CASIA OpenIR  > 毕业生  > 硕士学位论文
基于GMM的声音信号分类器研究
其他题名Sound Classificaition Based on GMMs
王志强
学位类型工学硕士
导师徐波
2003-06-02
学位授予单位中国科学院研究生院
学位授予地点中国科学院自动化研究所
学位专业模式识别与智能系统
关键词语音识别技术 模式识别 高斯混合模型(Gmms) Mfcc Sound Classification System Pattern Recognition Automatic Speech Recognition Text-independent Speaker Recognition Covariance-t
摘要声音信号的分类是模式识别技术的一个重要应用方向。近年来,它从语音识别技术等具体的模式识别应用技术中汲取了大量的有益经验,并根据自己的特 色,发展出了有自己特色的研究成果。与此同时,这一技术也存在着许多问题, 集中于怎样从声音信号中提取出表征信号本身的比较稳定的特征信息,以及怎 样改进训练方法使其针对现有的特征改进训练方法得到推广性更强的模型。本 文将详细介绍作者攻读硕士学位期间在基于高斯混合模型(GMMs)的声音信号 分类器方面的一些工作。 我们首先研究了今年来基于声纹的说话人识别技术,特别是文本无关的说话 人识别。在分析了对说话人识别效果影响的一些因素的基础上,建立了一个文 本无关的说话人识别系统。并针对现有建模技术对距离度量的不合理性,以及 对数据间共有信息的忽视,提出了基于方差共享的结构聚类方法,挖掘了相近 类中数据所包含的共有信息,更为有效地刻画了数据的分布状况。这个方法在 一定程度上解决了数据稀疏问题。 其次,基于GMM的说话人识别系统框架和技术,研制开发了油田管道声音信 号监测系统。该系统在实际使用中取得了较好的效果。在系统的开发过程中, 仔细分析了MFCC特征提取方法对声音信号特征的提取,并针对实际信号对该方 法作了改进。同时,针对实际工程的需要,实现了定点程序的声音信号分类系 统,并将识别过程中求取所有高斯组件的概率打分和作了改进,提高了识别速 度,识别率基本保持不变。 再次,在油田管道声音信号分类系统开发过程中,采集并建立了大量的现场 信号,为声音信号的进一步分析提供了数据保障。 本文提出的各种算法和以及改进都在实际信号中作了大量测试验证其效果, 并在试验后作了有针对性的分析。
其他摘要Sound classification system is one important application for pattern recognition. Recent years, sound classification system benefits a lot from the research in the field of pattern recognition such as automatic speech recognition, etc, and is becoming a reasonable selection as independent research aspect. However, the technique of sound classification has many questions to be solved, which focuses on how to get more robust feature from sound signal and how to improve the training process to get robust models based on the current technique of feature extraction. This paper will present in detail the author's research work on these problems in the courts of study as a master candidate. Firstly, we do some work to understand the speaker recognition technique based on speech signal, especially on text-independent speaker recognition. After the analysis of some factors in speaker recognition, we set up one text-independent speaker recognition system. We find some unreasonable question on distance measurement and the neglect of the mutual information between data of different sound classes. So based on these factors, we proposed a novel method, namely Covariance-tied Clustering Method, to mining the mutual information between data of different classes, and effectively take in much information of data distribution for distance measurement. This method can avoid data sparseness to some purpose. Secondly, based on the GMMs system frame of text-independent speaker rec6gnition, we developed the system of sound monitor for oil pipeline. This system gets better performance in the application. In the development of oil system, we analyzed the MFCC method of feature extraction for speech signal, and improve it for the oil sound. In the same time, in order to fit the realization in hardware, we achieved the integer program for the sound classification system, and modified the recognition process to save more time while the recognition rate is almost invariable. Thirdly, during the course of developing the sound monitor for oil pipeline, we collect the sounds of oil pipeline and constructed a corpus of actual sound and ensure the farther analysis of pipeline sounds.
馆藏号XWLW687
其他标识符687
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/6845
专题毕业生_硕士学位论文
推荐引用方式
GB/T 7714
王志强. 基于GMM的声音信号分类器研究[D]. 中国科学院自动化研究所. 中国科学院研究生院,2003.
条目包含的文件
条目无相关文件。
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[王志强]的文章
百度学术
百度学术中相似的文章
[王志强]的文章
必应学术
必应学术中相似的文章
[王志强]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。