Machine learning and statistical classification of birdsong links vocal acoustic features with phylogeny

doi:10.21203/rs.3.rs-2609155/v1

Download PDF

Article

Machine learning and statistical classification of birdsong links vocal acoustic features with phylogeny

https://doi.org/10.21203/rs.3.rs-2609155/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 01 May, 2023

Read the published version in Scientific Reports →

You are reading this latest preprint version

Birdsong is a longstanding model system for studying evolution, and has recently emerged as a measure of biodiversity loss due to deforestation and climate change. Here, we collected and analyzed high quality song recordings from seven species in the family Estrildidae. We measured the acoustic features of syllables and then used dimensionality reduction and machine learning classifiers to identify features that accurately assigned syllables to species. Species differences were captured by the first 3 principal components, corresponding to basic spectral features, spectral shape, and spectrotemporal features. We then identified the measured features underlying classification accuracy. We found that fundamental frequency, mean frequency, spectral flatness, and syllable duration were the most informative features for species identification. Next, we tested whether specific acoustic features of species’ songs predicted phylogenetic distance. We found significant phylogenetic signal in syllable spectral features, but not in spectral shape or spectrotemporal features. Results indicate that spectral features are more constrained by species’ genetics than are other features, and are the best signal features for identifying species from song recordings. The absence of phylogenetic signal in spectral shape and spectrotemporal features suggests that these song features are labile, reflecting learning-processes and individual recognition.

Biological sciences/Evolution

Earth and environmental sciences/Climate sciences

Biological sciences/Neuroscience

Biological sciences/Neuroscience/Computational neuroscience

Biological sciences/Neuroscience/Sexual behaviour

Biological sciences/Neuroscience/Social behaviour

Biological sciences/Biological techniques

Biological sciences/Biological techniques/Behavioural methods

Biological sciences/Computational biology and bioinformatics/Machine learning

Acoustic communication signals are a focus for understanding species evolution [1-7], mating strategies [8-12] and physiological mechanisms of behavior [13-16]. More recently, they have been used to quantify species’ use of habitat [17-20] and changes in biodiversity associated with deforestation and climate change [21-27]. Animals use species-specific sounds to exchange social information in complex environments and at distances that preclude the use of other sensory cues [5,7]. Many vertebrate species use vocalizations to convey species, sex, location, and behavioral state to receivers such as potential mates, competitors and family members [3,8,9,13,28-31]. Understanding the features of vocal signals that communicate specific social information requires a comparative analysis of species with spectrotemporal diversity and known relatedness in order to quantify how differences in the former may change across species based on the latter. In imitative vocal learners, such as songbirds, signal acoustics result from both genetic and experiential factors. Birdsong is a robust model behavior for studying mechanisms of plastic social communication; in turn, the evolution of song features is shaped by processes of cultural [8,32], sexual [33], and natural selection [9,34]. Signal processing theory predicts that the production mechanisms of signalers and the perceptual systems of receivers evolve to optimize communication [5,29]. The mapping of the social information onto specific features is currently unclear for most complex vocal signals.

The mating vocalizations (songs) of songbird species in the family Estrildidae (estrildid finches or waxbills/grassfinches) differ widely in spectral and temporal features, despite similarities in evolutionary origin, size, habitat, and diet [9,13,35-40] (Fig. 1 and 2). In each species, song syllables are acoustically complex, species-specific, and learned, providing a rich diversity of signal features for species and individual recognition. Across species, songs are composed of hierarchically-organized acoustic units; notes are grouped to form “syllables,” and sequences of syllables form song [39-46]. Quantitative studies of estrildid song have largely focused on one or two species [13,37-40,43-45,47], and either composite measures [48-51] or individual features, such as fundamental frequency [52,53].

Here, we measured 21 acoustic features of each song syllable, and used dimensionality reduction and machine learning classifiers, including multivariate, unsupervised and supervised models, to identify acoustic features that accurately assign song syllables to each of seven estrildid species. Results of these approaches were used to quantify and compare song acoustics across species and identify features that scale with phylogenetic distance to identify the acoustic features that convey species identity versus those that are shaped by learning processes. Identifying and understanding these features will inform comparative behavioral and neural research and produce a song-evolutionary framework for the study and interpretation of processes and mechanisms underlying communication perception, selectivity, and preference [8,13,54]. We quantified song syllable acoustics in 3 Australian species, 3 African species, and 1 Southeast Asian species, all with known phylogenetic relationships [55-58]. We predicted that spectral (e.g., basic measures such as fundamental frequency) and spectral shape features (e.g., measures describing the distribution of energy across frequencies such as kurtosis) would differ most across species and least within a species, consistent with previous spectral categorizations [13,31,36,40,59]. Specifically, species whose songs are characteristically broad-band should share more space along dimensions of spectral and spectral shape measures with each other than with tonal species, and vice-versa. We do not, however, expect to find clustering of species along spectrotemporal dimensions given the ubiquitous and extensive ability for spectrotemporal modulation in songbirds [60,61]. We also predicted that features known to be used by birds for sound discrimination in behavioral studies (e.g., fundamental frequency, Weiner entropy, time entropy, frequency modulation) [38,62-64] will be the most important in syllable classification. Finally, we tested the hypothesis that song evolves primarily through changes in spectral characteristics but not labile spectrotemporal ones. To test this, we performed a phylogenetic signal analysis on song features and predicted that spectral and spectral shape features but not spectrotemporal features would correlate with the known estrildid phylogeny.

Spectral, Spectral Shape, and Spectrotemporal Components Explain 72% of Variation in Song

Acoustic features were measured for the song syllables of 3 African species (red-cheeked cordon-bleu [Uraeginthus bengalus], CB; red-billed firefinch [Lagonosticta senegala], RF; gold-breasted waxbill [Amandava subflava], GW), 3 Australian species (double-barred finch [Stizoptera bichenovii], DF; long-tailed finch [Poephila acuticauda], LF; zebra finch [Taeniopygia guttata castanotis], ZF), and 1 Southeast Asian species (Bengalese finch [Lonchura striata domestica], BF). We conducted PCA using 21 feature values for each of the 15,081 syllables in our song dataset. The first 3 PC scores explained 72.28% of the variation in syllable acoustics and represented the song spectral (PC1), spectral shape (PC2), and spectrotemporal (PC3) features (Supplementary Table S1). Spectral features were singular frequency measures (e.g., fundamental frequency, mean frequency, dominant frequency). Spectral shape features described the shape of the frequency power spectrum (e.g., entropy, spectral flatness, skew). Spectrotemporal features included syllable duration and changes in spectral features over time (e.g., modulation index, dominant frequency slope).

The degree of shared and unique volume in PC space differed significantly across species comparisons (Fig. 3). Clusters in PC space represent single species (i.e., single color present; e.g., LF clusters around PC1 = -3, PC2 > 2.5; Fig. 3a) or mixed species (i.e., multiple colors present; e.g., BF, DF, LF, ZF cluster around PC1 = -2.5, PC2 = 0; Fig. 3a) syllables. Species differed in volume spread (e.g., GW syllables clustered in a small space and BF syllables were more variable as a species; Fig. 3b).

Species clustered into biogeographical phylogenetic groupings along the dimension of spectral components (Fig. 3a-b; Supplementary Fig. S1). Australian species (ZF, LF, DF) had low spectral values, characteristic of low frequency vocalizations. African species (RF, CB, GW), on the other hand, had high spectral values, characteristic of high frequency vocalizations. The BF spectral values ranged between the Australian and African species, with an intermediate median value and variance spanning the range of the other species. Syllables showed no clear phylogenetic clustering along spectral shape (Fig. 3a,c; Supplementary Fig. S1). With respect to spectral shape, ZF generally scored lowest, characteristic of vocalizations with a high spread in frequency and high entropy (e.g., noisy, broadband harmonic song). LFs, on the other hand, scored higher, characteristic of vocalizations with low harmonicity and low entropy (e.g., tonal song). All other species were intermediate between ZF and LF. Finally, species showed no clear phylogenetic clustering or differentiation in the spectrotemporal axis (Fig. 3b,c; Supplementary Fig. S1). BF syllables generally scored lowest in spectrotemporal component values, characteristic of vocalizations with short duration and down-sweeps. All other species scored similarly to each other.

