混合语音识别系统的一种新的简化神经网络结构

A Simplified Neural Network Architecture for a Hybrid Speech Recognition System

研究适用于隐马尔可夫模型 ( HMM)结合多层感知器 ( MLP)的小词汇量混合语音识别系统的一种简化神经网络结构。利用小词汇量混合语音识别系统中的 HMM状态所形成的规则的二维阵列 ,对状态观测概率进行分解。基于这种利用 HMM的二维结构特性的方法 ,实现了用一种由多个简单的 MLP所组成的简化神经网络结构来估计状态观测概率。理论分析和语音识别实验的结果都表明 ,这种简化神经网络结构在性能上优于 Fran-co等人提出的简化神经网络结构。

A simplified neural network architecture is presented. It is applicable to any small vocabulary hybrid speech recognition system that combines hidden Markov model (HMM) with multi-layer perceptron (MLP). By using the regular two-dimensional array of HMM states in a hybrid speech recognition system with small vocabulary size, the factorization of observation probabilities is performed. Based on this approach, by using the property of the two-dimensional structure possessed by the HMM, a simplified neural network architecture consisting of multiple simple MLPs, employed to estimate observation probabilities is achieved. The theoretical analysis and the results of the speech recognition experiments show that the simplified neural network architecture is superior to that of Franco, et al .