基于深度学习的孤立词语音识别系统设计

发布时间：2017-12-31 04:12

本文关键词：基于深度学习的孤立词语音识别系统设计　出处：《东北电力大学》2017年硕士论文　论文类型：学位论文

【摘要】：在人与人之间的自然对话中,语音对话是被使用最频繁的一种交流方式,因为它是最直接、最方便的一种表达方式。那么在人工智能领域中,语音识别也就成为了人机交互技术里最重要的部分之一,并随着现代科技的发展,各类语音识别软件逐步融入人们的生活,例如智能手机的语音拨号、发短信、查询等,以及汽车上的一些语音控制设备。在如今的互联网时代里,获得大量的语音数据已不成问题,那么如何高效的利用这些未经标注的而且大量的语音数据,就成为了当前语音识别领域的一个研究热点。与此同时,深度学习模型对数据有着强大的建模能力,并能直接对未标注的原始数据进行预处理,这就使其与语音识别的联系越来越紧密。在此背景下,本文以语音识别理论为基础,针对如何将深度学习理论应用于语音识别中这一问题展开研究,最后在此研究上设计了一个孤立词语音识别系统,主要有以下内容。(1)根据语音识别基本原理,在Matlab环境下,编程搭建一个非特定人的孤立词语音识别系统框架,这个系统框架中主要包括语音信号预处理、特征提取、基于HMM声学模型的训练和解码搜索这四个部分。实验中用训练库中语音数据,以提取传统的声学特征(MFCC)为基本特征,训练出基于HMM声学模型的模板库,再结合测试库中的语音数据进行解码搜索,检测此识别系统的识别率,并以此得到了一个合适的声学模型参数,以便下来的试验。(2)以深度学习为基础理论,提出了一种升维式的堆栈式深度自动编码器模型,利用此模型提取新的深层语音特征。并分别从网络的隐含层层数、各层节点数、各层节点的激活函数、有监督和无监督训练方式等多个方面构建深度自动编码器模型及其在语音特征提取中的具体应用。实验证明,新的深层特征比原始MFCC特征,使识别系统在词识别正确率和时间方面都有所提高,这就证明了深度学习模型在特征提取上对提高语音识别正确率的有效性。(3)本文基于以上研究成果,在Labview平台上对上述所研究的基于深度学习的孤立词语音识别系统进行了系统设计,其功能包括录音、播放、训练和识别。其中为了语音系统的识别正确率能达到95%以上,对模板库进行改进,设计为特定人的模板库,即在进行识别时,选择与个人对应的模板库对其语音进行解码搜索以得到识别结果。
[Abstract]:Between people and natural dialogue, dialogue is a way of communication is the most frequently used, because it is the most direct, an expression of the most convenient way. So in the field of artificial intelligence, speech recognition has become a human-computer interaction technology, the most important part of the one, and with the development of the modern science and technology, all kinds of voice recognition software gradually into people's lives, such as intelligent mobile phone voice dialing, texting, query, some voice control equipment and automobiles. In today's Internet era, get a large amount of speech data is not a problem, so how to efficiently utilize these data without voice but a lot of annotation, has become a hot research topic in the field of speech recognition. At the same time, deep learning model has a strong ability of data modeling, and can directly to the original unlabeled data into The pretreatment, which makes it more and more closely linked with speech recognition. Under this background, this paper is based on the theory of speech recognition, aiming at how deep learning theory is applied to the problem of speech recognition research, then research on the design of a speech recognition system, the main contents are as follows. (1) according to the basic principle of speech recognition, in the Matlab environment, the framework of speech recognition system programming to build a non specific people, the framework of the system includes voice signal preprocessing, feature extraction, the four part of the HMM acoustic model training and decoding search based on voice data in the training set. By experiment in order to extract acoustic features, the traditional (MFCC) as the basic characteristics of the training HMM acoustic model based on template library, combined with voice data test library for decoding search, detection of the identification systems of knowledge Don't rate, and thus a suitable acoustic model parameters, in order to test down. (2) with deep learning theory, puts forward a type of stack depth sunwy autocoder model, extraction of speech features new using this model. And from the network hidden layer number. The number of nodes in each layer, the activation function of nodes in each layer, and its application in speech feature extraction in the construction of many aspects of depth auto encoder model supervised and unsupervised methods. Experiments show that the new deep feature than the original MFCC feature, the recognition system can improve word recognition accuracy and time., which proves the validity of deep learning model accuracy to improve speech recognition in feature extraction. (3) this paper based on the above research results, based on the Labview depth study of isolated word learning based on the above research The speech recognition system is designed, its functions including recording, playback, training and recognition. In order to speech recognition system accuracy can reach more than 95%, the template library was improved, designed for specific template library, namely in the identification, selection and individual corresponding template library to decode the search in order to get the speech recognition results.

【学位授予单位】：东北电力大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN912.34

【相似文献】