摘要
对于电话手机语音的文本无关说话人确认,语音受到传输信道、话筒等的影响,失配问题尤为显著.为此提出一种GMM-UBM框架下基于失配信息子空间的说话人确认失配补偿方法,该方法利用失配信息子空间和基准信息模型获得训练和测试语音的相对失配信息,此失配信息是失配信息子空间中超矢量的一种线性组合,补偿更为准确,能够根据失配对不同语音类的不同影响,对不同的语音类进行相应的补偿.NIST06数据库上的实验表明,经过失配补偿的系统性能在EER和MinDCF有近50%的提高.
One of the major remaining challenges to improve performance in a text-independent speaker verification system is reducing the mismatch which includes transmission channel effects and transducer characteristics between training and testing. A mismatch information subspace （MIS） based channel compensation approach for speaker verification under the framework of GMM-UBM was proposed. The MIS supervectors were weighted to estimate mismatch information. With this algorithm, mismatch information can be compensated by different mixtures of GMM. Experiments on NIST06 SRE database show that the proposed approach can improve system performance by as much as 50% both in EER and MinDCF over a baseline system.
