摘要
该文提出了一种高斯混合模型(GMM)参数估计的改进算法.原始的特征向量先经Schmidt正交化消除各维间的相关性,再用数学形态学方法估计出各维概率分布中混合分量的真实个数,最后按真实的混合分量个数用EM算法对各维分别作标量GMM参数估计.该方法能缓解GMM传统参数估计算法引起的“不易扩展”的不便.实验结果表明,将其应用于说话人辨认,能在较大幅度提高训练速度的基础上相对传统GMM参数估计方法获得更高的识别率.
A modified algorithm for parameter estinlation of Gaussian mixture model (GMM) is proposed. The original feature veetor is first transformed with the Schmidt orthogonal proeedure to remove eorrelation between pairs of dimensions. Subsequently, mathematie morphology (MM) is adopted to evaluate the aetual number of mixture eomponents in the probability distribution of eaeh dimension approximated by the sealar GMM. The parameters of eaeh sealar GMM are finally estimated by EM algorithm based on the eorresponding number of the mixture eomponents. This algorithm ean alleviate the ineonvenienee eaused.by the traditional veetor-based parameter estimation algorithm. When applied to speaker identifieation experiments, the results show that a higher reeognition rate is aehievable compared with that obtained with eonventional methods, and the training time is also signifieantly redueed.
出处
《上海大学学报(自然科学版)》
CAS
CSCD
北大核心
2005年第5期475-480,共6页
Journal of Shanghai University:Natural Science Edition