Speech enhancement algorithm based on improved spectral subtraction

To improve the enhancement effect of noisy speech signals, this paper presents and analyses a new speech enhancement algorithm based on improved spectral subtraction. In contrast to the standard spectral subtraction algorithm, the new algorithm accurately estimates the noise according to that the amplitude spectral of narrowband white Gaussian noise obeys Rayleigh distribution, based on that all noise can be changed into Additive White Gaussian Noise (AWGN). This algorithm also adopts a new speech activity detection technology based on frequency band variance to detect speech activity. The emulational analyse indicates that the algorithm in this paper is better suit for noise elimination than standard spectral subtraction.