多特征和SVM改进的语音关键词识别系统  

Phonetic Keyword Recognition System Improved by Multi-features and SVM

作　　者：陈太波 张翠芳[1] CHEN Tai-bo;ZHANG Cui-fang(School of Information Science and Technology,Southwest Jiaotong University,Chengdu 611756,China)

机构地区：[1]西南交通大学信息科学与技术学院

出　　处：《小型微型计算机系统》2019年第11期2291-2296,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61503059)资助;四川省科技计划项目(2018GZ0008)资助

摘　　要：本文在HMM/Filler模型的基础上,通过对汉语的408个音节建立支持向量机SVM分类器提升了HMM系统的识别率,这408个SVM音节分类器有助于全面支持关键词的动态设定.针对SVM分类器训练数据的定长要求,本文通过对关键词特征帧进行等长化处理,使得对音节分类的SVM分类器得以训练.进一步,为了调节系统的融合性,本文提出概率融合公式,即通过设定融合参数λ来调整SVM分类器对HMM/Filler的作用程度.实验结果表明,当融合参数λ=0.3时,SVM与HMM/Filler融合的关键词识别系统效果最佳,此时融合系统相对于HMM/Filler基线系统识别率有着6.74%的提升,多特征训练的融合系统相对于单特征训练的HMM/Filler基线系统识别率提升近10%.Based on the HMM/Filler model,the recognition rate of the HMM system are improved by establishing SVM classifiers for 408 syllables in Chinese.These 408 SVM syllable classifiers help to fully support the dynamic setting of keywords.set.Aiming at the fixed length requirements of SVM classifier training data,this paper makes the SVM classifier for syllable classification training by equalizing the keyword feature frames.Further,in order to adjust the fusion of the system,this paper proposes a probability fusion formula,which is to adjust the effect of the SVM classifier on HMM/Filler by setting the fusion parameterλ.The experimental results show that when the fusion parameterλ=0.3,the keyword recognition system with SVM and HMM/Filler fusion is the best.At this time,the accuracy of fusion system has a 6.74%improvement compared with the HMM/Filler baseline system.The multi features training fusion system improves the recognition rate of the single feature training HMM/Filler baseline system by nearly 10%.

关 键 词：识别 隐马尔可夫模型HMM HMM/Filler模型 支持向量机SVM 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]