Personal Page

Thesis Title

All-Pole Spectral Envelope Modeling of Speech

Thesis Abstract

All-pole models pervade speech signal processing systems, including  speech coders, recognizers, and synthesizers. This dissertation provides some new insights and new methods of all-pole spectral envelope modeling of speech, particularly voiced speech. We propose the Minimum Variance Distortionless Response (MVDR) spectrum method for determining all-pole models of speech. In contrast to Linear Prediction (LP) all-pole spectral envelopes which tend to have sharp contours that overemphasize the large powered harmonics of voiced speech, the MVDR spectrum of suitable order yields a smooth contoured all-pole filter that provides an excellent spectral envelope model of voiced speech. In contrast to LP, the MVDR all-pole spectral envelope's modeling of voiced speech spectral powers improves as the

filter order increases, leading to a monotonically decreasing discrete log spectral distortion. We show that the MVDR all-pole spectrum of suitable order can model the powers of voiced speech spectral samples exactly. Several simulations are provided to demonstrate the efficacy of MVDR all-pole modeling of voiced speech, unvoiced speech, and mixed spectra.

Based on the insight of MVDR modeling, we also propose methods for determining lower order all-pole models suitable for speech coding contexts. First, we develop Reduced Order MVDR (ROMVDR) all-pole filters which are based on direct time-domain correlations of speech and attempt to translate the quality of high order MVDR spectral envelopes to lower order filters. It is shown that ROMVDR filters are

superior to LP filters in terms of modeling voiced speech, particularly in the neighborhoods of the perceptually important formants, and perform well in tasks germane to speech coding. Second,

we propose several methods for determining all-pole envelopes that match a set of speech spectral samples. Deserving of special mention is the Weighted All-Pole (WAP) method. The WAP filter, based upon a relationship between MVDR and LP spectra, performs well in speech spectral sample modeling, and does not suffer from sharp contour problems of other methods. This feature of smooth contoured spectral envelopes is shown to be more suitable for quantization from a high-rate vector quantization theory viewpoint.

Year of Graduation: 1999