Cosine similarity-based few-shot bioacoustic event detection with automatic frequency range identification in...

Area of Science:

• Bioacoustics
• Machine Learning
• Signal Processing

Background:

• Few-shot sound event detection (SED) in bioacoustics aims to reduce labeling effort by using minimal positive examples.
• Existing methods like prototypical networks struggle with small, weak, or noisy acoustic targets in Mel-spectrograms.
• Challenges include frequency overlap between target and background events and low signal-to-noise ratios.

Purpose of the Study:

• To develop an effective few-shot SED system for bioacoustics.
• To address the limitations of existing methods in handling challenging acoustic environments.
• To create a system that does not require extensive training data or pre-trained models.

Main Methods:

• Introduced an automatic frequency range identification algorithm for effective handling of small objects in Mel-spectrograms.
• Employed cosine similarity computation in a gliding-window manner for event detection after frequency range identification.
• The system is designed to operate without large datasets or reliance on pre-trained models.

Main Results:

• Achieved an F-score of 46.9% on the DCASE 2024 Task 5 dataset.
• Secured a top 3 position in the DCASE 2024 Task 5 competition.
• Demonstrated the system's potential in overcoming specific challenges in bioacoustic SED.

Conclusions:

• The proposed automatic frequency range identification and cosine similarity-based SED system is effective for bioacoustics.
• The method shows promise for real-world applications requiring efficient analysis of acoustic recordings with limited labeled data.
• This approach offers a viable alternative to data-intensive methods in bioacoustic monitoring.