【摘要】 说话人识别的关键在于如何为集合中的每一个人建立一个能表征该说话人个性特征的声学模型,建模方法将会严重影响系统的性能。基于当今与文本无关的话者识别的主流模型———高斯混合模型(Gaussian Mixture Model,GMM)的基础上,从声学的角度剖析了男女发音的差别,以增加说话人之间的差异性为出发点,引入竞争性思想和通用背景模型(Universal Background Model,UBM),提出了具有区分性的GMM的建模方法,克服了传统GMM需要大量训练样本的局限性和UBM将说话人强制服从统一分布的弱点。最后实验的对比结果表明,具有区分性的GMM相比传统的高斯混合模型在识别率上有所提高。更多还原

【Abstract】 Characteristic modeling plays an important role in technology of Speaker Recognition.The modeling method will seriously impact on the performance of speaker recognition system.This article is based on the main model of Text-independent Speaker Recognition,analyses gender differences in pronunciation and increases the differences between the speakers as a starting point,then introduces the idea of competition and UBM into Gaussian Mixture Model(GMM) at the time of characteristic modeling.Therefore,a new approach of characteristic modeling is dipartite which is presented.The method overcomes the limitation that there are plentiful training samples for traditional GMM and the shortcoming that the distributions of all speakers are unified for UBM.Finally,it may build dipartite models for every speaker.The experiment shows that the model based on dipartite algorithm has higher performance than traditional GMM algorithm.更多还原