强噪声环境下语音识别及VUI系统设计与实现
发布时间：2018-09-18 13:31
【摘要】：伴随着人机交互技术的快速发展,语音用户界面(Voice User Interface,VUI)逐步成为国内外的研究热点。借助VUI系统,改变传统的键盘输入模式,代之以语音输入的方式,人机交互更加的便捷和人性化。然而实际应用中环境噪声复杂,VUI往往会遇到识别和训练环境不相匹配的情况,从而使得语音识别率较低。因此,本文将经验模态分解(Empirical Mode Decomposition,EMD)、希尔伯特-黄变化(Hilbert-Huang Transform,HHT)以及双麦克风噪声干扰对消技术相结合。提高了VUI系统在强噪声环境下的识别率,从而给飞机辅助维修设备提供可靠的人机交互。本文的主要研究内容如下:第一,针对VUI系统在国内外的研究现状和发展趋势,分析了目前民用航空辅助维修的需求,阐述了存在的实际问题和需要改进的方面。第二,以往的语音端点检测算法一般是利用语音信号的短时能量、短时平均过零率等时域特征参数。能量计算方法不尽合理,且在低信噪比情况下识别效果较差。本文研究了基于EMD和Teager能量算子的语音端点检测技术,该方法结合EMD和Teager能量算子在表征非线性非平稳信号上的优势,EMD分解语音信号实现初步去噪,然后利用Teager能量来代替短时能量进行端点检测。第三,在去噪处理方面,传统的方法是用单一麦克风获取带噪语音,然后进行小波变换、谱减等。考虑到飞机维修现场的噪声频域分布更广、幅度更大等特性,本文引入双麦克风自适应噪声对消技术,一路麦克风采集带噪语音,一路麦克风采集背景噪声,利用递归最小二乘(Recursive Least Square,RLS)自适应算法,在时域上对消两路信号,最大程度去除噪声成分,保留有效语音,最终实现信噪比的提高。第四,详细阐述了基于以上两种技术的VUI系统各模块实现过程以及相互之间的通信方式。该设计采用客户端-服务器端(C/S)结构,有效地利用了客户端和服务器端的负载。通过对隐马尔科夫模型(Hidden Markov Mode,HMM)的10次自适应训练,从语音模板、噪声门限值和二次识别语音库这三个方面进行改进,对语音信号进行测试实验,给出了本文所设计VUI的识别率测试结果。分析表明,该VUI系统具有更强的抗噪性能,在识别率测试上较以往的VUI系统有5%左右的提高。
[Abstract]:With the rapid development of human-computer interaction technology, voice user interface (Voice User Interface,VUI) has gradually become a research hotspot at home and abroad. With the help of VUI system, the traditional keyboard input mode is changed and replaced by voice input mode, so the human-computer interaction is more convenient and humanized. However, in the practical application, the environment noise is complex and the VUI often meets the situation that the recognition and training environment do not match each other, which makes the speech recognition rate lower. Therefore, empirical mode decomposition (Empirical Mode Decomposition,EMD), Hilbert-Huang variation (Hilbert-Huang Transform,HHT) and dual microphone noise cancellation are combined in this paper. The recognition rate of VUI system in strong noise environment is improved, thus providing reliable man-machine interaction for aircraft auxiliary maintenance equipment. The main contents of this paper are as follows: firstly, in view of the research status and development trend of VUI system at home and abroad, the requirements of civil aviation auxiliary maintenance are analyzed, and the existing practical problems and aspects need to be improved are expounded. Secondly, the previous speech endpoint detection algorithms usually use the time domain characteristic parameters such as the short time energy of speech signal, the short time average zero crossing rate and so on. The energy calculation method is not reasonable, and the recognition effect is poor in the case of low signal-to-noise ratio (SNR). In this paper, the speech endpoint detection technology based on EMD and Teager energy operator is studied. This method combines the advantages of EMD and Teager energy operators in the representation of nonlinear non-stationary signals. Then the Teager energy is used to replace the short-time energy for endpoint detection. Thirdly, in the aspect of denoising, the traditional method is to use a single microphone to acquire noisy speech, then wavelet transform, spectral subtraction and so on. Considering the characteristics of the aircraft maintenance site, such as wider distribution of noise frequency domain and larger amplitude, this paper introduces the dual-microphone adaptive noise cancellation technology, one way microphone to collect noisy voice, the other way microphone to collect background noise, By using the recursive least squares (Recursive Least Square,RLS) adaptive algorithm, two channels of signals are eliminated in time domain, the noise components are removed to the maximum extent, the effective speech is retained, and the signal-to-noise ratio (SNR) is improved. Fourthly, the realization process and communication mode of each module of VUI system based on the above two technologies are described in detail. The design adopts the client-server (C / S) structure, and makes effective use of the load of the client and server. Based on the 10 times adaptive training of Hidden Markov Model (Hidden Markov Mode,HMM), the speech signal is tested from three aspects: speech template, noise threshold and second recognition speech corpus. The recognition rate test results of the VUI designed in this paper are given. The analysis shows that the VUI system has better anti-noise performance, and the recognition rate is about 5% higher than that of the previous VUI system.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN912.34
