| 跟着信息和互联网家当的疾速成长,信息社会对智能化的程度提出更高的请求。语音辨认技巧是一种快捷、便利的信息交流措施,它以语音为研究对象,终究目标是完成人性能够天然顺畅的语音通讯,从而为人类生涯供给方便。与传统的语音辨认办法比拟,支撑向量机具有更好的泛化机能和较高的辨认效力,今朝已普遍应用于形式辨认范畴。经由近几十年的成长,支撑向量机实际获得很年夜的完美,在原始算法的基本上,提出了一些改良的算法。本文就支撑向量机分类算法停止了研究,重要任务以下：(1)起首扼要引见了语音辨认的根本道理和办法。在剖析了传统语音辨认办法存在的缺乏的地方后,引入了以后风行的一种机械进修办法—支撑向量机,其次重点论述了支撑向量机的统计学实际。支撑向量机其实质可转换成二次计划成绩来求解,在线性弗成分时,经由过程非线性映照,将原始样本映照到高维核空间,选用恰当的核函数,便可获得高维空间的分类函数,从而完成线性可分。(2)针对传统的一对一支撑向量机算法在猜测阶段存在的缺陷,本文对该算法停止了改良。在分类辨认阶段,将得票较低的种别先剔除失落,不消盘算由这些种别组成的二分类器的决议计划函数值,进步算法的辨认效力。最初在分歧辞汇量、分歧信噪比的韩语语音库下停止了试验,到达延长猜测时光的目标。(3)支撑向量机在小样本、信噪比拟高的情形下有较高的辨认效力,然则在年夜范围样本、噪声情况下的成果就不尽人意。为懂得决这些成绩,本文彩用K比来邻算法先对练习样本停止删减,样本删减准绳是删除练习样本中的噪声点和离群点,使分类超立体尽量简略,进而进步练习速度。删减完成以后再用支撑向量机停止后续的练习和辨认任务。试验成果注解,经由删减以后,练习样本集和支撑向量的数量都年夜年夜削减,支撑向量机的练习速度显著加速,同时还坚持了较高的辨认率。 Abstract: With the rapid development of Internet and information industry, the information society to the intelligent degree of higher request. Speech recognition technique is a fast and convenient information exchange method, its voice as the research object, the goal is to complete all humanity can natural smooth voice communication, so as to supply convenient for human life. With the traditional speech recognition methods in comparison, SVM has better generalization performance and higher recognition efficiency, has been widely used in pattern recognition field. Through decades of development, the support vector machine to get the perfect real big, basically in the original algorithm, puts forward some improved algorithm. In this paper, the support vector machine classification algorithm is studied, an important task of the following: (1) this paper introduced the basic principles and methods of speech recognition. In the analysis of the lack of local traditional speech recognition methods are introduced, after a popular machine learning methods, support vector machine, and then it focuses on the actual statistical support vector machine. Support vector machine (SVM) its essence can be converted to a quadratic program results to solve, online of Eph composition, through the process of nonlinear mapping, the original sample mapping to high-dimensional kernel space, choose an appropriate kernel function, obtain the classification function in high dimensional space, thus completing the linearly separable. (2) according to the traditional one to a support vector machine algorithm in this paper has the defects of speculation stage, improved the algorithm. Identify the stage in the classification, will lower a vote not to eliminate needless loss calculation by these species are composed of two classifier decision function value, the algorithm identifies efficiency progress. The initial differences in vocabulary differences, the signal-to-noise ratio of the Korean speech database from the test, to prolong the time of the target of speculation. (3) the support vector machine to identify the effectiveness of higher than under the condition of the high in the small sample, but the signal-to-noise in the large-scale sample, under noise results. To understand these results, the color with k nearest neighbor algorithm to the training samples to stop cut, sample cut criterion is to delete your practice in a sample of noise points and outliers, and make the classification ultra stereo as far as possible simple, and improve the training speed. Cut and then after the completion of the support vector machine to stop the following exercises and identification task. The test results through the cut, the number of notes, training sample set and the support vectors are greatly reduced, the support vector machine practice speed significantly accelerated, but also adhere to the high recognition rate. 目录: |
