Speech modeling via model reduction

Using model reduction, a new approach for low-order speech modeling is presented. In this approach, the modeling process starts with a relatively high-order (full-order) autoregressive (AR) model obtained by some classical method. The AR model is then reduced using the a state projection method, operating in the state space. The model reduction yields a reduced-order autoregressive moving-average (ARMA) model which interestingly preserves the key properties of the original full-order model such as causality, stability, minimality, and phase minimality. Line spectral frequencies (LSF) and signal-to-noise ratio (SNR) behavior are also investigated. To illustrate the performance and the effectiveness of the proposed approach, some computer simulations are conducted on some practical speech segments.