A novel perceptual distortion minimization speech enhancement algorithm

We propose a two-stage speech enhancement algorithm to minimize perceptual distortion caused by residual noise and speech distortion. Noise reduction was first applied to noisy speech, then a perceptual enhancement filter was applied to the residual noise designed to minimize speech distortion keeping the sum of speech distortion and residual noise below the masking threshold of the clean speech. When compared with the conventional state-of-the-art noise reduction algorithms, the proposed algorithm tends to achieve significant improvements in terms of objective quality measures. Furthermore, in a subjective preference evaluation with 24 listeners, the average mean opinion score of the enhanced speech signal is increased by up to 0.36 when compared to the conventional algorithms.