Tonality over a broad frequency range is linked to vocal learning in birds

Abstract

Many birds emit tonal song syllables even though the sound sources generate sound with rich upper harmonic energy content. This tonality is thought to arise in part from dynamically adjusted filtering of harmonic content. Here, we compare tonality of song syllables between vocal learners and non-learners to assess whether this characteristic is linked to the increased neural substrate that evolved with vocal learning. We hypothesize that vocal learning ability is correlated with enhanced ability for generating tonal sounds, because vocal production learners might also have an enhanced ability to articulate their vocal tracts and sound source for producing tonality. To test this hypothesis, we compared vocal learners and non-learners from two groups (186 passerines and 42 hummingbirds) by assessing tonality of song syllables. The data suggest that vocal learners in both clades have evolved to sing songs with higher tonality than the related, non-vocal learning clades, which is consistent with stronger roles for broadband dynamic filtering and adjustments to the sound source. In addition, oscine songs display higher tonality than those of hummingbirds. A complex interplay of vocal tract biomechanics, anatomical differences of the sound source as well as increased motor control through vocal learning facilitates generation of broad tonality.

1. Introduction

Vocal production requires coordinated activity of the motor systems driving respiration, the vocal organ and upper vocal tract structures. Together this motor action gives rise to highly complex vocal behaviours. In birds, the most diverse vocal behaviour is song, which is used in reproductive contexts and is therefore thought to be under strong sexual selection [1]. Vocal learning, the ability to imitate sounds acquired during development, appears to be a key adaptation that allowed some taxa to master large repertoires of song and highly complex sound features [1–3]. Vocal learning is a prominent aspect of vocal development in oscine passerines, some hummingbirds and in parrots [4–6].

All birds investigated so far produce most of their syringeal vocal signals as harmonic signals with a fundamental frequency and multiple upper harmonics [7,8]. Many birds are able to centre most of the acoustic energy on a single harmonic, most often the fundamental frequency (= 1st harmonic) [9]. The amplitudes of the other harmonics decrease with increasing harmonic number. Here, we quantify the decay of harmonic energy from fundamental to second harmonic and refer to the difference as tonality. Many songbirds have achieved a high degree of tonality, which can be attributed to two major factors: the vibratory behaviour of the sound source and upper vocal tract filtering (figure 1).

Figure 1.

Tonality of vocalizations is determined by the vibratory behaviour of the syringeal vocal folds (lateral and medial labia) (a) and by the upper vocal tract filter properties (b). (a) The energy content across the harmonic spectrum is determined by the oscillatory motion of the labia (a complex pattern of motion determines the open and closed time of a labial pair, top panel). For example, a changing ratio of open and closed time, as is observed for different fundamental frequencies, also changes the tonality of the generated sound (as indicated schematically in the lower panels). (b) The tracheal tube constitutes a relatively static resonance, but dynamic adjustments in the volume of the oropharyngeal-oesophageal cavity (OEC) enable birds to match resonances to the fundamental frequency of sounds and thus maintain tonality across a broad range. If the resonance frequency of the upper vocal tract can follow dynamically the F0 by adjusting the volume of the oropharyngeal-oesophageal cavity, it is possible to filter out upper harmonics across a broad frequency range and, thus, to generate tonal frequency sweeps (as indicated in the frequency over time panel and sketches of morphological correlates of extreme positions). Thus, tonality is high across the entire F0 range and not just in a small spectral range of F0. The more precisely changes in resonance frequency match the changes in fundamental frequency, the more consistent tonality can be achieved (inset). X-ray images were taken from [7]. (Online version in colour.)

