Discriminative training is quickly becoming standard in speech recognition systems today. Most of the discriminative training results for large scale tasks so far use the Maximum Mutual Information (MMI) framework, but recent results for the Minimum Classification Error (MCE) framework suggest that MCE too yields good improvements in recognition accuracy on large scale tasks. In this talk, I will summarize the MCE framework, and present recent results for large scale speech recognition tasks obtained using both the MIT Galaxy system and the NTT Communication Science Labs SOLON system.