Study On Bandwidth Extension Of Speech Signal

Nowadays, there are two kinds of speech signals, narrowband speech and wideband speech. In the current public service telephone network, the speech signal is narrowband, whose intelligibility and naturalness are poor. Without any change of current network and devices, higher band signal can be built up from narrowband speech via bandwidth extension. The wideband speech, which consists of the rebuilt higher band speech and the original narrowband speech, gives better quality.In this thesis, several methods for bandwidth extension are described, including generation of excitation signal and extension of spectral envelope. A bandwidth extension system based on GMM model is introduced, and a new method on generation of excitation is presented. Unprecedentedly, wavelet transform is applied to bandwidth extension in this thesis. A method based on wavelet transform maximum to reconstruct wideband speech is proposed.The main tasks in this thesis are as follows:(1)The processing and model of speech signals production are studied briefly. The source-filter model and linear prediction of speech are also introduced. And algorithms of bandwidth extension are described,(2) The feature coefficients of speech signal are introduced. The linear prediction is described in detail. And an application of wavelet transform for speech signal is discussed.(3) Based on the source-filter model, the presented methods for bandwidth extension are briefly discussed. Several methods for generation of the excitation signal as well as extension of spectral envelope are discussed.(4) A bandwidth extension system based on GMM model is introduced, and a new method on generation of excitation is given. In addition, a new bandwidth extension algorithm based on wavelet transform maximum is proposed.(5) The performances of the proposed algorithms are evaluated by experiments and the results of objective and subjective tests are given.