基于遗传算法的声学模型拓扑结构优化  

Genetic algorithm based optimization of acoustic model topologies

作　　者：包希日莫[1] 高光来[1] 张璟[2] 

机构地区：[1]内蒙古大学计算机学院,呼和浩特010021 [2]西安理工大学计算机科学与工程学院,西安710048

出　　处：《计算机工程与应用》2014年第14期5-8,63,共5页Computer Engineering and Applications

基　　金：国家自然科学基金(No.61263037)

摘　　要：针对当前创建语音识别系统时只能采用经验式或启发式方法选择声学模型拓扑结构的情形,提出了一个基于标准遗传算法的声学模型拓扑结构优化算法。与以往的类似应用相比,该算法具备同时优化模型状态数与各状态高斯核数和摒弃高斯核均匀分配的特点。连续数字串TIDigits语料上的以贝叶斯信息准则为目标函数的实验表明,与传统方法创建的基线系统相比,模型拓扑优化的系统能够以较低的复杂度获得较高的识别率,这说明该算法是声学模型拓扑结构优化的有效工具。Aiming at the current situation of selecting acoustic model topologies empirically or heuristically, an acoustic model topology optimization algorithm based on standard Genetic Algorithm（GA）is proposed. Compared with the previ-ous similar algorithms, it is an automatic one with the characteristics of optimizing model state number and kernel num-bers simultaneously and rejecting uniform allocation of Gaussian kernels. Experiments using Bayesian Information Criterion （BIC）as the objective function on TIDigits corpus show that, the speech recognition system whose model topologies are optimized using the GA based algorithm is able to obtain higher recognition performance with smaller number of parame-ters compared with the baselines built in the conventional way, which indicates that the GA based algorithm is an effective tool of optimizing acoustic model topologies.

关 键 词：隐马尔可夫模型 遗传算法 语音识别 声学模型拓扑结构 贝叶斯信息准则 

分 类 号：TN912.34[电子电信—通信与信息系统]