Quantitative assessment of vocal development in the zebra finch using self-organizing neural networks.

To understand the mechanisms of song learning by songbirds it is necessary to have in hand tools for extracting, describing, and quantifying features of the developing vocalizations. The extremely large number of vocalizations produced by juvenile zebra finches and the variability in these vocalizations during the sensorimotor learning period preclude manual scoring methods. Here we describe an approach for classification of vocalizations produced during sensorimotor learning based on self-organizing neural networks. This approach allowed us to construct probability distributions of spectrotemporal features recorded on each day. By training the network with samples obtained across the course of vocal development in individual birds, we observed developmental trajectories of these features. The emergence of stereotypy in sequences of song elements was captured by computing the entropy in the matrices of first- and second-order transition probabilities. Self-organizing maps may assist in classifying large libraries of zebra finch vocalizations and shedding light on mechanisms of vocal development.