深度神经网络音素识别系统及其隐藏层研究

发布时间：2018-03-06 23:09

本文选题：深度神经网络　切入点：语音识别　出处：《天津大学》2014年硕士论文　论文类型：学位论文

【摘要】：经过近半个世纪的积累和酝酿,语音识别技术在最近几年达到了大规模商业应用水平。而最近五年来,深度神经网络(Deep Neural Network,DNN)的出现给语音识别带来了新的突破。尽管深度神经网络在语音领域取得了巨大的成功,但其内在的机制及原理目前仍不清楚。本研究通过初步探索深度神经网络的隐藏层的功能与作用来研究这一问题,探索深度神经网络隐藏层与语音生成感知机理(音素的发音位置)之间的关系。本工作首先分别构建了汉语、英语、日语三种语言的深度神经网络音素识别系统,为了后续实验有可对比性,三个网络的结构配置均相同,采用目前同用的7层,2048节点/层的结构。根据数据库的可靠程度,选取英语和日语的深度神经网络进行隐藏层研究。由于神经网络的某一隐藏层无法被单独研究,所以本研究采用去掉一层或替换一层的方法,通过识别结果的变化,来反推被去层或被换层的作用。我们分别对英语和日语进行了去层实验,并用日语网络隐藏层对应替换英语隐藏层进行了换层实验。从分析结果可以看出,神经网络的不同隐藏层主要负责的音素可以大致按照发音位置进行分类:较低层主要负责后元音,较高层主要负责前元音;神经网络的第二层对发音约束位置靠前的辅音更敏感,而其余的辅音更依赖中间层。换层实验说明上述结论在不同的语言间存在共通性。本工作作为对深度神经网络隐藏层的初步研究,只是作了简单的定性分析,在同类领域中,本研究提出一种新颖的方法来分析神经网络的隐藏层,第一次将研究重点放在深度神经网络的每一隐藏层上;第一次将人的发音机理加入深度神经网络,并通过实验方法探索两者间的关系。要想探究DNN表征机制,其内部隐藏层的功能,还需要未来的进一步研究。
[Abstract]:After nearly half a century of accumulation and brewing, speech recognition technology has reached the level of large-scale commercial applications in recent years. The emergence of Deep Neural Network (DNN) has brought a new breakthrough in speech recognition, despite the tremendous success of the Deep Neural Network in the field of speech. However, its intrinsic mechanism and mechanism are still not clear. This study studies this problem by exploring the function and function of the hidden layer of the deep neural network. This paper explores the relationship between the hidden layer of depth neural network and the mechanism of phoneme perception (phoneme pronunciation location). Firstly, a depth neural network phoneme recognition system for Chinese, English and Japanese is constructed, which is composed of three languages: Chinese, English and Japanese. In order to be comparable in subsequent experiments, the three networks have the same configuration, using the same 7-layer 2048 node / layer structure, according to the reliability of the database. Since a certain hidden layer of neural network can not be studied separately, the method of removing one layer or replacing one layer is adopted to identify the change of the result. We have carried out delamination experiments on English and Japanese, and replaced English hidden layers with Japanese network hidden layers. From the results of the analysis, we can see that, The phonemes of different hidden layers of neural network can be classified according to the pronunciation position: the lower level is mainly responsible for the post vowel, the higher level is mainly responsible for the front vowel; The second layer of the neural network is more sensitive to the consonant in the front of the articulation constraint. The rest of the consonants depend more on the middle layer. The experiment of changing layers shows that the above conclusions are common among different languages. As a preliminary study of the hidden layer of deep neural networks, this work has only made a simple qualitative analysis in the same field. In this study, a novel method is proposed to analyze the hidden layer of neural network, which is focused on each hidden layer of the deep neural network for the first time, and the mechanism of human pronunciation is added to the deep neural network for the first time. In order to explore the mechanism of DNN representation, the function of the inner hidden layer still needs further study in the future.
【学位授予单位】：天津大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP183;TN912.34

【相似文献】