【作者】 陆茵；

【导师】 李世作；

【作者基本信息】 广西大学 ， 控制理论与控制工程， 2007， 硕士

【摘要】 语音识别是一个复杂的非线性过程，基于线性系统理论的语音识别方法如隐马尔可夫模型(HMM)等技术的局限性逐渐凸现。随着人工神经网络的非线性理论研究和应用的逐渐深入，基于ANN的语音识别方法，逐渐成为研究的焦点。本文主要就前馈神经网络(BP神经网络)的原理及其在语音识别中的应用进行如下研究：1、如何有效的提取语音信号特征关于该问题首先从如何有效的检测语音信号的起止点进行研究，分析讨论了传统端点检测方法的优缺点，并对其进行了改进；然后，本文在研究基于线性预测倒谱和非线性MEL刻度倒谱特征的基础上，研究了LPCC和MFCC参数提取的算法原理及提取算法，并推导了一阶差分倒谱特征参数的提取算法。2、关于BP神经网络在语音识别中的应用主要研究了BP神经网络的原理，分析讨论了标准BP算法的优缺点及改进方法，并对传统的变学习率方法进行了改进。3、关于遗传算法在神经网络中的应用主要探讨了用遗传算法来优化神经网络的拓扑结构及权值的问题。该部分介绍了遗传算法的原理以及用它来优化神经网络拓扑结构、权值的步骤，同时还分析了遗传算法的主要参数对优化性能的影响。本文最终研究构造了一个基于遗传算法与BP神经网络的汉语语音识别模型，并完成了基于VC++6.0实验软件平台的程序设计与开发。针对非特定人的孤立词识别，识别率可以达到95％以上。更多还原

【Abstract】 Speech recognition is a complex nonlinear process. Up to now, most speech recognition method based on conventional linear system theory, such as Hidden Markov Model (HMM) is hard to have a breakthrough. Recently, with the development of nonlinear-system theories about artificial neural networks (ANN), the recognition method based on ANN became the focus of the research.The paper mainly stresses on theoretical studies and applications in speech recognition of feedforward neural networks (BP networks).1. How to extract speech-feature efficientlyAbout this queston, this paper firstly researches how to efficiently detect the endpoint of speech signal. In addition, the paper analyzed the advantages and disadvantages of traditional detection method and their improved method. Then, the paper discusses cepstral features based on linear predictive and non-linear Mel scal, and induces LPCC and MFCC extraction algorithm and their one rank coefficient is also presented.2. About the application of BP neural networks in speech recognitionThis paper studies the design principle of BP neural network and analyzes the advantages and disadvantages of traditional BP algorithm and its improved method. Then, the paper improved traditional method with dynamic learning rate to train network.3. About the application of Genetic Algorithm in neural networkThis paper mainly studies on those problems that optimizing neural networks topology structure and weights with genetic algorithm. In this part, the paper also introduces the theory of genetic algorithm and the progress of optimizing neural networks topology structure and weights with them. What’s more, the paper also analyzes the affect of genetic algorithm’s mainly parameters on optimizing performance.In the end, this paper designs a Chinese speech recognition model based on genetic algorithm and BP neural network. The design and exploitation of software for experiments is also completed based on VC++ 6.0. Experiment result show better recognition performance and particular application advantages are achieved by the method for speech recognition based on GA-BP. Especially for the Isolated Word Recognition (IWR), the recognition accuracy is better than 95%.更多还原