Abstract
In this paper, we present several features of minimum variance distortionless response (MVDR) based all-pole filters which are suitable for modeling all types of speech. In particular, we demonstrate how the MVDR all-pole spectrum, based upon time-domain correlations, can provide high quality spectral envelope modeling of voiced speech. Simulation results are included showing that the MVDR all-pole spectrum's modeling of voiced speech harmonics improves as the model order increases, leading to a monotonically decreasing spectral distortion. Furthermore, we show how the MVDR all-pole envelope can be enhanced by using forward-backward linear prediction. In addition, low order (10-14) MVDR based all-pole filters are examined and compared with other all-pole spectral envelopes. The reduced order MVDR all-pole spectrum is shown to compare favorably with linear prediction (LP) and LP cubic spline spectral envelopes in terms of spectral modeling and complexity.