The competitive forward-backward algorithm (CFB)

We present a novel neural network algorithm to train HMM models, called Competitive Forward Backward algorithm (CFB). It focuses on the minimization of the misclassification rate, rather than the classical maximization of the likelihood of each model. The essence of the CFB algorithm is the application of LVQ neural network classification technique into the Baum Welch algorithm. This algorithm is introduced for the first time in this work. Some initial experiments have shown that greatly outperforms the Baum Welch, and can be applied successfully to speech recognition