Celp Algorithm Based On Arma Model Research

Speech coding technology is the important area of common concern among the communication and information technology. So far, the Code Excited Linear Prediction(CELP) is a hybrid coding technology that is used most widely. It has advantages of high quality synthetic speech, excellent resistance to noise and multiple forwarding capability etc, so it has been used in low-rate speech coding widely, representing the direction of development of the low-rate speech coding. Based on the CELP algorithm model, many speech standards are made. And the all-pole model were used in all of them. In fact, the speech signal includes zeros, the feature of the speech signal can’t only be reflected by AR model. Therefore, the thesis studies on the ARMA model. The main contents of the thesis are shown as follows:(1) The codec principle of CELP algorithm is considered. FS-1016, with a rate of4.8kb/s, is used as an example, linear prediction, perceptual weighting and codebook searching are eloborated intensively. FS-1016is simulated in Matlab, and we can see that, the time waveform of input speech can be restored by synthetic speech roughly, but the situation in the lower energy areas is so worse.(2) The model identification and judging model’s order for time series are considered, the way for judging the order of the speech signal is obtained. The type of the model can be identified by analyzing the correlation function graph, the order of the model can be judged by using the AIC criterion. Actually, the speech signal is a type of time series with short-term stationarity. Considering the correlation function graph and the feature of the speech signal, ARMA model is selected to research it. At last, the order of zeros and poles are judged by averaging the best zero-pole.(3) The model parameters and power spectrum are considered. Initial parameters are solved by using moent estimation and inverse function. Parameters estimation of ARMA model are obtained by using the conjugate gradient method in the least square method. According to the simulated results, the power spectrum of ARMA model is more accurate than that of AR model, which is more suitable to reflect the feature of speech signal.(4) ARMA model is used in CELP. The order of voiced and unvoiced speech is different in CELP. The linear prediction filter is consisted of parameters of ARMA model. Simulation results shows that, the time waveform of the synthetic speech of ARMA-CELP is much closer to the input speech, and the speech quality is better.