Research On Speech Denoising Real-time Processing Algorithm

With the progress of science and technology, production and living place for human expansion, the impact of noise on human beings have become more and more serious. So speech denoising becomes increasingly important to improve voice quality in noisy environments. In recent years, due to the rapid development of DSP technology, speech denoising technology has been the focus gradually shifting from hardware to improve the transition to the improvement of the algorithm. Adaptive speech denoising technology is an important way to highlight the advantage of its non-voice based on any model, loss of voice characteristics, the effect of significantly de-noising, which is widely used.The research on adaptive filtering algorithms is one of the most attentional directions in the field of adaptive signal processing. The goal that researchers try to achieve is to find an adaptive filtering algorithm with higher convergence speed, lower steady-state error and better reform factor. This paper aims at improving the LMS adaptive algorithms, then reseach the theory analyses and computer simulations of variable-step-size algorithms.We introduce traditional speech denoising method such as spectral subtraction, wavelet transform and subspace method, compare the advantage and shortcoming of these methods, and then the basic adaptive noise cancellation system was given. The theories and filtering performance of the LMS adaptive algorithm are analyzed, and the effects of the LMS algorithm eliminating the representative noise are verified by simulations. The results of simulations indicate that the LMS algorithm is effective in eliminating the noise, but the algorithm still have a defect that slowly convergence. Therefore, the LMS algorithm has been improved. First of all, some improved algorithms of variable-step-size algorithms in time domain are researched from the following aspects, such as basing on sigmoid function and error self-correlation estimation. Computer simulations indicate that both of the improved algorithms have favorable convergence speed, steady-state error and reform factor. Secondly, based on the classical LMS adaptive algorithm, the BLMS algorithm and the FBLMS algorithm were derived and studied, using the MATLAB simulation software, after simulation analysis, we have a result that computational complexity and computing time of the FBLMS algorithm reduced than the classical LMS adaptive algorithm. Therefore, the several improved algorithms in this article solve the problem that LMS algorithm convergence is so slowly,this conducive to the realization of real-time algorithm, simulation results show that speech denoising for speech to receive good results.In the end, the conclusions are summarized,and the unsolved problem and further research ideas are indicated.