摘要
在针对电话语音的语种识别系统中,训练语音和测试语音之间存在不同说话人的个性差异带来的干扰,是影响系统识别性能的一个重要因素.基于此,本文首先对当前语种识别系统中消除此影响的方法进行研究,对比分析它们各自的优缺点,选择将锚模型方法引入语种识别系统中,该方法将语料映射至说话人无关的锚超矩阵进而消除说话人相关信息.针对锚超矩阵的选择存在语种混淆和信息冗余等问题,本文并提出一种结合支持向量机的锚模型训练算法,该方法下得到的锚超矩阵更具语种区分性,并去除了混淆信息的影响,增强了矩阵的紧致性.实验结果表明,新方法下的锚模型映射方法能有效提高基线系统的识别性能,并降低了语种识别系统训练和识别时的计算量.
In language identification on the telephone conversation speech, there is a disturbance in the recognition rate caused by speaker variability between test and train utterances, which is a key factor in increasing the system＇s performance. To tackle this problem, firstly we study the different ways to eliminate this influence, contrast and analyze the advantages and disadvantages of them, then introduce the anchor model into the language identification system, in which the influence is eliminated by a projection of utterance into the speaker independent anchor super matrix. Then in order to solve the problems of languages mixture and information waste, a training algorithm of anchor model based on support vector machine is proposed, by which a language discriminative and information compact anchor super matrix is constructed, then we choose the projection again to achieve the goal of avoiding the impartibility sample in high dimensionality. The results indicate that new language discriminative anchor model trained by support vector machine is better for improving the performance and reducing the system＇s calculation of training and testing than the original baseline model.
出处
《小型微型计算机系统》
CSCD
北大核心
2013年第4期837-842,共6页
Journal of Chinese Computer Systems
基金
国家"八六三"高技术研究发展计划重点项目(2008AA011002)资助
关键词
语种识别
锚模型
支持向量机
区分性
language identification
anchor model
support vector machine
discriminative
