Joint Optimization of Deep Neural Network-Based Dereverberation and Beamforming for Sound Event Detection in...

Area of Science:

• Signal Processing
• Machine Learning
• Acoustics

Background:

• Sound event detection (SED) in multi-channel environments is challenging due to noise and reverberation.
• Existing methods often process dereverberation and beamforming separately, limiting overall performance.

Purpose of the Study:

• To propose and evaluate a joint optimization framework for DNN-supported dereverberation and beamforming in CRNN-based SED.
• To enhance the accuracy and robustness of SED systems in adverse acoustic conditions.

Main Methods:

• Utilizing deep neural networks (DNNs) for weighted prediction error (WPE) dereverberation and minimum variance distortionless response (MVDR) beamforming.
• Jointly training dereverberation, beamforming, and SED modules using a single loss function.
• Employing focal loss to address data imbalance issues in training deep learning models for SED.

Main Results:

• The proposed joint optimization significantly improves SED performance in noisy and reverberant environments.
• Experimental results demonstrate the effectiveness of the integrated DNN-based dereverberation and beamforming approach.
• Focal loss effectively mitigates training difficulties caused by class imbalance.

Conclusions:

• Joint optimization of DNN-supported dereverberation and beamforming is a promising approach for robust multi-channel SED.
• The proposed method offers substantial performance gains over traditional, separate processing techniques.
• This framework provides a unified solution for enhancing audio signal processing in complex acoustic scenarios.