Classification of budgerigar vocalizations

Parrots are capable of an astounding degree of vocal learning and rich social interactions, making them an important model system for understanding the neural basis of communication. Here we devise a method for automatic classification of the vocalizations of the budgerigar, a small Australian parrot. Pairs of adult male birds were recorded, producing vocal elements including warble, contact call and alarm call. The most complex of these vocalization types is warble, which contains continuous vocal gestures with varying durations. Acoustic features in warbles can resemble vocalizations in other categories as well as cage noises due to birds' movement. These properties pose challenges to many schemes designed for birdsong syllable classification. Inspired by the fact that neurons in higher auditory areas respond to complex acoustic features, we devised a new approach that utilizes feature detectors to categorize the vocal elements into the three vocalization types and cage noises. We find that this feature detector method performs remarkably well on budgerigar vocalizations, producing over 90% accuracy in identifying elements compared to human annotated categorization.