A generalized network based on multi-scale densely connection and residual attention for sound source localization...

Area of Science:

• Acoustics
• Signal Processing
• Machine Learning

Background:

• Sound source localization and detection (SSLD) is challenging due to diverse sound events and variable source locations.
• Existing SSLD methods struggle with integrating multi-scale spatial information and efficient feature propagation.

Purpose of the Study:

• To propose an effective SSLD method integrating multi-scale dense connections (MDC) and residual attention (RA) mechanisms.
• To introduce a dual-path attention (DPA) unit for recalibrating feature maps within the proposed architecture.

Main Methods:

• Developed a Multi-Scale Densely Connected (MDC) block to capture information across various receptive fields.
• Explored three Residual Attention (RA) blocks to improve information flow between network layers.
• Incorporated a Dual-Path Attention (DPA) unit for feature map recalibration within MDC and RA blocks.

Main Results:

• Individual components (MDC, RA, DPA) were validated for their effectiveness.
• The proposed SSLD method outperformed four other methods on a development dataset.
• The approach demonstrated strong generalization capabilities when tested against SELDnet and SELD-TCN on five additional datasets.

Conclusions:

• The proposed MDC and RA mechanisms, along with the DPA unit, significantly improve sound source localization and detection.
• The method shows robust performance and generalization across diverse acoustic environments and datasets.