Empirical risk optimisation: neural networks and dynamic programming

The authors propose a novel system for speech recognition which makes a multilayer perceptron and a dynamic programming module cooperate. It is trained through a cost function inspired by learning vector quantization which approximates the empirical average risk of misclassification. All the modules of the system are trained simultaneously through gradient backpropagation; this ensures the optimality of the system. This system has achieved very good performance for isolated-word problems and is now trained on continuous speech recognition. >