Difficulty in setting reject_threshold for reject prediction.

[rhythmize]

User specific GMMs(Gaussain Mixture Models) are trained for each of the new training user and a Generalized UBM(Universal Background Model) is trained for all the users. UBM is used along with GMM for reject prediction (when a user is not trained but still tries to predict). Although the system is working fine without UBM. The only issue with UBM is the evaluation of reject_threshold which results in the unexpected results.

P.S. : Any suggestions is appreciated.