... Master¹

Supported by the National ScienceFoundation

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... here.²

As noted above, other Bayesian models and representations may be more appropriate when $(u,v,r)$ does not capture source knowledge well, and we describe those in future work.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... sources³

This allows us to include the very likely possibility that only one source is active at one time. This also allows us to see the previous generation DUET and DASSS systems a subset of the current system; the current system becomes the old system by setting the probability of all non-NULL two source combinations to zero.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... frequencies.⁴

The bin corresponding to zero frequency must not be included due to the division by zero needed above.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... training.⁵

Experiments showed that this may make the results appear overly optimistic, however.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.