Introduction

Vocal production learning, a complex and intriguing phenomenon, has independently evolved across various taxa, including elephants, bats, pinnipeds and cetaceans1. However, vocal learning is particularly apparent in avian species varying from songbirds2 to parrots, hummingbirds and some other bird species3,4. While songbird vocal learning has been well-characterized through model species like the zebra finch (Taeniopygia guttata5) and white-crowned sparrow (Zonotrichia leucophrys6), comparative studies exploring vocal learning processes in parrots remain scarce3. Moreover, direct comparisons of vocal production learning between parrots and songbirds are currently lacking with regards to vocal production complexity7. The complexity of vocal production learning is often discussed in relation to repertoire size, but less so in terms of the complexity and accuracy with which these sounds are imitated. Nevertheless, certain birds may imitate a wide variety of sounds including human speech,8,9,10 environmental sounds11 but also artificial sounds like chainsaws12.

Imitating sounds from different species, an ability known as allospecific vocal imitation, allows birds to enhance the complexity of their vocal repertoire13. This increased vocal variation can function to deter predators by reproducing allospecific alarm calls when disturbed at the nest, such as in spotted bowerbirds (Ptilonorhynchus maculatus14), to imitate the vocalizations of host species during brood parasitism (cuckoos15) or to create complex songs to attract mates as a form of sexual selection in European starlings (Sturnus vulgaris16). The quality of this allospecific vocal imitation is usually compared between one species and the model sounds, but not across different species imitating the same model sound17,18,19,20.

Specifically, a comparison between species from different clades and with different vocal systems could give insight into the role of cognitive abilities and/or physiological structures in vocal imitation accuracy. For example, songbirds (such as starlings) and parrots (like budgerigars; Melopsittacus undulatus) are both vocal production learners with extensive repertoires but also have a distinct syrinx morphology (Fig. 1). Starlings have independent control over the syringeal muscles at both sides of the syrinx which allows them to produce two sounds at the same time21. In contrast, parrots do not have this independent control over the syrinx’ muscles (and lack a second sound producing source) and thus lack this biphonation ability22,23,24. Moreover, songbirds and parrots also differ in the neurological structures underlying vocal production learning25. Vocal learning is neurologically controlled by a well described song system in songbirds. In parrots, similar neurological structures have been found in an area called the “core” region. However, the nuclei surrounding the core region in the so called “shell” region, which is unique to parrots, have been suggested to play an important role in vocal imitation. Variation in relative size between these core and shell regions has been hypothesized to be related to variation in the complexity of vocal production learning between parrot species25.

Fig. 1

Anatomical differences between syrinxes of parrots and starlings. (a) Schematic overview of the syrinx of a budgerigar as a representative of parrots. Abbreviations: SS, m. syringealis superficialis; SP, m. syringealis profundus; LL, lateral labia; TY, tympanum; LTM, lateral tympaniform membrane; BC, bronchial cartilage. Based on Abdel-Maksoud et al.28 and partly modified from Larsen & Goller24. (b) Schematic overview of a starlings’ syrinx. Abbreviations: TY, tympanum; LL, lateral labia; ML, medial labia; MS, intrinsic syringeal muscles; P, pessulus; BC, bronchial cartilage; MTM, median tympaniform membrane. Note the lateral (LL) and medial labia (ML), that are part of the syrinx of the starling and play an important role in biphonation, are different in the budgerigar’s syrinx without the presence of the medial labia. Modified from Prince et al.21.

This study presents a dataset of ten different species all imitating the exact same, complex, sound model: R2-D2 the renowned droid from Star Wars. We analyzed allospecific vocal imitation accuracy in eight different vocal units. These iconic sounds were created by Ben Burtt using an ARP 2600 modular synthesizer in combination with Ben Burtt’s own voice26. The vocabulary of R2-D2 can be distinguished in two types of sounds: monophonic and multiphonic sounds. Whilst monophonic sounds are relatively simple synthesized bleeps and clicks, the multiphonic sounds were generated using the ring modulator of the ARP 2600. A ring modulator is a signal processing device that takes two input signals—a carrier and a modulator—and multiplies them to produce a complex output. This multiplication creates new frequencies known as “sidebands,” which are the sum and difference of the frequencies of the input signals. For example, if the carrier signal has a frequency of 1,500 Hz and the modulator signal is 400 Hz, the output would consist of two sine waves: one at 1900 Hz (sum of the two frequencies) and one at 1100 Hz (difference). These sidebands reflect the combined spectral information from the inputs. If the modulator signal is more complex, the output produces mirrored spectral patterns (i.e. spectral contours; ten Cate & Honing27: Fig. 2) around the carrier frequency. See Fig. 2f for an example of such a mirrored multiphonic sound.

