Adaptive Noise Reduction for Sound Event Detection Using Subband-Weighted NMF.

Area of Science:

• Audio Signal Processing
• Machine Learning for Acoustics

Background:

• Real-world sound event detection is challenged by non-stationary and time-varying noise.
• Existing methods struggle with adaptive noise reduction in complex acoustic environments.

Purpose of the Study:

• To develop an adaptive noise reduction technique for robust sound event detection.
• To enhance sound event detection performance in noisy conditions using a novel NMF approach.

Main Methods:

• Employed robust non-negative matrix factorization (NMF) for adaptive noise dictionary learning.
• Developed a supervised source separation framework using learned noise and pre-trained event dictionaries.
• Extended NMF to a weighted form to prioritize target sound event components and suppress noise.

Main Results:

• Achieved comparable results to state-of-the-art Convolutional Recurrent Neural Networks (CRNNs) on the DCASE 2017 dataset.
• Demonstrated significant noise reduction capabilities with the weighted NMF method.
• Improved F-score by 5% compared to the unweighted NMF approach for rare sound event detection.

Conclusions:

• The proposed weighted NMF method offers an effective solution for adaptive noise reduction in sound event detection.
• This approach significantly enhances the separation of target sound events from interfering noise.
• The method shows strong potential for improving the reliability of sound event detection systems in real-world applications.