad scale.

Conclusion

The use of field playbacks to assess whether vocal differences between isolated populations constitute barriers to reproduction has a long history (e.g., Lanyon 1978). Because song is an important barrier to reproduction in birds, perhaps particularly so in clades with innate song, evaluating song divergence is an important component of assessing the taxonomic rank of allopatric populations. Our results show that playback experiments and acoustic trait analyses do not provide interchangeable inference, and they demonstrate that it is feasible to conduct playback experiments on a broad scale. We argue that this approach provides valuable data for assessing species limits using the BSC, and we advocate for the continued use of playback experiments on tropical birds to help flatten the latitudinal gradient in taxonomy. In sum, we encourage the further use of field playback experiments to measure species recognition between allopatric songbird populations, an approach that has been termed “behavioral systematics” (Pegan et al. 2015).