In this paper we describe an alternative estimation procedure for training Hidden Markov Models (HMM) called Competitive Forward-Backward (CFB) algorithm, which is aimed at minimizing the number of recognition errors. CFB integrates the LVQ3 neural network classification technique into the Baum-Welch training algorithm, and improves significantly the performance of HMM systems over previously reported procedures, such as Baum-Welch and Corrective Training.