Characterizing the dynamics of syllable emergence in the juvenile songbird*

As juveniles, songbirds, such as the Bengalese finch, learn to sing. Their songs are built from sequences of stereotypical acoustic units known as syllables. The dynamics of emergence of these syllables in the process of learning is not understood. During learning, syllables can be created through syllable differentiation, corresponding to an increasing difference over time between the acoustic structure of pairs of pre-existing syllables. Syllables can also emerge de-novo, with seemingly no relation of similarity to other syllables in the repertoire. To characterize the dynamics of emergence, the acoustic structure of songs is usually classified using manual labeling and quantified by computing scalar acoustic measures such as pitch, amplitude and spectral entropy. To go beyond such human-defined features, we use low-dimensional latent representations that can be learned from the syllables' spectrograms to analyze the vocal repertoire of songbirds at different developmental points. We classify syllables in each bird's repertoire using unsupervised clustering methods and analyze the dynamics of the latent representations in the time periods prior to and following syllable differentiations and creations. This allows us to describe the relation between the spectral structure of syllables and the timing of their appearance within different song contexts. This new understanding of the dynamics underlying the emergence of new syllables should facilitate the search for neural correlates of learned motor repertoires in the songbird.