【作者】 叶勇；

【导师】 王琦；

【作者基本信息】 北京邮电大学 ， 电子与通信工程， 2015， 硕士

【摘要】 语音识别技术是实现计算机与人之间进行语言交流的关键技术之一,语音识别技术在交互方面具有极高的应用价值。随着关键词检测、语音合成、声纹识别等技术的进步,人们对语音技术的需求逐渐加深,语音识别产品正在走进我们的工作生活,逐渐深入到各个领域。本文主要是对汉语语音识别系统中关键词技术进行研究,通过两类不同的实际应用场景对关键词检测技术进行验证,主要的工作成果如下：1、针对智能电话客服的应用深入研究了基于垃圾模型(Garbage Model)关键词检测技术。完成了端点检测技术的改进、垃圾语料建模选取、以及确定垃圾模型和关键词模型网络权重值,再此基础上与人合作实现了关键词检测引擎,通过对接国内某电脑生产商维修预约客服电话系统,实现对关键词检测技术的验证,该系统中典型的关键词有硬盘问题、开机故障、系统崩溃等。通过话术引导、合理增加关键词长度或者合并关键词显著提高关键词检出率。2、面向语音文档检索应用需求,系统研究了基于音节网络(Syllable)的关键词检测技术,完成了对保险业和旅游业的的声学模型的训练,引入语言学知识提高关键词识别效果。借助保险业和旅游业的客服录音记录进行测试表明：可依据预先设定的关键词实现语音文档检索功能。更多还原

【Abstract】 Speech recognition is one of the key technologies to realize communications between equipment and human. In this field, Speech recognition technology has high value. With the progress of keyword spotting, speech synthesis and voiceprint recognition, and the increasing of demands, voice recognition products are walking into our work, lives and other fields.This research focuses on keyword spotting of Chinese speech recognition system. The keyword spotting technology has been verified and tested by two different practical applications. The main work and achievements are as follows:1. In this paper we considered the keyword spotting technology which is based on garbage model by intelligent customer service phone system. We achieved the improvement of endpoint detection technique, the selection of garbage language material, and confirming the network ratio of keyword model and garbage model. Based on the above work, we have realized the keyword spotting engine. In addition, by telephone customer service of a nation PC manufacturer, the keyword spotting technology has been verified. This system can detect some keywords associated with PC failure, such as "hard disk problem","start-up problem" and "system breakdown". Keywords detection rate also can be improved significantly through dialogue guide, increasing keyword length or merging keywords.2. Facing the demand for spoken document retrieval, we focused on the keyword spotting technology which is based on syllable network. The results of keyword spotting have been improved by training acoustic models of specific areas and introducing the linguistic knowledge to keyword spotting system. By testing the "Insurance" and "tourism" customer service record, we can draw the conclusion that keyword spotting system can achieve the function of speech document retrieval.更多还原