| 基于模型融合的母语与非母语语音识别 |
| |
| 引用本文： | 曾定,刘加. 基于模型融合的母语与非母语语音识别[J]. 电子测量技术, 2009, 32(6): 81-83,115 |
| |
| 作者姓名： | 曾定 刘加 |
| |
| 作者单位： | 清华大学电子工程系,清华信息科学与技术国家实验室(筹),北京,100084 |
| |
| 基金项目： | 国家自然科学基金委员会与微软亚洲研究院联合资助项目60776800，国家高技术研究发展计划(863计划):项目2006AA010101、项目2007AA042Z23、项目2008AA02Z414 |
| |
| 摘 要： | 母语与非母语英语发音方式通常存在固有的差别，这导致基于母语发音训练的语音识别模型不能适应非母语说话人。一种有效的方法是建立模型的补偿机制，来容忍母语与非母语说话人之间的发音变化。分析了中国人受母语的影响带来的英语发音变化，针对音素变化和声音变化，分别采用多发音字典和模型融合技术，实现了中国人说英语的语音识别率提高了15％，但母语英语的语音识别率下降不到1％。
|
| 关 键 词： | 语音识别 非母语 模型融合 多发音字典 |
Native and non-native speech recognition based on acoustic model merging |
| |
| Zeng Ding,Liu Jia. Native and non-native speech recognition based on acoustic model merging[J]. Electronic Measurement Technology, 2009, 32(6): 81-83,115 |
| |
| Authors: | Zeng Ding Liu Jia |
| |
| Affiliation: | (Tsinghua National Laboratory for Information Science and Technology, Department of Electronic Engineering, Tsinghua University, Beijing 100084) |
| |
| Abstract: | The inherent differences between native and non-native language pronunciation can lead to non-native language rate of decline using the model trained with native language speech. The confusions between Native and non-Native speaker lead the rate of decline. , which need to create a new model to tolerance this change. Set up on baseline Native English recogntion system, the character of Chinese people speaking English is firstly analyzed in this paper. We propose to analyze and model the phonetic and acoustic confusuons separately, using pronunciation dictionary and acoustic model merging technology to create a new model, with a significant 15% absolute WER reduction on the Chinese English, which only sacrifics 1% recognition rate on the native English. |
| |
| Keywords: | speech recognition non-native model merging pronunciation dictionary |
| 本文献已被 维普 万方数据 等数据库收录！ |
