采用HDPHMM符号化器的语音查询样例检测方法  

Query-by-example Spoken Term Detection by Applying the HDPHMM Tokenizer

作　　者：曹建凯[1] 张连海[1] 

机构地区：[1]解放军信息工程大学信息系统工程学院,河南郑州450001

出　　处：《信号处理》2017年第5期703-710,共8页Journal of Signal Processing

基　　金：国家自然科学基金资助项目(61673395;61403415;61302107)

摘　　要：提出一种基于层级狄利克雷过程隐马尔科夫模型(HDPHMM)符号化器的无监督语音查询样例检测(Qb E-STD)方法。该方法首先应用一个双状态层隐马尔科夫模型,其中顶层状态用于表示所发现的声学单元,底层状态用于建模顶层状态的发射概率,通过对顶层状态假设一个层级狄利克雷过程先验,获得非参贝叶斯模型HDPHMM。使用无标注语音数据对该模型进行训练,然后对测试语音和查询样例输出后验概率特征矢量,使用非负矩阵分解算法对后验概率进行优化得到新的特征,然后在此基础上,应用修正分段动态时间规整算法进行检索,构成Qb E-STD系统。实验结果表明,相比于基于高斯混合模型符号化器的基线系统,本文所提出的方法性能更优,检索精度得到显著提升。This paper presents a study of hierarchical Dirichlet processing hidden Markov model （HDPHMM） approach for unsupervised query-by-example spoken term detection （QbE-STD）. First a hierarchical hidden Markov model is applied, in which the top layer states are used for representing the finding acoustic units, bottom layer states are used for modeling the emission probability of top layer states. We can get a nonparametric Bayesian model HDPHMM when imposing a hierarchical Dirichlet processing prior on the top layer states. After the model is trained by unlabeled speech data, it outputs posteriorgram feature vector for test utterance and query term. The posteriorgram feature is optimized by non-negative matrix factorization al- gorithm. Then the detection is performed by modified SDTW algorithm. Experimental results show that the proposed method outperforms the baseline system based on Gaussian mixture model tokenizer, and improve the detection precision obviously.

关 键 词：无监督 语音查询样例检测 层级狄利克雷过程 非负矩阵分解 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]