Speech enhancement for linear-predictive-analysis-by-synthesis coders

Speech coding techniques commonly used in low bit rate analysis-by-synthesis linear predictive coders (LPAS coders) create a model that emphasizes the important features of a speech signal. The utilization of these coding methods for speech enhancement is shown. Specifically, the speech signal will be modeled as the output of a cascade of an adaptive formant filter and an adaptive pitch filter, driven by a white Gaussian process with a time changing variance. The parameter and signal estimation method, which is based on the Expectation Maximization (EM) algorithm and implements this speech signal model, is investigated. The proposed approach yields better performance both in SNR and subjective impression than do speech enhancement methods which use only AR speech signal parameters.