Species’ Syllables Differentially Overlap in Acoustic PC Feature Space along Cladistic Groupings

Species boundaries were generated in the acoustic PC space using support vector machine (SVM) algorithm (Fig. 4a,b). We obtained the Jaccard similarity index as a measure of the proportion of overlap between species in terms of the total space occupied by both (Supplementary Fig. S2). The greatest amount of shared volume (> 30%) occurred between CB and BF (37.02%) and between CB and RF (31.14%), while the smallest amount of overlap (< 5%) occurred between GW and DF (0.13%), GW and LF (4.13%), and GW and ZF (4.66%).

We also calculated both asymmetrical pairwise overlap indices for each species (i.e., what proportion of species X overlaps with species Y, and vice-versa; Fig. 4c). The largest overlaps were between RF in BF (92.02%) and RF in CB (90.47%), and the smallest overlaps were between GW in DF (0.31%) and DF in GW (0.21%). Spearman correlation showed asymmetrical feature space overlap between species; large overlap of species X in species Y did not indicate large overlap of species Y in species X (asymmetrical), and these asymmetrical indices of overlap were not correlated (Spearman rank correlation: r_s = 0.055, N = 21, P = 0.81).

The overlap results highlight two key findings regarding phylogenetic relationships. First, species tend to overlap more with close relatives than distant relatives. This is shown by the higher overlap percentages (hotter colors in Fig. 4c) among Australian species (upper left corner) and African species (lower right corner). Second, a considerable amount of each species’ volume overlapped with that of BF song (hotter colors going across the BF row in Fig. 4c). This result reflects the broad range of acoustic features in BF syllables, which spanned the acoustic ranges of both the African and Australian species.

Six Acoustic Features Yield 93% Classification Accuracy of Syllables to Species

Results from the feature selection procedure showed that 6 features were most informative for explaining species’ differences in song: the minimum-, maximum-, and mean- fundamental frequency, spectral flatness, mean frequency, and duration of a syllable (see Fig. 2c-e and Supplementary Fig. S3).

Once we determined the 6 acoustic features that were most important for discrimination, we used those features to train and test a random forest model for syllable classification. The 6-feature model classified species syllables with 92.89% accuracy. Comparing the 6-feature model performance with the 21-feature model using their weighted F₁ Scores showed no difference in performance (21-feature model: M ± SD = 0.97 ± 0.031; 6-feature model: M ± SD = 0.95 ± 0.050; two-tailed independent samples t test: t₁₂ = 1.01, P = 0.33).

Similar to the hypervolume overlap analysis, the random forest analysis showed asymmetrical patterns of misclassification between species (Fig. 5c); misclassification of species X syllables as species Y was not indicative of misclassification of species Y syllables as species X (asymmetrical). Asymmetrical misclassifications were not significantly correlated (Spearman rank correlation: r_s = 0.38, N = 21, P = 0.091). Misclassification was highest for ZF (7.09%), CB (14.70%), and RF (49.63%) syllables, while misclassification was lower than 5% for all other species.

We conducted under-sampling and resampling procedures and ran the random forest classification on the under- and re-sampled datasets to compare classification performance with these datasets to that of our original dataset (Supplementary Fig. S4). Results from the under-sampled data showed that model performance significantly decreased with reduced sample sizes (F_1UndSamp (M ± SD) = 0.84 ± 0.088; two-tailed independent samples t test: t₁₂ = 3.02, P = 0.011). For the 6 species that were under-sampled (i.e., all except RF), classification accuracy was lower with the under-sampled dataset than with the full dataset (Supplementary Fig. S4). For RF song, syllable classification was slightly (5%) better with the under-sampled dataset. Results from our resampled data suggest no change in model performance using the balanced sample sets (F_1SCUT (M ± SD) = 0.93 ± 0.032; two-tailed independent samples t test: t₁₂ = 0.84, P = 0.42). In most species, classification accuracy changed only slightly (< 3%) with under- and over-sampling consistent with their SCUT resampling (e.g., over-sampled species decreased in misclassifications, under-sampled species increased in misclassification). We found inconsistent change with the resampling of GW and LF songs following resampling, but these were minor (1.3% increase and 0.7% decrease in misclassification, respectively). In contrast, results showed a modest decrease in misclassification of CB syllables (from 15–9%) and a large decrease in misclassification of RF syllables (from 50–17%). The large improvement in RF syllable classification suggests that sample size affects model predictions more for this species than for others. Overall, the results of sample size manipulations showed that the accuracy of random forest classification was robust to differences in sample sizes, with the exception of one species.

In order to test the relationship between acoustic PC space overlap and syllable misclassification, we measured the Pearson correlation on the log-transformed indices for each pairwise species comparison. We found a moderate correlation between acoustic space overlap and syllable misclassification (Pearson correlation: r = 0.57, N = 42, P < 0.001). The results show a relationship between the pair-wise quantification of similarity (i.e., PC volume overlap) and the multi-species quantification of similarity (i.e., random forest misclassification, wherein syllables may be classified as any of the seven species), such that greater overlap in acoustic component space is indicative of greater syllable misclassifications between species (Fig. 5d). For example, the highest misclassification of RF syllables occurs with the other African species (i.e., CB and GW) and with the BF (Fig. 5c), consistent with high overlap in PC space between RF and these species (Fig. 4c).

Spectral Features Correlate with Phylogeny

Comparing the topologies of the dendrograms for acoustic and phylogenetic distance revealed that species clustered along spectral features consistent with their geographical distributions (i.e., Australian, African, and Southeast Asian; FMI = 1.00, P < 0.001; Fig. 6b). Post-hoc FMI tests on the acoustic features that load significantly onto spectral component (PC1) found significant similarity between biogeographical groupings and syllable starting dominant frequency (FMI = 0.73, P = 0.042). No other acoustic feature showed significant shared topology with the phylogeny (all P > 0.05; see Supplementary Table S1).

Clusters along spectral shape (FMI = 0.46, P = 0.15; Fig. 6c) and spectrotemporal (FMI = 0.52, P = 0.13; Fig. 6d) components did not match the species phylogeny. Within spectral shape features (PC2), tonal species (LF, RF, CB, GW) formed one cluster, while DF and BF formed a second cluster, and ZF formed its own cluster. The distinction between these clusters may involve harmonic banding (tonal vs. BF-DF clusters) and amount of nosiness (e.g., entropy, flatness) between harmonics (highest in ZF vs lower in all other species). Within spectrotemporal features (PC3), the BF cluster likely resulted from the large proportion of short duration syllables in the song repertoire of this species, while the RF cluster is due to the large proportion of longer-duration syllables in their song repertoire (as opposed to smaller proportions of song repertoire represented by very long syllables in LFs and CBs). All other species (intermediate between these two extremes) clustered together.

Results from the standard Mantel test demonstrate a significant correlation between phylogenetic distance and spectral (P = 0.006), but not spectral shape (P = 0.645) or spectrotemporal (P = 0.310), components. The Brownian motion EM-Mantel (one-tailed) test detected a significant phylogenetic signal in spectral components (r = 0.75, P = 0.015). The result suggests acoustic spectral relationships have evolved as predicted by the evolutionary time between species. Despite significant and robust omnibus results, post-hoc Mantel tests on the acoustic features that load significantly onto the spectral component (PC1) did not detect significant correlations between phylogenetic distance and any of the individual features (all P ≥ 0.058; see Supplementary Table S1).

