HMM-based warping in neural networks
Yonggui Gao, Tai-Yi Huang, D.-W. Chen · International Conference on Acoustics, Speech, and Signal Processing · 2002
A speech recognition method using an integration of multilayer neural network and hidden Markov model (HMM) techniques which can treat the time sequence signal with temporal variations is described. As an efficient implementation of the method, a pretwist procedure, which can compensate the recognition error caused by the alignment, is proposed. The HMM-based warping method leads to networks which can respond in a more flexible way to variations in the temporal structure of speech. As an experimental result, 91.2% recognition accuracy was obtained on the vocabulary of the Chinese final vowel. This performance is much better than that of the neural network or HMM alone.>