Hidden Markov and Gaussian mixture models for automatic call classification.

Area of Science:

• Bioacoustics
• Marine Mammal Research
• Computational Biology

Background:

• Automatic classification of animal sounds offers efficiency for large datasets.
• Marine mammal communication analysis benefits from objective and consistent methods.
• Previous research has not applied HMMs or GMMs to marine mammal call classification.

Purpose of the Study:

• To evaluate the effectiveness of Hidden Markov Models (HMMs) and Gaussian Mixture Models (GMMs) for classifying Northern Resident killer whale calls.
• To compare automated classification results with established perceptual (human) classifications.
• To introduce novel computational methods for marine mammal bioacoustic analysis.

Main Methods:

• Utilized a dataset of 75 Northern Resident killer whale calls.
• Employed cepstral coefficients as acoustic features for analysis.
• Applied Hidden Markov Models (HMMs) and Gaussian Mixture Models (GMMs) for automated classification.
• Compared model outputs against human-derived perceptual classifications into seven call types.

Main Results:

• Both HMMs and GMMs achieved over 90% agreement with human perceptual classification.
• The HMM method demonstrated over 95% agreement in certain cases.
• These computational methods show significant promise for classifying marine mammal vocalizations.

Conclusions:

• HMMs and GMMs are effective tools for the automated classification of killer whale vocalizations.
• These methods provide a scalable and reliable alternative to manual analysis of bioacoustic data.
• The study highlights the potential of machine learning in advancing marine bioacoustics research.