Absence of song suggests heterogeneity of vocal-production learning in hummingbirds

Hummingbirds have been recognized, along with songbirds and parrots, as capable of learning songs. However, it is still unclear whether singing can be treated as a homologous trait within the family, analogous to songbirds. Therefore, we systematically compared the information about hummingbird vocalizations in the literature in a phylogenetic framework. In general, songs were emitted by perching males in a reproductive context, while calls were mainly in agonistic contexts. Singing was ancestral in most of the nine major hummingbird clades, but has been lost at least once in the mountain gem clade and twice in the bee clade. This evolutionary loss of singing might suggest heterogeneity of vocal-production learning.


Introduction
For communication, birds use a range of species-specific sounds uttered in specific behavioral contexts, arbitrarily divided into two mutually exclusive subgroups: songs and calls (Spector 1994;Catchpole and Slater 1995). In most studies referring to temperate songbirds, the term "song" is used to the elaborate vocalizations emitted by males in the context of reproduction to defend their territory and attract females (Catchpole and Slater 1995). In studies mainly referring to tropical birds, "song" is the vocalization that, among other functions, primarily serves species-specific recognition (Vielliard 1987;Mathevon et al. 2008). In contrast, "call" has always been a broader term, often used for all non-singing vocalizations, such as those used to maintain contact, beg for food, and alarm conspecifics (Catchpole and Slater 1995). The definition of song is of particular interest because there are numerous examples of song learning in songbirds, but very few examples of call learning (Zann 1990). Therefore, song in songbirds is usually considered a learned behavior. Although hummingbirds are less studied, there is compelling evidence of song learning (e.g.: Jarvis et al. 2000;Gahr 2000; Araya-Salas and Wright 2013; Johnson and Clark 2020); however, a direct comparison to songbirds and within hummingbirds may come across a homology 1 3 problem given by the possibility that different authors may have relied on different definitions of song. We investigated this problem in here first, then analyzed the evolution of singing in hummingbirds under the light of the most recent phylogeny (McGuire et al. 2014).

Singing in hummingbirds is functionally equivalent to songbirds
First, we gathered information from the literature on the vocalizations of hummingbirds to investigate which are classified as songs or calls and in which context they occur. In our survey, we reviewed 74 publications, from which we extracted 98 vocalizations described as songs and 125 as calls. This comprised 78 species (22% of the 363 species) and three subspecies (Table S1) representing 47 genera (42% of the 112 genera). The hummingbird family (Trochilidae) was traditionally subdivided into two subfamilies (Phaethornithinae and Trochilinae) (Hoyo et al. 1999); however, the most updated phylogeny suggests nine clades (eight tribes and one subfamily) (McGuire et al. 2009). Here, we follow the subdivisions suggested by McGuire et al. (2014) based on a multilocus (six genes and five loci) phylogenetic estimate which recognized nine clades. We sampled eight of these clades: topazes (three species, 75% of the total of 4 species in the clade); hermits (12 species, 30.8% of the total); mangoes (9 species, 32.1%); coquettes (7 species, 10.4%); Patagona (single species clade); mountain gems (3 species, 17.6%); bees (17 species, 48.6%) and emeralds (26 species, 22.6%) (McGuire et al. 2014;Winkler et al. 2020). Then we grouped the references by phylogenetic clades and applied a text mining method using the packages "dplyr" (Wickham et al. 2020), "tidytext" (Silge and Robinson 2016) and "wordcloud" (Fellows 2018) in R (R Core Team 2018). We obtained word clouds for seven of the nine major clades except for the little-studied Patagona and Brilliant clades. Details on the literature survey and word mining are in the Supplementary Material.
The most commonly used terms to describe the behavioral context of songs in all clades were "perch(ed)", "male(s)" and "display" and of calls were "agonistic" and "aggressive", except in hermits, whose calls are emitted mainly in flight ("flight" and "flying", most frequently used words) (Fig. 1, right column). We know from songbirds that perching songs emitted by males are frequently associated with reproductive contexts (Catchpole and Slater 1995), and this seems to be the same for hummingbirds. Although inter-species comparisons of hummingbird show strikingly different levels of temporal-spectral complexity in their songs (Fig. 1, left column), song appears to be functionally equivalent within the family.