Airflow across the syringeal constriction sets laterally positioned labia (i.e. avian vocal folds) into vibration [10,11]. Complex, non-sinusoidal oscillation characteristics generate harmonic frequencies, the amplitudes of which decrease with increasing harmonic number (figure 1a) [12–15]. The specific vibratory characteristics impact the energy distribution across the spectrum, and, typically, high-frequency sounds contain less energy in the upper harmonics than low-frequency sounds. Whereas this mechanism appears to be the predominant vibratory regime for song production, additional regimes of vibratory behaviour of the labia can generate different energy distributions across the frequency range. Among these are nonlinear phenomena, which are characterized by aperiodic oscillatory behaviour with broad distribution of energy across the frequency spectrum [16]. The use of nonlinear phenomena has been documented in bird songs, but typically nonlinear elements comprise only very short segments and generally are not a major aspect of bird songs.

Once sound is generated in the syrinx, it passes through the vocal tract. The spectral properties of the source signal are further shaped by the vocal tract filter properties. Investigations in songbirds suggest that many species match the first vocal tract resonance to the fundamental frequency of the song [7,17]. The position of the syrinx at the tracheobronchial junction provides birds with a longer effective vocal tract than similar-sized mammals, and birds therefore can more easily match vocal tract resonances to source-generated frequencies [18]. Birds modify harmonic content during the passage through the upper vocal tract, consisting of the tracheal tube, larynx, oropharyngeal-oesophageal cavity (OEC) and the beak [17,19]. While the tracheal tube constitutes a mostly static filter [20], the resonant frequencies of the OEC can be adjusted by changes in geometry or volume [17,21,22]. The volume of the OEC constitutes the main resonating chamber, which can either be static and not change its volume, or it can be dynamically regulated through movements of the hyoid skeleton (figure 1b) [17]. A low volume of the OEC results in higher resonant frequencies, while an increased volume shifts the resonance to lower frequencies [7].

Dynamic adjustments of the filter properties can lead to different acoustic effects. For example, in human speech and some songbird species, such as the zebra finch (Taeniopygia guttata), upper harmonic content is emphasized and dynamically altered [17,23–25], and the resulting formant structure conveys information. Alternatively, several species of songbirds have been shown to dynamically adjust OEC volume to produce resonance frequencies (F1 in figure 1b) that track the fundamental frequency (F0 in figure 1b) of the song and thus filter out energy in harmonic overtones [7,17,26]. For sounds with rapidly modulated fundamental frequencies, the resonance frequency tracks the changing fundamental frequency, which enhances the difference in relative intensity between neighbouring harmonics, usually fundamental frequency (F0) and second harmonic (2F0) (figure 1a,b), resulting in increased tonality. The accurate dynamic adjustment of the resonance frequency to the changing fundamental frequency requires precise changes in the volume of and opening to the OEC [22,27], which in turn requires complex motor control of the hyoid musculoskeletal system [15].

Some of the most complex motor patterns of respiration and the vocal organ, the syrinx, have been attributed to vocal production learning (i.e. the ability to imitate acquired sounds) [8,28]. The mechanism to modulate fundamental frequency depends on motor regulation of vocal fold tension and several other parameters [8]. Thus, an expanded range of fundamental frequency requires more sophisticated motor control at the syringeal level. Additionally, it may require complex and dynamic patterns of the upper vocal tract structures that modify sound as it travels from the source (syrinx) to the mouth where it is emitted, but little is known about this latter aspect. The ability to produce tonal sounds while rapidly modulating fundamental frequency has been studied extensively in regard to its evolutionary importance. A trade-off between frequency bandwidth and trill rates was attributed to performance limits in dynamic adjustments of the upper vocal tract [28–33]. Playback studies showed that tonal syllables elicit stronger responses from males than the same syllables with increased energy in upper harmonics [34,35], thus illustrating the perceptual importance of tonal song quality.

The circuitry for song production and learning has been well characterized for respiratory and syringeal control [4,6,36] and is much more elaborate than in non-learning birds [37,38]. It is less well understood how motor control of upper vocal tract structures is linked to the song control network, but spatial proximity between the terminal field of song system output and jaw pre-motor neurons suggests a direct connection [39]. It is therefore possible that upper vocal tract motor control is linked to the elaborate forebrain circuitry in taxa that learn to sing, thus suggesting the potential for more elaborate control of filter properties than is possible in non-learning taxa. Interestingly, young sparrows tutored with song syllables with high upper harmonic content still sing mostly tonal syllables as adults [40], suggesting that tonality develops irrespective of the harmonic structure of acquired models during song ontogeny. If vocal learners employ a more dynamic vocal tract filter and/or sound source, we expect that they will demonstrate greater tonality across their entire fundamental frequency range.