本文编号：2248071
[Abstract]:With the rapid development of human-computer interaction technology, voice user interface (Voice User Interface,VUI) has gradually become a research hotspot at home and abroad. With the help of VUI system, the traditional keyboard input mode is changed and replaced by voice input mode, so the human-computer interaction is more convenient and humanized. However, in the practical application, the environment noise is complex and the VUI often meets the situation that the recognition and training environment do not match each other, which makes the speech recognition rate lower. Therefore, empirical mode decomposition (Empirical Mode Decomposition,EMD), Hilbert-Huang variation (Hilbert-Huang Transform,HHT) and dual microphone noise cancellation are combined in this paper. The recognition rate of VUI system in strong noise environment is improved, thus providing reliable man-machine interaction for aircraft auxiliary maintenance equipment. The main contents of this paper are as follows: firstly, in view of the research status and development trend of VUI system at home and abroad, the requirements of civil aviation auxiliary maintenance are analyzed, and the existing practical problems and aspects need to be improved are expounded. Secondly, the previous speech endpoint detection algorithms usually use the time domain characteristic parameters such as the short time energy of speech signal, the short time average zero crossing rate and so on. The energy calculation method is not reasonable, and the recognition effect is poor in the case of low signal-to-noise ratio (SNR). In this paper, the speech endpoint detection technology based on EMD and Teager energy operator is studied. This method combines the advantages of EMD and Teager energy operators in the representation of nonlinear non-stationary signals. Then the Teager energy is used to replace the short-time energy for endpoint detection. Thirdly, in the aspect of denoising, the traditional method is to use a single microphone to acquire noisy speech, then wavelet transform, spectral subtraction and so on. Considering the characteristics of the aircraft maintenance site, such as wider distribution of noise frequency domain and larger amplitude, this paper introduces the dual-microphone adaptive noise cancellation technology, one way microphone to collect noisy voice, the other way microphone to collect background noise, By using the recursive least squares (Recursive Least Square,RLS) adaptive algorithm, two channels of signals are eliminated in time domain, the noise components are removed to the maximum extent, the effective speech is retained, and the signal-to-noise ratio (SNR) is improved. Fourthly, the realization process and communication mode of each module of VUI system based on the above two technologies are described in detail. The design adopts the client-server (C / S) structure, and makes effective use of the load of the client and server. Based on the 10 times adaptive training of Hidden Markov Model (Hidden Markov Mode,HMM), the speech signal is tested from three aspects: speech template, noise threshold and second recognition speech corpus. The recognition rate test results of the VUI designed in this paper are given. The analysis shows that the VUI system has better anti-noise performance, and the recognition rate is about 5% higher than that of the previous VUI system.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN912.34
【参考文献】
相关期刊论文 前1条
1 张德祥;吴小培;吕钊;郭晓静;;基于经验模态分解和Teager峭度的语音端点检测[J];仪器仪表学报;2010年03期
，本文编号：2248071
本文链接：https://www.wllwen.com/kejilunwen/wltx/2248071.html
