ANIMAL-SPOT enables animal-independent signal detection and classification using deep learning

Area of Science:

• Bioacoustics
• Machine Learning
• Deep Learning

Background:

• Manual bioacoustic signal identification is time-consuming and error-prone.
• Existing machine learning approaches are difficult for biologists to adapt to specific datasets.
• Large volumes of bioacoustic data require automated analysis methods.

Purpose of the Study:

• To present an animal-independent, open-source deep learning framework (ANIMAL-SPOT) for bioacoustic signal identification.
• To address common bioacoustic tasks: target signal detection, species classification, and call type categorization.
• To make deep learning accessible to a broader audience in bioacoustics research.

Main Methods:

• Developed an animal-independent, open-source deep learning framework named ANIMAL-SPOT.
• Investigated three signal identification tasks: target signal vs. background noise, species classification, and call type categorization.
• Validated the framework on diverse animal datasets and a public bioacoustic corpus.

Main Results:

• Achieved high accuracy (mean 97.9%) and AUC (95.9%) for target signal detection on unseen recordings.
• Obtained strong segmentation accuracy (95.4%) and F1-score on the BirdVox-Full-Night dataset.
• Demonstrated high multi-class classification accuracy for species (96.6%) and call types (92.7%).
• Outperformed a baseline system in multi-species classification with an Unweighted Average Recall (UAR) of 89.3%.

Conclusions:

• ANIMAL-SPOT provides an effective and accessible solution for bioacoustic signal identification.
• The framework's animal independence and ease of use lower barriers for biologists.
• Automated analysis using ANIMAL-SPOT enhances the efficiency and scalability of bioacoustic research.