We set out to test the hypothesis that vocal learning facilitates higher tonality (i.e. the ratio between the intensity of the first and second harmonic, F0 and 2F0; figure 1a,b) across the frequency range in the songs of vocal learning than in non-learning groups. We first compare tonality in songs of closely related passerines (suboscines versus oscines), for which song learning ability has been tested at least in some species. Most suboscine passerine groups are thought to develop song innately, with clear demonstrations for Tyrannidae [37,41] and Thamnophilidae [42]. We then extend these analyses to hummingbirds, where song learning is present in some species but not in others, while for most species, we do not know how song develops. If we do not find a difference between vocal production learners and taxa with innate song development, the hypothesis is clearly falsified. As laid out above, in the case of a difference, the results need to be interpreted carefully as to whether upper vocal tract filtering dynamics provide a functional link between tonality and vocal production learning.

2. Methods

From the vocal learning oscines, we investigated the songs of 10 species of Passerellidae and three species each of seven additional oscine families. Ten species of Tyrannidae and three species of additional seven suboscine families were used as non-learning species for comparison (electronic supplementary material, table S1). Altogether, we analysed songs of 62 species from 16 families. Among the passerellid and tyrannid species, we matched species in terms of body size and habitat and geographical distribution (North America). All selected species from the other families are native to North and South America. For all species, we chose three recordings from three different males for analysis. We restricted our analysis to male song because female song has not been recorded as much and is not available for such a broad comparison.

As an independent test group, we chose hummingbirds, because they contain both learning and non-learning species [4]. We compared seven species of each, with three individuals per species, comprising 42 individuals overall (electronic supplementary material, table S2). The vocal learning species are Calypte anna [43], Aphantochroa cirrochloris and Glaucis hirsutus [6], Lampornis clemenciae [44], Colibri coruscans and Colibri thalassinus [45], and Calypte costae [46]. The non-learning species are Archilochus colubris and Selasphorus sasin [4], Selasphorus rufus, Selasphorus calliope, Selasphorus flammula, Selasphorus platycercus and Archilochus alexandri [47]. Only a small number of hummingbird species has been identified as vocal learners or non-learners, the learning status of most species is still unknown [46].

Additionally, we quantified tonality in songs of hummingbird species that have not been identified as vocal learners or non-learners (electronic supplementary material, table S3). We analysed recordings of 25 species from all eight hummingbird clades [48] with three individuals per species, except for the clade of bee hummingbirds, since most of the species with known learning status (learners and non-learners) are in this clade. The data were used to predict if these species could be grouped with the confirmed learning or non-learning species.

The recordings were obtained from the online database www.xeno-canto.org and analysed using Praat (v. 6.0.43) [49]. Files with sample rates of at least 44.1 kHz or higher were chosen to include high-frequency harmonic content in the analysis. All files were manually inspected for sufficient quality to qualify for the following analysis, with low background noise and no interference from other birds. The recording quality ratings on xeno-canto were used as guides for the file selection, but were still inspected before analysis. The ratings are entered by the uploaders and are therefore subjective (audio clarity rating guidelines for uploaders: A, loud and clear; B, clear, but bird a bit distant, or some interference with other sound sources; C, moderately clear, or quite some interference; D, faint recording, or much interference; E, barely audible). Most files (226/312) were chosen from the highest quality tier ‘A’, some files from tier ‘B’ (72/312). In a few cases, recordings of sufficient quality from tiers ‘C’ or lower were chosen (14/312), if recordings of higher quality tiers were not available to reach three recordings per species. Degree of tonality was calculated as the difference in amplitude between the fundamental frequency and the second harmonic using power spectra. High tonality indicates in part high ability to dynamically adjust filter properties. Less tonality indicates a more static or less dynamic vocal tract filter.