Fig. 2

Spectrograms of R2-D2 model sounds. These sounds consist of elements (single sounds) and units (combination of one or more elements). Spectrograms of monophonic units are shown on the left side of the figure. (a) Unit “H”. (b) Unit “M”. (c) Unit “N”. (d) Unit “P”. Note that unit “M” and “N” look similar, but are different in the number of frequency modulations (i.e. maximum fundamental frequencies are 3.0 and 2.7 kHz for unit “M” and “N”, respectively). On the right side, the multiphonic units are shown: (e) Unit “C”. (f) Unit “X”. (g) Unit “Y”. (h) Unit “W” (see Movie S1 for sound examples).

Vocal imitations of nine parrot species and European starlings were analyzed to answer three key questions (1) How do parrots and starlings differ in their imitation capabilities? (2) How do different parrot species compare in their ability to replicate these complex sounds? (3) What are the differences in imitation accuracy between multiphonic and monophonic sounds? Our study gives insights in a broader vocal imitation framework. Understanding the nuances of allospecific vocal imitation in birds provides valuable insights into the cognitive and morphological mechanisms driving vocal learning accuracy. By comparing imitation accuracy between two distinct vocal systems—the syrinx of parrots and starlings—and species with different cognitive abilities, we highlight key differences in adaptability and learning diversity across species.

Results

Starlings and parrots imitate monophonic and multiphonic units differently

We used citizen science methods to collect 107 videos of nine different parrot species and eight videos of European starlings (115 videos in total) imitating R2-D2 sounds uploaded by parrot owners on various social media platforms (YouTube, TikTok and Instagram). These sounds are always produced in a specific sequence consisting of different units that are made up of multiple elements (Fig. S1). Since the birds imitated these sequences in the same order as R2-D2 we could confirm that the birds were imitating R2-D2 sounds and not producing natural calls or imitating other sounds. We annotated each sound on an elemental and unit level (Fig. S1) in Praat29. A total of 2130 imitations of 98 different units were imitated by the birds. We selected the most frequently imitated monophonic units (units: “H”, “M”, “N” and “P”; Fig. 2a–d) and multiphonic units (units: “X”, “Y”, “C” and “W”; Fig. 2e–h). This resulted in a comparison of 419 monophonic and 463 multiphonic units imitated by birds (n = 103; 95 parrots and 8 starlings, see Table S1 for number of individuals per species) to the original model units.

We analyzed for each individual whether it imitated two sounds at the same time or not, i.e. biphonic sound imitation, while imitating multiphonic sounds. None of the parrot species were able to imitate the two mirrored contours at the same time whereas most of the starlings used biphonation when imitating the mirrored spectral contours (Fig. 3). Spectrograms show the biphonic sound imitation of starlings, reproducing the two mirrored spectral contours created by ring modulation (Fig. 4d, e). Six out of eight starlings imitated these mirrored contours when imitating multiphonic sounds. On the other hand, parrots always copied only the fundamental frequency of multiphonic units and never imitated the mirrored contours (Fig. 4d, f). There were no apparent differences in monophonic sound imitation between both groups (Fig. 4a–c).

Fig. 3

Parrots did not imitate multiphonic sounds using biphonation. Bar diagrams showing the percentage of parrot and starling individuals that imitated sounds using biphonation. Colors correspond to whether biphonation was used or not.

Fig. 4

Spectrograms comparing the model sounds to imitations of starlings and parrots. (a) Example of a monophonic unit: “N”. (b) Imitation of unit “N” by a starling. (c) Imitation of unit “N” by a budgerigar. (d) Example of a multiphonic unit: “X”. (e) Imitation of unit “X” by a starling. (f) Imitation of unit “X” by a budgerigar (see Movie S2 for sound examples).

Starlings imitate multiphonic units more accurately

To quantify this difference, we compared imitation accuracy between species by calculating dissimilarity scores for each unit using the dynamic time warping algorithm of Luscinia30. Dissimilarity scores were calculated based on weightings of several spectro-temporal parameters (Table S2 and methods for more details). A lower dissimilarity score indicates a more accurate imitation of the R2-D2 model sound unit.

Imitation accuracy between species varied both between and within monophonic and multiphonic units (Fig. 5). Multiphonic units were in general more difficult to imitate accurately (mean: 8.758; range: 1.338–13.581) than monophonic ones (mean: 0.992; range: 0.103–7.791). Variation was also found in how accurately each unit was imitated within both types of units. For example, unit “P” was more accurately imitated than unit “H”, and unit “C” was more difficult to imitate than unit “W”, within the monophonic and multiphonic categories, respectively (Figs. 2 and 5). Dissimilarity scores also varied within species on an individual level (Fig. S2).

