Effects of finite-precision conversion on linear predictive coefficients

Two novel methods are presented for improving the finite-precision conversion of linear prediction coefficients in the form of reflection coefficients. The first method is based on the improvement of the dynamic range of the intermediate variables in the Schur (1986) recursion using the block exponential format. For the second method, improvement is achieved by changing the statistical properties of the quantisation noise incurred during conversion. This is achieved with the addition of low-power uniform uncorrelated noise during the conversion from a lattice filter structure to direct-form filter structure. Statistical tests performed over a 1000 test frames reveal a significant reduction in both the variance and mean of the error compared with the standard fixed-point LPC filter