Reshaping Bioacoustics Event Detection: Leveraging Few-Shot Learning (FSL) with Transductive Inference and Data...

Area of Science:

• Bioacoustics
• Machine Learning
• Signal Processing

Background:

• Bioacoustic event detection is challenging due to the scarcity of labeled data for training supervised learning models.
• Existing methods struggle with limited recordings and few labeled examples, hindering accurate animal sound classification.

Purpose of the Study:

• To develop and evaluate a few-shot learning (FSL) method for bioacoustic event detection that overcomes data limitations.
• To improve the recognition and classification of animal sounds in natural habitats using minimal labeled data.

Main Methods:

• A few-shot learning (FSL) approach combining transductive inference and data augmentation (SpecAugment on Mel spectrograms).
• Transductive inference iteratively refines class prototypes and feature extractors to capture key patterns.
• Data augmentation techniques are employed to increase the volume and diversity of training data.

Main Results:

• Significant F-score improvements of 27% on the DCASE-2022 dataset and 10% on the DCASE-2021 dataset compared to state-of-the-art methods.
• The proposed method demonstrates robust performance across various animal species, recording conditions, and durations.
• All components of the FSL method contributed to substantial performance gains.

Conclusions:

• The developed FSL method effectively addresses the challenge of limited labeled data in bioacoustics.
• This approach offers a promising solution for accurate and adaptable animal sound detection in real-world scenarios.
• The method's ability to generalize across different acoustic environments and species highlights its practical utility.