基于anchor模型的说话人检索技术

本文研究了基于anchor模型的说话人检索技术,提出了基于SCV分量方差和基于广义似然比聚类的模型剪枝方法,对基于序数比较的相似测度进行了改进,使SCV各分量的数值和比值都参与到识别中来,提高了检索性能.通过实验印证了将传统的说话人检索中的模型训练过程转变为计算语音相对于anchor说话人模型距离的映射过程,所带来的计算量和存储量的优势,为说话人检索在大型语音库和嵌入式系统的应用提出了新的思路.

Speaker Indexing Using Anchor Models

This paper introduces the technique of anchor modeling in the applications of unsupervised on-line speaker indexing. This new method is specifically designed to lower the complexity of the modeling phase, compared to classical techniques. The speaker space is built and refined by pruning the number of models needed based on SCV components' variance, and find a set of virtual anchor speaker models, which is the most representative, by merge the closest speaker models. Improved discriminate ability by suggesting a new metrics based on sequential comparing. The system we built is superior in computing and story with is some times the most important factor in large audio databases and embedded systems.