多类分类支持向量机在嵌入式语音识别系统中的研究

发布时间：2018-01-04 07:15

本文关键词：多类分类支持向量机在嵌入式语音识别系统中的研究　出处：《太原理工大学》2012年硕士论文　论文类型：学位论文

【摘要】：语音识别作为一门交叉学科,在人类智能化和信息化的道路上有着不可忽视的作用。近些年,嵌入式已经成为了信息领域的研究热点。在嵌入式系统中应用语音识别技术成为了语音识别发展的新方向。语音识别技术的关键是解决多类分类问题,基于统计学习理论的支持向量机方法因为其在解决分类问题上面的独特优势,已经成为了语音识别领域的研究热点。支持向量机方法来源于统计学习理论,克服了传统语音识别方法(人工神经网络、隐马尔科夫模型)的不足,在有限样本多类分类问题中得到了广泛的应用。支持向量机方法本来是解决二分类问题的,研究者在其基础上推广出了多种多类分类的方法,一对余组合分类法、对一组合分类法、决策有向无环图组合分类法、纠错输出编码多类分类、超球多类分类等方法。本文采用DM6446开发板作为嵌入式语音开发平台,其软硬件性能完全能满足语音识别技术的嵌入式开发。针对决策有向无环图支持向量机算法和纠错输出编码支持向量机算法,搭建嵌入式交叉编译平台。在搭建好的平台上,将两种支持向量机算法移植到DM6446开发板。两种多类分类支持向量机算法分别经过两种不同的语音库进行语音识别实验,均得到了较高的识别率。但由于嵌入式系统自己的局限性,算法所需的训练时间较长,通过样本预选取算法处理训练样本后再进行语音识别实验,在保证良好识别率的情况下,极大地缩短了训练所需时间,达到了预期的效果。基于Java语言的超球支持向量机算法是多类分类方法的一个新思路。在Java程序设计中,Java虚拟机处于核心地位,正是由于虚拟机的存在,保证了Java语言在各种平台都可以不加修改的运行。本文针对Java VM虚拟机,搭建嵌入式交叉编译环境,成功的将虚拟机移植到嵌入式开发板,并验证了Java类包的正确性。运用超球支持向量机算法进行语音识别实验,得到了良好的识别结果。
[Abstract]:Speech recognition is a cross subject, plays an important role in human intelligence and information on the road. In recent years, embedded system has become a hot research field of information. The application of speech recognition technology in embedded systems has become the new direction of development of speech recognition.
The key technology of speech recognition is to solve the multi class classification problem, based on statistical learning theory and support vector machine method because of its unique advantages in solving classification problems above, it has become a research hotspot in the field of speech recognition. Support vector machine method in statistical learning source theory, to overcome the traditional speech recognition method (artificial neural network. The hidden Markov model) problems, has been widely used in the multi class classification problem of limited samples. SVM is to solve two classification problems, the researchers based on the promotion of a variety of multi class classification method, a classification method of combination of Yu, a combination classification method, decision making acyclic graph combination classification method, error correcting output encoding multi class classification, hyper sphere multi class classification method.
This paper uses the DM6446 development board as the embedded speech development platform, the software and hardware performance can meet the development of embedded speech recognition technology. Aiming at the decision directed acyclic graph support vector machine algorithm and error correcting output encoding algorithm of support vector machine, build the embedded cross compiler platform. In the platform, the two kinds of support vector machine algorithm ported to DM6446 development board. Two kinds of multi class classification algorithm of support vector machine respectively through two different speech database for speech recognition experiments were obtained with high recognition rate. But due to the limitation of the embedded system, the long training time required by the algorithm, through the sample pre selection algorithm of training samples after speech recognition experiments, in order to ensure the good recognition rate, greatly shorten the training time needed to achieve the desired results.
Based on the Java language of the hyper sphere support vector machine algorithm is a new idea of multi class classification method. In the design of Java program, the Java virtual machine is the core, it is because of the existence of the virtual machine, to ensure that the Java language can be run without modification in various platforms. According to the Java VM virtual machine. Build the embedded cross compiler environment, the success of the virtual machine is transplanted to the embedded development board, and verifies the correctness of the Java package. The use of hyper sphere support vector machine algorithm for speech recognition experiment, get good recognition results.

【学位授予单位】：太原理工大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP368.1;TN912.34

【引证文献】