Modelling correlation of an ensemble
Francis J. Smith, J. Ming, P. Hanna, Darryl Stewart · 2003
Two methods have been derived for modelling of uncertain correlation in an ensemble of quantities. One is designed for an unordered ensemble and is valid for Gaussian and non-Gaussian distributions, the second is designed for a Markov sequence, and is valid for a Gaussian distribution. When applied to speech the results for the first method based on an unordered ensemble demonstrate that the traditional assumption in speech that the correlation between frames can be neglected is invalid. The correlation is on average is close to 80% of the maximum possible correlation between the frames. This application demonstrates that this method can be used effectively to solve problems when correlation in an ensemble is important, but is unknown.