Dynamic time warping and sparse representation classification for birdsong phrase classification using limited...

Area of Science:

• Bioacoustics
• Machine Learning
• Ornithology

Background:

• Manual annotation of birdsong phrases is time-consuming and limits behavioral and population studies.
• Limited data, due to scarce recordings or rare vocalizations, poses a challenge for automated classification.
• Developing automated methods is crucial to reduce manual annotation efforts and expand research scope.

Purpose of the Study:

• To develop an automated birdsong phrase classification algorithm designed for limited data scenarios.
• To classify up to 81 phrase classes of Cassin's Vireo using minimal training samples (1-5 per class).
• To enhance the accuracy and efficiency of birdsong analysis for scientific research.

Main Methods:

• The algorithm employs dynamic time warping (DTW) to improve phrase similarity by accounting for individual bird variations and segmentation inconsistencies.
• A two-pass sparse representation (SR) classification is utilized, where the second pass refines decisions when initial classifications conflict.
• The SR classifier identifies phrases by finding sparse linear combinations of training feature vectors.

Main Results:

• The proposed algorithm achieved high classification accuracies of 94% (manually segmented) and 89% (automatically segmented) for Cassin's Vireo phrases.
• These results were obtained using only five training samples per class on unseen individuals.
• The developed classifier outperformed traditional methods like DTW, support vector machines, and a basic SR classifier.

Conclusions:

• The developed automated birdsong phrase classification algorithm is highly effective, even with limited training data.
• This approach significantly reduces the need for manual annotation, making birdsong analysis more accessible for behavioral and population studies.
• The algorithm demonstrates robust performance in classifying complex vocalizations across different segmentation types.