Stabilizing discrete spectral modeling of audio signals

We investigate the numerical instability of discrete spectral modeling of audio signals. It is identified that the autocorrelation matrix is ill conditioned when the discrete harmonic spectra lie only in a low-frequency region of the entire signal bandwidth. A simple but efficient method that redefines the sampling frequency and associate harmonic frequencies is proposed to improve the numerical stability. Simulation results using real audio a signals indicate its superior stabilizing ability as well as improved accuracy in discrete spectral modeling.