Fig. 5

Imitation accuracy of monophonic and multiphonic units per species. Mean dissimilarity scores per unit for each species are shown. Dissimilarity scores varied between each unit. Colors and shapes indicate the different species. Scores in the white and grey background correspond to monophonic and multiphonic units, respectively. Sizes of dots are related to the number of imitated units per species. Note: not all species imitated all units due to sample size differences.

To test whether starlings imitated multiphonic sounds more accurately than parrots we created a Linear Mixed Model (LMM). We found that dissimilarity scores were explained by an interaction between group (starlings vs. parrots) and monophonic or multiphonic units (LMM: − 3.240 ± 0.293, t = − 11.042, p < 0.001; Table S3). A Post-Hoc test revealed that starlings imitated multiphonic units more accurately than parrots (Tukey: 2.771 ± 0.284, t = 9.759, p < 0.001; Fig. 6 and Table S4). Furthermore, multiphonic units were imitated less accurately than monophonic units by both parrots (Tukey: 2.630 ± 0.284, t = 9.271, p < 0.001; Fig. 6) and starlings (Tukey: 3.099 ± 0.286, t = 10.820, p < 0.001; Fig. 6) confirming that multiphonic units are more difficult to imitate than monophonic ones. Parrots and starlings did not differ in imitation accuracy for monophonic units (Tukey: − 0.469 ± 0.328, t = − 1.431, p = 0.482; Fig. 6 and Table S4).

Fig. 6

Starlings imitate multiphonic units more accurately than parrots, no difference in monophonic sound imitation. Violin plots showing the distribution of the data. Horizontal lines and boxes correspond to medians and interquartile ranges. Each dot represents an individual imitating a sound unit. Dissimilarity scores were combined for all monophonic and all multiphonic units separately. Colors correspond to dissimilarity scores of parrots or starlings. Stars indicate significance levels (*** P < 0.001).

Parrot species vary in imitation accuracy

Interestingly, the larger brained parrots, African greys (Psittacus erithacus, mean: 1.871; range: 0.872–7.791) and Amazon parrots (Amazona sp., mean: 1.198; range: 0.621–2.038), imitated the monophonic units the least accurate (Fig. S3). In contrast, smaller brained species like budgerigars (mean: 0.909; range: 0.103–2.797), cockatiels (Nymphicus hollandicus, mean: 1.029; range: 0.267–2.769) and also starlings (mean: 0.839; range: 0.383–1.992) imitated monophonic units the most accurately. Moreover, budgerigars (Tukey: 1.400 ± 0.329, t = 4.258, p = 0.004; Fig. S3) and cockatiels (Tukey: 1.613 ± 0.353, t = 4.573, p = 0.001; Fig. S3) imitated monophonic units significantly more accurate than African greys, and cockatiels imitated more accurately than Amazon parrots (Tukey: 1.232 ± 0.340, t = 3.623, p = 0.041; Fig. S3). Imitation of monophonic units by all other parrot species varied between the scores of African greys and budgerigars. Multiphonic units were imitated the least accurate by rainbow lorikeets (Trichoglossus moluccanus, mean: 12.233; range: 11.444–13.022) and caiques (Pionites sp., mean: 12.931; range: 12.332–13.581). Budgerigars (mean: 8.845; range: 3.590–12.868) and cockatiels (mean: 8.109; range: 4.323–13.474) had the lowest dissimilarity scores for multiphonic units amongst the parrots, similarly to the monophonic results (Fig. 5). All other parrot species’ imitation accuracy varied in between the scores of budgerigars & cockatiels and lorikeets & caiques. In summary, parrot species varied in how accurately they imitated R2-D2 sounds with smaller brained parrots imitating these sounds more precisely than larger brained ones, and starlings imitated multiphonic sounds more accurately than parrots.

Dendrograms confirm similarities between bird species and R2-D2

Finally, to corroborate our results we created dendrograms to hierarchically cluster species based on similarities between imitations (Fig. 7). As expected, starlings clustered together with the R2-D2 model in a dendrogram based on all multiphonic units, thus validating our previous results. In contrast, when comparing monophonic units, cockatiels’, budgerigars’ and parrotlets’ imitations were more similar to the model, whereas starlings split off at the more basal end of the dendrogram and are thus less similar to the model (Fig. 7). Furthermore, African greys clustered at the most basal end of the monophonic dendrogram with the largest distance away from the model and therefore corroborating our earlier result of African greys imitating monophonic units less accurately. No overlapping results in species clustering were found between the monophonic and multiphonic dendrograms.

