文章摘要
吴延渠,曾以成,蒋阳波.MMCE算法在FAGMM中的应用[J].声学技术,2010,(1):83~86
MMCE算法在FAGMM中的应用
Application of MMCE algorithm to FAGMM
投稿时间:2009-01-09  修订日期:2009-03-26
DOI:
中文关键词: 因子分析高斯混合模型(FAGMM)  改进的最小分类错误(MMCE)算法  FAGMM+MMCE模型
英文关键词: factor analyzed Gaussian mixture model(FAGMM)  modified minimum classification error (MMCE) algorithm  FAGMM+MMCE model
基金项目:湖南省自然科学基金(08JJ5031)
作者单位E-mail
吴延渠 湘潭大学光电工程系, 湖南湘潭, 411105  
曾以成 湘潭大学光电工程系, 湖南湘潭, 411105 yichengz@xtu.edu.cn 
蒋阳波 湘潭大学光电工程系, 湖南湘潭, 411105  
摘要点击次数: 1137
全文下载次数: 1617
中文摘要:
      提高说话人模型的识别性能一直是语音识别领域的一个重要课题。因子分析高斯混合模型(FAGMM)是因子分析方法与高斯混合模型(GMM)结合而成的多维概率统计模型,能更好地表征语音特征矢量的相关性,然而模型参数过多导致不能实现很好的分类。把改进的最小分类错误(MMCE)算法应用于该模型,形成一种新的FAGMM+MMCE模型,能解决前述问题,而且克服了传统的最小分类错误(MCE)算法在系统训练时不灵活、训练速度慢的缺点。实验结果表明,在30个说话人的识别应用中,本模型的识别率随着高斯混合数的增加而提高,较传统的MCE算法,识别率平均提高了3%,训练时间也平均节省了20%,说明该方法是有效的。
英文摘要:
      To improve performances of speaker models is a significant research subject in the field of speech recognition.The factor analyzed Gaussian mixture model(FAGMM) is a multi-dimensional probability statistical model through the combination of factor analysis and GMM,and can reflect the intraframe correlation of feature vectors well.However,it has too many model parameters to classify.In this paper,a modified minimum classification error (MMCE) algorithm is applied to the model,which forms a new FAGMM+MMCE model.The new model not only realizes the optimized classification of FAGMM model parameters,but also has the better flexibility and the faster training speed over the traditional MCE algorithm.The experimental results show that the identification rate of the new model continuously increases with Gaussian mixture number,and compared with the conventional MCE,it increases by an average of 3%.Besides,the training time also reduces by an average of 20% with a 30 speaker population.Such proves the new model to be effective.
查看全文   查看/发表评论  下载PDF阅读器
关闭