Perceptual artifacts in speech noise suppression

Most of the state of the art speech noise suppression gains rely on the frequency decomposition of the signal. This decomposition is usually done using the fast Fourier transform. These algorithms may be successful in suppressing the noise but often distort the speech that is of interest. The noise suppression gain is rapidly changing in time. When the gain is applied to the noisy signal, it modulates the signal and generates artifacts. In this paper, we will show that by using a frequency decomposition based on the human perception the artifacts can be made imperceptible. We will use this proposed frequency decomposition with the traditional Wiener Gain for noise suppression and Ephraim-Malah suppression rule to show that the resulting speech is natural sounding with minimal artifacts.