We used computational models and direct measures of song acoustics to determine the features that accurately classify vocal communication signals by species, and to identify features that correlate with relatedness. By comparing the results of multiple approaches, we developed a framework for understanding how vocal acoustics convey social information. Results showed that: 1) song syllables can be described and compared by their spectral, spectral shape, and spectrotemporal features; 2) simple spectral features map onto phylogeny; 3) complex features vary within species and overlap between species; and 4) syllables can be accurately classified by species using six acoustic features. Previous neurophysiological and behavioral studies suggest that songbirds are sensitive to complex acoustics and manipulations in the relationship between acoustic features [38, 62, 65]. Generalization of findings to different species, however, has been criticized given the limited number of species and acoustic features typically studied, and the limited consideration of interactions between acoustic features in previous research [66]. Our paper aims to provide a standardized approach for future researchers to quantify the acoustic similarities between species, identify the acoustic features that optimize species-level classification, and determine which acoustic relationships may be associated with phylogeny. This song-evolutionary framework may be used to test the role of acoustic and phylogenetic relationships on species recognition in comparative studies.

Overall, our results show that estrildid songs can be meaningfully characterized in reduced acoustic dimensions from their spectral, spectral shape, and spectrotemporal components (Fig. 3). Hypervolume analyses showed that species occupy specific regions of acoustic feature space and exhibit variable amounts of overlap with each other, representative of the acoustic similarity in their song syllable PC features (Fig. 4) [3, 67]. In comparing species relationships along these components, we found that lower overlap in the three dimensions correlates with greater differentiability of syllables by their raw acoustic features (Fig. 5d). Phylogenetic analysis showed that clustering in spectral features follows cladistics groupings: Australian species cluster in the lower range of PC1 (i.e., these species generally produce lower pitch songs), African species in the higher range, and the Bengalese finch spans between these (Fig. 3a; Supplementary Fig. S1). Clustering in PCs 2 and 3, however, does not follow cladistic grouping.

Ensemble tree procedure for feature selection and model optimization showed that fundamental frequency (minimum, maximum, and mean), mean frequency, spectral flatness, and syllable duration are the features most important for classifying song by species. Most of these features (i.e., fundamental frequency; “entropy,” but see note on spectral flatness below) have been previously established for use in calculating similarity between songs for measurements of song imitation, variability, and change over time and experimental manipulations [48]. Spectral flatness is a measure of noisiness in the frequency domain of a signal, similar to spectral entropy [68] and Wiener entropy [48]. Mean frequency has previously been found to vary along genetic lines (i.e., across species [38] and strains [59]), but not rearing conditions when cross-fostered. These findings suggest that mean frequency could be used in classification by allowing variability in song within species ranges while still preserving species identities [69]. Previous studies with estrildids have not used duration as a diagnostic feature for species identity, despite its usefulness in categorizing vocalizations in song- [70] and non-song-learning [71] birds. Estrildid studies, however, have found that duration could be used in conjunction with other features to differentiate between syllables that otherwise overlapped in measures of entropy and fundamental frequency [72]. Fundamental frequency has also been shown to be used in categorical discrimination of vocalizations in estrildids [70]. Previous studies have also attributed variations in neural [52] and behavioral [70] response to these acoustic features in songbirds. Overall, the features identified as most important in song classification in our study parallels those previously identified in acoustic discrimination and preference research. Machine learning methods may be informative in further identifying the acoustic features that elicit species-recognition and neural tuning [73].

Phylogenetic signal analysis provides a measure of the amount of trait variation that can be explained by evolutionary relatedness [37] and has been found in species-level analyses of vocalizations across taxa, including birds [74–78], anurans [79], and mammals [80]. Acoustic features that identify species from field recordings allow measurement of biodiversity, migration, and habitat use [17–20]. Our analyses detected phylogenetic signal in spectral, but not spectral shape or spectrotemporal, components. Despite song being a learned individual behavior, previous studies suggest that spectral features of song are strongly heritable and evolve in populations via natural selection. Spectrotemporal characteristics are proposed to be less constrained due to the vocal production organ (i.e., the syrinx), which may provide a ubiquitous and flexible mechanism on which processes of cultural transmission (e.g., song learning) and sexual selection impose demands for faster rates of change [3, 81, 82]. Our results match previous studies in Regulus, a songbird genus that expresses phylogenetic signal in frequency features (PC1 in our study), and for which frequency bandwidth and modulation (related to PCs 2 and 3 in our study, respectively) were previously established as learned acoustic parameters [83]. Changes in song frequency features may be constrained by pleiotropic or polygenic traits that affect vocal production organs (e.g., body size affecting fundamental and dominant frequencies through syrinx size) [78] and other morphological features implicated in sound modification (e.g., the beak) [84]. These constraints would impart evolutionary rates on song frequency changes that correlate with time and reflect phylogenetic relationships (as predicted under Brownian motion), as opposed to faster rates of change – both between and within species – typically found in song features that freely shift under cultural transmission [85]. Although significant differences exist in syringeal morphology across birds [78], song learning in oscine birds – including estrildids – has been facilitated by an adaptable capacity for vocal production by conserving a uniform syringeal morphology that can meet the demands of culturally transmitted song [81] (but see [61, 78] for discussions on constraints).

Previous electrophysiological studies have found neural tuning to spectral shape (PC2) and spectrotemporal (PC3) modulations – but not to basic frequencies (PC1) – in regions critical for song learning (e.g., caudomedial nidopallium [38], and L3 [62]. Neurophysiological studies have characterized spectrotemporal tuning in auditory neurons and its role in song learning and perception [38, 65, 86, 87]. Moore and Woolley [38] found that neural selectivity for the spectrotemporal modulations is shaped by vocal learning during development, while basic spectral tuning measured from tone receptive fields is not. This study supports the hypothesis that basic spectral features are more conserved across individuals of the same species than are complex features [3, 61]. The vocal acoustics driving PC1 were more similar between closely related species than between distantly related species (Fig. 3a; Supplementary Fig. S1), suggesting that spectral features can be used to identify species and predict relatedness. This relationship between spectral features and phylogeny may function in species recognition, without visual or direct contact between signalers and receivers [88].

Significant phylogenetic signal using the EM-Mantel procedure suggests that divergence in those features (i.e., PC1) may be accounted for by the simple “random walk” of traits as a function of the evolutionary time between species (i.e., Brownian motion). The absence of phylogenetic signal, however, does not necessarily mean that traits evolve independently of evolutionary constraints. Spectral energy distribution (PC2) and spectrotemporal (PC3) features, which did not express significant phylogenetic signal under Brownian motion, may instead be evolutionarily labile and change under alternative selection pressures and processes within species [3, 61, 81, 82], specifically those important for vocal communication (e.g., individual recognition, mate attraction, intrasexual competition, and cultural evolution) [13, 69, 78, 89]. In this case, we would expect species to cluster along PCs 2 and 3 independent of their phylogenetic relatedness, and this result can be seen in the polyphyletic clustering of the species as well as in the divergence within the Australian (and to a lesser extent the African) species along PCs 2 and 3 (Fig. 3c; Supplementary Fig. S1).

Songbird vocalizations offer a diverse system for studying the effects of learning and evolutionary processes on behavior. Communication signals allow for coding, transmittance, and interpretation of social information between senders and receivers, and behavioral and neural studies in various animal taxa support this theory [1, 13, 31, 65, 90–92]. We therefore expect acoustic differences between species’ songs to scale with differences in perceptual mechanisms and to – in turn – affect processes of behavioral response, reproductive success, and ultimately speciation in estrildid finches [3, 6, 9, 13, 30, 31, 61, 93, 94]. The machine learning and statistical models used here identified the acoustic dimensions that best characterize and distinguish species’ songs. We also found that spectral features such as fundamental frequency correlated with phylogenetic distance, while spectral shape and spectrotemporal features did not. These findings suggest that basic spectral characteristics and more complex characteristics are generated by different mechanisms. Selection may conserve spectral features that identify the species of a signaler [84], while features such as spectral shape and spectrotemporal modulations may vary across conspecific individuals and overlap between species [3]. These more complex and variable features within a species may function to aid in individual recognition and competition for mates [95–97]. Further studies on the vocal acoustics in multiple suites of closely related species will further disentangle the effects of genetics and learning on vocal communication and guide our understanding of the mechanisms driving communication production and perception. Future studies should incorporate measures of syntax (the temporal sequencing of syllables) into quantitative analyses of species difference in song acoustics.

