Pole-Zero Modeling and Its Applications to Speech Processing.

摘要

Autocorrelation Pole-Zero modeling identifies the parameters of a rational transfer function H(z) whose short time-lag autocorrelations either exactly match (Autocorrelation Partial Realization--APR or closely approximate(Autocorrelation Prediction--AP) those of a given spectrum. As a result, the spectrum of the H(z) obtained from either method approximates the gross structure of the given spectrum. APR uses the Pade approximation to determine the denominator coefficients of H(z). In contrast, (AP) uses only Linear Prediction (LP) to determine both the denominator and numerator coefficients. Therefore, once the autocorrelation function of the given spectrum is known, AP uses only linear operations and no Fourier Transformations to determine the parameters of H(z). Moreover, the resulting rational transfer function is guaranteed to be minimum phase and consequently stable. A dynamic filtering process, based on Wiener filtering and Autocorrelation Prediction, was developed to suppress the background noise from degraded speech.