采用高斯概率分布和支持向量机的说话人确认

Speaker Verification Based on Gaussian Probability Distribution and SVM

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 text-independent speaker verification research, the probability distribution against the universal background model （PD-UBM） is calculated. And the score of each UBM Gaussian mixture is adopted as the input feature of the support vector machine （SVM） during the training and testing process. The proposed PD-UBM algorithm with linear kernel function can obtain the same or better performance as the generalized linear discriminant sequence （GLDS） kernel system. Furthermore, if the scores of the Gaussian mixture models （GMM-UBM） , the GLDS and the PD-UBM are combined, the significant improvement of the system can be achieved. In 2006, on NIST 1conv4w-1conv4w speaker recognition evaluation （SRE） corpus, the fusion system obtained 25% relative improvement equal error rate （ERR） of over the GMM-UBM system.