面向电话语音的蒙古语关键词检测技术的研究

Mongolian language is a kind of influential language in the world, which is used in many countries such as China, Republic of Mongolia and Russia. In China, the used Mongolian is called "traditional Mongolian", which is different from "Cyril Mongolian" used in Republic of Mongolia. These two kinds of Mongolian languages have the same speaking, but they are different in writing. Under the circumstance, the Mongolian language possesses an outstanding position in safety and strategy of our country. In addition, more and more voice resources of Mongolian language have been formed with rapidly increasing, which involves in education, culture, film, television and other fields. These are the precious culture resources for the Mongolian people. And these voice resources need to be developed and utilized further. This project will research all key issues in Mongolian speech keyword spotting. And the Mongolian speech is achieved by telephone, which is the study object in this project. The detailed research contents include the decoding of the Mongolian speech recognition system, lattice optimization, indexing, the keyword spotting model, the confidence measure calculation approach, the processing for words out of vocabulary, query expansion, the conversion of grapheme to phoneme. Finally, a Mongolian keyword spotting system will be developed, which could achieve requirements for practical application. We will adopt the advanced experience of other languages. By analyzing the characteristics of the Mongolian language, all key issues of this project could be resolved. And the precision of the keyword spotting system would be improved as much as possible. The academic value of this project are very important, which can maintain the safety of our country and the stability of minority nationality regions. Meanwhile, It is very meaningful for promoting and developing the culture of minority nationality.

蒙古语是一个跨多国、多地区的语言，在国际上是有广泛影响的一种语言文字，使用者分布在中国、蒙古国和俄罗斯等国家。中国和蒙古国使用的蒙古语言文字是"语同文不同"，因此安全战略地位十分突出。另外，蒙古语语音资源应用越来越普遍，数量急剧增加，已形成了宝贵的民族文化资源，有待于进一步开发利用。本项目以蒙古语电话语音为对象，对语音关键词检测技术所涉及到的蒙古语语音识别系统的解码、网格数据优化及索引建立、关键词的检测模型和置信度计算方法、集外词处理、关键词查询扩展、蒙古文字母到音素的自动转换等一系列关键问题进行研究，并搭建一个基本能达到应用要求的蒙古语关键词检测系统。我们将借鉴其它语言的先进经验，并结合蒙古语的特点，突破一系列难点来提高系统检测的准确度。本项目研究的蒙古语语音关键词检测技术不仅具有重要的学术价值，并对维护国家安全及边疆少数民族地区的稳定，繁荣和发展少数民族文化具有重要意义。

蒙古语是一个跨多国、多地区的语言，在国际上是有广泛影响的一种语言文字，使用者分布在中国、蒙古国和俄罗斯等国家。本项目以蒙古语电话语音为对象，围绕语音关键词检测技术所涉及到的蒙古语语音识别系统的声学模型和语言模型的建立、解码、网格数据优化及索引建立，关键词的检测模型和置信度计算方法，集外词处理，关键词查询扩展，蒙古文字母到音素的自动转换，蒙古文校正等一系列关键问题进行了研究。课题组根据蒙古文的构词特点，提出了将词干和后缀分割识别的蒙古语LVCSR方法，并重新建立了声学模型和语言模型，为国际上其它黏着语的相关研究提供了新的思路和方法。采用深度学习方法研发了蒙古语大词汇量连续语音识别系统，识别正确率达到了90%以上。结合蒙古语的发音和构词特点，提出了基于词混淆网络和音素混淆网络的关键词检测方法，解决了集内词和集外词检测问题，研发了基本能达到应用要求的蒙古语关键词检测系统。研发了蒙古文字母到音素的转换系统，转换系统的词误识率为16.32%，音素误识率仅为3.37%，基本达到了实用要求。研发了西里尔蒙古文与传统蒙古文相互转换系统，西里尔蒙古文到传统蒙古文转换正确率达到了95%以上，并且对外服务的翻译量已超过100万条。建立了1300小时的蒙古语语音库和对应标注库、4GB的蒙古文文本库、6.5万条的蒙古语发音词典。在IEEE Transactions on Audio, Speech and Language Processing 、ICASSP、COLING著名期刊和国际会议上发表了5篇学术论文，申请国家发明专利4项，其中2项已获得授权。本项目培养毕业了2名蒙古族博士，4名硕士。本项目取得的这些成果对于蒙古文信息化工作具有重要意义，并对维护国家安全及边疆少数民族地区的稳定，繁荣和发展少数民族文化具有重要推动作用。