Research Of Speech Enhancement Algorithm Under Non-stationary Environments

In many practical situations, speech processing systems are confronted with high ambient noise levels, and their performance degrades drastically. Thus, there is a strong need to improve the performance of these systems in noisy conditions by developing enhancement algorithms.In this paper, we focus on single channel speech enhancement in non-stationary environments. This algorithm uses improved minimum statistics method for noise estimating to overcome the deficiency of the conventional voice activity detectors, which are only efficient for stationary environments and can't track the varying noise with constant threshold. A sub-window approach is utilized in this method to speed up the performance of tracking. Following, auditory masking properties are introduced into the subtractive-type enhancement process. It allows for an automatic adaptation in time and frequency of the parametric enhancement system with the masking threshold, and finds the best tradeoff between noise reduction and the speech distortion. Objective and subjective evaluation of the proposed algorithm is performed with several noise types in the Noisex-92 database. The results show that this method leads to a significant reduction of background noise and the unnatural structure of the residual noise. Furthermore, another method is researched primarily which is based on the auditory filter banks, and the design of filter banks is also discussed in this paper.