The Research On Blind Source Separation Method For Speech Signals

Blind signal separation(BSS)is a class of signal processing method,which derived in the later last century.It has great practical value in the modern communication,speech signal processing,mechanicbreakdown diagnosis,image processing,et al.Speech signal processing plays more and more important role in speech recognition,telephone communication,human-computer interaction and military monitor.The realization of the functions above all needs speech with high signal to noise ratio or pure speech signal.Normal noise reductionmethods can't gain specific objective's speech with high noise and speech background.BSS can deal with such problems.The paper discussed speech signal's feature and the way to gather speech signals.We analyzed the character of fixed-point algorithm,improved the natural gradient algorithm,studied the separation algorithms for convolutive mixing model and improved the underdetermined blind separation based on two-step method.The specific work is done as follows.1.Compared with other signals,speech signal has its own features.Before the mixing speech is separated,the feature of the speech itself should be analyzed,for example,what demand should be paid attention to when the speech signals are collected.The preprocess for the speech is analyzed thoroughly based on the features of speech and demand for the algorithms.Such work makes great contribution to understanding the algorithm and decreasing the computing amount.2.There are many algorithms with high performance for determined blind separation,in which the fixed-point algorithm have many strong points.We analyzed the difference in the algorithmsbased onkurtosisand entropy on the speech signal and we also analyzed the performance under the noisy condition.After studying the natural gradient algorithm,we found the natural gradient algorithm had high separation accuracy but a big iteration which would lead a low separation speed.In order to overcome such problems above,two types of variable step-size natural gradient algorithms are proposed to improve iteration efficiency so that the amount of iterations reduced.At last,the performance of the separation based the algorithms above was analyzed.3.We studied the algorithms for the convolutive mixing model in the time and frequency domain.Because the online de-convolutive method in the time domain has a big amount of iteration and low defense to noise,an improved method was proposed to overcome such problems above.After the algorithms in the frequency domain were studied carefully,we found that such algorithms could decrease the amount of computation greatly.Because speech signal has the amplitude correlation performance,we realized the blind algorithm based on the amplitude correlation performance.At last,an objective evaluation was done between the time and frequency domain algorithms.4.Nowadays the underdetermined blind separation is a hot point and the phenomenon that speech signals are underdetermined is normal.Because the classical Bofill potential underdetermined separation algorithm couldn't evaluate the mixing matrix accurately and had a weak defense to the noise,we improved two aspects.Through simulation and analysis;we prove the improved method has better performance in noise defense ability andeparation accuracy.