基于LM算法的神经网络语音识别  

Neural network speech recognition based on LM algorithm

作　　者：葛玲[1] 贾志成[1] 夏克文[1] 王霞[1] 

机构地区：[1]河北工业大学信息工程学院,天津300132

出　　处：《计算机工程与设计》2006年第14期2534-2536,2539,共4页Computer Engineering and Design

基　　金：国家自然科学基金项目(60377020)

摘　　要：由于语音识别中采用标准BP算法存在的训练速度慢、容易陷入局部极小等问题,提出一种基于稳定、快速的Leve-nberg-Marquardt算法的神经网络语音识别方法,主要包括语音信号预处理、特征提取、网络结构优化设计、网络学习训练和语音识别等过程。其中网络隐含层节点数的选取采用黄金分割优选法。试验仿真表明,LM算法明显提高了网络训练速度,减少了训练时间,其效果优越于标准BP算法。For the defects of standard BP algorithm used in speech recognition, such as very slow training speed, very easy to falling into local minimization, and so on, a new method of neural network speech recognition is presented based on a stable and fast Levenberg- Marquardt algorithm, which includes following processing steps, speech signal preproeessing, characteristic extracting, optimization design of network structure, network training and speech recognizing. Besides, an optimization algorithm based on the principle of golden section is adopted to design the number of hidden layer nodes in neural network. The simulation experiments shows that the Levenberg-Marquardt algorithm is superior to that of standard BP, which obviously quickens training speed and decreases training time, and the application effect is notable.

关 键 词：神经网络 语音识别 标准BP算法 LEVENBERG-MARQUARDT算法 黄金分割优选法 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]