To make sure that compression of files to the MP3 format did not decrease the precision of the measurements [50], we tested the effects of compression with one synthetically generated sound, consisting of frequency sweeps, as well as song recordings of three passerellid and one tyrannid species recorded in Utah (Marantz PMD 661; Sennheiser MKE 600; 44.1 kHz sample rate). The compression reduces the frequency range, such that measurements of second harmonics were not possible for fundamental frequencies above 10 kHz. We then measured harmonic energy as described above for the uncompressed and compressed versions of the same files for fundamental frequencies below 10 kHz. The most noticeable change occurring during compression was a time shift. Once we corrected for this shift, the two files yielded very similar data (differences in frequency measurements: mean ± 1 s.d. = 3 ± 14 Hz; difference between measurements of tonality between the two files: mean difference ± 1 s.d. = 0.06 ± 0.79 dB; n = 37). These data show potential error margins, and we only interpret differences substantially greater than these values as biologically meaningful differences.

The tonality of each bird was sampled across the entire range of fundamental frequencies present in each recording. Measurements were made manually for each recording, since song repertoires can include multiple syllables that overlap in fundamental frequencies, vary in structure, and often do not cover the entire spectrum between the lower and upper limits of a bird's frequency range. Frequency modulated song syllables were prioritized over more static syllables, when syllables occupied the same frequency range, to predominantly measure syllables with dynamic filter adjustments in the birds' vocal range. Noisy or inharmonic syllables where fundamental frequencies or harmonics could not be clearly identified were not measured. Within syllables, sample points were selected in steps between 200 and 500 Hz (with time in between steps being dependent on the frequency modulation rate) to collect 10–15 data points per bird. The short on- and offsets of syllables were excluded from measurement, as they are generally of lower amplitude and may contain transient adjustments of sound source and vocal tract that do not represent optimal performance during the audible parts of a syllable. On- and offset times vary but are generally very short (1–2 ms) and probably do not play a large role in the assessment of tonality by listening birds. The mean of all differences across the frequency range was calculated for each individual to assess tonality. If tonality remains high throughout a call/syllable with modulated F0, we infer that the filter was adjusted to follow the fundamental frequency (figure 1b) whereas more variable expression of tonality may indicate a more static filter setting.

To evaluate whether tonality may be enhanced by the use of dynamic and static filters, we calculated the coefficient of variation for all measurements of each individual. Lower values indicate a greater matching between F0 and F1, indicating the ability to adjust the vocal tract filter dynamically. Additionally, the F0 range and F0-centre mean (difference between the highest and the lowest measured F0), and the average tonality (mean ΔdB value) were calculated. A mixed linear model (RStudio software, package: lme4, function: lmer) was calculated with family and species as random factors (linear mixed model fit by REML). T-tests use Satterthwaite's method; formula: ΔdB ∼ Group + (1 | family) + (1 | species)).

Body mass is an additional variable that explains part of the variation in sound frequency and song frequency range [51] and influences the dimensions of the vocal tract and therefore filter characteristics. To assess to what degree body size differences between the respective groups in the two comparisons may influence the outcome, we computed regressions between body mass and centre frequencies of the range of the song repertoire of each individual. Mean body mass data were taken from the literature [52,53].

3. Results

(a) . Tonality

First, we tested whether tonality was predictive for vocal learning in passerines. Tonality was significantly higher in vocal learners than in non-learners (figure 2a). Even when including species and family as random effects, a linear mixed model rates the difference in filtering efficiency as highly significant (electronic supplementary material, table S4: linear mixed model in R; p < 0.001). The mean difference between fundamental and second harmonic for vocal learners was 36.05 dB and for non-learners 23.68 dB, respectively. In both groups, there was substantial variation between species (figure 2a), and the ranges of the two groups show considerable overlap. Because maintaining tonality across a broad frequency range requires dynamic filtering, the respective frequency ranges also need to be considered. Despite having larger frequency ranges (vocal learners: 3.269 kHz; non-learners: 2.014 kHz; t-test: p = 0.0003) and overall higher fundamental frequency of vocal repertoires (vocal learners: 4.313 kHz; non-learners: 3.413 kHz; t-test: p < 0.001), oscines achieve greater tonality than suboscines.

