COMPARISON OF DISTANCE MEASURES IN DISCRETE SPECTRAL MODELING

We present a general adaptive approach for discrete spectral modeling by minimizing different spectral distances in an adaptive filtering context. By comparing the steady-state error for several spectral distance measures for real speech, we study the performance of these important distance measures. We also present a fast converging algorithm for the COSH distance that is shown to yield a more accurate estimate of the spectral envelope than the Itakura-Saito (I-S) distance measure.