Ethics statement

All animal handling and use procedures were conducted under the guidelines and approval of the Institutional Animal Care and Use Committee (IACUC) of Columbia University. The study reported here was conducted in compliance with the recommendations in the ARRIVE guidelines.

Experimental Model and Subjects

We analyzed the acoustics of songs recorded from males in each of 7 estrildid species. Species were selected based on differences in song acoustics and phylogenetic relatedness [55, 78, 98] (Fig. 1). The songs of 3 African species (red-cheeked cordon-bleu [Uraeginthus bengalus], CB; red-billed firefinch [Lagonosticta senegala], RF; gold-breasted waxbill [Amandava subflava], GW), 3 Australian species (double-barred finch [Stizoptera bichenovii], DF; long-tailed finch [Poephila acuticauda], LF; zebra finch [Taeniopygia guttata castanotis], ZF), and 1 Southeast Asian species (Bengalese finch [Lonchura striata domestica], BF) were studied. All birds were housed in our lab in mixed-species colonies maintained at 27˚C, 30% humidity, and a 14 − 10 light/dark cycle with ad libitum access to food and water. Sound recording chambers were maintained at the same conditions.

Song Recording and Syllable Extraction

Adult (> 120 days post-hatch) male finches were individually placed in sound-attenuating chambers (MAC-1, Industrial Acoustics Company) equipped with microphones (MKE 2–60, Sennheiser) for song recording using Sound Analysis Pro Recorder (v2011; 44100 Hz sampling rate) [48]. Individuals were continuously recorded for 1–3 days, and audio files of singing were stored for analysis. For each species, 20 song motifs recorded from each of 10 birds were analyzed (i.e., 200 motifs per species), except when limited by availability. Songs of 4 CBs (80 total motifs for CB) and 3 RFs (for which one male only produced 8 motifs; 48 total motifs for RF) were analyzed. For GWs, in which motif structure is not clearly defined, we used uninterrupted bouts of song syllables (as described in[40]) with inter-syllable intervals (ISI) < 2500 ms (ISI, Median ± IQR = 842.38 ± 579.37 ms; Number of syllables per bout, Median ± IQR = 8 ± 5 syllables). Hereafter, procedures described for “motifs” apply to these bouts.

Song motifs were bandpass filtered (250–8000 Hz) and root-mean-square power matched (65 dB) prior to analysis. Motifs were segmented into syllables using the warbleR library [68] for R (v. 4.0.3) [99]. Syllable boundaries were visually inspected and adjusted as needed to ensure accurate segmentation. The complete dataset consisted of 15,081 song syllables with the following number of syllables for each species: BF: 3993; CB: 883; DF: 4322; GW: 1765; LF: 2821; RF: 207; ZF: 1090.

Data Preparation

Acoustic measurements were made from each syllable using the specan function in warbleR, which relies on the seewave package [100]. We measured 21 features to quantify spectral, temporal, and spectrotemporal acoustics of each syllable (Supplementary Table S1). Due to the diversity of syllable types, including syllables with minimal acoustic structure (i.e., resemble white noise), 3 of the 21 features could not be measured from some syllables (6.22% of all syllables per feature); warbleR did not detect a fundamental frequency for these syllables. In order to include these syllables in our analyses, estimates of these missing data (i.e., minimum-, maximum-, and mean- fundamental frequency) were imputed through regularized iterative principal component analyses (riPCA) for each species using the missMDA package [101] for R. riPCA provides estimates of the missing values by taking into account the similarities between syllables (within species) and the relationships between acoustic features [102]. Projection of the imputed data dimensions onto the incomplete data’s PCA dimensions suggested no significant deviation from the reference dimensions, and therefore no undue influence of the imputed data on the completed dataset. We therefore use the imputed data set in all subsequent analyses.

Principal Component and Volume Analysis

Given the high number of measured features and variation in multiple features [103], we performed principal component analysis (PCA) on the 21 acoustic features for the 15,081 syllables in our song dataset using the prcomp function in R [99]. PCA projects the distribution of syllable features into low-dimensional space and reveals the major underlying axes of variation in syllable acoustic structure. We used a feature loading criterion [104] (L_Crit, Eq. 1) to identify the raw acoustic features that significantly loaded to each principal component (PC; Supplementary Table S1) as follows:

$${L}_{Crit}= \sqrt{\frac{1}{f}},$$

where f is the number of raw acoustic features used in PCA. We found that the first 3 PCs met Kaiser-Guttman criterion [105] (i.e., Eigenvalues > 1), included significant loadings by all 21 of the raw features, and explained 72.28% of the variance in the data. We used these first 3 PCs in subsequent analyses. Using L_Crit, we explored the qualitative relationship between acoustic features that significantly loaded to each PC.

We next quantified and compared the acoustic space occupied by each species’ song syllables. We used the SVM method in the hypervolume package [106] for R to determine the boundaries of each species’ song in 3-dimensional PC space. SVM is a nonparametric classification algorithm that estimates the nonlinear one-class boundaries (e.g., species X vs. not species X) in the space of the original data [107]. Importantly, SVM generates the boundaries of acoustic space for each species that includes all data points, regardless of distribution of points within the space (Fig. 4a-b and Supplementary Fig. S2). This is in contrast to volumes estimated based on central tendencies or centers of mass [49, 107, 108].

Using these species boundaries, we quantified species volume and pairwise overlap between volumes using the Jaccard similarity index calculated by the hypervolume_overlap_statistics function [106]. The Jaccard similarity index is a measure of the proportion of shared space (i.e., intersection volume) to total space (i.e., union volume) occupied between two species. In order to quantify the proportion of shared-to-unique volume in acoustic feature space between species (e.g., what proportion of species X’s distribution is within species Y, and vice-versa) we calculated directional overlap indices between each species pair. For these indices, we divided the shared volume by the focal species’ volume. We used Spearman correlation to test whether the overlap between species was symmetrical (i.e., if large overlap of species X in species Y was indicative of large overlap of species Y in species X). Because of the random initialization of parameters during the SVM algorithm [107], within-species volume boundaries varied slightly across computations. Therefore, SVM volume boundaries and overlap indices were computed as above, stored, and averaged across 100 iterations.

Feature Selection and Random Forest Classification

After quantifying acoustic similarity between species using volume overlap in low-dimensional space, we tested which acoustic features were most relevant for discriminating one species’ syllables from others. We used a random forest algorithm in order to test the accuracy of syllable classification to a species using the 21 acoustic features [69, 109]. Random forest classification was conducted using the TreeBagger function in MATLAB [110], which implements Breiman’s algorithm [111], to create 1000 decision trees. Each decision tree in the random forest receives a subsample of syllables and acoustic features in the dataset, and learns a flowchart-like series of binary decisions to assign a species label to each syllable (see Fig. 5a). Importantly, species’ syllables were sampled for training with probabilities calculated by taking the average number of syllables per motif, per species. This standardization controls for the effects of unbalanced sample size on model performance by allowing the random forest to train and test itself using syllable proportions that are reflective of the natural occurrence in species song and not of the sampling effort in our study. Each decision tree in the random forest then casts a “vote” on the species label for each syllable, and the syllable is labeled according to the majority vote. TreeBagger uses two-thirds of the input data in order to build decision trees and train the classification model, and then tests model performance using the remaining one-third of the data.

We next tested the importance of each acoustic feature for accurately classifying syllables to species (Fig. 5) using TreeBagger with specified parameters to optimize feature selection. In this process, each tree was given the complete set of acoustic features to use in its decisions and sampled species’ syllables with equal probabilities. Splitting decisions at each node of the tree were made after accounting for predictive association between features. Specifically, the algorithm took into account the interaction between features as well as all other alternative feature splits possible at each node. Feature importance was computed by measuring the change in prediction error (“delta error”) that results if values of a given predictor are randomized (Fig. 5b and Supplementary Table S1). As such, predictors that are more important to maintaining classification accuracy would yield higher delta error values. TreeBagger calculates a single importance (i.e., delta error) value for each predictor across all decision trees. Like SVM, TreeBagger relies on the subset of data used to train the classification model. We therefore employed a custom script (SONg-TUUL) to compute, store, average, and plot the bootstrapped model metrics for 100 iterations of feature selection. We rank-ordered the importance scores for features and identified the inflection point [105] to find the features that were most important for distinguishing the syllables of species’ song. Our analysis showed that the first 6 rank-ordered features scored significantly higher than the rest of the acoustic features.

