Novel speech spectra all-pole modelling based upon selective even-samples linear prediction

In this paper, a novel linear predictive method (SELP) for speech spectra modelling is proposed. This method allows one to develop an all-pole filter which combines p+1 consecutive even preceding samples of the speech signal x(n) into p pairs for linear extrapolation. In addition, a weighting selective scheme is employed to obtain high signal-to-error ratio (SER). Comparing to the traditional LPC modelling, the proposed filter´s order is then raised to 2p+2 when both filters are with p-normal equations. In order to demonstrate the proposed method usefulness, this new model is simulated at 22.05 kHz speech spectra. Experimental results show that at least 10 dB SER improvement is obtained with p=5 when compared with that of LPC modelling