多模型语音识别算法在智能客服程序中的应用  

The application of multiple-model speech recognition algorithm on intelligent customer service program

作　　者：于梦吟 方坤玉 孙昕 陈素婷 董梁 YU Mengyin;FANG Kunyu;SUN Xin;CHEN Suting;DONG Liang(Guangdong Tobacco Shanwei Co.,Ltd.,Shanwei 516600,China;School of Data Science and Engineering,South China Normal University,Shanwei 516625,China)

机构地区：[1]广东烟草汕尾市有限公司,广东汕尾516600 [2]华南师范大学数据科学与工程学院,广东汕尾516625

出　　处：《中国高新科技》2024年第15期28-30,共3页

摘　　要：文章介绍了一种多模型语音识别算法,为待识别的每一个类属训练了包含多个隐马尔可夫模型(HMM)的识别器,每个HMM在训练时会逐步适合未能正确识别的类属样本,得到的新模型对这些样本的识别能力渐次提高,最终合成得到的多模型识别器可以更好地识别分布较为分散的序列样本。这一算法被应用在一款智能客服程序中,对具有不同方言特征的问询短语、短句进行识别。实验结果表明,多模型语音识别算法在多方言、小词汇量语音识别任务中的表现优于单模型识别器和一些常用的语音识别接口。A multiple-model speech recognition algorithm is presented in this paper,with which a classifier comprising of a number of Hidden Markov Models(HMMs)is trained for each class in a multiple-class identification problem.The approach employs a training strategy that lays more emphasis on samples that are not correctly identified in earlier attempts.By combining the decisions of the composite HMMs,the multiple-model classifier could well recognize samples with a rather spread-out distribution.The algorithm is applied to an intelligent customer service program to identify phrases and short sentences articulated in different dialects.The experimental results show that the multiple-model classifier outperforms single-model classifier and some speech recognition interfaces in the cases of multiple-dialect and small-vocabulary.

关 键 词：多模型分类器 隐马尔可夫模型 语音识别 智能客服程序 

分 类 号：TD52[矿业工程—矿山机电]