摘要
在应用最小分类错误(MCE)准则对识别说话人的高斯混合模型(GMM)训练中,采用一个权重函数来确定说话人模型参数调整量的权值,使得比较近的竞争说话人模型的权值大,比较远的竞争说话人模型的权值小。并采用梯度概率递减算法来实现损失函数的最小化,有效提高了说话人识别的速度和精度。
In the paper,Minimum Classification Error(MCE) criterion is used to train Gaussian Mixture Model(GMM) for Speaker Recognition.A weight function is determined to adjust the weight of speakers to assign a greater weight for more similar speaker.Algorithm of gradient probability degression is utilized to minimize lost function.The presented method has improved speed and accuracy of speaker recognization.
出处
《自动化与仪器仪表》
2010年第6期21-23,共3页
Automation & Instrumentation
关键词
说话人识别
高斯混合模型
最小分类错误
Speaker Recognition
Gaussian Mixture Model
Minimum Classification Error