Non-causal linear prediction of voiced speech

Noncausal linear prediction for voiced speech and an error-minimization-based algorithm for iteratively determining the predictor coefficients are introduced. The noncausal component is included to model the shape of the glottal excitation more accurately. It is shown that with the inclusion of the noncausal component, the prediction error signal is much closer to an impulse train than the error signal generated using a traditional, purely causal, linear prediction filter. Furthermore, the prediction gain achieved using the noncausal component is significantly higher than that achieved by a causal linear predictor of comparable order. This demonstrates the potential of the algorithm to decrease the bit rate needed for coding the excitation in linear prediction based speech coding algorithms, while improving the quality of the speech