Title: From Detection to Direction: An Overview of Sound Event Localization and Detection
Presenter: Jun Wei Yeow and Ee-Leng TAN
Part I: Overview of Sound Event Localization and Detection (SELD) (30 minutes)
- Introduction to SELD and its applications
- History of SELD and its component tasks (Sound Event Detection and Sound Source Localization)
- Recent advances and challenges in SELD
- Publicly available SELD datasets
Part II: Core Technical Components of SELD (60 minutes)
- Spatial audio formats used for SELD, including First Order Ambisonics, microphone array signals, and binaural recordings.
- Contemporary feature extraction techniques that capture spatiotemporal cues needed for robust event detection and localization.
- Deep learning architectures designed for SELD, including convolutional recurrent networks (CRNNs), transformer-based models, and multi-branch or multi-task setups.
- Training strategies, such as multi-task learning (joint DOA and event classification), data augmentation for spatial audio, and domain adaptation techniques.
- Benchmark datasets and metrics, including a deep dive into the DCASE Challenge series as well as evaluation criteria such as localization errors, detection accuracies, and combined SELD scores.
Coffee Break (30 minutes)
Part III: Advanced and Emerging Topics (60 minutes)
- Semi-supervised and weakly labelled learning approaches.
- Robustness to reverberation, overlapping events, and unseen acoustic scenes.
- Multi-modal SELD systems that integrate complementary modalities, such as video recordings or motion sensors.
- Complementary performance using acoustic scene classification (ASC)
Coffee Break (30 minutes)
Part IV: Real-Time Implementation of SELD (40 minutes)
- Real-time constraints and considerations
- Lightweight models suitable for real-time and edge applications.
- Discussion and Q&A session