Rival training: efficient use of data in discriminative training

We evaluate a simple extension of the corrective training algorithm for reestimation of the acoustic parameters, using | in addition to misrecognized sentences | also a selection of correctly recognized sentences for discrimination. Our approach (called \rival training") is implementationally much less expensive than lattice{based discriminative training methods, since we apply a \hard" threshold criterion to select a subset of sentences for which a single competitor is used for discrimination. Still, significant performance gains are obtained compared to maximum likelihood and corrective training even for triphone models with 61 densities per mixture (on a digit string and a large vocabulary isolated word recognition task). Further, the hard selection scheme may be used to accelerate the training process due to faster convergence and by restricting the training process to a xed subset of training utterances.