MSFDnet: A Multi-Scale Feature Dual-Layer Fusion Model for Sound Event Localization and Detection.

Area of Science:

• Acoustics
• Machine Learning
• Signal Processing

Background:

• Existing Sound Event Localization and Detection (SELD) methods struggle with long-duration, dynamic audio due to insufficient feature extraction and underutilization of multi-task feature complementarity.
• This limitation restricts system performance in complex acoustic environments.

Purpose of the Study:

• To propose a novel SELD model, MSDFnet, designed to overcome the limitations of current methods in dynamic audio scenarios.
• To enhance the accuracy of sound event detection and localization by improving feature extraction and fusion strategies.

Main Methods:

• Developed MSDFnet, incorporating a Multi-Scale Feature Aggregation (MSFA) module for capturing diverse spatial features.
• Implemented a Dual-Layer Feature Fusion (DLFF) strategy to strengthen the relationship between Sound Event Detection (SED) and Direction of Arrival (DOA) features.
• Evaluated the model on the DCASE2020 Task 3 dataset.

Main Results:

• MSDFnet achieved competitive scores on the DCASE2020 Task 3 dataset, including ER20 of 0.319, F20 of 76%, LEcd of 10.2°, LRcd of 82.4%, and SELDscore of 0.198.
• The model demonstrated excellent performance in complex audio scenarios.
• Ablation studies confirmed the significant contributions of the MSFA and DLFF modules to the overall performance.

Conclusions:

• MSDFnet effectively addresses limitations in existing SELD methods for dynamic audio.
• The proposed MSFA and DLFF modules significantly enhance sound event localization and detection accuracy.
• The model shows strong potential for real-world applications in complex acoustic environments.