《家用电器中语音识别技术的DSP实现.doc》由会员分享，可在线阅读，更多相关《家用电器中语音识别技术的DSP实现.doc（72页珍藏版）》请在金锄头文库上搜索。
1、分类号 学校代号 10462 UDC 密级 学 号 200101001 Z ZH HE EN NG GZ ZH HO OU U I IN NS ST TI IT TU UT TE E O OF F L LI IG GH HT T I IN ND DU US ST TR RY Y 硕士学位论文 家用电器中语音识别技术的家用电器中语音识别技术的 DSP 实现实现 化学发光免疫定量测定仪的研制化学发光免疫定量测定仪的研制 学学位位申申请请人人 吴吴晓晓平平 导导师师姓姓名名及及职职称称 崔光照崔光照 教授教授 路康路康 副教授副教授 专专业业名名称称 电电机机与与电电器器 学学科科门门类类 工工学学
2、 论论文文提提交交日日期期 2004 年年 5 月月 20 日日 论论 文文 评评 阅阅 人人王王世世卿卿 段段银银田田 Zhengzhou Institute of Light Industry DSP Realization of Speech Recognition Technology In Home Electric Appliances Mas D Candidate Wu Xiaoping Supervisor Prof Cui Guangzhao A Prof Lu Kang Major Electrical Machinery and Electric Equipment M
3、ayMay 20042004 I 摘摘 要要 随着数字化技术的迅猛发展 语音识别技术也日趋完善 并且在工业 军事 交 通 医学 民用等领域都得到了广泛的应用 目前小词汇量的语音识别系统的研究已 经比较成熟 并且各种各样的产品也相继涌现 在这种背景下 如何将语音识别技术 运用到家电中 实现利用语音控制家电 是智能化信息家电中的一个重要的研究课题 与研究方向 论文首先介绍了语音识别技术的基本概念 讨论了语音识别系统的组成和实现的 技术 在此基础上阐明了本课题的课题背景 研究任务及研究意义 其次 论文详细分析了构成语音识别系统的四个组成部分 包括语音信号数字化 与预处理 语音的端点检测 特征提取与模
4、式匹配 着重介绍了实现端点检测的短时 平均能量与短时平均过零率分析 语音信号的线性预测分析及在此基础之上的倒谱特 征参数 以及实现模式匹配的常用的矢量量化技术 动态时间规整技术和隐马尔可夫 模型 接着 根据提出的语音识别系统的构成 论文介绍了在 MATLAB6 5 上实现了采用 动态时间规整算法的识别系统的仿真分析 仿真结果表明 语音信号的线性预测分析 较好的描述了语音帧之间的声学相似性 采用倒谱特征参数的识别系统具有良好的识 别效果 动态时间规整解决了由于说话人发音速度不均匀而引起的语音信号变化 对 短时的语音信号具有良好的匹配效果 课题的研究成果是在 TI 公司推出的 TMS320VC54
5、02 DSK 开发板上实现了一个小 词汇量的语音识别系统 该系统实现了从 0 9 十个汉语数码的实时语音识别 系统 系统完成了语音信号的端点检测 倒谱特征参数的求解以及动态时间规整的匹 配 软件由 C 语言编写 并通过了调试 实验结果表明 语音识别系统的性能达到预 期的目的 能够满足实时要求 根据家用电器中的语音控制命令的特点 提出了识别 系统在家电中的应用方案 论文的最后部分对研究出的语音识别系统提出了一些改进方案 如采用隐马尔可 夫模型与矢量量化技术的具体方法 在高噪音环境中的语音识别方案 改进的动态时 间规整算法以及说话人识别技术 为下一步的工作指明了方向 所讨论的语音识别系统具有一定的
6、识别能力 但是真正要面向实用还有很多的工 作要做 对本系统的研究为下一步的开发和研究具有实用性的语音识别系统作了许多 重要的尝试和探索工作 II 关键词 关键词 语音识别 动态时间规整 矢量量化 隐马尔可夫模型 倒谱特征参数 说 话人识别 ABSTRACTABSTRACT With the rapid development of digitization technology the speech recognition technology becomes perfect day by day also it has been widely used in the field of ind
7、ustry military traffic medical science civil etc The research of little vocabulary recognition system has been relatively skilled at present and various products emerge in succession Under this background how to apply speech recognition technology to the home electric appliances implementing sound c
8、ontrol is an important subject and direction to research in intelligent information household appliances This thesis had introduced the basic conception of the recognition technology of the pronunciation at first had discussed composition of the speech recognition system and technology realization a
9、lso had elucidated subject background research tasks and research meanings of the subject on this basis Secondly the various components of the speech recognition are laboured in the thesis The recognition system is made of four parts including the digitization and the preprocessing of the speech sig
10、nals the capsheaf measurement the extraction of character and the mode match The short time average energy and passing zero rate measure the extreme point The linear predication analysis of the speech signals and cepstrum character coefficient is introduced The vector quantization VQ technology the
11、dynamic time warping DTW technology and the hidded Markov model HMM are widely used for the mode match Then according to the provided components of the speech recognition simulation analysis of the recognition system with the DTW algorithm is realized on the MATLAB6 5 The simulation results indicate
12、d that the linear predication analysis of the speech signals better describes the acoustic similarity between the speech frames and the recognition system using the cepstrum character coefficient has good discernment results The speech signals change because of the different sound speed which is exa
13、ctly resolved by the DTW algorithm for it has excellent match effect for the short time speech signals A little vocabulary speech recognition system is realized on the TMS320VC5402 DSK produced by TI The real time speech system can recognize ten Chinese numbers from zero to III nine It completes the
14、 capsheaf measurement of speech signals the calculation of the cepstrum character coefficient and the mode match of DTW algorithm The software is compiled by C langage and debugged successfully The experiment results show that the speech recognition system can achieve the anticipated goal and meet t
15、he real time request According to the properties of speech control the schemes of speech recognition are given in the field of of home electric appliances Some improvement schemes to the speech recognition system is put forward in the last part of the thesis such as the detailed ways using the HMM a
16、nd VQ technology the speech recognition schemes in the high noise environment the improved DTW algorithm and speaker s recognition technology They point out the direction for the next work The discussed speech recognition system has certain recognition capability but a lot of works still should be done for the practical Many important tries and exploring work research of had been finished for further application of speech recognition system Key Words Speech Recognition Dynamic Time Wrapping DTW
