Cross-Referencing Self-Training Network for Sound Event Detection in Audio Mixtures

Area of Science:

• Artificial Intelligence
• Machine Learning
• Signal Processing

Background:

• Sound event detection is crucial for audio tagging, identifying sound categories and temporal boundaries.
• Deep neural networks have advanced sound event detection but require extensive labeled data.
• Supervised methods dominate, demanding costly data collection and annotation.

Purpose of the Study:

• To propose a semi-supervised method for sound event detection to reduce reliance on labeled data.
• To explore a student-teacher scheme balancing self-training and cross-training for pseudo-label generation.
• To investigate post-processing techniques for enhancing sound event detection performance.

Main Methods:

• A semi-supervised approach using a student-teacher model for pseudo-label generation from unsupervised data.
• Implementation of a scheme balancing self-training and cross-training within the student-teacher framework.
• Application of post-processing methods to refine sound event interval extraction from network predictions.

Main Results:

• The proposed semi-supervised method achieved significant performance improvements on the DCASE2020 challenge dataset.
• Evaluations on validation and public evaluation sets of the DESED database demonstrated superior results.
• The approach outperformed existing state-of-the-art systems in semi-supervised sound event detection.

Conclusions:

• The developed semi-supervised method offers an effective alternative to fully supervised approaches for sound event detection.
• The student-teacher scheme combined with post-processing enhances detection accuracy while minimizing labeling effort.
• This research contributes to more efficient and scalable sound event detection systems.