Figure 2.

(a,b) Tonality (ddB = dBF0-dB2F0) is greater in vocal learners than non-learners (a, passerines; b, hummingbirds), but a substantial overlap in range exists between the datasets. (c) Measurements for hummingbird species of unknown learning status are dispersed across the range outlined by those of known learning status (shown at the ends). (d) Oscine songs (green) exhibit greater tonality than those of suboscines (red) throughout the frequency range (total n = 2089). Each data point represents a ΔdB value and the fundamental frequency at which it was measured. Lines are smoothed conditional means. (RStudio; ggplot2: geom_smooth (method = ‘gam’). (Online version in colour.)

Next, we repeated the analysis for hummingbirds who are distantly related with passerines. The boundary between learners and non-learners is less clearly defined than in passerines, however evidence for vocal learning capability exists in some species. The results display similar differences between vocal learners and non-learners as in passerines (figure 2b). The vocal learning hummingbirds show overall significantly higher tonality than non-learners (vocal learners: 26.77 dB; non-learners: 12.91 dB; t-test: p < 0.001), and the difference in tonality is very close to the difference measured in passerines. Non-learning hummingbirds tend to have larger frequency ranges than learning species (learners: 2.904 kHz ± 0.221; non-learners: 4.729 kHz ± 0.280). The F0-centre means are higher in learning hummingbirds (learners: 6.302 kHz ± 0.354; non-learners: 5.301 kHz ± 0.217).

To explore the full range of tonality in hummingbirds, we also quantified tonality in the songs of species of unknown status regarding vocal production learning. Additionally, this analysis probes to what degree tonality of song may allow inferences of possible status regarding vocal production learning. The observed levels of tonality vary broadly (figure 2c). The observed range exceeds that of the species with known status (figure 2c), which can, in part, be attributed to syllables in which the second harmonic is louder than the fundamental frequency. Because the ranges of learners and non-learners overlap substantially, a clear prediction of learning status is not possible for many species. The sampled individuals in the groups of Brilliants and Hermits show similar general trends in tonality, coefficients of variation (CV) and frequency range as the confirmed vocal learners and could provide clues about their learning status.

(b) . Tonality across frequency range

Next, we explored tonality across the frequency range of a given vocal repertoire to test the hypothesis that vocal learners display enhanced ‘ability’ to do so. The CV for vocal learners indicate significantly more consistent tonality for vocal learners than non-learning taxa in passerines (figure 3a; vocal learners: 0.199; non-learners: 0.552; t-test: p < 0.001). The greater consistency in tonality in oscines is even more notable, because the frequency range of individual species is significantly broader than that of suboscines. The CV values for suboscines are more scattered and, especially for large frequency ranges, are consistently higher than those of oscines (figure 3a). To rule out the possibility that higher tonality in oscines is connected to differences in frequency range and mean fundamental frequency, we tested tonality across the entire frequency range of species. The relative amplitude differences between fundamental and second harmonic are consistently higher over the entire observed frequency range in oscines than suboscines (figure 3c).

Figure 3.

(a) Coefficients of variation for tonality (one value for each individual) are higher in non-vocal learning songbirds, despite the lower frequency ranges (in kilohertz) (left panel; violin plots for overall comparison, right panel). Low coefficient of variation values indicates more even filtering across the vocal range of a bird and thereby suggest the use of a dynamic filter, rather than a static filter. Two outliers were excluded in this figure to show more detail on the remaining data points (non-learning songbirds with CVs of 7.68 and 3.69, and frequency ranges of 0.815 and 1.34 kHz). (b) Similarly, in hummingbirds, coefficients of variation tend to be lower in vocal learners than in non-vocal learners. (c,d) Coefficients of variation (c) and frequency ranges (d) for hummingbirds of unknown learning status are highly variable across groups (for comparison, data of confirmed learners and non-vocal learners are depicted in the most left and most right panels). (Online version in colour.)

