A fixed point computation of partial correlation coefficients in linear prediction

This paper introduces a new method for the computation of the partial correlation coefficients from the autocorrelation sequence. Derived from Levinson's algorithm, it involves the calculation of the crosscorrelations between the inputs and the outputs of the successive models which are formed. Using these, a fixed-point implementation is found for real time speech analysis on a 16-bits microprocessor. The method also gives an approximation of the impulse response of the system, which may be useful for identification with moving average models.