Acoustic Denoising using Dictionary Learning with Spectral and Temporal Regularization.

Area of Science:

• Signal Processing
• Acoustics
• Biomedical Engineering

Background:

• Speech enhancement is challenging in extremely noisy environments, such as during Magnetic Resonance Imaging (MRI) scans.
• Existing methods often require prior knowledge or specific assumptions about noise characteristics, limiting their applicability.

Purpose of the Study:

• To develop an advanced speech enhancement algorithm for data acquired in highly noisy conditions.
• To improve the accuracy and robustness of speech signal recovery from corrupted data.

Main Methods:

• A dictionary learning approach using complex nonnegative matrix factorization with intra-source additivity (CMF-WISA) was employed.
• Dictionaries for noise and speech+noise were learned to decompose the noisy spectrum into estimated speech and noise components.
• Spectral and temporal regularization terms were incorporated into the CMF-WISA cost function to refine noise modeling.

Main Results:

• The proposed algorithm demonstrated significant improvements in speech enhancement compared to traditional Least Mean Squares (LMS) filtering.
• Objective and subjective assessments validated the superior performance of the new method.
• The algorithm effectively reduced noise without requiring prior knowledge of noise waveform periodicity.

Conclusions:

• The developed dictionary learning method offers a robust solution for speech enhancement in extremely noisy environments.
• This technique provides a significant advancement over existing methods, particularly for challenging data like MRI audio.
• The algorithm's ability to perform without specific noise assumptions enhances its practical utility.