Modeling speech signals in the time-frequency domain using GARCH

In this paper, we introduce a novel modeling approach for speech signals in the short-time Fourier transform (STFT) domain. We define the conditional variance of the STFT expansion coefficients, and model the one-frame-ahead conditional variance as a generalized autoregressive conditional heteroscedasticity (GARCH) process. The proposed approach offers a reasonable model on which to base the estimation of the variances of the STFT expansion coefficients, while taking into consideration their heavy-tailed distribution. (C) 2004 Elsevier B.V. All rights reserved.