基于神经网络的多方言口音汉语语音识别系统研究

发布时间：2018-08-20 13:53

【摘要】：语音识别技术随着互联网和其他移动终端性能的提升，越来越受到工业生产和日常生活的青睐，因此如何提供一个识别性能高，鲁棒性强的语音识别系统显得越来越重要。但是，自从语音识别技术应用在中文识别上，就有一个不能忽视的现象——口音问题，它在很大程度上降低了识别系统的性能，尤其是我国是一个多口音地区国家，，口音问题，尤其是多口音问题，成为中文语音识别的一个关键问题。我国是个多口音地区国家，除了标准的普通话，还有其它七个主要方言——官话，吴，粤，湘，客家，闽以及赣。方言与标准普通话差异大，方言之间变化大。生活在这些方言地区的人们都是把普通话作为第二语言学习的，导致一个最直接的结果就是生活在这些地区的人们，在用普通话进行表达的时候，会存在很大程度上的口音变化。结果就是当前在标准普通话数据上训练的模型并不能针对特定方言的语音识别。带口音的普通话语音识别的困难主要体现在两个方面，因为我国多口音现象的存在产生了不同口音地区发音的变异带来的模型不匹配问题，同时导致了训练口音相关模型和难以获得大批量的不同口音地区的语音语料数据的矛盾。本文采用口音分类和提升特定口音声学模型性能相结合的方法，针对这两个主要问题，不仅能够为地区未知的带口音普通话测试数据选择合适的声学模型，解决模型不匹配的问题，同时通过多层级适应性的网络（MLAN）提升特定口音的模型识别能力，能够进一步更好地解决多口音问题中的模型不匹配以及特定口音数据稀疏导致的的建模难题，进而提高识别率。论文介绍的MLAN系统充分利用了神经网络的区分性学习能力和交叉数据域的适应能力，通过第一级网络，将更大数据量的标准普通话数据和特定方言口音数据的共性适应性地被提取，再由第一级网络前馈特定口音数据训练第二级网络，并对标准普通话数据前馈，使得标准普通话数据被适应上了特定口音的个性特征。这种架构不但提升了特定口音数据的共性表现能力，也对大量的标准普通话数据进行了个性化的适应，极大地增加了含有特定方言口音特征的训练数据。目前在广州、重庆地区数据上的实验表明：在基线GMM-HMM系统模型的基础上本文提出的改进系统所带来的相对CER下降分别为23.03%和21.21%，性能提升效果相当明显。很好地验证了对口音未知测试数据进行口音分类的必要性和MLAN框架的优越性。本文提出这种系统架构具有很好的扩展性和适应性，除了能很好的应对多种口音的语音识别问题，同时也适合更加复杂交叉领域和更加细致分类的情形，比如多种语言，复杂噪声条件等。
[Abstract]:With the improvement of the performance of the Internet and other mobile terminals, speech recognition technology is becoming more and more popular in industrial production and daily life. Therefore, how to provide a speech recognition system with high recognition performance and strong robustness is becoming more and more important. However, since the application of speech recognition technology in Chinese recognition, there is a phenomenon that can not be ignored-accent problem, which greatly reduces the performance of the recognition system, especially in China, which is a country with multiple accents. In particular, the problem of multiple accents has become a key issue in Chinese speech recognition. China is a multi-accented country, in addition to standard Mandarin, there are seven other major dialects-Mandarin, Wu, Guangdong, Hunan, Hakka, Fujian and Jiangxi. The difference between dialect and standard Putonghua is great, and the change between dialects is great. People living in these dialects learn Putonghua as a second language, and one of the most direct results is that people living in these areas are speaking in Mandarin. There will be a significant change in accent. The result is that the current model trained on standard Putonghua data is not specific to the speech recognition of a particular dialect. The difficulty of Putonghua speech recognition with accent is mainly reflected in two aspects, because the existence of multi-accent phenomenon in China has resulted in the model mismatch caused by the variation of pronunciation in different accent areas. At the same time, it leads to the contradiction between the relevant models of training accent and the difficulty of obtaining large quantities of speech data from different accent areas. In this paper, we combine accent classification with improving the performance of specific accent acoustic models. Aiming at these two main problems, we can not only select appropriate acoustic models for the unknown area of Mandarin with accent test data, but also improve the performance of specific accent acoustic models. To solve the problem of model mismatch, and improve the model recognition ability of specific accents through multi-level adaptive network (MLAN), It can further solve the modeling problem caused by model mismatch in multi-accent problems and sparse data of specific accents, thus improving the recognition rate. The MLAN system introduced in this paper makes full use of the discriminative learning ability of the neural network and the adaptability of the cross data domain. The generality of standard Putonghua data with larger amount of data and specific dialect accent data is extracted adaptively, and then fed forward by the first level network to specific accent data training the second level network, and feedforward the standard Putonghua data. The standard Putonghua data are adapted to the personality characteristics of specific accents. This architecture not only improves the common expression ability of specific accent data, but also adapts to a large number of standard Putonghua data, and greatly increases the training data with specific dialect accent features. The experiments in Guangzhou and Chongqing show that the relative CER decrease of the improved system based on the baseline GMM-HMM system model is 23.03% and 21.21% respectively, and the performance improvement effect is quite obvious. The necessity of accent classification based on unknown accent test data and the superiority of MLAN framework are well verified. This paper proposes that this architecture has good scalability and adaptability. It can not only deal with speech recognition problems with multiple accents, but also be suitable for more complex cross-domain and more detailed classification situations, such as multi-language. Complex noise conditions, etc.
【学位授予单位】：中国科学院深圳先进技术研究院
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP183;TN912.34

【共引文献】