The CV data in hummingbirds show a similar difference between vocal learners and non-learners. The vocal learning species show low CV values, while the non-learners are scattered and have much higher values (vocal learners: 0.351; non-learners: 1.341; t-test: p < 0.001; figure 3b). Species of unknown learning status, again span a broad range of CV values (figure 3c). In contrast to passerines, the frequency ranges of learners are not greater than those of non-learners (figure 3d).

(c) . Effects of body size on mean fundamental frequency and peak resonance frequency

The complex interactions between sound source and filter characteristics [18,27] and the above differences between vocal learners and non-learners raise the question to what degree body size affects the various acoustic parameters, including expansion of the fundamental frequency range and vocal tract filter properties and vibratory behaviour of the sources (i.e. increased open time; figure 1). The relationship between mean frequency of the song repertoire and body mass is significant for the passerines in our dataset but explains only a small fraction of the variation (linear regression for log centre frequency over log body: R² = 0.05, p = 0.0016). The relationship is highly non-significant for the hummingbirds (R² = 0.001; p = 0.97). Interestingly, among passerines the residuals of the regression are distinctly positive for oscines (mean residual = 0.0631 ± 0.01) and negative for suboscines (mean residual = −0.0631 ± 0.019), indicating that oscines tend to sing higher frequencies than suboscines for the same body mass.

4. Discussion

These data provide strong support for our hypothesis that greater tonality is linked to vocal learning. A consistent difference in tonality between vocal learning and non-learning taxa in both passerines and hummingbirds provides critical, independent evidence supporting this hypothesis. Vocal learning evolved independently in these groups and different selective regimes related to feeding mechanics must have acted on upper vocal tract biomechanical systems in these two groups. Compared to suboscines, oscines also employ a greater F0 range and are able to maintain high tonality across the entire F0 range. These results indicate that oscines make use of tonality increasing mechanisms more broadly than suboscines. Tonality can be increased by several active mechanisms, most prominently active control of the upper vocal tract filter and adjusting vocal fold settings to achieve increased open time of vocal fold vibration (figure 1).

The filter characteristics of the upper vocal tract are determined by multiple resonances, which arise from different air spaces (figure 1b). First, the air column in the trachea constitutes a fairly fixed resonance, because tracheal length may not be changed much through active processes in passerines [20] (but see [51]). Second, the OEC provides a second chamber, whose resonance peaks can be dynamically adjusted [7,21,26]. To consistently filter out upper harmonic content across a broad frequency range, a dynamic filter is required. Therefore, birds employing dynamic filtering should display lower CV, while those using static filters should score high values, if they traverse a sufficiently broad frequency range in their song repertoire. If we assume that all birds are capable of a certain degree of active upper vocal tract control, our data show clear differences in tonality, suggesting differences in broad filtering efficiency between species and, overall, between vocal learners and non-learners.

The result that tonality can be high across broad frequency ranges is an important finding in regard to vocal performance of individual species. In passerine vocal learners, the ability to strongly modulate the sound source for generating an increased fundamental frequency range is paired with the ability to strongly modulate the vocal tract filter. These data are, to our knowledge, the first confirmation that tonality is achieved over a broad frequency range, although the ability to maintain tonality has been assumed in the song performance literature [29,31,54]. If tonality is an important indicator of the singing bird's quality, as suggested by limited playback experiments (e.g. [35,36]), skills for dynamic and flexible adjustment of the vocal tract filter are advantageous to achieve consistent filtering. Our data show that many oscines do indeed achieve high tonality over a broad frequency range, whereas suboscines do so less efficiently. To what degree selective pressure for tonality actually constrains the frequency range of individual species (for review, e.g. [32]) remains to be determined.

