基于模糊分类器及多层前馈神经网络混合结构的说话人辨认  

Hybrid architecture based on fuzzy classifier and multiplayer feed-forward neural network for speaker identification

作　　者：张玲华[1] 杨震[1] 郑宝玉[1] 

机构地区：[1]南京邮电大学通信与信息工程学院,江苏南京210003

出　　处：《通信学报》2005年第11期68-75,共8页Journal on Communications

基　　金：江苏省"青蓝工程"跨世纪学术带头人专项基金资助项目(QL003YZ)南京邮电大学科研发展基金资助项目(2001院17)

摘　　要：提出了基于模糊超椭球聚类算法的说话人辨认新方法。该算法首先将某一类的训练数据分成若干子类, 对每一子类在其中心周围定义具有超椭球区域的模糊规则。实验表明,该系统可以较快的聚类速度取得与HMM 相当的识别效果。进一步的研究表明,基于模糊超椭球聚类算法的说话人辨认系统与传统的基于HMM的识别方法存在一个共同的缺点,即抗噪性能较差。为此,通过引入多层前馈神经网络(MLFNN)与模糊超椭球分类器构成混合模型,使系统的识别性能和抗噪能力显著提高。A novel method for speaker identification was proposed which was based on a fuzzy classifier with hyperellipsoidal regions. First, the training data for each class were divided into several clusters. Then, for each cluster, a fuzzy rule with a hyperellipsoidal region was defined around a cluster center. The evaluation experiments had been conducted to compare the fuzzy hyperellipsoidal classifier with the HMM. It was found that the former classifier can achieve a comparable speaker identification performance to the latter with higher clustering speed. Further research showed that both fuzzy hyperellipsoidal classifier and the HMM worsened the recognition ability when the test data contained noise. To overcome this problem, a hybrid architecture based on fuzzy classifier and multilayer feed-forward neural network （MLFNN） was developed for speaker recognition. The experimental results showed that the new method can achieve a much better identification performance and robustness to the additive rioise.

关 键 词：说话人辨认 模糊 超椭球分类器 多层前馈神经网络 

分 类 号：TN912.3[电子电信—通信与信息系统]