Noise Prewhitening Using Supervised Nonnegative Matrix Factorization and Wiener Filtering

Noise prewhitening converts nonwhite noise in noisy speech into white noise, a preprocessing step to improve the performance of subsequent processing tasks. This is done by estimating and flattening the power spectral density (PSD) of noise. However, accurate estimation of PSD can be challenging, especially when dealing with nonstationary noise. We propose a spectro-temporal flattening filter based on Wiener filtering to address these challenges. This filter incorporates prior knowledge about speech and noise, learned through nonnegative matrix factorization (NMF). The ability of NMF to learn and represent the complex structures of speech and noise allows the filter to adapt to various real-world noises. The experimental analysis, using five types of environmental noise, demonstrates the effectiveness of the proposed method in flattening noise in noisy speech.