Noise-Power Estimation Based on Ratio of Stationary Noise to Input Signal for Noise Reduction

Abstract

This paper proposes a method for estimating noise-power spectrum and reducing stationary noise components in noisy speech signal. Noise reduction generally suppresses the stationary noise by applying a multiplicative gain calculated from the estimated noise-power spectrum. However, the accuracy of the noise-power estimation is degraded by the presence of superimposed speech. The proposed method aims to maintain the accuracy of the noise-power estimation in a period of speech. The method first estimates the power ratio of noise to input signal for each frequency bin by assuming the noise amplitude spectrum to be constant in a short time period which cannot be applied to the amplitude spectrum of speech. The method then improves the estimation accuracy by compensating for the errors caused by time variations in the stationary noise. Simulation results demonstrate the accuracy improvement of the noise-power estimation for the stationary noise and the better performance in terms of noise reduction.