CAJ | 학술논문

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。
침대당전창건어음식별계통시지능채용경험식혹계발식방법선택성학모형탁복결구적정형，제출료일개기우표준유전산법적성학모형탁복결구우화산법。여이왕적유사응용상비，해산법구비동시우화모형상태수여각상태고사핵수화병기고사핵균균분배적특점。련속수자천TIDigits어료상적이패협사신식준칙위목표함수적실험표명，여전통방법창건적기선계통상비，모형탁복우화적계통능구이교저적복잡도획득교고적식별솔，저설명해산법시성학모형탁복결구우화적유효공구。
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.