论文部分内容阅读
在采用支持向量机的说话人确认中,将语音特征参数相对于通用背景模型各高斯分量的概率分布作为支持向量机输入,在线性核函数的情况下,系统能取得与广义线性判别式序列核函数(GLDS)几乎相同的识别率,同时该高斯概率分布算法能够与混合高斯背景模型、广义线性判别式序列核函数的得分进行融合,进一步提高识别性能.在2006年 NIST SRE 1conv4w-1conv4w 数据库上,融合后的系统相对于基线的混合高斯模型最多有25%的等错误率下降.
In the speaker recognition using SVM, the probability distribution of the speech feature parameters relative to the Gaussian components of the universal background model is input as a support vector machine. In the case of a linear kernel function, the system can obtain the probability distributions consistent with the generalized linear discriminant sequence kernel (GLDS), and the Gaussian probability distribution algorithm can be fused with the mixed Gaussian background model and the score of the generalized linear discriminant sequence kernel function to further improve the recognition performance.In the 2006 NIST SRE 1conv4w-1conv4w database , The fusion system has a maximum error rate of 25% relative to the baseline mixed Gaussian model.