说话人确认中语音段差异建模相关问题的研究

发布时间：2018-08-13 18:04

【摘要】：随着说话人确认技术发展至今,复杂信道下的长时语音段的说话人确认技术已经日趋成熟,为其实用化奠定了基础。其中,建立在高斯混合模型-通用背景模型的基础之上的全局段差异建模,建模方式简便,在应用于说话人确认系统的过程中,结合高效的后端非说话人差异补偿模型,如概率线性区分性分析模型(Probabilistic Linear Discriminant Analysis, PLDA),自提出以来就成为了当前说话人领域的主流技术。全局段差异建模的思想在于将特定的语音段中所包含的异于所有语音段共同包含的共性信息(主要为发音内容),即该语音段的段差异(Session Variability),用一个固定长度的低维向量进行表达,即段差异向量。段差异向量中包含了发音内容以外的差异,包括非说话人差异(主要为信道差异)和说话人差异。在段差异向量之上,需要对其进一步建模消除说话人无关的差异对说话人辨认的影响,因此,全局段差异建模的核心问题在于提取语音段中的段差异以及在后端对非说话人差异进行补偿,从而提取到语音段中有利于说话人判别的说话人信息。本文针对段差异的建模和说话人差异的提取及判别方面进行了一系列的研究,主要内容如下：第一,通过全局段差异建模虽然能够获得关于语音段中的段差异的整体上的表达,但是却忽略了存在于语音段中的细节性的差异,因此我们提出利用局部段差异建模的方法提取存在于语音段中的不能被全局段差异建模所表达出的局部段差异,用于说话人确认。在局部段差异建模中,我们分别针对每一个高斯和每一个声学特征的维度中包含的段差异进行建模提取,即高斯局部差异建模和维度局部差异建模。进一步的,在维度局部差异建模中,我们提出对声学特征的维度进行不同方式的绑定,然后再提取绑定的维度组合中的局部段差异。由于局部段差异模型与全局段差异模型分别从不同的角度对语音中的段差异进行建模,二者之间存在一定的互补能力,使得我们可以在系统和模型两个层面分别融合,从而获得比单独的模型更优的性能。第二,在全局段差异建模中,当用于测试和模型训练的语音段在文本上保持一致时,全局段差异建模对于语音的段差异差异拥有良好的建模能力,如长时语音段的文本无关以及短时语音段的文本相关的说话人确认。而当存在文本不匹配的问题时,例如针对短时短时语音段的文本无关的说话人测试中,由于全局段差异建模无法针对文本进行段差异建模,导致说话人的判别受到文本差异的影响。我们在局部段差异建模的思想基础之上,借助语音识别中的基于深层神经网络的声学模型对语音帧以音素状态进行聚类,并以此为基础提取音素相关的局部段差异。在该建模方式中,我们分别采用单音素和三音素声学模型,在针对不同的音素提取到的局部段差异向量之上,根据测试用的语音段中包含的音素对局部段差异向量进行挑选并用于说话人判别,由此进行音素相关的说话人判别,解决短时语音段中的文本不匹配问题。在音素之上,我们进一步探索利用语音识别器的识别结果,以词为对象进行段差异提取并判别说话人,完善基于发音内容的段差异提取及说话人判别的研究。第三,当前基于全局段差异模型的后端信道补偿的主流技术PLDA是一个线性的概率模型,在本文中,我们针对后端信道补偿模型进行了一系列的改进研究。首先,对于PLDA,我们提出一个与现有的说话人得分计算模型等价的建立在自适应的说话人模型上的得分计算模型,在此模型基础之上,针对多语音段注册说话人的任务中,不同说话人注册语音段数量不同以及同一个说话人的不同语音段之间不同程度的重叠的现象,提出利用说话人因子的先验分布参数进行模型注册以取代传统的后验分布参数,以此来解决这两个问题对于得分计算带来的影响。此外,我们在说话人自适应模型的基础之上,引入信道自适应,在每一组测试中,将PLDA模型自适应到测试语音段的信道空间之中,进而计算得分,通过这种方式,我们可以考虑到不同测试的具体信息,从而提升系统的性能。其次,我们引入深层神经网络用来提取存在于全局段差异向量中的非线性的深层说话人信息,用于说话人判别,达到改善系统性能的目的。
[Abstract]:With the development of speaker recognition technology, long-term speaker recognition technology in complex channels has become more and more mature, which lays the foundation for its practicality. Among them, global segment difference modeling based on Gaussian mixture model-general background model is simple and convenient, and it can be applied to speaker recognition system. Cheng Zhong combines efficient back-end non-speaker difference compensation models, such as Probabilistic Linear Discriminant Analysis (PLDA), which has become the mainstream technology in the field of speakers since it was proposed. Session Variability is expressed by a fixed length low-dimensional vector, i.e. the segment difference vector. The segment difference vector contains differences beyond the pronunciation content, including non-speaker differences (mainly channel differences) and speaker differences. On the basis of segment difference vector, further modeling is needed to eliminate the influence of Speaker-Independent differences on speaker recognition. Therefore, the key problem of global segment difference modeling is to extract segment differences and compensate for non-speaker differences at the back end, so as to extract speakers who are helpful to speaker discrimination in speech segments. This paper focuses on the modeling of segment differences and the extraction and discrimination of speaker differences. The main contents are as follows: Firstly, global segment difference modeling can obtain the overall expression of segment differences in speech segments, but it neglects the detail differences in