基于模糊聚类决策树的分布式语者识别算法

摘要

In order to solve the problems of additive noise and high computational complexity in speaker identification and to improve the robustness and anti-noise ability of the large scale speaker identification algorithm,a distributed speaker identification algorithm with fuzzy clustering decision tree has been presented,which divides training data into several parts,and builds fuzzy clustering decision trees for these parts.For testing data,fuzzy decision trees has been employed,which are built in the previous step to decide which leaf node the people's speech belongs to.The speaker is identified by using the Mel-Frequency Cepstral Coefficients and the Gauss mixture model identification method on the selected leaf nodes.The process of fuzzy clustering on training data mainly includes four parts,i.e.extracting feature data from the corresponding layer,calculating the mean and standard deviation of the feature data,using Lloyd algorithm to get the separation vector,clustering to get the nodes of the next layer.The experimental result shows that compared with the traditional hard clustering algorithm,the proposed algorithm has improved the accuracy and classification efficiency of speaker identification,with the good anti-interference ability to the additive noise.%为解决大规模语者识别问题中普遍存在的加性噪声、高计算复杂度等问题,提高大规模语者识别算法的抗噪性和鲁棒性,利用模糊聚类决策树,提出了一种分布式语者识别算法.该算法将训练数据等分成几个部分,对这几个部分分别使用基于模糊聚类的决策树算法进行训练;对于输入的测试样本,用建好的决策树进行分类,判断它属于哪棵树的哪个叶节点;在该选定的叶节点上使用梅尔频率倒谱系数和高斯混合模型识别方法识别该语者身份.对训练数据进行模糊聚类的过程主要包括四个步骤:根据相应的层提取语音特征;计算特征数据的均值和标准差得到信任间距集合;对集合使用Lloyd算法得到分隔向量;以分隔向量为基础进行聚类分组得到下一层的节点.实验结果表明,与传统的硬聚类算法相比,该算法能够提高语者识别的准确率和分类效率,对加性噪声具有良好的抗干扰能力.