Impact Noise Suppression Using Speech Spectral Phase Estimator

抄録

This paper proposes an impact noise suppression method using a speech spectral phase estimator. We firstly estimate a speech and a noise spectral amplitudes in a zero-phase (ZP) domain. Next, we estimate the noise spectral phase as a slope of an observed spectral phase. Subtracting the estimated noise spectral amplitude with phase from the observed spectrum, and taking the spectral phase of it, we have the estimated speech spectral phase. Taking the IDFT of the estimated speech spectral amplitude with phase gives the estimated speech signal in time domain. When the analysis frame does not include the noise, we should not perform the noise suppression for avoiding any speech degradation. Hence, we introduce the Teager-Kaiser energy operator to detect the noise-existence frame with a high accuracy. Simulation results show that the proposed impact noise suppressor improves SNR, PESQ, and STOI scores in comparison to conventional methods.