【作者】 陈俊彬；

【导师】 钟映春； 翁衡；

【作者基本信息】 广东工业大学 ， 控制工程（专业学位）， 2016， 硕士

【摘要】 随着人口老龄化的日趋严重、以及心脑血管疾病的年轻化,使得越来越多人丧失行动能力,加上护理人员的缺乏,社会需要加速研发更为智能的护理床。护理床采用语音控制来取代单纯的按钮控制,能够为失能患者带来很大的方便。然而,在养老院以及医院,通常都会有多张护理床处在同一室内,环境相对嘈杂,若是多人同时发送语音命令则容易相互干扰。如果采用非特定人语音控制系统,那么任何人都能够对护理床发送语音命令,将对患者的安全产生严重影响,有可能导致患者二次伤害,这制约着护理床的推广应用。针对这个问题,本文基于Android平台,开发了一套融合了声纹识别的护理床语音控制系统。首先,本文综述了课题的背景知识、研究意义,并介绍了护理床的研发现状,分析了声纹识别以及语音识别的研究现状。本文指出了,尽管语音信号的处理识别技术发展至今已经相当成熟了,然而在目前市场上,能够针对特定患者的声纹特征进行语音识别的护理床控制系统还是缺乏。其次,本文基于用户体验原则,对护理床语音控制系统的需求作了分析,对各种开发平台进行了比对,选择了Android开发平台,然后进行语音控制系统的概要设计。第三,本文探讨常见的音频格式的特点,在Android上采用WAV格式做语音特征提取。然后分析了提取音频文件的数据,并对它做预处理,研究了确定其参数的方法。紧接着是对语音数据进行MFCC特征提取。接着,解决了在Android上编程时,遇到NaN数据的问题,实现在Android平台上的MFCC特征提取。第四,深入探讨了声纹识别以及语音识别的各种算法基础上,选择了GMM作为声纹识别的算法,并在Matlab上做实验进行验证,以及选择HMM和DTW进行Matlab对比实验,实验样本采用个人录制的语音所提取的MFCC特征。实验也证明了,在合理设置阈值的前提下,GMM适合作为控制系统的声纹识别算法,用于屏蔽非护理床用户的语音。也证明了在少量训练样本、识别对象为孤立词的情况下,采用DTW识别效果以及识别时问都优于采用HMM。而后,在Android平台上实现了基于GMM的声纹识别和基于DTW的语音识别。第五,本文根据概要设计、界面设计和流程设计,详细阐述了在Android平台上设计护理床语音控制系统的过程。然后,对语音控制系统进行性能试验。设置阈值后,当声纹识别的召回率为72.5%时,非护理床用户的语音有1.25%的可能致护理床动作。通过了声纹识别后,指定护理床用户的语音识别率为97.13%,而非指定护理床用户的语音识别率为14.29%。指定护理床用户从读入语音到护理床产生动作的过程,需要时间1-2秒。最后,对本课题的工作作了总结和展望,并提出了课题今后应该努力的方向。更多还原

【Abstract】 With the seriousness of the aging of the population and the youth of cardiovascular and cerebrovascular diseases, more and more people lose the ability to act, besides, with the lack of the nursing staff, more intelligent care bed need to be developed by society. Care beds controlled by speech instead of a simple button, can bring great convenience for patients with disability. However, in gerocomium and hospitals, there are usually more than one care bed in the same room, and the environment is relatively noisy, it is easy to interfere with each other if many people send speech commands at the same time. If the non-specific human speech control system is applied, anyone can control care bed by sending speech commands, this would have a serious impact on the safety of the patients, and it is possible to cause secondary damage to the patients, which restricts the popularization and application of care bed. To solve this problem, this paper developed a set of speech control system of care bed that have the function of voiceprint recognition based on Android platform.First of all, this paper reviews the background and significance of the research of this topic and the development status quo of care bed, and analyzes the development status quo of voiceprint recognition and speech recognition. The paper points out that control system of care bed that can recognize the characteristics of voiceprint of a particular patient is lacking in the current market, although the processing and recognition technology of speech signal has been quite mature so far.Secondly, Android development platform is selected based on the experience of user with the analyses of the demand of voice control system of care bed and comparison of various development platforms, and then summary design of the speech control system is made.Then, the paper discusses the characteristics of the common audio format, and the experiment of speech recognition is carried out with WAV format on the Android platform, then the data of the audio file extracted is analyzed, and the method of the determination of the parameters is studied. And then the speech data were extracted by Mel transform coefficients (MFCC). At the end of this chapter, the method of feature extraction of MFCC is realized on the Android platform, and the solution to the problem of NaN data is put forward when the program is designed.And then, various algorithms about voiceprint recognition and speech recognition are studied, and GMM is selected as the algorithm of voiceprint recognition. The paper makes experiment validation on Matlab, and chooses the HMM and DTW algorithms as a comparison. And the paper adopts MFCC feature extracted by individual recorded voice as the experimental samples. The experiments have proved that GMM algorithm used for shielding non-care bed user’s voice is suitable for voiceprint recognition of control system under the premise of setting the threshold value reasonably. It is also proved that the recognition effect and the recognition time of the DTW are better than that of the HMM in the case of a small number of training samples and the identification of the isolated words. At the end of this chapter, the algorithms of GMM and DTW are also implemented on the Android platform, and the problem and method of processing of covariance matrix approximation "strange" is proposed in the GMM program design.Then, according to the summary design, starting from with the interface and program, this paper expounds the design process of care bed control system of speech in two aspects on the Android platform, after that, making performance tests on speech control system that have being realized. After setting the threshold value, when the recall rate of voiceprint recognition is 72.5%, speech of non-care bed users has a probability of 1.25% of causing a care bed action. After voiceprint recognition, speech recognition rate of care bed users is 97.13%, however, speech recognition rate of the non-care bed users is 14.29%.At last, the paper makes the summaries and prospects of this project, and puts forward the direction of the subject in the future.更多还原