采用竞争选择神经网络的语音识别方法

A New Optimal Shortening Algorithm of the Channel Impulse Response for DMT Modulation Systems

本文在Ｇｅｒｓｈｏ的ａｓｙｍｐｔｏｔｉｃ理论基础上，推导了向量聚类的等失真原则，并提出了一种竞争选择学习算法（ＣＳＬ），这种算法可以避免局部最优，把它应用到隐马尔柯夫模型（ＨＭＭ）的聚类中，可以起到很好的效果。结合安全拒识措施，本文提出用并行、自组织、层次神经网（ＰＳＨＮＮ）将ＨＭＭ输出的每一个模板的记分再分类，使识别率明显提高。

In discrete multitone(DMT)modulation systems,a cyclic prefix(CP)is inserted before every symbol block being transmitted,to ensure the independence of each subchannel and alleviate the interference between consecutive symbol blocks The length of the CP is equal to the memory length of the impulse response of the effective channel Using a long CP reduces the throughput of the system largely To avoid using a long CP,an finite impulse response filter is used to shorten the length of the effective channel impulse response An algorithm used to calculate the coefficients of the optimal shortening impulse response filter(SIRF)was given in However,this algorithm requires that the length of the SIRF must be smaller than or equal to the memory length of the target impulse response In this paper,we modify this algorithm and make it suitable for calculating the coefficients of the SIRF with arbitrary length