Research On Speech Enhancement Algorithm For Dictionary Learning Based On Noise Power Estimation Threshold

The traditional speech enhancement algorithm has a good suppression effect on stationary noise,but it is difficult to have a good effect on non-stationary noise.In recent years,sparse representation and dictionary learning algorithms have provided new methods for non-stationary noise removal.Based on sparse representation and dictionary learning algorithm,this paper proposes a method to optimize the speech enhancement process by using the power spectrum of noisy speech signals.Firstly,on the basis of phase optimized K-means and singular value decomposition algorithm and phase optimized orthogonal matching tracking algorithm,a method is introduced to obtain the spectral reduction threshold by using the power spectrum calculation of noisy speech,which can solve the error caused by manual estimation of spectral reduction threshold of this algorithm in the speech denoising stage,so that the selection of spectral reduction threshold is self-adaptive.The method can be optimized in different degrees according to the actual situation,so as to improve the quality of speech.The experimental results show that the spectral reduction threshold based on the power spectrum calculation can provide the optimal speech enhancement results,and can effectively remove the steady-state noise and steady-state noise.In the-10 d B noise environment,compared with the contrast algorithm,the improved algorithm can improve the segmental signal-tonoise ratio by 1-2d B,and improve the perceived speech quality evaluation and shortterm intelligibility by 2.1 and 0.11 points,respectively.Secondly,aiming at the problem that the search for the best atom in the orthogonal matching tracking algorithm requires a large amount of computation and takes too long time,a firefly algorithm is proposed to improve the search for atoms in the orthogonal matching tracking algorithm.The application scenario of the algorithm is speech signal.The calculation method of the original step factor needs to be adjusted so that the improved orthogonal matching tracking algorithm can not only guarantee the accuracy of voice signal reconstruction,but also converge faster.In the speech enhancement stage,spectral reduction threshold based on power spectrum calculation is also used to optimize the enhanced speech.Experimental results show that the modified orthogonal matching tracking algorithm of firefly algorithm runs 58.07 times faster than the original orthogonal matching tracking algorithm,and the obtained signal residual error is reduced by 26.05% and the mean square error is reduced by 25.95%.The improved orthogonal matching tracking algorithm also scored 2.5% to 15% higher than the original algorithm in the objective evaluation of speech enhancement applications.