Deep audio embeddings for vocalisation clustering.

Area of Science:

• Bioacoustics
• Machine Learning
• Animal Communication

Background:

• Non-human animal communication analysis often involves manual transcription of vocal sequences into discrete units, forming species-specific vocal repertoires.
• Expert-driven vocal repertoire characterization can be time-consuming and prone to bias, necessitating automated solutions.

Purpose of the Study:

• To develop and evaluate a novel machine learning method for automated vocal repertoire characterization in non-human animals.
• To improve the efficiency and objectivity of analyzing animal vocalizations using deep representation learning.

Main Methods:

• Utilized a convolutional auto-encoder network to learn abstract representations of animal vocalisations.
• Evaluated the learned representations against expert-labeled vocalization types across 8 datasets from 6 species (birds and marine mammals).
• Compared the auto-encoder approach with state-of-the-art methods for vocal representation and clustering.

Main Results:

• The auto-encoder method significantly improved the relevance of vocalisation representations for repertoire characterization.
• Demonstrated superior performance compared to existing methods across diverse datasets and species.
• The approach requires minimal parameter tuning for effective application.

Conclusions:

• Deep representation learning with auto-encoders offers a powerful and efficient tool for bioacoustic analysis and vocal repertoire characterization.
• The developed method reduces bias and labor associated with traditional analysis techniques.
• A Python package is released to facilitate the use of auto-encoders for bioacoustics research, aiding in vocal repertoire browsing and annotation.