In order to test the efficacy of the classification model using only the 6 most important features discovered (“6-feature model”), we conducted 100 iterations of random forest training and testing using just these features as first described for the 21 features (“21-feature model”). We compared performance of the 6-feature model to the 21-feature model using the weighted F₁ Scores for each [112]. The F₁ Score takes into account model sensitivity (the proportion of all syllables produced by birds of species X that were correctly classified as belonging to species X) and precision (the proportion of all syllables classified as belonging to species X that were actually produced by birds of species X) and yields an index value between 0 and 1 for each species. We calculated the weighted average of this score across species (using number of syllables per species in our dataset as the weight) to obtain the model-wise F₁ Score for each of the random forest classification models. We tested the difference between the 6-feature and 21-feature models’ F₁ Scores across iterations using a Student’s independent samples t-test.

Because the number of syllables in the datasets for each species differed, we tested the effect of sample size on model performance. For this, we generated syllable datasets in which sample sizes were matched across species by under-sampling and resampling, and then compared their model-wise F1 scores from the random forest classification to that of our originally-sampled data, using the 6-feature model for all three datasets. First, we generated an under-sampled dataset by randomly selecting 207 syllables without replacement from the full dataset for each species other than RFs to match the RF sample size. Second, to test whether the model – and specifically the low-sample species – would improve classification with increased sample size, we generated a resampled dataset by applying a multi-class synthetic minority oversampling technique (SMOTE) with cluster-based under-sampling using the scutr package[113] in R; the complete procedure is referred to as SCUT [114]. The SCUT procedure balances multi-class data by first dividing the total sample size (15081 syllables) by the number of classes (7 species) to calculate a balanced target sample size of 2154 syllables for each species. Species that have sample sets larger than the target sample size are under-sampled by iteratively removing same-species syllables with the smallest distance to its neighbors. Species that have sample sets smaller than the target sample size are over-sampled through SMOTE procedure [115]. Briefly, SMOTE artificially grows sample sets by creating simulated values along the feature distance vector between nearest neighbors in the original dataset. As such, simulated values are confound to the original data’s region in multidimensional feature space, and these “synthetic” values prevent classification models from over-fitting on the (repeated) minority class samples as is done following traditional over-sampling techniques (e.g., resampling with replacement) [115].

Transformation and Association between Random Forest Misclassification and PC Volume Overlap

Because accurate syllable classification depends on the discriminability of each species’ syllable acoustics, we tested whether random forest performance using the discovered features correlated with the amount of overlap between species in acoustic PC space. Similar to the PC volume overlap indices, we calculated the asymmetrical pairwise misclassification indices from random forest as “the number of syllables belonging to species X that were incorrectly classified as species Y.” In order to test the association between the acoustic similarity of species’ song syllables in PC space and the discriminability of those same syllables based on the most important acoustic features discovered, we plotted and conducted a Spearman test on the log-transformed values of these indices. To account for indices equal to 0 (i.e., random forest misclassifications between GWs and DFs), we added a constant following Stahel’s quantile recommendation[116, 117] before log-transforming. Stahel’s recommended adjustment calculates the constant by dividing the squared first quantile by the third quantile of the data (excluding 0s). The transformed hypervolume overlap (T_HY) and random forest misclassification (T_RM) indices were calculated as follows:

$${T}_{HY}= \text{log}(H+ 0.05389)$$

$${T}_{RM}= \text{log}(R+ 0.0004216)$$

where H is the original hypervolume overlap index and R is the original random forest misclassification index between species. Finally, we plotted and used the transformed pairwise PC volume overlap and random forest misclassification indices for species X in species Y, and the pairwise transformed PC volume overlap and random forest misclassification indices for species Y in species X in our analysis.

Acoustic Feature Distances and Phylogenetic Analyses

To identify the acoustic features of song that correlate with evolutionary relatedness, we tested whether species distances along acoustic dimensions mapped on to the species phylogeny. To test for broad cladistic groupings of species by phylogeny, we used the Fowlkes-Mallows Index (FMI) [118] for comparing the hierarchical clustering (topology) in the phylogeny to the that of the acoustic distances. In order to acquire the hierarchical clustering in acoustic features, we first calculated the pairwise acoustic distances between species for each of our acoustic features (i.e., PCs 1–3 as well as our 21 acoustic features). The distributions of acoustic feature values were z-scored and binned spanning the minimum and maximum numbers of the data in increments of 0.01 in order to standardize the range and frequencies across features and species: 2501 bins for raw features and 1901 bins for PCs. We then calculated the pairwise distances between each species for each acoustic feature by calculating the sum of squared differences (hereafter “acoustic distances”) between the histograms, and used these values to produce the distance matrices for each acoustic feature. Acoustic distances were normalized to range between 0 (no distance) and 1 (most distance) and used to construct dendrograms using the hclust function for hierarchical clustering in R [99]. We used the FM_index function in the dendextend package [119] for R for FMI cluster analysis. We used k = 3 to test whether species will group into their biogeographical clades (i.e., Australian, African, and Southeast Asian clusters) when clustered by acoustic distances. FMI tests the null hypothesis that the acoustic feature dendrograms are random reshufflings of the phylogenetic tree and provides an index for similarity (0 [dissimilar] to 1 [similar]). We used this index as well as the provided mean and variance for the expectancy value in the pnorm function of R in order to calculate the one-tailed p-value under the null hypothesis [118, 119] (i.e., trees are dissimilar; rejected at α = 0.05 when the observed FMI is significantly greater than the critical value).

In addition to biogeographical phylogenetic groupings, we wanted to test if the acoustic components that explained the greatest amount of variance in syllables (i.e., PCs 1–3) suggest phylogenetic constraint (i.e., predicted by relatedness under models of evolution) [16, 82, 120] over lability (e.g., due to selection pressures not accounted for by the selected model of evolution) [6, 77]. Testing for phylogenetic signal allows us to detect patterns in acoustic feature distributions that correlate with phylogenetic relatedness [54, 121]. We produced a pruned phylogenetic tree for the species (see Fig. 1; Fig. 6a) extracted from a multi-locus estrildid phylogeny [55]. We used the acoustic distance matrices and the pruned phylogenetic tree in the EMMantel function (Brownian model; 1,000 permutations) provided by [122]. By incorporating an evolutionary model to the Mantel test (EM-Mantel), we tested the probability of obtaining the observed pairwise acoustic distances if these features evolved under the specified model of evolution (i.e., Brownian motion) [122]. The Brownian motion EM-Mantel procedure tests the hypothesis that the observed species acoustic feature relationships are greater than or equal to (i.e., one-tailed) those expected to arise as a simple function of the evolutionary time between species.

Phylogenetic signal testing via the Mantel test is a two-step procedure [122]. First, we conducted the standard Mantel test on the phylogenetic distance and the acoustic distance matrices of species. If the standard Mantel test detected a significant correlation between phylogenetic distance and acoustic distance, we ran the EM-Mantel test for phylogenetic signal detection on those features. We applied false discovery rate (FDR) correction on all tests involving multiple features. We first conducted the two-step EM-Mantel procedure on the acoustic feature PCs. We then conducted the two-step procedure on all raw acoustic features that significantly loaded onto the PCs that had statistically significant EM-Mantel results.

Data Availability

The analysis code and datasets used here are available from the corresponding author on request.

