Study On The Technique Of Bandwidth Extension In Speech Communication System

Nowadays,the quality of voice transmitted through communication systems is less satisfactory. Because of the interruption of transmission and the limited of link Layer, the narrowband speech which bandwidth is limited to 0KHz~4KHz will be heard in both ends of the call. Narrowband speech can simply be used for basic communication, but it is not enough to meet the requirement of high quality speech communication. Even though the technique of wideband coding-decoding has been implemented and becomes popular, the entire systems could not be replaced by a brand new wideband system in the short term. As an optimal method, the bandwidth extension(BE) not only can enhances the quality of the bandwidth limited speech, but also becomes a hot research topic in speech processing.In this thesis, a prototype of speech bandwidth extension system has been designed and implemented in the framework of linear source-filter model. The whole BE system consists of four subsystems: the feature extraction, the reconstruction of spectral envelope, extension of excitation and the Analysis-By-Synthesis(ABS) of speech. The main contents of the thesis are as follow:(1)Corresponding methods like feature extraction to preprocess and analyze the speech signal as well as the linear source-filter model will be introduced following the study of the mechanism of speech. The procedure for computing the pitch period, linear cepstrum coefficient and Mel frequency cepstrum coefficient of speech will also be described.(2)With the summary of existing BE methods, the excitation generation approach based on spectral shift will be proposed as a new approach. Moreover, a new strategy of spectral envelope extension based on the grouped audio feature and Gaussian Mixture Model will be proposed too. The new approach and the new strategy will be composed as a brand new prototype of bandwidth extension system.(3)Experiments have been set up to demonstrate the performances of the existing BE methods and the new proposed methods. The results of the objective evaluation have shown that the proposed methods have achieved better performance. The speech synthesized by new BE methods has better quality than the one reconstructed by existing BE methods.