A Stochastic Speech Model Supporting W-Disjoint Orthogonality

In previous work, we have successfully used an ideal joint sparseness assumption: W-Disjoint Orthogonality (WDO). This assumption, that the time-frequency representations of the sources have disjoint support, is satisfied in an approximate sense by many signals of practical interest, including speech. Here we discuss results derived from a stochastic model of speech signals that justify the WDO hypothesis. If the magnitude of the timefrequency components of the source signals have Laplacian priors, a subset of their maximum á posteriori (MAP) estimates are guaranteed to satisfy the WDO assumption.