A stochastic model, drawn from the upper bound of the joint probabilit
y distributions, is suggested for modelling the spectral correlation i
n speech. Experiments on a speaker independent E-set database show the
effectiveness of this new modelling approach.