Doppler-variant Modeling of the Vocal Tract: Stability Considerations

A common technique to deploy linear prediction to non-stationary signals is time segmentation and local analy-sis. In [1], the temporal changes of linear prediction coef?cients (LPCs) are modeled as a Fourier series. This allows analysis and optimization of larger speech segments, i.e., virtually global analysis. Possibly resulting non-minimum-phase prediction error polynomials are subject to all-pass ?ltering. We show that introducing the stabilizing ?lter does not deteriorate the overall predictor performance.