计算机应用与软件2016，Vol.33Issue(9)：159-162,4.DOI:10.3969/j.issn.1000-386x.2016.09.038
一种改进的有监督训练的 TV 语种识别方法
AN IMPROVED LANGUAGE IDENTIFICATION METHOD USING SUPERVISED TOTAL VARIABILITY
摘要
Abstract
Traditional GMM-TV (Gaussian mixture model-total variability)system is benefited from its good recognition effect and excellent recognition efficiency,and has been widely used in language identification (LID).However the training process of load matrix T is unsupervised,this leads to its classification space not being optimised the best.Existing supervised-TV (S-TV)algorithm,through stitching a vector with tag information on mean super vector,makes the training process of T matrix become a supervised process,but it only achieves a little performance gain while introduces the problem of load matrix’s freedom.In this paper we propose an improved S-TV method which puts a regularisation item into the objective function for solving the freedom problem and meanwhile greatly improves its classification effect. The improved system achieves excellent effect in the experiment on 30s dataset of NIST LRE2009,the equal error rate (EER)reduces to 4.96% from 5.40% and the fusion system’s EER has even reached 3.86%.关键词
语种识别/TV 系统/有监督训练/载荷矩阵Key words
Language identification/TV system/Supervised training/Load matrix分类
信息技术与安全科学引用本文复制引用
张翼飞,腾潇琦..一种改进的有监督训练的 TV 语种识别方法[J].计算机应用与软件,2016,33(9):159-162,4.基金项目
北京市科委项目（Z141100006014002）。 （）
