A practical guide for generating unsupervised, spectrogram-based latent space representations of animal...

Area of Science:

• Bioacoustics
• Computational Ethology
• Machine Learning in Ecology

Background:

• Manual analysis of animal vocalizations is time-consuming and requires specialized expertise.
• Objective and generalizable methods are needed for pattern detection, categorization, and similarity quantification in acoustic data.
• Neighbourhood-based dimensionality reduction of spectrograms offers a promising computational approach for analyzing animal calls.

Purpose of the Study:

• To demonstrate the utility of neighbourhood-based dimensionality reduction for analyzing animal vocalizations.
• To obtain meaningful latent space representations of meerkat (Suricata suricatta) calls that align with their known taxonomy.
• To provide practical guidance and examples for researchers using this method.

Main Methods:

• Utilized a dataset of manually annotated meerkat vocalizations.
• Applied neighbourhood-based dimensionality reduction techniques to spectrograms of the vocalizations.
• Generated latent space representations to visualize and analyze call structures.

Main Results:

• The method successfully produced latent space representations reflecting the established taxonomy of meerkat call types.
• Analyzed the strengths and weaknesses of the dimensionality reduction approach for bioacoustic analysis.
• Demonstrated applications including the classification of ambiguous calls and detection of mislabeled data.

Conclusions:

• Neighbourhood-based dimensionality reduction is an effective and conceptually simple method for analyzing animal vocalizations.
• The approach offers objective insights into call structure and facilitates taxonomic classification.
• Accompanying example code empowers researchers to apply this technique to their own bioacoustic studies.