Harmonic content of the emitted sound is determined by aspects related to the sound source and the upper vocal tract filter. For example, different labial profiles and dynamic oscillatory movements generate variable spectral slopes at the sound source (e.g. [55–57]). Generally, we expect a higher harmonic content for pulse tone sounds than for those generated by modal vibration [56,57]. Since pulse tone and modal vibration differ in predominant frequency range in which they occur, filter properties required to achieve tonality may differ depending on fundamental frequency. The observation that filtering across a broad frequency range is consistent in some species suggests that OEC volumes are appropriately adjusted and match the frequency range of the vocal repertoire. Constraints at the low (limits on active expansion of the OEC volume) and high fundamental frequency end (minimal OEC volume generates resonance frequency lower than the sound frequency) may differ but have been well integrated by vocal learners leading to a greater efficiency.

In hummingbirds, light gas experiments confirmed that vocalizations are generated by flow-induced vocal fold vibrations [58]. Little is known about upper vocal tract filtering in this group. Film material of singing individuals show throat movements, which appear overall smaller than those observed in singing passerines (M.F., T.R. & F.G. 2022, personal observation). Our data support the notion that the dynamic filter may be limited in hummingbirds. The CV increases with increasing frequency range (figure 3b), suggesting that dynamic adjustments are either less prominent or less effective than in oscine songbirds.

Hummingbird vocal anatomy and feeding biomechanics are very different from those of passerines, both of which affect aspects of upper vocal tract filtering. Hummingbird tracheae are relatively shorter than in other birds and are associated with a high, extra-thoracic position of the syrinx [57]. Generally, the position of the syrinx at the lower end of a long trachea, as found in most birds, provides an exceptional advantage to sound production [18]. To what degree the high position of the syrinx in hummingbirds affects upper vocal tract filtering needs to be investigated. Furthermore, the unique nectar feeding strategy of hummingbirds with associated adaptations in morphology and biomechanics of the lingual and hyoidal complex for tongue extension and fast licking rates [58,59] (for review, see [60]) may limit the range over which dynamic adjustments of the OEC are possible. This biomechanical difference could contribute to the generally lower tonality of hummingbirds compared to that of passerines (figures 2 and 3).

In general, physical attributes of birds set a range within which resonance peaks can occur. Interestingly, our data show that the highest tonality tends to occur at the centre frequency of the vocal repertoire. This suggests that peak resonances are matched to the used frequency range, independent of body size. Very small body size may preclude resonances for low fundamental frequencies, which would result in relatively low amplitude and, thus, make upper harmonic content more prominent. Zebra finches, for example use complex emphasis of upper harmonic content for many of their harmonic stack syllables, indicating that resonance peaks are well above the fundamental frequency [18,25,61,62]. Similarly, within the species in our datasets, we find a trend towards an emphasis of second harmonics in syllables with the lowest fundamental frequency. This occurs more frequently in suboscines, which is consistent with their tendency to use lower frequencies than the oscines (figure 2d). For very low fundamental frequencies, many birds use a closed-beak strategy combined with a large-volume resonating chamber [63–65].

When assessing the ability to generate tonal sounds broadly, we also need to consider the perceptual relevance of tonality. For fundamental frequencies above 6 kHz, the upper harmonics should be well above the sensitivity of avian hearing [66]. While this is true for passerines and hummingbirds, it applies more to hummingbird songs, because their fundamental frequency ranges often extend to higher frequencies than those of passerines (e.g. figure 2) [67,68]. Upper harmonics may not fall within the hearing range (e.g. [68]), although some high-frequency perception occurs in hummingbirds with very high fundamental frequency of song [69]. Therefore, selective pressure for tonality may not be as strong for high fundamental frequencies.

