Introducing the Software CASE (Cluster and Analyze Sound Events) by Comparing Different Clustering Methods and Audio...

Area of Science:

• Bioacoustics
• Computational Ecology
• Machine Learning

Background:

• Unsupervised clustering is vital for classifying animal sounds in ecology and conservation.
• Current methods often use one-dimensional feature extraction, which inadequately characterizes frequency-modulated vocalizations.
• Assessing the reliability of clustering results for vocalizations remains a challenge.

Purpose of the Study:

• To address limitations in current bioacoustics clustering techniques.
• To introduce a novel multidimensional feature extraction method for animal vocalizations.
• To evaluate various unsupervised clustering and classification methods for improved accuracy.

Main Methods:

• Tested established and novel unsupervised clustering algorithms (community detection, affinity propagation, HDBSCAN, fuzzy clustering).
• Employed classifiers like k-nearest neighbor, dynamic time-warping, and cross-correlation.
• Developed and applied a multidimensional data transformation procedure for acoustic features.

Main Results:

• Multidimensional feature extraction significantly enhances clustering applicability, especially for frequency-modulated vocalizations, compared to one-dimensional methods.
• The study identified strengths and weaknesses of different clustering and classification combinations.
• The CASE software facilitates automated parameter optimization and reliable result verification.

Conclusions:

• Multidimensional characterization and clustering of animal vocalizations hold substantial potential for advancing bioacoustics research.
• The CASE software provides a comprehensive tool for applying and comparing multiple clustering algorithms.
• CASE enables automated parameter optimization, improving the efficiency and reliability of bioacoustic analysis.