论文部分内容阅读
说话人聚类研究如何将一段语音中同一说话人的语音聚合.提出一种基于结合广义似然比与归一化交叉似然比两种距离测度的聚类算法.算法首先提取每一段语音信号的MFCC特征,并建立高斯混合模型,最后采用基于结合广义似然比与归一化交叉似然比两种距离测度的层次化策略对语音信号进行聚类.在算法中,贝叶斯判据用以确定聚类结束的条件.实验表明,该算法提高了系统的综合性能,较好的解决了无监督说话人聚类问题.结合两种距离测度比单独使用任何一种距离测度的系统性能提高了6%.并且,通过改进更新类间距的方式,聚类速度相比传统高斯混合模型聚类方法提升6倍.
The speaker clustering researches how to aggregate the speech of the same speaker in a speech segment.This paper proposes a clustering algorithm based on a distance metric between generalized likelihood ratio and normalized cross-likelihood ratio.The algorithm first extracts each segment of speech signal , And establish Gaussian mixture model.Finally, the speech signal is clustered based on the hierarchical strategy based on the combination of generalized likelihood ratio and normalized cross-likelihood ratio.In the algorithm, the Bayesian criterion Which is used to determine the condition of the end of clustering.Experiments show that this algorithm improves the overall performance of the system and solves the problem of unsupervised speaker clustering.Combining the performance of the two distance measures than using any kind of distance measure alone Increase by 6% .And, by improving the way of updating the class spacing, the clustering speed is 6 times higher than the traditional Gaussian mixture model clustering method.