Next: Distribution Assumptions Up: Summary of Variations Previous: Summary of Variations

Training Criterion

The parameters of the recurrent network are estimated with a discriminative training criterion. This leads to a mechanism for estimation of the posterior probability for the phones given the data. Standard HMM training utilises a maximum likelihood criterion for estimation of the phone model parameters. The recurrent network requires substantially fewer parameters because discriminative training focuses the model resources on decision boundaries instead of modeling the complete class likelihoods.