Few-Shot Emergency Siren Detection.

Area of Science:

• Machine Learning
• Audio Signal Processing
• Deep Learning

Background:

• Supervised deep learning for sound event detection requires large, labeled datasets, which are costly and time-consuming to acquire.
• Existing methods face challenges in data collection and labeling for robust emergency siren detection systems.

Purpose of the Study:

• To propose and evaluate a few-shot metric learning workflow for efficient emergency siren detection.
• To minimize the reliance on large, labeled audio datasets for training sound event detection models.

Main Methods:

• Utilized prototypical networks trained on diverse data sources (public, synthetic) for few-shot learning.
• Implemented a knowledge transfer approach for identifying ambulance sirens using only a few instances for prototype computation.
• Investigated the impact of filtering techniques on background noise reduction for siren recordings.

Main Results:

• Achieved Area Under the Precision-Recall Curve (AUPRC) scores of 0.86 (unfiltered) and 0.91 (filtered).
• Outperformed a baseline convolutional model, demonstrating the effectiveness of the few-shot learning approach.
• Validated the technique's reliability in real-world scenarios with various sensor placements.

Conclusions:

• Few-shot metric learning offers a viable and efficient solution for emergency siren detection.
• The proposed workflow demonstrates robustness and effectiveness, even in noisy, real-world conditions.
• This research provides valuable insights for developing advanced in-car emergency vehicle detection systems.