Fig. 7

Dendrograms of imitation accuracy of model units. Hierarchical representations of similarity of model unit imitations between species. Left shows a dendrogram based on monophonic units while the right dendrogram is based on multiphonic units. Distances were calculated as Euclidean distances between all species and R2-D2.

Discussion

The present study compared vocal imitation accuracy of monophonic and multiphonic sounds between nine parrot species and European starlings using R2-D2 sounds as the model. In general, monophonic units were better imitated than multiphonic units by all bird species. However, the starlings showed particularly accurate imitation of multiphonic units due to their biphonation ability, specifically compared to all parrot species. Some units were easier to copy than others, both within the monophonic and multiphonic sound categories, which could be explained by some units potentially being more similar to the natural vocalizations of our tested species. Amongst the parrots, the smaller brained budgerigars and cockatiels were more accurate at imitating both monophonic and multiphonic sounds than the larger brained parrots, contradictory to results from studies on variation in core and shell nuclei size25.

Starlings imitated monophonic sounds similarly to parrots, but were significantly more accurate at imitating multiphonic sounds than parrots. Interestingly, a physiological explanation seems most convincing due to the anatomical differences between starlings and parrots. The syrinx of starlings contains two vibrating sound sources called the lateral labia and medial labia (Fig. 1). These labia can be independently controlled at both sides of the syrinx by the syringeal muscles thus making it possible for starlings to produce two pitches at the same time22. In contrast, parrots lack the independent control of these labia at both sides of the syrinx and therefore lack the biphonation ability that starlings have while imitating sounds22,23,24. Differences in perception or cognitive abilities can also be excluded as explanation for the observed differences in imitation accuracy since starlings did not differ in how accurately they imitated monophonic sounds from parrots. Although biphonation made it possible for starlings to imitate both the fundamental frequency and the mirrored spectral contours of the multiphonic units not all starlings imitated the two mirrored spectral contours. Two out of eight starlings did not copy the two spectral contours of any of the multiphonic sounds. We hypothesize that these two individuals did not receive sufficient training to accurately reproduce the fundamental frequency and the mirrored spectral contours. Additionally, starlings learn environmental sounds that are frequent and similar to their own vocalizations and thus we could assume that the starlings heard the R2-D2 sounds frequently enough but have not received enough training to produce two sounds at the same time14.

Parrot’s imitation accuracy of monophonic units showed an unexpected pattern. Chakraborty et al.25 hypothesized that parrots with relatively larger shell nuclei compared to core nuclei to have more complex vocal learning abilities. This relative size difference is larger in parrots with larger brains (like macaws, African greys and amazon parrots) which are also well known for their speech imitation abilities.8 Therefore, we expected these species to be more accurate at imitating monophonic sounds. However, in our study we found that parrots with larger brains, and also relatively large shell nuclei, imitated monophonic sounds significantly less accurate than budgerigars and cockatiels that have smaller shell regions and larger core regions. Parrots with smaller brains, however, have a smaller repertoire of imitated sounds11 as expected by the results of Chakraborty et al.25. Our results could therefore be explained by a trade-off between the capacity to learn allospecific sounds versus the degree of imitation accuracy. Larger brained parrots may have a higher capacity to learn more sounds11 but are less accurate at imitating the sounds whereas smaller brained parrots focus more on the accuracy of the few sounds11 they have learned by practicing each imitated sound likely more often than parrots with significantly larger imitation repertoires. These different strategies could be related to different functions of allospecific imitation. Larger brained parrots may use allospecific imitation to socially bond with group members and therefore learn more sounds in captivity31. In contrast, smaller brained parrots, and starlings, could use allospecific imitation in mate attraction with more accurate imitations being more attractive to a potential partner, similar to satin bowerbirds18. Previous studies found effects of imitation of male budgerigars on female mate choice suggesting some tentative evidence for this hypothesis32. However, budgerigars are also able to imitate speech at similar rates of parrots with larger shell regions11,25 but do so by frequency modulating their ‘‘warble’’ song and thus not producing completely new imitated sounds but modified natural vocalizations, making it potentially easier to imitate speech for a smaller brained parrot10. Alternatively, male budgerigars are known to imitate calls of females during pair bonding which could potentially contribute to differences in the accuracy of imitating sounds across species33. However, this has been suggested to occur in all imitating parrots and little is known about variation of the degree of imitation accuracy during pair bonding between different species7.