Singing is ancestral, widespread, but not ubiquitous
Since the term "song" was used consistently in the analyzed publications, we used these references to investigate the evolution of singing in the hummingbird family. Seventy-one of the 78 included species and three subspecies were described as singing and 7 species (9%) as non-singing (Fig. 2, green and red circles). The sex of singing species was unknown in 54%, only males sang in 38%, and both sexes sang in 8% (six species, Fig. 2, venus symbol). Singing females are scattered throughout the phylogeny, thus, it is probably overlooked rather than rare. Likewise, non-singing species may be underrepresented since it is not attractive to report negative observations such as "non-singing". Being aware of these limitations, we, nevertheless, estimated the probabilities of singing occurring in the clades of the phylogeny (McGuire et al. 2014) using the maximum-likelihood method on the equal rate (ER) model and all different rates (ARD) model provided by the R package "ape" (Paradis and Schliep 2019). The ARD model had a better fit to the data compared to the ER model (LR = 7.12, df = 1, p < 0.01). Thus, gains and losses of singing behavior within hummingbirds might have happened at different rates. The ARD model estimated higher transition rates for gains (0.034 ± 0.016) than for losses (0.003 ± 0.002). We obtained the posterior probabilities of the ARD model for singing to estimate the ancestral conditions of the hummingbird family and the clades within the family using the R package "phylotools" (Revell 2012). Support values above 70% are in general thought to be acceptable. Details on ancestral character estimation are provided in the Supplementary Material. In the hummingbird family as a whole, the probability of singing as an ancestral trait (92.6%) is higher than the non-singing (7.4%). Within the family, singing is probably, except for bees (69.1%), the evolutionary ancestral condition in topazes (94.2%), hermits (92.2%), mangoes (92.8%), coquettes (92.1%), mountain gems (90.9%) and emeralds (92.2%). Within bees, the ARD model estimated at least two evolutionary losses of song in the ancestor of the genus Archilochus (96.4%) and the common ancestor of the genera Selasphorus and Atthis (100%). As species in the genus Atthis sing, the model estimated a recovery (100%) in ancestors of this genus. Regardless, further loss of song occurred within mountain gems (Fig. 2, pie charts).
The presence of song in hummingbirds appears to correlate with vocal-production learning (Fig. 2, white circles, triangles and squares), as is also the case in songbirds (Kuhl et al. 2020), and evolutionary loss of song with degenerate development of the putative vocal control system (Fig. 2, crossed squares), a phenotype similar to that of female songbirds unable to produce learned songs (Gahr 2000) (Table 1).

