Fig. 7: Classification scores from a machine-learning algorithm (ML; a, b) and similarity scores from Sound Analysis Pro (SAP; c, d).

a, b A machine-learning algorithm was trained on independent sets of zebra finch song recordings to discriminate between ‘same’ and ‘different’ dialect from the perspective of an individual female in Generation 2 given her experiences in a rearing aviary. In the training data set ‘same’ is represented either by the songs of the set of 8 fathers (Generation 1) or the set of 10 peer members (Generation 2) in the rearing aviary; ‘different’ is represented by the respective songs from an aviary of another population type (domestic D or wild-derived W, by males that will not be encountered in the social or breeding experiment). The 40 males that a female will encounter in the social and breeding experiment (20 of the same song dialect, shown in blue; 20 of a different song dialect, in red) are then classified by ML as either ‘same’ or ‘different’ with complementary confidence scores that add up to one. Note that each male contributes 4 data points (2 ‘same’ and 2 ‘different’) because he encounters four types of females (DD, DW, WD, WW) from different rearing aviaries. c, d Similarity scores from SAP using the same representation as in (a) and (b) (similarity to the songs of the peers or fathers of a female’s rearing aviary, which the focal male never met, such that any similarity is indirect). The machine-learning algorithm (a, b) achieves much clearer differentiation compared to the traditional SAP software (c, d). Males that were cross-fostered within population (DD or WW; a, c) are discriminated with slightly higher confidence than DW or WD males (b, d; see the crosses that mark the group means).