CASIA OpenIR  > 毕业生  > 硕士学位论文
共振峰参数提取和语音识别系统鲁棒性的研究
杨芳
Subtype工学硕士
Thesis Advisor刘迎建
2002-05-01
Degree Grantor中国科学院研究生院
Place of Conferral中国科学院自动化研究所
Degree Discipline模式识别与智能系统
Keyword语音识别 鲁棒性 共振峰的跟踪 基本轨迹 语音学 谱减法 倒谱均值减 Speech Recognition Robust Speech Recognition Formant Tracking Elemental Track Phonetics Spectral Subtraction
Abstract本论文的研究内容包括两部分:l、鲁棒性参数-共振峰参数的提取和跟踪问题 的研究;2、传统的抗噪方法对识别系统的影响。 我们首先通过对汉语拼音的共振峰的观察总结出元音的共振峰分布的特殊规 律,即: 1)发音段稳定的局部表现出很强的关联性。 2)音与音之间的过渡段表现很强动态的变化特征。 以此规律为基础,借鉴英语共振蜂跟踪算法,我们提出了一套适用于拼音单元 音和双元音的共振峰跟踪算法,这个算法有以下的特点: 1)以语音线性分析为基础。 2)以点为基础通过约束条件产生共振峰的基本轨迹。 3)在以线(基本轨迹)为基础,以类似于Viterbi算法的动态规划跟踪算法为 工具,应用前后帧的相关性和引入虚拟基本轨迹的方法得到最终的跟踪结果。 以上算法得到较满意的结果,为以后共振峰参数在汉语的语音识别和抗噪问题 的解决,提供了研究的基础。 语音识别系统抗噪问题的研究,我们采用前端数字信号处理的方法来提高识别 器的鲁棒性。针对两类失真情况:通道的变化,及噪声环境的影响,分别采用倒谱 均值减和改进后谱减法来处理。最后,我们将两种方法结合起来,取得了一定的效 果。
Other AbstractThis paper contains two parts: 1. FORMANT tracking of the mandarin. 2. The conventional methods to improve the robustness of the recognition system. We firstly draw two conclusions of the FORMANT distribution of the mandarin by the observation: 1) It shows the connection in the region of stable pronunciation. 2) It shows the fluctuations in the region changing from one phone to another phone. Secondly, we proposed our method of optimal mandarin FORMANT tracking base on the conclusions drew above, the theory of mandarin phonetics and the methods of tracking in English speech. The original aspects of our FORMANT tracking method are followings: 1) It is based on the LPC analysis. 2) Construct the FORMANT elemental tracks by the dot. 3) Based on the elemental tracks, use the dynamic programming algorithm to tracking the tacks of the FORMANT. In this course we propose the virtual elemental tracking and relationship between the speech frames to help the performance of the algorithm. In the experiment of the FORMANT tracking of the phone and diphone, we get the promising results. All these efforts provide a base for the future work such as applying the FORMANT feature in the speech recognition and robust speech recognition. As to the conventional methods to improve the robust, we use signal procession methods in the front end of the speech recognition to improve the robust of recognition system. We use the algorithm of Cepstral Mean Subtraction combined with improved Spectral Subtraction method to the recognition system, and the result is promising.
shelfnumXWLW644
Other Identifier644
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/6865
Collection毕业生_硕士学位论文
Recommended Citation
GB/T 7714
杨芳. 共振峰参数提取和语音识别系统鲁棒性的研究[D]. 中国科学院自动化研究所. 中国科学院研究生院,2002.
Files in This Item:
File Name/Size DocType Version Access License
硕士生学位论文-644.pdf(3411KB) 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[杨芳]'s Articles
Baidu academic
Similar articles in Baidu academic
[杨芳]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[杨芳]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.