基于HMM与神经网络的声学模型研究

An Acoustic Model Based on HMM/ANN

神经网络能依靠权值进行长时间记忆和知识存储,但是对输入模式的瞬时相应的记忆能力比较差;而隐马尔科夫模型的短时记忆的能力比较强,但是假定的前提又与实际情况不符.因此,采用HMM和ANN的混合模型来取双方之长,并在这种混合模型的基础上,对神经网络从结构设计、训练、到训练后期的结构调整进行了全程的优化;应用隐节点剪枝算法,并利用广义的Hebb规则重新确定网络的参数.实验表明,这种混合模型在语音识别中取得了良好的效果.

The Artificial Neural Network （ANN） can depend on weight values to store memory and knowledge for a long time. However it possesses a weak memory, not being suitable to store the instantaneous response to various input modes. The Hidden Markov Model （HMM） is better in instantaneous memory,but the presupposition precondition is not according with the real situation. So we design a hybrid HMM/ANN model to overcome the flaws of using either of them. And basing on this model, we make a global optimization for ANN in structure design, training and structure adjustment in the later period of training. We propose an algorithm to prune hidden nodes in a trained neural network,and utilize the generalized Hebbian algorithm to reconfigure the parameters of the network. Some experiments show that the hybrid model has a good performance in speech recognition.