Statistical modelling of speech signals

The Gaussian and Laplacian models of speech signals are investigated in this paper. We use different hypothesis tests to compare these two models. The Gaussian model has been widely used while our experimental results show that the probability density functions (PDFs) of speech signals are more like the Laplacian distributions. Based on the fact that the KLT and DCT have been excessively used in speech signal processing, the distribution of speech components in both decorrelated domains are also investigated. All the results illustrate that the speech signals follow Laplacian distributions both in the time domain speech samples and in the KLT or DCT (excluding DC) domains. The distribution of speech signals in uncorrelated domains can be assumed as a multivariate Laplacian.