【摘要】 说话人识别的本质就是模式分类。传统分类器算法中参数模型方法的主要缺点是预先假定的概率分布函数形式不一定符合待分类的数据。非参数模型方法,如PNN分类器,可以有效地克服参数模型的缺点,但其巨大的内存开销与低的分类速度使得PNN作为大量和高维的数据样本分类几乎不可行。FCM虽具有良好的模糊聚类能力,但无法直接给出概率分类结果。该文提出的FCM-PNN分类器,在FCM聚类的基础上,以贝叶斯置信度为基础,利用PNN进行概率分类。它结合了FCM聚类和PNN概率分类的优势,同时克服了传统参数模型分类和FCM聚类的局限性。实验结果证实了FCM-PNN分类器具有分类精度高、速度快及揭示细节的能力。更多还原

【Abstract】 This paper presents a new probabilistic classifier,called FCM-PNN classifier,for data sets in high dimention-al spaces such as speaker recognition.The new classifier produces probabilistic classification with Bayes confidence measure which is highly desirable with the rendered data.In the traditional parametric methods,the predefined probabili-ty function may be inappropriate to the presented data well.Non-parametric methods,such as PNN,can overcome the difficulty of the parametric methods.However,the huge storage and the slow evaluation time of the PNN algorithm make it impractical in data classification with large training sets.FCM is efficient in clustering but it can not produce proba-bilistic classification directly.Based on the trained FCM algorithm,the FCM-PNN classifier performs the probabilistic classification using PNN algorithm.This combined use of FCM and PNN speeds up the PNN evaluation significantly and increases its accuracy as well.The proposed FCM-PNN classifier has been used to speaker recognition resulting in much better recognition quality than the PNN algorithm alone.Compared to the non-probabilistic classification,the prob-abilistically classified high dimentional data results in more informative rendering with more details.These demonstrate that the FCM-PNN classifier is a fast,accurate and probabilistic classifier for high dimentional data.更多还原