The stroke segments:' are proposed to be used as the basic features for handwritten Chinese character recognition. In this way, it is possible to overcome the difFiculties of unstable stroke information caused by ...The stroke segments:' are proposed to be used as the basic features for handwritten Chinese character recognition. In this way, it is possible to overcome the difFiculties of unstable stroke information caused by stroke Joinings. The techniques of data pre-processing and stroke segment extraction have been described. In extracting stroke segment, not only the characteristics of the stroke itself, but also its absolute positions as well as relative positions with other strokes in the character have been taken into account.The primitive features for recognition were extracted under these comprehensive considerations.展开更多
The Nondictionary Chinese Segmentation algorithm tased on corpus and word frequency is presented, and the Natural Hierarchcal Network is utilized to represent and aquisite the information of language usage. The comput...The Nondictionary Chinese Segmentation algorithm tased on corpus and word frequency is presented, and the Natural Hierarchcal Network is utilized to represent and aquisite the information of language usage. The computing and space complexity of the algorithm is linear. The rate of correct abstract reaches 98 percent. The algorithm gives a new way to solve the problem of new words abstracting.展开更多
提出一种基于稀疏神经网络的说话人分割方法,利用稀疏的单隐层神经网络提取语音的超矢量特征中说话人因子特征,然后通过K均值聚类得到每帧语音的标号来分割不同说话人,在稀疏网络的训练过程中引入了dropout技术以克服过拟合问题.在TIMI...提出一种基于稀疏神经网络的说话人分割方法,利用稀疏的单隐层神经网络提取语音的超矢量特征中说话人因子特征,然后通过K均值聚类得到每帧语音的标号来分割不同说话人,在稀疏网络的训练过程中引入了dropout技术以克服过拟合问题.在TIMIT语音数据库构成的多说话人语音数据上的实验结果表明:通过增加稀疏网络中隐层节点的个数可以提高说话人分割的效果,与贝叶斯信息准则(Bayesian information criterion,BIC)方法和稀疏自编码网络方法相比,所提基于稀疏神经网络的说话人分割方法的性能有明显提高.展开更多
文摘The stroke segments:' are proposed to be used as the basic features for handwritten Chinese character recognition. In this way, it is possible to overcome the difFiculties of unstable stroke information caused by stroke Joinings. The techniques of data pre-processing and stroke segment extraction have been described. In extracting stroke segment, not only the characteristics of the stroke itself, but also its absolute positions as well as relative positions with other strokes in the character have been taken into account.The primitive features for recognition were extracted under these comprehensive considerations.
文摘The Nondictionary Chinese Segmentation algorithm tased on corpus and word frequency is presented, and the Natural Hierarchcal Network is utilized to represent and aquisite the information of language usage. The computing and space complexity of the algorithm is linear. The rate of correct abstract reaches 98 percent. The algorithm gives a new way to solve the problem of new words abstracting.
文摘提出一种基于稀疏神经网络的说话人分割方法,利用稀疏的单隐层神经网络提取语音的超矢量特征中说话人因子特征,然后通过K均值聚类得到每帧语音的标号来分割不同说话人,在稀疏网络的训练过程中引入了dropout技术以克服过拟合问题.在TIMIT语音数据库构成的多说话人语音数据上的实验结果表明:通过增加稀疏网络中隐层节点的个数可以提高说话人分割的效果,与贝叶斯信息准则(Bayesian information criterion,BIC)方法和稀疏自编码网络方法相比,所提基于稀疏神经网络的说话人分割方法的性能有明显提高.
