Noise reduction of speech signals via Gabor expansion

This paper applies Gabor expansion to passive noise reduction of speech signals. The Gabor expansion with Gaussian prototype function has a property that its time and frequency product is minimal. The property makes the noise-only segment more clearly and the noise estimate more robustly. When a Gabor component contains noise only, the constant weight is applied to the component, and when a Gabor component contains both noise and voice, weights estimated via likelihood ratios are applied to the Gabor coefficients. The noise-reduced signal is obtained via synthesis of the weighted Gabor coefficients.