Discussion
Hummingbird songs are mainly emitted by males perched on exposed twigs probably because, in several species, males aggregate into singing assemblies (leks) (Stiles Lekking species often have vocal dialects that, in some of them, cannot be explained by genetic differentiation Ornelas 2014, 2019), but rather by cultural transmission. Thus, vocal dialects are also evidence for The outer circle indicates the prevalence of singing. Note that singing behavior is heterogeneously distributed within the hummingbird family. The alternate rates (ARD) model was used to estimate the ancestral character for each node. The area of pie represents the proportion given by the Bayesian posterior probability that the ancestral condition in the node is associated with one of two alternative conditions' presence (green) or absence (red) of singing. Singing behavior was gained, lost, and regained within the hummingbird family. The venus symbol indicates species with singing females. The type of evidence of vocal-production learning (VPL) is also represented: the presence of dialect (circle), the abnormal development of song under experimental acoustic isolation (triangle), the presence of a putative vocal control system (VCS) (square), and vestigial VCS (crossed square). Note that Anna's hummingbird accumulates multiple pieces of evidence for VPL vocal-production learning (Kroodsma and Baylis 1982;Araya-Salas and Wright 2013) (Table 1). It is, therefore, possible that males of a lek use dialects to acoustically distinguish neighbors, which are most likely to hold singing territories with stable boundaries within the leks, from newcomers, which are probably seeking a singing territory. Thus, territorial males can use this information to modulate their aggression during territorial responses, according to the risk of the threat. This hypothesis predicts two selective benefits: (1) males can increase their detectability and consequently their mating success and (2) decrease the costs of a fight, thus increasing their survival odds. The reduced aggression toward familiar individuals is called the "dear enemy effect" and has been shown in songbirds (Briefer et al. 2008), but this effect is awaiting experimental confirmation in hummingbirds.
In bees, the evolution of singing is negatively correlated to the production of sounds with the wing during dive displays (wing trills) as they seem functionally equivalent (Clark et al. 2018). Dive display is ancestral (Clark and Feo 2010;Clark et al. 2018), whereas dive or homologous behaviors seem rare outside the bees (Clark et al. 2018). Furthermore, their relatively small body sizes may lower the energetic requirements for dives and modified wings may facilitate maneuverability (Payne 1984;Clark et al. 2011). Interestingly, bumblebee hummingbird (Atthis heloisa) and wine-throated hummingbird (Atthis ellioti) which inhabit forests, pine-oak woodlands, and neighboring shrubby areas (Arizmendi et al. 2020;Thurber et al. 2020) are non-diving and regained singing (Clark et al. 2018). Overall, the expansion of bees to North America and consequent occupation of open habitats combined with their anatomical specializations might have favored the enhancement of visual displays and associated mechanical sounds.
Most of the bees species that lost the song occupy open habitats, for example, Allen's hummingbird (Selasphorus sasin) is a riparian breeder that often perches conspicuously on leafless branches (Clark and Mitchell 2020). The predation of adults has been, nevertheless, seldom observed and predation does not seem a significant risk to hummingbirds Pieces of evidence of VPL are the occurrence of similarities in the song of a certain population that is attributed to natural intraspecific vocal imitation ("Dialect"); the abnormal vocal development under acoustic deprivation observed under experimental conditions ("Abnormal song under acoustic isolation"), and the presence of brain nuclei related to VPL, the vocal control system (VCS) ("VCS present"), which is vestigial in two species ("VCS vestigial")  (Miller and Gass 1985). It is, therefore, possible that some bees avoided perched song because, particularly in open habitats, learned songs that are, in general, more variable and less susceptible to habituation, made them more conspicuous to predators. A relaxed predatory pressure seems to precede the evolution of vocal-production learning because predators, likely, habituate more easily to the constancy of innate calls than to the variation of learned songs (Jarvis 2006;Nowicki and Searcy 2014). Predatory risks must be investigated comparatively between singing and non-singing bees to elucidate this possibility.

Conclusion
In summary, singing is ancestral in most of the hummingbird clades and commonly associated with lekking behavior whereas evolutionary losses happened especially in bees, a clade that also evolved sophisticated visual displays such as the dive. We especulate that gains, losses and regains of singing and maybe vocal-production learning seem to have happened under distinct evolutionary pressures. It is important to bear in mind that such a literature survey can be biased toward singing species. Particularly species of clades in which singing is ancestral need to be investigated to confirm our findings. Although recent studies have shed light on the ontogeny of vocal-production learning in some hummingbirds Clark 2020, 2022), experimental studies that tackle vocal-production learning are still scarce and the related neurobiological correlates need confirmation; the brain areas suggested by connectivity, histology and gene expression to control song of hummingbirds (Jarvis et al. 2000;Gahr 2000) have not yet been confirmed using electrophysiological or lesion approaches. Nevertheless, the diversity in the vocal communication of hummingbirds including songs of different complexity levels which can be combined with a variety of visual displays or predominately visual offers a unique opportunity to investigate the mechanistic convergence, and perhaps the phylogenetic constraints, of vocal-production learning. Furthermore, comparative studies involving little-studied songbird and parrot taxa would reveal whether a similar heterogeneity exists in the other two vocal learner bird groups.