In this paper, new approaches to noise spectrum estimation and spectral gain control are proposed for noise spectral suppressors. First, the speech absent frames are detected by using spectral entropy. In the speech absent frames, a weighting factor used in estimating the noise spectrum is modified so as to emphasize effect of the noisy speech signal. Next, a spectral gain is more reduced by multiplying a factor in order to suppress effects of the noise in the speech absent frames. Furthermore, in the speech present frames, in order to reduce signal distortion, the spectral gain is controlled to be unity based on an SNR calculated by using a ridgeline spectrum. Finally, the original noisy speech is added to the estimated speech in some ratio. This ratio is controlled by the long term averaged SNR of the estimated noise and the noisy speech. Computer simulations by using speech signals, the white noise, the car noise and the bubble noise, which are available in public, have been carried out for the conventional methods and the proposed method. The proposed method can improve a segmental SNR and speech quality compared to the conventional methods. Especially, it is useful for the bubble noise.
展开▼