Imitation accuracy could also not be related to the capacity of complex vocal learning and shell region size but instead the rate of learning could be more related to differences in shell size. Peach-fronted conures (Eupsittula aurea) for example have been shown to have more rapid vocal modification abilities compared to budgerigars which have smaller brains than conures34. Furthermore, another possible explanation could be that the natural vocalizations of smaller brained parrots are more similar to R2-D2 units and thus easier to copy. However, vocal repertoires have not been extensively studied in all parrot species in the present study7. Variation in imitation accuracy is also found between other bird species. Superb lyrebirds (Menura novaehollandiae17) and satin bowerbirds (Ptilonorhynchus violaceus18) are known to be incredible accurate imitators while the brown thornbill (Acanthiza pusilla19) and Icterine warbles (Hippolais icterina20) are less accurate at imitating sounds. Interestingly, the male bowerbirds imitate sounds to incorporate in their songs to attract females and the most accurate imitators are more attractive to female bowerbirds18 whereas brown thornbills and Icterine warbles imitate alarm calls of other species19,20. This suggests that vocal imitation accuracy varies depending on its function. While variation in imitation accuracy across species may partly reflect differences in cognitive or motor constraints35, there is growing evidence that vocal imitation has been shaped by direct selection pressures in several species, including sexual selection, where accurate imitation may enhance mating success or social cohesion36.

Multiphonic sounds were more difficult to imitate than monophonic sounds regardless of having biphonation abilities or not. Starlings imitated the mirrored spectral contours of the multiphonic sounds, however, it was still difficult to accurately imitate the sounds indicated by the higher dissimilarity scores. Moreover, large variation between imitation accuracy was found between different multiphonic units indicating that some units were easier to copy than others. Unit “W” consisted of only one rising mirrored formant and had the lowest dissimilarity scores. In contrast, unit “C” consisted of four rising and falling mirrored spectral contours and showed the least accurate dissimilarity scores of all sounds (Fig. 2). Unit “X” and “Y” both had two rising and falling mirrored spectral contours and scores fell in between the previous two units suggesting that the amount of mirrored spectral contours is related to the complexity of the sounds. Similarly, variation in imitation accuracy was found between monophonic units but to a smaller degree than with the multiphonic units. Units with more rising and falling pitches (like unit “H”) were more difficult to imitate than units with a flatter pitch contour (like unit “P”; Fig. 2).

One of the drawbacks of the present study is that the amount of training for each individual is unknown. Individual variation within each species was large, which could be an indication of different levels of auditory experience. However, most individuals within a species showed similar dissimilarity scores while a few had (higher) scores deviating from the rest suggesting that most individuals received similar amounts of training. Furthermore, it is also unknown how they were trained, i.e. only hearing playbacks of R2-D2 sounds versus also receiving positive rewards like food. We only have anecdotal stories of video descriptions mentioning that one of the parrots started to imitate the model sounds after listening to 3 h of playbacks. In general, knowledge about the processes behind vocal learning of imitated sounds is largely lacking37. Furthermore, usually only one species is compared with the respective model sound that is being imitated17,18,19,20. The use of citizen science enabled us to compare imitation accuracy across ten different species with each other showing the potential citizen science has to increase the volume and diversity of data collection. Our study helped contributing to compare imitation accuracy across species, however the difference in learning capacity between species is still unknown. Similarly to Schachner et al.,38 we collected videos from social media platforms to conduct our study, however, citizen science could also be used to directly work together with companion parrot owners to conduct more detailed studies to have a deeper understanding behind the learning processes of allospecific imitation and the evolution of vocal complexity, like initiatives as the bird singalong project39 and many parrots project40.

In conclusion, we showed that starlings are able to more accurately imitate multiphonic sounds than parrots due to differences in syrinx anatomy. Parrots and starlings did not differ in imitation accuracy of monophonic sounds. However, variation in imitation accuracy was found between parrot species which could potentially be related to differences in brain size. Furthermore, more studies are needed as to how parrots and starlings learn to imitate these complex sounds. What are the learning mechanisms behind vocal imitation and how do they differ from other forms of vocal production learning? Finally, other questions can be answered with the dataset we have in which we know the source in future studies: are differences in imitation accuracy related to difference in relative core and shell region size? Are there perceptual biases for certain acoustic dimensions that differ between species and could these be related to natural vocalizations? Which aspects of the model are copied (in an absolute fashion) and which are relatively imitated (e.g. transposed pitch but maintained pitch contour)? Are there temporal differences in imitating R2-D2 sounds between species? Answering these questions will contribute to our understanding of the evolution of complex communication, vocal learners, and getting a better understanding of the precursors of human language and music in animals27.

Methods

Data collection

We collected