| This dissertation is concerned with statistical modelling of speech signals and its applications in Voice Activity Detector (VAD) and speech enhancement.; The statistical modelling of speech signals is investigated first. The probability distribution function (PDF) of speech signals is studied. Previous experimental results show that the PDF of long time speech signals could be approximated by a Laplacian distribution in time domain. A series of tests in different domains are performed where the Laplacian Modelling is compared with the Gaussian Modelling.; In this thesis, a new approach of VAD is developed based on the Laplacian statistical modelling of speech signals and Gaussian modelling of noise signals. The decision rule is derived by applying a Bayesian hypothesis test in the KLT domain. A two-state Hidden Markov Model is employed here to make use of the strong correlation in the consecutive occurrences of speech frames, as well as silence frames.; Then, the mentioned statistical modelling is applied to develop a new algorithm for speech enhancement. (Abstract shortened by UMI.)... |
