The effect of musical residual noise is still a major problem for a wavelet-based algorithm in speech enhancement. In this investigation, we propose to utilize a minimum-mean-square-error (MMSE) estimator with homogeneous-frame-merged (HFM) approach to reduce the effect of musical residual noise. Initially, a fixed-length frame is employed to calculate the power of noisy wavelet coefficients (WCs). Hence, merging homogeneous frames in a segment is performed to estimate the power of speech WCs for each subband, which fact enables the gain factors to vary smoothly over homogeneous frames. Therefore, the effect of musical residual noise is reduced. A merged frame with arbitrary shape for estimating the power of speech WCs is obtained. Experimental results show that the proposed approach can improve the performance of the MMSE estimator and reduce the effect of annoying residual noise in enhanced speech.
Relation:
International Journal of Electrical Engineering 15:311-321
