基于BP神经网络的语音识别系统毕业设计论文.doc

摘要 随着计算机技术与人工智能的快速发展与广泛应用，语音识别越来越受到人们的关注和重视。目前常见的语音识别技术主要包括，基于矢量量化(Vector Quantization, VQ)的识别技术、动态时间规整(Dynamic Time Warping, DTW)、隐马尔可夫模型(Hidden Markov Models, HMM)、人工神经网络(Artificial Neural Network, ANN)等等。 本文讨论了语音信号的处理技术，包括语音信号预处理、信号特征提取，分析论述了BP神经网络模型思想。在Matlab 7.0环境实现了语音信号的预加重、分帧加窗、端点检测等基本信号处理过程，提取出线性预测分析系数(Linear Prediction Coding, LPC)、Mel频率倒谱系数(Mel Frequency Cepstrum Coefficient, MFCC)等特征数据。 针对语音识别，设计了三层的BP神经网络，对影响神经网络的关键参数进行了分析和调整。从统计角度上，分析对比了LPC和MFCC特征参数对应的BP神经网络的语音识别性能。 关键词： Abstract With speedy development and comprehensive application of computer technology and artificial intelligence, people began to focus on speech recognition. Up to now, there have been several speech recognition technologies being familiar by us, including Vector Quantization, Dynamic Time Warping, Hidden Markov Models, Artificial Neural Network, and so on. This paper would go into details about some processing technologies of speech signal, including pre-process of speech signal and feature extraction, would relate and analyze the principle of the model of feedforward neural network based on back-propagation algorithm (BPNN). In Matlab2007 environment, those experiments have implemented some basic operation of signal processing, such as pre-emphasis, enframing, windowing and endpoint checking of speech signal, extracted corresponding Linear Prediction Coding Coefficient (LPC) and Mel Frequency Cepstrum Coefficient (MFCC) from speech’s data, and constructed BP neural network. A endpoint checking algorithm having good performance is given in the paper. In the process, all the parameters are analyzed and setted properly in order to fetch the actual speech segment from the original speech accurately. A three-layer BPNN is designed for speech recognition system; the critical parameters of BPNN is studied. Statistically, it compares LPC’s BPNN with MFCC’s BPNN on the speech recognition ability. In the end, a strategy of reducing similar frame is proposed to solve the problem of the non-st