DNN-HMM acoustic model as compared with conventional acoustic models based on the GMM-HMM, the only difference is replaced with DNN GMM to model the probability of observing input speech signal. DNN has the following advantages as compared with the GMM:
1, DNN distribution of the acoustic features need not be assumed to obey;
2, DNN input frames may be continuous stitching, it is possible to make better use of the context information;
3, DNN training process can be used stochastic optimization algorithms to achieve, rather than using the traditional batch optimization algorithm, so when training large-scale data can be very efficient training, obviously, the larger the training data, resulting the more accurate acoustic model, the more help to improve speech recognition performance;
4, in the sound producing pattern classification, the distinction of formula the DNN than GMM model is also more appropriate model for this production.