Bird song comparison using deep learning trained from avian perceptual judgments.

Area of Science:

• Animal communication
• Neurobiology of learning
• Bioacoustics

Background:

• Bird song analysis is crucial for understanding animal communication and learning.
• Current song similarity assessments rely on human perception or intuition-based algorithms.
• Validated comparisons of complex song syllables are essential but challenging.

Purpose of the Study:

• To compare and externally validate existing song similarity algorithms.
• To develop a novel deep learning method for avian-based song similarity judgments.
• To establish a robust framework for validating methods in animal behavior data.

Main Methods:

• Utilized a novel automated operant conditioning system to collect zebra finch song syllable similarity decisions.
• Compared established software (Raven, Sound Analysis Pro, Luscinia) against avian judgments.
• Developed and trained a deep learning model on the avian decision dataset.

Main Results:

• Existing algorithms perform better than chance but do not fully emulate avian similarity assessments.
• The novel deep learning method significantly outperforms established methods in accuracy.
• A modified deep learning approach accommodating inconsistent avian decisions achieved the strongest performance.

Conclusions:

• The developed deep learning method provides a more accurate assessment of bird song similarity.
• The avian decision dataset serves as a valuable resource for validating new similarity methods.
• This approach can be extended to validate similarity judgments in other species and behavioral data.