Understanding how birds react to each other’s songs can shed light on their behaviour and communication in the wild. But such studies require a way to generate realistic-sounding synthetic birdsong. To achieve this, a team from the University of Buenos Aires has used mathematical modelling to synthesize a fake birdsong that’s credible enough to prompt wild birds to respond.
The researchers – Roberto Bistel, Ana Amador and Gabriel Mindlin – studied the rufous-collared sparrow, a highly territorial songbird. The sparrow’s song comprises a brief theme composed of 2–4 notes, followed by a trill, with the entire song usually lasting about 2 s. During the breeding season, the male sparrow sings roughly three times each minute, with each bird singing its own unique theme. If the sparrow hears song from another bird of the same species, it typically responds by increasing its rate of singing.
To create realistic artificial birdsong, the team used a mathematical model based on the physics of avian sound production. The model describes the dynamics of the two pairs of syringeal labia (located at the junctions between the bronchi and the trachea) that modulate airflow and create the sound waves. These sound waves are then filtered as they pass through the bird’s trachea, oro-oesophageal cavity and beak.
By defining the muscle activity and subsequent filtering during birdsong production, the low-dimensionality model can generate sounds with temporal and acoustic properties that emulate a given song. For their study, the researchers used the model to generate 11 synthetic themes based on previous in-the-field recordings from rufous-collared sparrows.
In the field
Lab-based tests of songs generated by the model showed that the synthetic songs generated similar neural responses to those evoked by the bird’s own song. But to fully assess the degree of realism, the team moved on to studying the behaviour of birds in the wild. This involved playing both synthetic and real birdsong to male sparrows and recording their behavioural response, classified as the number of songs that the auditory stimulus elicited.
Working in a roughly 0.4 km2 area within Parque Pereyra Iraola (a UNESCO Biosphere reserve) in Buenos Aires, Argentina, the researchers first identified the locations of birds with moderate singing activity, studying a total of 26 individuals in 15 sites. During each 13 min test, they recorded natural sound for 2 min, repeatedly played a song for 1 min, and then continued the recording for 10 min. In total, the birds were exposed to 256 tests, 90 with real songs, 81 with synthetic songs and 85 with songs from three different bird species.
To quantify the birds’ responses, the team processed the recorded audio files with a noise reduction filter and a band-pass filter to focus only on the sparrow’s frequency range and then computed corresponding spectrograms. Two independent observers inspected the spectrograms to count the number of songs per minute produced by each individual. As every bird sings a unique song, the team could confirm each bird’s identity in the recordings.
At the start of each test, the birds sang three to four songs per minute. While the recording was played, the singing performance increased significatively. Once the playback stopped, the original singing rate gradually returned. The researchers note that there was no significant difference between the birds’ responses to real and synthetic songs in the field. Songs from other species, however, failed to evoke a response.
Bistel, Amador and Mindlin say that the model provides a valuable tool for investigating a wide range of biological questions. “This work paves the way for manipulating auditory stimuli in an interpretable way, allowing to address a series of questions that can greatly benefit from the flexibility in the generation of acoustic stimuli permitted by our physical model,” they write.
One potential application could be to test the performance hypothesis, which proposes that two attributes of singing (the frequency of syllable production and the spectral range of syllables) are reliable indicators of the quality of the singer. Another option is to study how wild birds can learn via playback, through the use of automatic audio players.
Pitch of birdsong is determined by body size
“The success of our approach instils confidence in the hypotheses underpinning the model and provides a valuable tool for investigating a wide range of biological questions,” the researchers conclude. They plan to continue their fieldwork, further evaluating behavioural responses to synthetic songs generated by the model.
“On one hand, we plan to modify certain parameters in the model to identify which acoustic characteristics make a song more intimidating to other males, and which make it more attractive to the females,” Mindlin tells Physics World. “We hope that the interpretability of the model will allow us to understand the anatomical or physiological features underlying these acoustic properties.”
The study is described in Physical Review E.
