一种基于并行覆盖前馈优先神经网络的说话人识别方法

Research on Speaker Recognition Based on the Parallel Coverage of Priority Ordered Neural Network

为实现对说话人特征空间多聚类区的有效识别,提出一种基于并行覆盖前馈优先级网络(PCPONN)的说话人识别方法。该方法以LBG算法生成每个说话人特征空间初始的聚类中心,对本类样本按聚类中心分类后,用前馈优先级神经网络(PONN)对每个聚类区进行并行覆盖。相关实验证明,PCPONN符合说话人特征空间点的分布特点,得到更好的稳定性和更高的识别率。

In order to realize the effective coverage of multiple clusters area in the speaker＇s feature space,a method of parallel coverage of priority ordered neural network （PCPONN） is put forward. Based on the initial clusters center generated by the LBG in every speaker＇s feature space, every sample can be classified, so every clusters area can be parallel covered by the PONN. The relative experiment results show PCPONN is consistent with the distribution of speaker＇ s feature point, so has better stability and higher correct recognition.