Dynamic Bayesian Networks incorporating a discrete noise variable for speech recognition

摘要

The model trained on speech at one SNR level is inappropriate for testing under various noise conditions. To improve the robustness of the recognizer, it is necessary to increase the types of speech to adapt to various test conditions. In order to enhance the performance of the baseline Dynamic Bayesian Network (DBN) which is subjected to training set under different noise conditions, this paper provides DBN incorporating a discrete noise variable for speech recognition. The experimental results show this model can deal with the mixed training set and get a fair performance in comparison with that trained on training set containing only one SNR level.