Acknowledgements

This research was supported by grants from the U.S. National Science Foundation (IOS-1656825) and National Institutes of Health (DC009810) to SMNW, fellowships from the City University of New York to MR, a National Research Service Award (F31DC020904) to JAE, and a Humboldt Foundation Research Award through the University of Bielefeld, Germany to MEH. Artwork credit to A.E. Johnson.

Author contributions

MR, JAE, MEH, and SMNW conceived the experiments and analyses. MR and JAE conducted the experiments and analyses. MR and SMNW wrote the manuscript text. All authors reviewed and edited the manuscript.

Additional information

Supplementary Information accompanies this paper at: http://www.nature.com/srep.

Competing financial interests: None.

Endler, J. A. Signals, signal conditions, and the direction of evolution. The American Naturalist 139, S125-S153 (1992).
Romer, H. Ecological Constraints for the Evolution of Hearing and Sound Communication in Insects. Evolutionary Biology of Hearing, 79-93 (1992).
Slabbekoorn, H., and Smith, T.B. Bird song, ecology and speciation. Philos T R Soc B 357, 493–503 (2002).
Mendelson, T. C. & Shaw, K. L. Sexual behaviour: Rapid speciation in an arthropod. Nature 433, 375-376, doi:10.1038/433375a (2005).
Bradbury, J. W. & Vehrencamp, S. L. Principles of Animal Communication. (Sinauer Associates, 2011).
Mason, N.A., Burns, K.J., Tobias, J.A., Claramunt, S., Seddon, N., and Derryberry, E.P. Song evolution, speciation, and vocal learning in passerine birds. Evol 71, 786-796 (2016).
Chen, Z. & Wiens, J. J. The origins of acoustic communication in vertebrates. Nat Commun 11, 369, doi:10.1038/s41467-020-14356-3 (2020).
Beecher, M. D., and Brenowitz, E. A. Functional aspects of song learning in songbirds. Trends in ecology and evolution 20, 143-149 (2005).
Catchpole C. K., and Slater P. J. B. Bird song: biological themes and variations. 2nd edn, (Cambridge University Press, 2008).
Seddon, N., Merrill, R. M. & Tobias, J. A. Sexually selected traits predict patterns of species richness in a diverse clade of suboscine birds. American Naturalist 171, 620-631, doi:10.1086/587071 (2008).
Riebel, K. . Song and female mate choices in zebra finches - a review. Adv Stud Behav 40, 197–238 (2009).
Ackermann, H., Hage, S. R. & Ziegler, W. Brain mechanisms of acoustic communication in humans and nonhuman primates: an evolutionary perspective. Behav Brain Sci 37, 529-546, doi:10.1017/S0140525X13003099 (2014).
Woolley, S. M. N., and Moore, J. M. Coevolution in communication senders and receivers: vocal behavior and auditory processing in multiple songbird species. Ann N Y Acad Sci 1225, 155-165 (2011).
Wilkins, M. R., Seddon, N. & Safran, R. J. Evolutionary divergence in acoustic signals: causes and consequences. Trends in Ecology & Evolution 28, 156-166, doi:10.1016/j.tree.2012.10.002 (2013).
Podos, J. & Patek, S. N. Acoustic Signal Evolution: Biomechanics, Size, and Performance. Animal Signaling and Function: An Integrative Approach, 175-203 (2015).
Demery, A.-J.C., Burns, K.J., and Mason N.A. Bill size, bill shape, and body size constrain bird song evolution on a macroevolutionary scale. Ornithol 138, 1-11 (2021).
Frouin‑Mouy, H. et al. Acoustic and visual cetacean surveys reveal year‑round spatial and temporal distributions for multiple species in northern British Columbia, Canada. Sci Rep 12 (2022).
Wood, C. M. et al. Using the ecological significance of animal vocalizations to improve inference in acoustic monitoring programs. Conserv Biol 35, 336-345 (2020).
Fournet, M. E. H. et al. Some things never change: multidecadal stability in humpback whale calling repertoire on Southeast Alaskan foraging grounds. Sci Rep 8 (2018).
Lewis, R. N., Williams, L. J. & Gilman, R. T. The uses and implications of avian vocalizations for conservation planning. Conserv Biol 35, 50-63 (2020).
Araujo, M. B., Pearson, R. G., Thuiller, W. & Erhard, M. Validation of species-climate impact models under climate change. Global Change Biol 11, 1504-1513, doi:10.1111/j.1365-2486.2005.01000.x (2005).
Betts, M. G. et al. Forest degradation drives widespread avian habitat and population declines. Nat Ecol Evol 6, 709-+, doi:10.1038/s41559-022-01737-8 (2022).
Forzieri, G., Dakos, V., McDowell, N. G., Ramdane, A. & Cescatti, A. Emerging signals of declining forest resilience under climate change. Nature 608, 534-+, doi:10.1038/s41586-022-04959-9 (2022).
McDowell, N. G. et al. Pervasive shifts in forest dynamics in a changing world. Science 368 (2020).
Morrison, C. A. et al. Bird population declines and species turnover are changing the acoustic properties of spring soundscapes. Nat Commun 12, 6217, doi:10.1038/s41467-021-26488-1 (2021).
Stewart, P. S. et al. Global impacts of climate change on avian functional diversity. Ecol Lett 25, 673-685, doi:10.1111/ele.13830 (2022).
Wiersma, Y. F. A review of landscape ecology experiments to understand ecological processes. Ecol Process 11 (2022).
Simmons, Andrea Megela. in Acoustic Communication (eds Andrea Megela Simmons, Richard R. Fay, & Arthur N. Popper) 1-14 (Springer New York, 2003).
Searcy W. A., and Nowicki S. The Evolution of Animal Communication: Reliability and Deception in Signaling Systems., ( Princeton University Press, 2005).
Dooling, R. J., Brown, S. D., Klump, G. M., and Okanoya, K. Auditory perception of conspecific and heterospecific vocalizations in birds: Evidence for special processes. J Comp Psychol 106, 20–28 (1992).
Hauber, M. E., Campbell, D. L., & Woolley, S. M. N. The functional role and female perception of male song in Zebra Finches. Emu-Austral Ornithology 110, 209-218 (2010).
Lynch, A. in Ecology and evolution of acoustic communication in birds (ed D. E. Kroodsma, and Miller, E. H.) 181-197 (Comstock Publishing Associates, 1996).
Catchpole, C. K. in Acoustic communication in birds Vol. 1 (ed D. E. Kroodsma, and Miller, E. H.) 297–319 (Academic Press, 1982).
Wiley, R. H., and Richards, D. G. Physical constraints on acoustic communication in the atmosphere: Implications for the evolution of animal vocalizations. Behav Ecol Sociobiol 3, 69-94 (1978).
Payne, R. B. in Handbook of the Birds of the World Vol. 15 (ed Andrew Elliott Josep del Hoyo, David A. Christie) 234-377 (Lynx Edicions, 2010).
Payne, R. B., Payne, L. L., Woods, J. L., and Sorenson, M. D. Imprinting and the origin of parasite–host species associations in brood-parasitic indigobirds, Vidua chalybeata. Animal behaviour 59, 69-81 (2000).
Zann, R. A. Variation in the Songs of Three Species of Estrildine Grassfinches. Emu-Austral Ornithology 76, 97-108 (1976).
Moore, J. M., and Woolley, S. M. N. Emergent tuning for learned vocalizations in auditory cortex. Nature neuroscience 22, 1469-1476 (2019).
Zann, R. A. The zebra finch: a synthesis of field and laboratory studies., ( Oxford University Press, 1996).
Payne, R. B., and Payne, L. L. Song mimicry and species associations of west African indigobirds Vidua with Quail‐finch Ortygospiza atricollis, Goldbreast Amandava subflava and Brown Twinspot Clytospiza monteiri. Ibis 136, 291-304 (1994).
Hosino, T., and Okanoya, K. Lesion of a higher-order song nucleus disruptsphrase level complexity in Bengalese finches. NeuroReport 11, 2091-2095 (2000).
Woolley, S. M. & Rubel, E. W. Bengalese finches Lonchura Striata domestica depend upon auditory feedback for the maintenance of adult song. J Neurosci 17, 6380-6390, doi:10.1523/JNEUROSCI.17-16-06380.1997 (1997).
Honda, E., and Okanoya, K. Acoustical and syntactical comparisons between songs of the white-backed munia (Lonchura striata) and its domesticated strain, the Bengalese finch (Lonchura striata var. domestica). Zoological Science 16, 319-326 (1999).
Glaze, C. M., and Troyer, T. W. Temporal Structure in Zebra Finch Song: Implications for Motor Coding. J Neurosci 26, 991-1005 (2006).
Troyer, T. W., Brainard, M. S., and Bouchard, K. E. Timing during transitions in Bengalese finch song: implications for motor sequencing. J Neurophysiol 118, 1556–1566 (2017).
Ondracek, J. M., and Hahnloser, R. H. R. in Insights from Comparative Hearing Research (ed C. Köppl, Manley, G. A., Popper, A. N., and Fay, R. R. ) 347-388 (Springer, 2013).
Lachlan, R. F., Van Heijningen, C. A., Ter Haar, S. M., and Ten Cate, C. Zebra finch song phonology and syntactical structure across populations and continents—a computational comparison. Front. Psychol. 7 (2016).
Tchernichovski, O., Nottebohm, F., Ho, C.E., Bijan, P., and Mitra, P.P. . A procedure for an automated measurement of song similarity. Anim Behav 59, 1167-1176 (2000).
Wohlgemuth, M.J., Sober, S.J., and Brainard, M.S. Linked control of syllable sequence and phonology in birdsong. J Neurosci 30, 12936-12949 (2010).
Chen, Y. & Sakata, J. T. Norepinephrine in the avian auditory cortex enhances developmental song learning. J Neurophysiol 125, 2397-2407, doi:10.1152/jn.00612.2020 (2021).
Rodriguez-Saltos, C. A. et al. Song preferences predict the quality of vocal learning in zebra finches. Sci Rep 13, 605, doi:10.1038/s41598-023-27708-y (2023).
Sober, S. J., Wohlgemuth, M. J., and Brainard, M. S. Central Contributions to Acoustic Variation in Birdsong. J Neurosci 28, 10370-10379 (2008).
James, L. S. & Sakata, J. T. Learning Biases Underlie "Universals" in Avian Vocal Sequencing. Curr Biol 27, 3676-3682 e3674, doi:10.1016/j.cub.2017.10.019 (2017).
MacLean, E.L., et al. How does cognition evolve? Phylogenetic comparative psychology. Anim Cogn 15, 223-238 (2012).
Olsson, U., and Alström, P. A comprehensive phylogeny and taxonomic evaluation of the waxbills (Aves: Estrildidae). Mol Phylogenet Evol 146, 106757 (2020).
Sorenson M. D., and Payne R. B. . A single ancient origin of brood parasitism in African finches: implications for host-parasite coevolution. Evolution 55, 2550–2567 (2001).
Sorenson, M. D., Balakrishnan C. N., and Payne R. B. Clade-limited colonization in brood parasitic finches (Vidua spp.). Syst Biol 53, 104–153 (2004).
