ARMA filter design for music analysis/synthesis

Preliminary results that demonstrate the advantages of using a recently proposed method to design autoregressive moving average (ARMA) filters for music analysis/synthesis are presented. This approach uses homotopy continuation methods to provide the theoretical assurance that the desired fixed-order optimal ARMA model will consistently be computed. Previous ARMA estimation techniques fail to produce acceptable estimates of these undermodeled spectra. Equation error methods exhibit significant bias, and gradient-based or direction set methods may diverge or converge to local minima. These models will be implemented in efficient pipelined VLSI circuits and used in a code excited linear prediction (CELP)-like framework to produce high-quality synthesized music at low cost