Hybrid HMM-NN for speech recognition and prior class probabilities

During the last years, speech recognition technologies have started their migration from research laboratories to real word applications gaining market shares. Although this shows that paradigms like Neural Networks have reached a high level of accuracy in modeling speech, it must be realized that there is still room for improving recognition performances exploiting the feedbacks coming from the applicative fields. In these cases, in fact, precious application dependent speech material can be recorded, and used to train the acoustic models in order to improve the behaviour of the recognizer on target dictionaries. The best results can be achieved when an iterative, refining process is set up. Unfortunately, speech corpora coming from the field are seldom phonetically balanced and this can cause the performances of the Neural Network to get worse, wasting the benefits of the refining process. In this paper, the problem of Prior Probability normalization has been faced and a method for Prior Probability normalization has been investigated, with the important characteristic of being applicable simply through a modification of the biases at the end of the training phase (therefore on trained nets). An experimentation on several languages is reported, showing the Prior Probability normalization seems quite useful to improve recognition accuracy and to get rid of some undesired effects of training data-bases not perfectly phonetically balanced.