Improving PLDA Speaker Verification Using Unlabeled In-domain Data: Towards Speaker Verification on Internet Audios
As the information security on network is being widely concerned,automatic speaker recognition technology could be used to find terror speeches containing some specific speakers.In this paper,we propose an unsupervised approach to improve the performance of an existing NIST-SRE-domain i-vector/PLDA system for the internet speaker verification application with a set of non-speaker-label audios collected from the internet.A speaker factor vector in i-vector space is extracted for each audio with the existing background models.Then a SVM classifier is applied on these speaker factor vectors to do speaker recognition.Speaker factor vectors of those given unlabeled in-domain data are used as the negative samples to train speaker-dependent SVM models.Experiments are conducted on NIST SRE 2010 condition-1,condition-2 task and an internet test-set.Results on the internet test-set shows that the propose approach achieves a relative performance improvement of about 50%in both EER and minDCF over the baseline i-vector/PLDA system.
speaker recognition i-vector/PLDA domain mismatch
HUANG Houjun YUAN Qingsheng ZHOU Ruohua BAO Xiuguo YAN Yonghong
Institute of Acoustics,Chinese Academy of Sciences,Beijing,China In Institute of Information Engineering,Chinese Academy of Sciences,Beijing,China;National Computer Institute of Acoustics,Chinese Academy of Sciences,Beijing,China;National Computer network Emergency
国内会议
武汉
英文
264-268
2015-05-26（万方平台首次上网日期，不代表论文的发表时间）