Finding, visualizing, and quantifying latent structure across diverse animal vocal repertoires

Area of Science:

• Bioacoustics
• Computational Biology
• Animal Communication

Background:

• Animal vocalizations vary greatly in complexity and require expert analysis.
• Human interpretation of animal sounds can be subjective and prone to perceptual biases.
• Understanding complex animal communication is crucial for ecological and evolutionary studies.

Purpose of the Study:

• To develop and apply novel computational methods for analyzing complex animal vocalizations.
• To overcome limitations of human-based characterization of animal communication.
• To enable high-powered comparative analyses of vocal acoustics across species.

Main Methods:

• Spectrogram-based projection of animal vocalizations into low-dimensional latent spaces.
• Application of methods to diverse vocal datasets from over 20 species (mammals, birds, primates, humans).
• Analysis of vocalizations using both discrete sequence and continuous latent variable approaches.

Main Results:

• Latent projections effectively uncover complex features in vocalization data.
• Methods provide visually intuitive and quantifiable insights into acoustic structures.
• Successful disentanglement of spectro-temporal structure and observation of long-timescale organization.

Conclusions:

• Computational analysis of vocalizations offers an objective approach to studying animal communication.
• The developed methods facilitate comparative bioacoustic analyses across a wide range of species.
• This work advances our ability to understand the complexity and organization of animal vocal behavior.