Polyphonic sound event localization and detection based on Multiple Attention Fusion ResNet.

Area of Science:

• Acoustics and Signal Processing
• Machine Learning for Audio Analysis

Background:

• Sound event localization and detection are crucial in diverse applications.
• Polyphony and noise interference pose significant challenges to accurate sound event prediction and localization.

Purpose of the Study:

• To develop an advanced deep learning model for robust sound event localization and detection.
• To address the challenges of polyphony and noise interference in complex acoustic environments.

Main Methods:

• A Multiple Attention Fusion ResNet architecture based on ResNet34 is proposed.
• Gated Channel Transform is integrated to enhance residual blocks, improving contextual information capture and noise reduction.
• Split Attention and Coordinate Attention mechanisms are employed to manage cross-channel information and focus on spatial sound event characteristics.

Main Results:

• The proposed model demonstrates significant performance improvements over state-of-the-art methods.
• Superior performance is achieved in environments with multiple polyphonic sounds and directional noise interference.
• Competitive results are obtained in scenarios with single polyphonic sound interference.

Conclusions:

• The Multiple Attention Fusion ResNet effectively enhances sound event localization and detection capabilities.
• The model's attention mechanisms are key to mitigating interference from polyphony and noise.
• This research offers a promising solution for real-world sound analysis challenges.