基于子空间的说话人自适应技术研究

发布时间：2018-02-17 05:22

本文关键词： 连续语音识别说话人自适应流形学习本征音正交局部保持投影正则化方法特征参数归一化　出处：《解放军信息工程大学》2014年硕士论文　论文类型：学位论文

【摘要】：训练与测试数据之间关于说话人的失配制约着连续语音识别系统的实用化。如何利用少量的自适应数据,增加声学模型与测试数据之间的匹配程度,一直是连续语音识别研究的重点和难点问题。子空间方法通过对高维空间中的低维流形结构进行建模,不仅可以对高维空间进行降维,有效地避免维数灾难问题,而且可以发现数据本身的结构特点,提高模型参数估计的稳健性。本文研究如何利用子空间技术获得更为实用化的说话人自适应技术,主要内容如下：针对本征音算法在自适应数据量较少时,易出现过拟合导致系统性能下降的问题,提出了正则化本征音说话人自适应方法。该方法通过对目标函数引入适当的正则化因子,构造新的目标函数进行优化,从而估计出更优的说话人因子,提高解的稳定性。在NIST LRE2003评测集上进行的语种识别实验表明,改进算法与基线系统相比,在测试语料为短语音段时,系统性能有一定的提升,且测试语料越短,性能提升越明显。在微软语料库上进行的中文连续语音识别实验表明,在自适应数据较为充足时,正则化本征音自适应方法略微降低了系统的性能,但在自适应数据不足时,正则化本征音自适应方法可以有效的提高系统的稳健性。针对本征音这类线性子空间方法无法精细描述非线性子空间内在结构的问题,提出了正交拉普拉斯说话人自适应方法。该方法通过正交局部保持投影算法对说话人子空间进行分析,在去除声学无关信息的基础上,进一步发现这些信息的内在结构。并分别给出了该方法适用于语种识别和连续语音识别的系统框架和实现步骤。在NIST LRE 2003评测集上进行的语种识别实验证明正交拉普拉斯算法能够有效的提升特征的区分性。在微软语料库上进行的中文连续语音识别实验进一步证明该方法优于本征音说话人自适应方法。针对模型层的说话人自适应影响解码速度的问题,提出了特征空间本征音自适应方法。该方法借鉴RATZ算法,采用高斯混合模型对特征空间中的说话人信息进行建模,同时充分利用估计参数之间的相关性,减少估计参数的数量,在对特征空间精确建模的同时,降低了算法对自适应数据量的需求。在基于微软语料库的中文连续语音识别实验中,特征空间本征音自适应方法在自适应数据量极少时仍能取得较好的性能,同时配合说话人自适应训练能够进一步降低词错误率。
[Abstract]:The mismatch between the training and test data about the speaker restricts the practicability of the continuous speech recognition system. How to use a small amount of adaptive data to increase the matching degree between the acoustic model and the test data, Subspace method can not only reduce the dimension of high-dimensional space, but also avoid the problem of dimensionality disaster by modeling the low-dimensional manifold structure in high-dimensional space. Moreover, the structural characteristics of the data can be found and the robustness of model parameter estimation can be improved. In this paper, we study how to use subspace technology to obtain more practical speaker adaptive technology. The main contents are as follows: when the adaptive data amount is small, the intrinsic sound algorithm is prone to the problem that over-fitting results in the deterioration of system performance. A regularization eigen-speaker adaptive method is proposed in this paper. By introducing a proper regularization factor to the objective function, a new objective function is constructed and the better speaker factor is estimated. The experiment of language recognition on NIST LRE2003 evaluation set shows that compared with the baseline system, the improved algorithm can improve the performance of the system when the test corpus is phrasal segment, and the shorter the test data is, the better the performance of the improved algorithm is. The performance improvement is more obvious. The Chinese continuous speech recognition experiment on Microsoft corpus shows that when the adaptive data is more adequate, the regularized intrinsic tone adaptive method slightly reduces the performance of the system, but when the adaptive data is insufficient, the regularized intrinsic tone adaptive method reduces the performance of the system slightly, but when the adaptive data is insufficient, The regularization eigentone adaptive method can effectively improve the robustness of the system. The linear subspace method such as eigensound can not accurately describe the problem of structure in nonlinear subspace. An orthogonal Laplace speaker adaptive method is proposed, in which the speaker subspace is analyzed by orthogonal local preserving projection algorithm, and the acoustic-independent information is removed. Furthermore, the internal structure of these information is found. The system framework and implementation steps of this method for language recognition and continuous speech recognition are given respectively. The experimental results of language recognition based on NIST LRE 2003 prove that the orthogonal method is orthogonal. Laplace algorithm can effectively improve the distinction of features. The experiment of Chinese continuous speech recognition in Microsoft corpus further proves that this method is superior to the intrinsic speaker adaptive method. The problem of human adaptation affecting decoding speed, Based on the RATZ algorithm, Gao Si hybrid model is used to model the speaker information in the feature space, and the correlation between the estimated parameters is fully utilized to reduce the number of the estimated parameters. At the same time, the requirement of adaptive data is reduced. In the experiment of Chinese continuous speech recognition based on Microsoft corpus, The eigenspace eigensound adaptive method can achieve good performance even when the adaptive data is small, and the speaker adaptive training can further reduce the word error rate.
【学位授予单位】：解放军信息工程大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN912.34

【相似文献】