一种改进的PRSVM语种识别方法  

Improved PRSVM Method for Language Recognition

作　　者：金恬[1] 宋彦[1] 戴礼荣[1] 

机构地区：[1]中国科学技术大学电子工程与信息科学系,安徽合肥230027

出　　处：《小型微型计算机系统》2011年第5期1017-1020,共4页Journal of Chinese Computer Systems

基　　金：安徽省自然科学基金项目(090412056)资助

摘　　要：传统的PRSVM方法存在以下问题:音素识别器的符号化结果与原语音存在不一致;向量空间维数高,稀疏.针对以上问题,先改用更适合噪声环境下连续电话语音的音素识别器,并采用词图结构改善解码效果,再分别用全局和局部两种隐含语义分析策略改进区分性训练问题.实验表明,本方法不但有效,而且大大减少了运算量.在NIST2007语种识别评测30秒、10秒和3秒任务中,本方法比基线系统性能有显著提高,等错误率分别相对降低了22.3%、14.7%和12.2%.Although Language Recognition based on Phone Recognition followed by Support Vector Machine（PRSVM） can achieve good performance,there are several problems： the inconsistency of phone recognition is very serious due to the noise;the vector space is high-dimensional and sparse.To tackle these problems,a language recognition method with lattice and latent semantic analysis is proposed.In this method,we apply an NN/HMM phone recognizer and improve the performance of Viterbi decoding by lattice,then a dimensionality reduction method mapping term space to topic space is applied based on singular value decomposition of term-document matrix in a global strategy and a local strategy respectively.Finally,we use a discriminative training method,support vector machine to do the classification.The experiments on NIST Language Recognition 2007 30、10 and 3 sec evaluation task show advantage of our proposed method,that the Equal Error Rates are reduced relatively about 22.3%、14.7% and 12.2%.

关 键 词：语种识别 词图 支持向量机 隐含语义分析 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]