Arnaiz-Villena A., Ruiz-del-Valle V., Gomez-Prieto P., Reguera R., Parga-Lozano C., and Serrano-Vela I. Estrildinae finches (Aves, Passeriformes) from Africa, South Asia and Australia: a molecular phylogeographic study. Open Ornithol J 2, 29–36 (2009).
Kagawa, H., Suzuki, K., Takahasi, M., and Okanoya K. Domestication changes innate constraints for birdsong learning. Behavioural Processes 106, 91-97 (2014).
Podos, J. A performance constraint on the evolution of trilled vocalizations in a songbird family (Passeriformes: Emberizidae). Evolution 51, 537-551 (1997).
ten-Cate, C. in Nature’s music: the science of birdsong (ed P. R. Marler, and Slabbekoorn, H.) 300-321 (Elsevier Academic Press, 2004).
So, N. L. T., Edwards, J. A., and Woolley, S. M. N. Auditory Selectivity for Spectral Contrast in Cortical Neurons and Behavior. J Neurosci 40, 1015-1027 (2020).
Vignal, C., Mathevon, N., and Mottin, S. Mate recognition by female zebra finch: Analysis of individuality in male call and first investigations on female decoding process. Behavioural Processes 77, 191-198 (2008).
Chen, Y., Clark, O., and Woolley, S. C. Courtship song preferences in female zebra finches are shaped by developmental auditory experience. Proc. R. Soc. B 284 (2017).
Woolley, S. M., Fremouw, T. E., Hsu, A., & Theunissen, F. E. . Tuning for spectro-temporal modulations as a mechanism for auditory discrimination of natural sounds. Nature neuroscience 8 (2005).
Burkett, Z., Day, N., Peñagarikano, O., Geschwind, D. H., and White, S. A. VoICE: A semi-automated pipeline for standardizing vocal analysis across models. Sci Rep 5 (2015).
Marler, P. R., and Slabbekoorn, H. Nature's music: the science of birdsong. (Academic Press/Elsevier, 2004).
Araya-Salas, M., and Smith-Vidaurre, G. . warbler: an r package to streamline analysis of animal acoustic signals. Methods Ecol Evol 8, 184-191 (2017).
Mann, D.C., Lahti, D.C., Waddick, L., and Mundinger, P.C. House finches learn canary trills. Bioacoustics, doi:10.1080/09524622.2020.1718551 (2020).
Vicario, D.S., Raksin, J.N., Naqvi, N.H., Thande, N., and Simpson, H.B. The relationship between perception and production in songbird vocal imitation: What learned calls can teach us. J Comp Physiol A 188, 897-908 (2002).
Beckers, G.J.L., Goossens, B.M.A., and Ten Cate, C. Perceptual salience of acoustic differences between conspecific and allospecific vocalizations in African collared-doves. Anim Behav 65, 605-614 (2003).
Wu, W., Thompson, J. A., Bertram, R., and Johnson, F. A statistical method for quantifying songbird phonology and syntax. J Neurosci 174, 147-154 (2008).
Brainard, M.S., and Doupe, A.J. Translating birdsong: Songbirds as a model for basic and applied medical research. Annu. Rev. Neurosci. 36, 489-517 (2013).
Krishnan, A., and Tamma, K. Divergent morphological and acoustic traits in sympatric communities of Asian barbets. R. Soc. open sci 3 (2016).
Slabbekoorn, H., and Smith, T.B. Habitat-dependent song divergence in the little greenbul: an analysis of environmental selection pressures on acoustic signals. Evolution 56, 1846-1858 (2002).
Feher, O., Ljubiac, I., Suzuki, K., Okanoya, K., and Tchernichovski, O. Statistical learning in songbirds: from self-tutoring to song culture. Phil. Trans. R. Soc. B 372 (2017).
McCracken, K.G., and Sheldon, F.H. Avian vocalizations and phylogenetic signal. P Natl Acad Sci USA 94, 3833-3836 (1997).
Arato, J., and Fitch, W.T. Phylogenetic signal in the vocalizations of vocal learning and vocal non-learning birds. Phil. Trans. R. Soc. B 376 (2021).
Amezquita, A. et al. Calls, colours, shape, and genes: a multi-trait approach to the study of geographic variation in the Amazonian frog Allobates femoralis. Biol J Linn Soc 98, 826-838 (2009).
Garcia-Navas, V. & Blumstein, D. T. The effect of body size and habitat on the evolution of alarm vocalizations in rodents. Biol J Linn Soc 118, 745-751 (2016).
Prum, R. O. Syringeal morphology, phylogeny, and evolution of the neotropical manakins (Aves, Pipridae). American Museum novitates no. 3043 (1992).
Forstmeier, W., Burger, C., Temnow, K., and Derégnaucourt, S. The genetic basis of zebra finch vocalizations. Evol 63, 2114-2130 (2009).
Päckert, M., Martens, J., Kosuch, J., Nazarenko, A.A., and Veith, M. Phylogenetic signal in the song of crests and kinglets (Aves: Regulus). Evolution 57, 616–629 (2003).
Podos, J. Correlated evolution of morphology and vocal signal structure in Darwin's finches. Nature 409, 185-188, doi:Doi 10.1038/35051570 (2001).
Grant, B.R., and Grant, P.R. Cultural inheritance of song and its role in the evolution of Darwin's finches. Evol 50, 2471-2487 (1996).
Sen, K., Theunissen, F. E. & Doupe, A. J. Feature analysis of natural sounds in the songbird auditory forebrain. J Neurophysiol 86, 1445-1458, doi:10.1152/jn.2001.86.3.1445 (2001).
Yanagihara, S. & Yazaki-Sugiyama, Y. Auditory experience-dependent cortical circuit shaping for memory formation in bird song learning. Nat Commun 7, 11946, doi:10.1038/ncomms11946 (2016).
van Buskirk, J. Evolution of song structure and note structure in American wood warblers. Proc. R. Soc. B 264, 755-761 (1997).
Price, J.J., and Lanyon, S.M. Reconstructing the evolution of complex bird song in the Oropendolas. Evol 56, 1514-1529 (2002).
Prather, J. F. Auditory signal processing in communication: perception and performance of vocal sounds. Hearing Research 305, 144-155 (2013).
Wilczynski, W., Rand, A. S., and Ryan, M. J. Evolution of calls and auditory tuning in the Physalaemus pustulosus species group. Brain, Behavior and Evolution 58, 137-151 (2001).
Ryan, M. J. Neuroanatomy influences speciation rates among anurans. PNAS 83, 1379-1382 (1986).
Cynx, J. Conspecific song perception in zebra finches (Taeniopygia guttata). Journal of Comparative Psychology 107 (1993).
Searcy W. A., and Andersson M. Sexual selection and the evolution of song. Ann Rev Ecol Syst 17, 507–533 (1986).
Woolley, S.M.N., Hauber, M.E., and Theunissen, F.E. Developmental experience alters information coding in auditory midbrain and forebrain neurons. Dev Neurobiol 70, 235-252 (2010).
Clayton, N.S. Assortive mating in zebra finch subspecies, Taeniopygia guttata guttata and T. g. castanotis. Phil. Trans. R. Soc. B 330, 351-370 (1990).
Mandelblat-Cerf, Y., and Fee, M. S. An Automated Procedure for Evaluating Song Imitation. PLoS ONE 9 (2014).
Laiolo, P., Rolando, A., Delestrade, A., and de Sanctis, A. Geographical variation in the calls of the choughs. Condor 103, 287–297 (2001).
R. version 4.0.3. Vienna, Austria: R Foundation for Statistical Computing (2020).
Sueur, J., Aubin, T., and Simonis, C. Seewave, a free modular tool for sound analysis and synthesis. Bioacoustics 18, 213-226 (2008).
Josse, J., and Husson, F. missMDA: a package for handling missing values in multivariate data analysis. J Stat Softw 70, 1-31 (2016).
Dray, S., and Josse, J. Principal component analysis with missing values: a comparative survey of methods. Plant Ecol 216 (2015).
Tachibana, R. O., Oosugi, N., and Okanoya, K. Semi-Automatic Classification of Birdsong Elements Using a Linear Support Vector Machine. PLoS ONE 9 (2014).
Holland, S. M. Data Analysis in the Geosciences: Principal Components Analysis. Retrieved from http://strata.uga.edu/8370/lecturenotes/principalComponents.html (2021).
Jackson, D.A. Stopping rules in principal components analysis: a comparison of heuristical and statistical approaches. Ecol 74, 2204-2214 (1993).
Blonder, B. hypervolume: high dimensional geometry and set operations using kernel density estimation, support vector machines, and convex hulls. R package version 2.0.12. (2019).
Blonder, B., Morrow, C.B., Maitner, B., Harris, D.J., Lamanna, C., Violle, C., Enquist, B.J., and Kerkhoff, A.J. New approaches for delineating n-dimensional hypervolumes. Methods Ecol Evol 9, 305-319 (2018).
James, G., Witten, D., Hastie, T., and Tibshirani, R. An introduction to statistical learning with applications in R. 2 edn, (Spinger, 2021).
Youngblood, M., and Lahti, D. Content bias in the cultural evolution of house finch song. bioRxiv, doi:10.1101/2021.03.05.434109 (2021).
MATLAB. version 9.9.0.1592791 (R2020b). Natick, Massachusetts: The MathWorks Inc. (2020).
Breiman, Leo. Random Forests. Machine Learning 45, 5-32, doi:10.1023/A:1010933404324 (2001).
Xie, J., Hilal, A. R., and Kulic, D. Driving maneuver classification: A comparison of feature extraction methods. IEEE Sens J 18, 4777-4784 (2018).
scutr: Balancing multiclass datasets for classification tasks v. 0.1.2 (2021).
Agrawal, A., Viktor, H.L., and Paquet, E. SCUT: Multi-class imbalanced data classification using SMOTE and cluster-based undersampling. Comm Com Inf Sc 1, 226-234 (2015).
Chawla, N.V., Bowyer, K.W., Hall, L.O., and Kegelmeyer, W.P. SMOTE: Synthetic minority over-sampling technique. J Artif Intell Res 16, 321-357 (2002).
Stahel, W.A. Statistische Datenanalyse: Eine Einführung für Naturwissenschaftler., (Vieweg Verlag, 2000).
Wunder, J., Brzeziecki, B., Żybura, H., Reineking, B., Bigler, C., and Bugmann, H. Growth-mortality relationships as indicators of life-history strategies: a comparison of nine tree species in unmanaged European forests. Oikos 117, 815-828 (2008).
Fowlkes, E.B., and Mallows, C.L. A method for comparing two hierarchical clusterings. J Am Stat Assoc 78, 553-569 (1983).
Galili, T. dendextend: an R package for visualizing, adjusting, and comparing trees of hierarchical clustering. Bioinformatics 31, 3718-3720 (2015).
De Kort, S.R., and Ten Cate, C. Response to interspecific vocalizations is affected by degree of phylogenetic relatedness in Streptopelia doves. Anim Behav 61, 239-247 (2001).
Münkemüller, T., Lavergne, S., Bzeznik, B., Dray, S., Jombart, T., Schiffers, K., and Thuiller, W. How to measure and test phylogenetic signal. Methods Ecol and Evol 3, 743-756 (2012).
Debastiani, V.J., and da Silva Duarte, L. Evolutionary models and phylogenetic signal assessment via Mantel test. Evol Biol 44, 135-143 (2017).

