| Information Theory-based training of Neural Networks applied to Hybrid Pattern Recognition Systems (2007) | |||||||||||||||
Abstract | |||||||||||||||
| Introduction While Neural Networks (NN) have been very successfully applied to static pattern recognition tasks like image classification, pure NN-speech recognition approaches often perform quite mediocre. So in many complex speech recognition tasks hybrid systems are used, that combine Neural Networks and traditional classification techniques. This is due to the fact that in general, Neural Networks cannot handle the difficulties of time alignment of the high dynamic nature of speech signals as well as the classical approaches with Hidden Models (HMM) do. There are two popular hybrid NN-HMM system philosophies used in speech recognition: 1. The Neural Network is used as probability estimator in the Hidden Markov Models. 2. Neural Networks are used to map the acoustic features on discrete labels. The dynamic time alignment of the produced labels is done by the Hidden Markov Models. A system overview is given in Fig. 1 2 System description The most hybrid s | |||||||||||||||
Publication details | |||||||||||||||
| |||||||||||||||