Machine Learning Reveals Cryptic Dialects and Cultural Mate Choice in Zebra Finches

In many animals, including birds, primates, and cetaceans, vocal communication is acquired through social learning. Individuals learn their vocalizations by imitating parents or conspecifics. This process rarely produces exact copies: small deviations accumulate over generations, a process known as cultural drift. In human societies, such drift gives rise to dialects; in birds, it can generate geographically restricted song dialects. However, when the songs of individuals vary greatly from one another, it has long been assumed that distinctive group-level song signatures cannot arise. The zebra finch (Taeniopygia guttata) represents such a case. During a brief learning period in adolescence, males crystallize their songs, after which they remain largely unchanged throughout life. Moreover, approximately 15% to 50% of song syllables arise through innovation, meaning that every male's song is essentially unique. Because earlier analyses failed to detect clear dialect differences among wild or captive populations, researchers generally concluded that zebra finch songs lack population-level signatures and that females do not choose mates based on song origin. But is this assumption correct?

The zebra finch (Taeniopygia guttata) （圖片來源：christoph_moning，採用 CC BY 4.0 授權）

To examine whether population-specific song patterns might nevertheless exist, researchers studied four zebra finch populations that had been separated for long periods. Two populations had been domesticated in captivity for roughly 100 generations, whereas two others originated from wild birds introduced 25 and 5 generations ago. Song recordings from each population were analyzed using Apple's machine-learning tool “Create ML Sound Classifier”.（https://developer.apple.com/machine-learning/create-ml/）The algorithm was trained to recognize which population a song belonged to based solely on acoustic data. Remarkably, the trained model identified the correct population for training songs with an accuracy of 93–97%, and when tested on songs from the next generation of birds that had not been included in training, it still achieved 85–95% accuracy. Although these differences are nearly impossible for human listeners to detect, the results reveal that subtle but consistent song differences exist among populations.

The next question was whether these subtle song differences influence mate choice. Researchers first placed unmated adults from different populations together in large aviaries and observed their natural pairing behavior. Zebra finches showed a strong tendency to form pairs with individuals from their own population, producing a clear pattern of assortative mating. However, this pattern could potentially arise from two different causes. One possibility is genetic: domesticated birds are larger (about 16 g) than wild-derived birds (about 12 g), so individuals might simply prefer partners with similar body size. The other possibility is cultural: birds might prefer partners whose songs resemble those heard during development.

To disentangle these alternatives, the researchers conducted a cross-fostering experiment. Eggs from different populations were exchanged immediately after laying, so that chicks were raised by foster parents from another population. As a result, the offspring inherited their genetic background from one population but learned songs from another. This design allowed the influence of genetic origin and cultural upbringing to be separated. If mate choice were genetically driven, birds should prefer partners with similar body size. If it were culturally driven, they should prefer partners whose songs resemble those learned during development.

The results strongly supported the cultural hypothesis. When the cross-fostered birds reached adulthood, females preferentially paired with males whose songs resembled those of the birds they had grown up with, even if the males differed in body size or genetic background. Conversely, males whose songs differed from those heard during the female's upbringing had significantly lower interaction and pairing rates. To quantify these interactions in detail, the researchers used an automated barcode tracking system that recorded the positions of individual birds every two seconds. Long-term monitoring confirmed that proximity and social interaction patterns were strongly influenced by song similarity learned during development.

The researchers then asked whether these cultural effects could persist across generations. Birds from the second generation were allowed to reproduce without further cross-fostering, and the pairing behavior of their offspring was examined. Even in this third generation, individuals continued to show mating preferences based on the cultural lineage established by earlier song learning. This indicates that culturally transmitted song differences can remain stable for at least two generations, effectively creating cultural separation between populations.

Further analyses revealed an additional intriguing pattern. Females tended to choose males whose songs resembled those of their adolescent peers rather than those of their fathers. This finding suggests that social learning during the juvenile period, particularly interactions with same-age companions, plays a major role in shaping mate preferences.

The researchers also considered other potential cultural signals, such as courtship dance movements or “distance calls.” However, these traits showed only weak correlations with song characteristics and did not predict mating patterns. Song dialects therefore appear to be the primary driver of assortative mating. These dialects had remained undetected for decades because traditional analytical methods focus on a limited set of measurable acoustic features, such as pitch or rhythm. Machine-learning approaches, in contrast, can analyze complex multidimensional patterns and detect subtle structures that conventional methods overlook.

Together, these findings demonstrate that zebra finch populations possess previously hidden song dialects. Although these differences are subtle to human perception, they play a significant role in mate choice. Cultural transmission of song thus has the potential to influence social structure and may even contribute to the emergence of reproductive isolation between populations.

Author: Sui-Ye You

Reference:

Wang D et al. (2022). Machine learning reveals cryptic dialects that explain mate choice in a songbird. Nature Communications.