BioCPPNet: automatic bioacoustic source separation with deep neural networks

Area of Science:

• Bioacoustics
• Machine Learning
• Signal Processing

Background:

• Overlapping vocalizations in bioacoustic recordings pose a significant challenge for data analysis.
• Existing methods struggle with complex mixtures across diverse species.
• Processing large volumes of acoustic data is crucial for ecological and behavioral studies.

Purpose of the Study:

• To introduce a novel, robust, and efficient machine learning architecture for bioacoustic source separation.
• To enable the separation of overlapping vocalizations from single-channel recordings across various taxa.
• To advance the processing of previously inaccessible bioacoustic data.

Main Methods:

• Developed BioCPPNet, a lightweight, modular U-Net-based architecture operating on raw audio waveforms.
• Utilized learnable or handcrafted encoders for feature extraction.
• Employed a permutation-invariant objective function trained on perceptual audio quality metrics.
• Evaluated performance using scale-invariant signal-to-distortion ratio (SI-SDR) and identity classification accuracy.

Main Results:

• BioCPPNet successfully separated concurrent vocalizations from macaques, bottlenose dolphins, and Egyptian fruit bats.
• Achieved state-of-the-art performance in end-to-end single-channel bioacoustic source separation.
• Demonstrated robust performance across different species and recording scenarios (open/closed speaker).

Conclusions:

• BioCPPNet offers a significant advancement in bioacoustic source separation technology.
• The developed method paves the way for processing vast amounts of previously unusable bioacoustic data.
• This work is a critical step towards deploying automated bioacoustic monitoring systems.