-
摘要: 在与文本无关的说话人识别研究中, 特征映射的方法可以有效减少信道的影响. 本文首先通过主成分分析的方法在模型域中估计出信道因子所在的空间, 然后通过映射的方法在特征参数域中减去信道因子的影响. 采用这种方法需要有信道信息标记的数据, 但是在特征映射时不需要对信道进行判决. 在NIST 2006年SRE 1conv4w-1conv4w数据库上, 采用本文推荐方法的系统相对基线系统在等错误率上降低了19\%.Abstract: In text-independent speaker verification research, feature mapping can reduce the bias by the channel. In this paper, the subspace of the channel is estimated by the generalized principal component analysis, then the bias of the channel is subtracted from the acoustic feature. The proposed algorithm requires labeled data in the training process but does not need the channel detection in the feature mapping process. In the NIST 2006 SRE 1conv4w-1conv4w corpus, the equal error rate (EER) of the proposed system can be down by 19\% against the baseline Gaussian mixture model (GMM) system.
点击查看大图
计量
- 文章访问数: 2031
- HTML全文浏览量: 29
- PDF下载量: 1663
- 被引次数: 0