一种新的HMM/SVM混合语音识别模型  

A Novel Hybrid SVM/HMM Speech Recognition Model

作　　者：高家宝[1] 来羽[2] 

机构地区：[1]河池学院现代教育技术中心,广西宜州546300 [2]中州大学开放教育学院,郑州450006

出　　处：《控制工程》2016年第11期1802-1807,共6页Control Engineering of China

基　　金：广西教育厅科研基金(201010LX454)

摘　　要：提出了一种新的基于隐藏马尔可夫(HMM)和支持向量机(SVM)的混合HMM/SVM模型。该模型利用HMM完成语音时间序列建模,计算得到信息输出概率,输入SVM中进行学习,输出语音分类信息,以完成识别决策。在此模型基础上,设计了一种基于并行结构蛙跳搜索算法(PSFL)优化SVM参数的方法以提升噪声环境下的语音识别效率。PSFL改进蛙跳搜索算法的循环主体,能够在寻优过程中维持个体多样性和提高收敛速度。实验结果表明,PSFL具有更优的收敛速度和优化性能,混合SVM/HMM模型在干净和噪声环境均能够获得很好的语音识别效率。A novel hybrid hidden Markov model （HMM） / support vector machine （SVM） model based on HMM and SVM is proposed in this paper. According to HMM and SVM, the proposed model uses HMM to build series model of speech signal stream and computes the output probability of speech signals to be recognized. Then the output probability is put into SVM for learning in order to obtain the speech classification information required for the correct recognition decisions. Moreover, a parallel shuffled frog leaping algorithm （PSFL） is presented to optimize the parameters of SVM for improving the speech recognition efficiency. PSFL improves the main portion of standard frog leaping search algorithm in order to maintain the individual diversity and improve the convergence speed. The experimental results demonstrate that, PSFL owns superior convergence speed and advanced optimization performance, and the hybrid HMM /SVM model can reach encouraging efficiency of speech recognition in clean and noisy environment respectively.

关 键 词：语音识别 支持向量机 隐藏马尔可夫模型 小生境技术:共享机制 蛙跳搜索 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]