speech segments. In this paper, we propose a local segment difference modeling method to extract the local segment difference which can not be expressed by global segment difference modeling for speaker recognition. Gauss Local Difference Modeling and Dimensional Local Difference Modeling. Furthermore, in Dimensional Local Difference Modeling, we propose to bind the dimensions of acoustic features in different ways, and then extract the local segment differences in the bound dimension combination. There is a complementary ability between the two, so that we can fuse them separately at the system and model levels to achieve better performance than a single model. Second, in global segment difference modeling, when the speech segments used for testing and model training are consistent in text, the global segment difference is achieved. Modeling has good modeling ability for speech segment differences, such as text-independent for long speech segments and text-dependent speaker verification for short speech segments. Based on the idea of local segment difference modeling, we use the acoustic model based on deep neural network in speech recognition to cluster the phoneme states of speech frames and extract the phoneme-related local segment differences. In the modeling method, we adopt the mono-phoneme and tri-phoneme acoustic models respectively. On the basis of the local segment difference vectors extracted from different phonemes, we select the local segment difference vectors according to the phonemes contained in the test phonemes and use them to distinguish the speakers. Thus, we can distinguish the speakers with phoneme correlation and solve the problem of short-term speech. On the basis of phonemes, we further explore the use of speech recognition results to extract segment differences and identify speakers, improve the research of segment differences extraction and speaker discrimination based on pronunciation content. Thirdly, the backend channel compensation based on global segment differences model is the main channel. Streaming technology PLDA is a linear probability model. In this paper, we make a series of improvements to the backend channel compensation model. Firstly, for PLDA, we propose an equivalent score calculation model based on adaptive speaker model. Aiming at the phenomenon of different number of speech segments registered by different speakers and different degree of overlap between different speech segments of the same speaker in the task of multi-segment registered speaker, a model registration method based on the prior distribution parameters of speaker factors is proposed to replace the traditional posterior distribution parameters. In addition, on the basis of speaker adaptation model, channel adaptation is introduced. In each group of tests, the PLDA model is adapted to the channel space of the test speech segment, and then the score is calculated. In this way, we can take into account the specific information of different tests, thereby improving the performance. Secondly, we introduce the deep neural network to extract the nonlinear deep speaker information which exists in the global segment difference vector, and use it to discriminate the speaker, so as to improve the performance of the system.
【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TN912.34

【相似文献】