Animal communication plays a crucial role in biology, yet the wide variability in vocalizations is not fully understood. Previous studies in birds have been limited in taxonomic and analytical breadth. Here we analyse an extensive dataset of >140,000 recordings of vocalisations from 10,124 bird species, representing nearly every avian order and family, under a structural causal model framework, to explore the influence of eco-evolutionary traits on acoustic frequency characteristics. We find that body mass, beak size, habitat associations, and geography influence acoustic frequency characteristics, with varying degrees of interaction with song acquisition type. We find no evidence for the influence of vegetation density, sexual dimorphism, range size and competition on our measures of acoustic frequency characteristics. Our results, built on decades of researchers’ empirical observations collected across the globe, provide a new breadth of evidence about how eco-evolutionary processes shape bird communication.

Acoustic signal characteristics of bird sounds

We accessed the recordings of bird vocalisations (both songs and calls) of all the species from the online repository xeno-canto. First, we downloaded the meta-data of all the bird recordings available on xeno-canto in December 2019, using the R package warbleR. We then cleaned the meta-data file by removing unidentified species, such as ‘Mystery mystery’ and non-avian sounds. For species that have a distinct song and call, we limited our analysis to songs only, based on the existing distinctions within Xeno-canto. For species without a clear distinction between songs and calls, such as raptors, we included all vocalizations in our analysis. For each species we then hierarchically downloaded the best available recording (refer Fig. S9 in manuscript), using the R package warbleR. The quality of the recordings in xeno-canto is ranked from A to E, with A representing the highest and E the lowest quality, with a few recordings having “no score”. We selected the highest available rank, without requiring a minimum number of recordings for each species. For example, if Copsychus malabaricus (White-rumped Shama) had recordings of quality A, B and C, we selected only and all those ranked as A. If another species, Copsychus saularis (Oriental Magpie-robin) had no recording of quality A, but only of quality from B to E, we chose B recordings.

We then analysed individual recordings for each species to detect and measure acoustic frequency values (refer Fig. S2 in manuscript). For each recording, we first automatically detected all acoustic signals therein, each between a duration of 1 – 5 sec - acoustic note, (refer Fig. S2b in manuscript). We used 0 and 22 kHz as our lower and upper limits of a frequency bandpass filter, to capture the entire range of the acoustic frequencies, with a Hanning window of default length of 512. We did not use any proportional reduction of amplitude envelopes through thinning. We only used converted Waveform Audio File Format (.wav) recordings that had a sampling rate of 44.2 kHz, after decompressing them from mp3 format. While this has been shown not to significantly bias any of the acoustic or similarity measurements, it has been shown to affect the precision of acoustic parameters, such as peak frequency. However, the negative effect of compression is assumed to be less problematic when comparing acoustic frequency values across species, as the differences between species are usually stronger, than within species.

For each acoustic note from a recording, we calculated the Signal-to-Noise Ratio (SNR) to discard low-quality (e.g., high background noise) selections, by following the recommendation from Araya-Salas et. al 2019, because background noise has been shown to bias most energy distribution-related parameters such as spectral entropy and affect the precision of most acoustic parameters. Here we used the default value of not using any lower and upper limits of a frequency bandpass filter, or a window length, as they were previously used to automatically detect the acoustic note in the previous step. We used a very small margin value of 0.04, adjacent to the start and end points of selection over which to measure noise. For each recording, we then selected the top 100 non-overlapping acoustic notes, i.e., those with the highest SNR, to measure acoustic characteristics of each species (refer Fig. S1c-d in manuscript ): (i) minimum frequency (ƒ_min), (ii) maximum frequency (ƒ_max), (iii) dominant frequency (ƒ_dom), (iv) standard deviation of ƒ_min (σf_min), (v) standard deviation of ƒ_max (σf_max), and (vi) interquartile range of acoustic notes (Δƒ).

We define ƒ_min as the start frequency of the amplitude peak containing the highest amplitude, ƒ_max as the end frequency of the amplitude peak containing the highest amplitude; ƒ_dom as the average dominant frequency measured across the spectrogram of the acoustic note; and Δƒ as the frequency range between the first quartile (frequency at 25% energy of the spectrum) and the third quartile (frequency at 75% energy of the spectrum). ƒ_min, ƒ_max, and ƒ_dom, forms the fundamental frequency measures, σf_min and σf_max characterize the variability found within minimum and maximum fundamental frequencies respectively, and Δƒ characterizes the frequency bandwidth. We also modelled Δƒ, measured as the difference between ƒ_min and ƒ_max, which we elaborate in Supplementary Information in the manuscript. Assuming the recordist manipulated their equipment and recording strategy such that the acoustic signal of the target species would be the clearest and loudest in the recording, we repeatedly estimated each of these six response variables under a series of thresholds - amplitude of every 5 dB between 5 and 50 dB (refer to Fig. S1e in manuscript). Median values of the estimates of the six response variables, from each of the threshold percentages (refer Fig. S1e, in manuscript) were used to estimate acoustic characteristics for each species, across all recordings (refer Fig. S1f, in manuscript). We used the R package warbleR to calculate the acoustic characteristics. We specifically chose these measures of bird vocalisation because they have all been previously linked to the evolutionary forces of physiology, inter-specific competition, distribution, and environmental variables (Fig. 1 & Table S1 in manuscript ). Additionally, they were also the easiest acoustic measures that could be consistently extracted from thousands of recordings of our focus bird species.

Zoogeographic distribution of species

We divided the zoogeographic distribution of our study species into 12 major zoogeographic realms (refer Fig S9 in manuscript). In addition to the 11 zoogeographic realms identified by Holt et al. 2013, namely, Afrotropical, Australian, Madagascan, Nearctic, Neotropical, Oceanian, Oriental, Palearctic, Panamanian, Saharo-Arabian, and Sino-Japanese, Oceanian, we also included species from Antarctica and surrounding islands under a new realm, South Polar, based on our data from xeno-canto. We listed a species under multiple zoogeographic realms if its distribution spanned multiple realms, either due to its natural or introduced distribution. We calculated the frequency values separately for species that occurred in more than one zoogeographic realm.

Species traits

We collated information on the morphology, behaviour, ecology, and geographical data of birds from AVONET and BirdBase. We used log_(e) transformed body mass (g) as a measure of body mass, results are presented as back-transformed values in the main text. We calculated beak size (log_(e) transformed) by multiplying beak length, width, and depth (mm³). As a measure of vegetation density, we used two categorical variables: (i) habitat density and (ii) primary habitat, both from AVONET. We estimated overall species richness as a proxy for competition for acoustic frequency space, by calculating the mean number of terrestrial vertebrates: amphibians and birds using global richness data from BiodiversityMapping.org. We extracted these values as means within a species’ breeding range, which we calculated as a circular area, equivalent to the species range size, centred around its centroid latitude and longitude. To account for plasticity in song acquisition, we divided the bird species into two major groups, (i) song learners – species that learn songs from adult tutors to use in mate choice and territorial displays, i.e., hummingbirds, parrots, and oscine passerines and (ii) innate songsters – species that do not learn their vocalisation. i.e., suboscine passerines and all remaining birds. Even though there are exceptions (e.g., a few species of Ant wrens and Wood creepers), in general, song learners tend to have more complex vocalisation. Additionally, song learners, show a slower rate of evolution and diversification due to their significantly greater variability in acoustic traits (by 54 – 79%) compared to innate songsters.