An empirical risk optimizer for speech recognition

The authors propose a new system for speech recognition which results in cooperation between a multi-layer perceptron and a dynamic programming module. It is trained through a cost function inspired from learning vector quantization which approximates the empirical average risk of misclassification. All the modules of the system are trained simultaneously through gradient back-propagation, which ensures the optimality of the system. This system has achieved very good performance for isolated-word recognition problems and was trained also on continuous speech recognition