【作者】 韩仲志；

【导师】 黄汉明；

【作者基本信息】 广西师范大学 ， 电路与系统， 2006， 硕士

【摘要】 独立分量分析(ICA)是一种高阶统计量信号处理方法,1995年以来逐渐被广泛接受,越来越多地被应用于与信号处理有关的领域。目前,独立分量分析是模式识别与信号处理等相关领域的一个重要的研究热点。本文旨在研究基于独立分量分析的盲信号处理(BSP)理论及其在语音识别预处理中语音信号盲源分离(BSS)和语音增强等方面的应用。目的是寻求一种稳健的语音识别预处理手段。本文在讨论独立分量分析理论,重点研究了被广泛应用的FastICA算法的基础上,做了如下三个方面的研究工作:探讨了用不同的线性ICA算法实现的混叠语音盲分离,并主观和客观等不同角度评价用ICA方法对这些混叠信号的分离效果和算法的分离性能,另外,本文实现了一类单信道的盲分离问题的简单解决方案,它可以只用一个信道就能从被噪声污染的信号中把较纯净的语音提取出来,而传统方法至少需要两个信道。基于高斯白噪声背景,本文比较了传统算法和基于ICA/BSS算法在语音消噪和增强方面的差异,指出了基于ICA方法在高噪声环境下的优越性,并且指出在较大的信噪比范围内基于此方法的消噪的效果具有一定的平稳性,接着,本文讨论了ICA方法对多种不同环境噪声的适应性问题。此外,文章还提出了一种基于ICA与小波的联合降噪方法,试验表明,这种方法能使得语音增强性能进一步提高。文章最后,用DTW的方法实现了一个简单的非特定人连接词智能家居语音识别系统,基于这个系统本文设计了一系列的试验,试验证明在恶劣的环境下,基于ICA的预处理方法能够使系统的识别性能得到很大的提高。总之,不论从混叠语音的盲分离,还是从语音增强效果,乃至系统识别率的提高等不同角度来看,独立分量分析都不失为一种有效的语音识别预处理方法。本文的研究将对噪声环境特别是高环境噪声下的语音分析与识别具有积极的意义。更多还原

【Abstract】 ICA is a higher-order statistics tool. From 1995, it has been more and more widely recognized and applied in diverse sub-fields of signal processing domain. Nowadays it is an important and indispensable research frontier in correlative fields of pattern recognition and signal processing. This paper is mainly engaged in the research of BSP theory based on ICA and its application in blind speech signal separation and speech enhancement for the preprocessing of speech recognition, attempting to secure some robust measurement of speech recognition.ICA theory is discussed in the paper, especially noted the Fast-ICA algorithms which has been widely recognized. Based on which, we did three aspects research as follows:This paper makes an attempt to apply these ICA algorithms in speech signal blind separation. For evaluating the performance of separating pure speech signal from mixed speech signal by ICA, we propose some subjective and some objective measures. In addition, we also implement an algorithm that can acquire pure speech signal from mixed speech signal by a single channel, traditionally ICA needing at least two signal channels.Upon comparing ICA with classical de-noising algorithms, we find out ICA precedes classical algorithms if intensive environmental noise presents, also ICA shows notable stabilization for vast SNR range. Although above results are driven from speech signal entangled with intense Gaussian whiten noise, their applicability also discussed in many other different noise environment. Furthermore, we propose a united de-noising algorithm based on ICA and wavelet. Tentative tests show that this united method holding more powerful speech enhancement capability.Finally, we implement a preliminary intelligent household conjunctive words speech recognition system, based on which we design a series of examinations. The results indicate that the preprocessing methods based ICA can enhance the recognition capability of the system.In conclusion,not only from the blind source separation of mixing speech, but also from the performance of speech enhancement or from the advance the recognition ratio of system, ICA is indeed an kind of efficiency methods for speech recognition preprocessing. The study of this paper will have positive meanings to speech analysis and recognition under noising environment, especially under intensive environment noise.更多还原