Mixed Excitation The Mvdr Speech Coding Technology

With the development of communication technology, speech plays an important role as one of the main methods for exchanging information in communication systems. Because of the contradiction between coding rate and quality of synthetic speech, the need for better coding method of speech is increasing especially when the bandwidth is limited. As a consequence the medium-low rate speech coding has been paid much attention in recent years.The accurate description of track model and excitation model is the key of the Medium-low rate speech coding. However the popular all-poles track model's parameters, which are calculated by LP method, can not provide a very accurate speech spectrum envelop. And medium-low rate speech coding methods have advantages and flaws in describing excitation, and can't describe all kinds of speech excitation perfectly. All of those have led to that the synthetic speech quality can't be improved further in Middle-low rate speech coding.In track modeling, this thesis has analyzed the base of LP method modeling spectrum envelops, the minimum mean square rule. This rule leads to that LP method can analyze formant frequency well, but over-emphasize the power in formant frequency, so short-time spectrum envelop can't be modeled perfectly, a sharp contour appears in formant frequency of spectrum envelop. Moreover, because there are some problems in LP method, the thesis has discussed apopular method in array processing — — minimum variance distortionless response andcompared the performances of the two methods in modeling spectrum envelop. Also in the base of the method of high-order minimum variance distortionless response, has discussed the method of order-reduced minimum variance distortionless response and the calculation method of minimum variance distortionless response coefficients.In the part of excitation models, this thesis has analyzed mixed excitation model and code-excited model, which are successful models in the fields of parameter coding and mixed coding. Mixed excitation model can extract the pitch better and describe the pitch wave better, so it has the capability of describing the excitation of sonant. Code-excited model aims at matching original wave, so the excitation of unvoiced can be described better, which mixed excitation can't achieve. This thesis has combined the two model, and designed the decoder indetail, which adopts CE/ME hybrid excitation model and MVDR method. The solutions of parameter calculation and bits distribution have also been discussed here.In the end, this thesis evaluates the above methods by experiment. The results are that minimum variance distortionless response method has better performance than LP method, and the 4kbps coder proposed in this thesis is better than 4.8kbps code-excited LP coder by ear.