Analysis of birdsong development and automated clustering of song syllables

During early development, young songbirds such as the Zebra Finch learn acoustically complex but stereotyped sequential behaviors which are termed "songs". Furthermore, zebra finches learn only one song in their lifetime, making the problem of developmental song analysis tractable. The manner in which the song develops from a relatively incoherent subsong (akin to "babbling") to a stereotyped sequence of syllables is remarkable and shows many similarities with human speech acquisition. To understand how a juvenile's song develops to mimic the song of a model (tutor) adult, we have developed methods for determining acoustic similarity based on low-dimensional embedding of high-dimensional sound spectrograms (frequency vs time representations) [1].

Available Project
To compare the songs of two birds is a complex problem with many intrinsic degrees of freedom. Automation of birdsong analysis would be generally very useful for researchers working in the domain of speech and language acquisition. We offer MSc/BSc projects to implement deep generative / discriminative-generative models to create useful low-dimensional embeddings of songbird vocalizations and that support song clustering and song similarity analysis.

[1] Kollmorgen S, Hahnloser RHR. Dynamic Alignment Models for Neural Coding. PLoS Comp Biol. 2014;10(3):1–19.