摘要
基于高斯混合模型(GMM)的序列评分和评分空间的映射方法可以有效地应用于数据稀疏及数据量较少的序列分类.然而在语音序列分类中,这种方法不能反映序列次序和时长差别.因此,本文提出了一种观测值次序均衡方法,即序列中观测值的评分与其出现的次序和它之前出现的所有观测值相关.同时,我们在次序均衡中引入了功能因子,对于不同目的的语音序列分类,可以通过调节功能因子而指定相应的评分函数.实验结果证明这种新的评分方法能有效提高基于GMM概率模型的语音序列分类性能.
Classification method based on GMM scoring and score space mapping is effectively used in sequence classification of sparse data. However, when applied to speech sequence, this method shows its limitation in discriminating sequences with different observation order and duration. Therefore in this work, we employ order compensation method and function factor for improvement. In our new method, score of an observed feature data is related to its present order and all its predecessors in the sequence. The new method can also be applied to different type of speech classification by using different function factor. Experiment result shows the new method can improve classification accuracy of speech sequence.
出处
《湖南大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2008年第11期79-82,共4页
Journal of Hunan University:Natural Sciences
基金
国家自然科学基金资助项目(60275005)
关键词
语音序列分类
GMM评分函数
次序均衡评分
功能因子
speech sequential classification
GMM scoring function
order compensation
function factor