All mechanisms to increase tonality (dynamic filtering; increased open quotient; closed-beak vocalization) seem to be associated with increased sound amplitude. Furthermore, the loudness of the emitted sound may depend in complex, interactive ways on relationships between the filter characteristics and the generated sound pressure levels [70], thus influencing how efficiently the generated sound is radiated at the mouth. Although data are available only for a few species (e.g. [71,–73]), increased song amplitude must provide a selective advantage to the singing individual. If we assume that hearing thresholds decrease with increasing frequency, song amplitude becomes even more important. The ability to generate loud sounds has also become more relevant as many environments are polluted with high levels of anthropogenic background noise.

It is interesting that the inverse relationship between body size (here body mass is used as a correlate of size) and tonality is similar for vocal learning and non-learning passerines. If dynamic vocal tract filtering was the dominant mechanism to increase tonality, we would expect a stronger relationship for non-learning taxa. Instead, our data show only a weak relationship for both groups and no difference in slope. Variation is substantially greater for small-bodied birds in both datasets, suggesting that the various resonances emerging from the different cavities may be more spread out across the frequency range than in larger birds. The difference in vertical intercept is consistent with the observation that suboscines in our sample tend to sing lower frequencies than oscines and indicates that for both datasets resonance peaks are somewhat matched to song frequency.

From an evolutionary perspective, our data provide a first test of whether or not the ability to generate tonal sounds across a broad frequency spectrum is linked to vocal learning. Unlike in parrots, hummingbirds have convergently evolved features of the vocal organ that are similar to those of songbirds [58,74], and their use of song is functionally also very similar. Furthermore, the independent evolution of vocal learning in passerines and hummingbirds enabled us to meaningfully compare vocal tract filtering in these two pairs. While the song development mode is only known for a few species of hummingbirds, the general trend observed in passerines is confirmed by these smaller datasets in hummingbirds. Interestingly, the data for hummingbird species of unknown song learning status suggest that vocal learning may be variably present in different clades. However, the predictive power is rather limited, given the large variation in tonality in both, vocal learners and non-learners.

How did the ability to control increased tonality evolve? It is possible that the broader frequency range of vocal learners drives this trend towards greater control over tonality. However, an ability for greater tonality must have evolved concomitantly with an increasing frequency range and repertoire expansion. The tendency towards broader and more consistent tonality in oscines compared to suboscines indicates that enhanced mechanisms for singing tonal songs must have evolved with this trend. For example, more elaborate dynamic filtering requires more neural control, and it is therefore likely that the control mechanisms for upper vocal tract filtering must be tightly linked to the dynamics of the elaborate and well-characterized vocal production circuitry of vocal learners [39,75].

In addition to neurally controlled active mechanisms for increasing tonality, there could also be structural adaptations. The efficient use of the OEC as an adjustable resonance chamber in oscines may have built on an exaptation in the epithelium of the oesophagus. Within the Passeriformes, oscines lost the cornulin gene, thereby gaining a more compliant oesophagus, which may have allowed them to use OEC expansion more effectively and across a broader frequency range than is possible for suboscines [76]. This change is not present in hummingbirds, which may also partially explain, why they do not achieve the high levels of sustained tonality across the full frequency range of vocal repertoires as is found in oscines. This scenario further illustrates that the evolutionary basis for elaborate and dynamic upper vocal tract filtering includes genetic changes to the involved tissues and modifications in their neuromuscular control. These evolutionary changes acted on biomechanical systems which were differentially constrained by necessities associated with other functions.

Data accessibility

All data used in this study are available from the Dryad Digital Repository https://doi.org/10.5061/dryad.rn8pk0pd4 [77] and in the electronic supplementary material [78].

Authors' contributions

M.F.: conceptualization, data curation, formal analysis, investigation, methodology, software, validation, visualization, writing—original draft, writing—review and editing; T.R.: conceptualization, data curation, investigation, methodology, validation, visualization, writing—original draft, writing—review and editing; F.G.: conceptualization, formal analysis, investigation, methodology, supervision, validation, visualization, writing—original draft, writing—review and editing.

All authors gave final approval for publication and agreed to be held accountable for the work performed therein.

Conflict of interest declaration

We declare we have no competing interests.

Funding

We received no funding for this study.