In this work we present a new HMM training procedure which aims explicitly at minimizing the recognition error and increase the discrimination between competing phonetic classes. We propose to minimize a function of the model parameters and the training data which can be interpreted as a temporal integration of a "frame recognition error". An iterative algorithm is proposed in which the parameters of the probability distributions associated with each state of the HMM are modified with the objective of reducing this error function. Experimental evaluation of the algorithm showed improved recognition performance relative to the values obtained with maximum likelihood training.