No competing interests reported.

RiveraetalSupplementalSongAcoustics20230217.docx

Download PDF

Journal Publication

published 01 May, 2023

Read the published version in Scientific Reports →

Editorial decision: Major revision
08 Mar, 2023
Reviews received at journal
25 Feb, 2023
Reviewers agreed at journal
22 Feb, 2023
Reviewers invited by journal
22 Feb, 2023
Editor assigned by journal
22 Feb, 2023
Editor invited by journal
21 Feb, 2023
Submission checks completed at journal
21 Feb, 2023
First submitted to journal
20 Feb, 2023

You are reading this latest preprint version

Machine learning and statistical classification of birdsong links vocal acoustic features with phylogeny

Status:

Journal Publication

Version 1

Abstract

Figures

Introduction

Results

Spectral, Spectral Shape, and Spectrotemporal Components Explain 72% of Variation in Song

Species’ Syllables Differentially Overlap in Acoustic PC Feature Space along Cladistic Groupings

Six Acoustic Features Yield 93% Classification Accuracy of Syllables to Species

Spectral Features Correlate with Phylogeny

Discussion

Methods

Ethics statement

Experimental Model and Subjects

Song Recording and Syllable Extraction

Data Preparation

Principal Component and Volume Analysis

Feature Selection and Random Forest Classification

Transformation and Association between Random Forest Misclassification and PC Volume Overlap

Acoustic Feature Distances and Phylogenetic Analyses

Declarations

References

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 1