Hybrid speech recognizers usually involve a frame-based classification followed by a segment alignment system, trained separately. The simplicity of such systems is counterbalanced by the lack of a global optimisation scheme for the whole system. In this paper we propose a discriminative training method for MLP/HMM hybrids based on the optimization of a global cost function at the phone recognition level. The MLP weights, usually updated according to the target values, are now updated according to the misclassifications present in the output of the system. Results are presented for the TIMIT phone recognition task and show that this method compares favourable with recent published results in this task. The global discriminative training method was also applied to a Portuguese speech database leading to promising results.