基于区分性Model Pushing的语种识别方法

提出一种区分性Model Pushing方法，将SVM训练出的支持向量沿最优分类面的法线方向进行适当移动，增大不同语种间的区分性，然后将移动后的支持向量反向应用于GMM。该方法既保留了SVM的区分性信息，又利用了GMM在短时语音上的优势，同时增加了目标与非目标的区分度。实验结果表明，区分性Model Pushing能有效地提高识别性能。...

Full description

Saved in:
Bibliographic Details
Published in电子技术应用 Vol. 38; no. 4; pp. 113 - 116
Main Author 刘伟伟 吉立新 李邵梅 徐文
Format Journal Article
LanguageChinese
Published 61906部队,江西鹰潭335000%国家数字交换系统工程技术研究中心,河南郑州,450002%61906部队,江西鹰潭,335000 2012
国家数字交换系统工程技术研究中心,河南郑州450002
Subjects
Pushing
区分性Model
语种识别
超平面法向量
高斯混合模型超矢量-支持向量机
语种识别
高斯混合模型超矢量-支持向量机
超平面法向量
区分性Model Pushing
Online AccessGet full text
ISSN0258-7998
DOI10.3969/j.issn.0258-7998.2012.04.034

Cover

More Information
Summary:提出一种区分性Model Pushing方法，将SVM训练出的支持向量沿最优分类面的法线方向进行适当移动，增大不同语种间的区分性，然后将移动后的支持向量反向应用于GMM。该方法既保留了SVM的区分性信息，又利用了GMM在短时语音上的优势，同时增加了目标与非目标的区分度。实验结果表明，区分性Model Pushing能有效地提高识别性能。
Bibliography:language recognition; discriminative Model Pushing; gaussian mixture model super vector-support vector machine; normal vector to hyperplanes
To improve the performance of short utterances in this model, this work proposes discriminative Model Pushing that moves the support vectors in the direction of the normal to the separation hyperplanes trained by SVM. Then the moved support vectors are pushed back to GMM. By this means, the discriminative information of SVM and the advantage of GMM in short utter- ances recognition are retained, and the diversity between targets and nontargets is enhanced. Experimental results demonstrate that the proposed discriminative Model Pushing produce improvement over the baseline GMM-UBM, GSV-SVM and Model Pushing.
Liu Weiwei, Ji Lixin, Li Shaomei, Xu Wen （1. National Digital Switching System Engineering＆Technological R＆D Center, Zhengzhou 450002, China; 2. Unit 61906, Yingtan 335000, China）
11-2305/TN
ISSN:0258-7998
DOI:10.3969/j.issn.0258-7998.2012.04.034