Objective analysis of song learning in birds: Toward automated techniques.

Abstract

It is demonstrated that indigo buntings sing bouts of ‘‘adult plastic’’ songs (APS) distinct from stereotyped songs (SS). Some distinguishing characteristics of APS include syllables of variable morphology and syllables not present in SS. The hypothesis that APS may play a role in adult song learning is tested by monitoring socially paired males housed together. Some yearlings changed their SS to more closely match their tutors’ SS, by replacing syllables or inserting new syllables. New syllables were developed in APS through transformation and combination of morphologically similar existing syllables, and syllables were transferred from APS into SS. To address the limitations of manual scoring of spectrographs, dynamic time warping (DTW) for template-based automated analysis of continuous recordings of birdsongs was evaluated. With laboratory recordings, the DTW algorithm employed identified syllables and syllable boundaries of SS and calls of indigo bunting and zebra finch with greater than 97% accuracy. APS constituents were identified with approximately 84% accuracy. Under these restricted recording conditions, DTW has general applicability to objective analysis of birdsongs, and dramatically decreases the effort required. Application of hybrid hidden Markov models may improve the performance for variable vocalizations such as APS, and under noisy conditions of field recordings.