Skip to Main Content
The main problem with pole-zero modeling of speech signals is its computational complexity, as a highly nonlinear optimization problem has to be solved for each analysis frame. We propose a method to save computation. At each new analysis frame, it is first decided whether the parameter estimation routine needs to be started. The decision is made by using statistical hypothesis testing techniques. Emphasis is laid on the discussion of various criteria based on the correlation and spectral properties of the residual signal.