MATLAB环境下的基于HMM模型的语音识别系统_语音识别_matlab_其他

<html xmlns="http://www.w3.org/1999/xhtml"> <head> <meta charset="utf-8"> <meta name="generator" content="pdf2htmlEX"> <meta http-equiv="X-UA-Compatible" content="IE=edge,chrome=1"> <link rel="stylesheet" href="https://static.pudn.com/base/css/base.min.css"> <link rel="stylesheet" href="https://static.pudn.com/base/css/fancy.min.css"> <link rel="stylesheet" href="https://static.pudn.com/prod/directory_preview_static/6250997e74bc5c01056b6b60/raw.css"> <script src="https://static.pudn.com/base/js/compatibility.min.js"></script> <script src="https://static.pudn.com/base/js/pdf2htmlEX.min.js"></script> <script> try{ pdf2htmlEX.defaultViewer = new pdf2htmlEX.Viewer({}); }catch(e){} </script> <title></title> </head> <body> <div id="sidebar" style="display: none"> <div id="outline"> </div> </div> <div id="pf1" class="pf w0 h0" data-page-no="1"><div class="pc pc1 w0 h0"><img class="bi x0 y0 w1 h1" alt="" src="https://static.pudn.com/prod/directory_preview_static/6250997e74bc5c01056b6b60/bg1.jpg"><div class="t m0 x1 h2 y1 ff1 fs0 fc0 sc0 ls0 ws0">　</div><div class="t m0 x2 h3 y2 ff1 fs1 fc0 sc0 ls0 ws0">　·470 ·　　　　　</div><div class="t m0 x3 h2 y3 ff1 fs0 fc0 sc0 ls0 ws0">计算机测量与控制 . 2004 . 12</div><div class="t m0 x4 h4 y4 ff2 fs3 fc0 sc0 ls0 ws0">(</div><div class="t m0 x5 h4 y3 ff2 fs3 fc0 sc0 ls0 ws0">5</div><div class="t m0 x6 h4 y4 ff2 fs3 fc0 sc0 ls0 ws0">)</div><div class="t m0 x7 h2 y3 ff1 fs0 fc0 sc0 ls0 ws0">　</div><div class="t m0 x3 h5 y5 ff3 fs3 fc0 sc0 ls0 ws0">Computer Measurement & Control 　</div><div class="t m0 x8 h6 y6 ff4 fs4 fc0 sc0 ls0 ws0">设计与应用</div><div class="t m0 x9 h2 y7 ff4 fs0 fc0 sc0 ls0 ws0">收稿日期 :2003 - 08 - 10</div><div class="t m0 x9 h2 y8 ff4 fs0 fc0 sc0 ls0 ws0">基金项目 :山西省自然科学基金资助项目</div><div class="t m0 x9 h2 y9 ff4 fs0 fc0 sc0 ls0 ws0">作者简介 :郭圣权</div><div class="t m0 xa h4 ya ff2 fs3 fc0 sc0 ls0 ws0">(</div><div class="t m0 xb h4 y9 ff2 fs3 fc0 sc0 ls0 ws0">1939 -</div><div class="t m0 xc h4 ya ff2 fs3 fc0 sc0 ls0 ws0">)</div><div class="t m0 xd h2 y9 ff2 fs3 fc0 sc0 ls0 ws0">,男 ,山西省定襄县人 ,教</div><div class="t m0 xe h2 yb ff1 fs0 fc0 sc0 ls0 ws0">授 ,主要从事导航、制导及语音识别在导航定位系统中的应</div><div class="t m0 xe h2 yc ff1 fs0 fc0 sc0 ls0 ws0">用。</div><div class="t m0 xe h2 yd ff4 fs0 fc0 sc0 ls0 ws0">文章编号 :1671 - 4598</div><div class="t m0 xf h4 ye ff2 fs3 fc0 sc0 ls0 ws0">(</div><div class="t m0 x10 h4 yd ff2 fs3 fc0 sc0 ls0 ws0">2004</div><div class="t m0 x11 h4 ye ff2 fs3 fc0 sc0 ls0 ws0">)</div><div class="t m0 x12 h2 yd ff2 fs3 fc0 sc0 ls0 ws0">05 - 0470 - 03 　　　　　　中图分类号 :TP31 　　　　　　文献标识码 :B</div><div class="t m0 x13 h7 yf ff3 fs5 fc0 sc0 ls0 ws0">MATLAB 环境下的基于 HMM 模型的语音识别系统</div><div class="t m0 x14 h8 y10 ff1 fs7 fc0 sc0 ls0 ws0">郭圣权 , 连晓峰</div><div class="t m0 x3 h4 y11 ff2 fs3 fc0 sc0 ls0 ws0">(</div><div class="t m0 x15 h2 y12 ff1 fs0 fc0 sc0 ls0 ws0">华北工学院 自动控制系 , 山西 太原 　030051</div><div class="t m0 x16 h4 y11 ff2 fs3 fc0 sc0 ls0 ws0">)</div><div class="t m0 x17 h2 y13 ff4 fs0 fc0 sc0 ls0 ws0">摘要 : 在 MATLAB 环境下利用语音工具箱 Voice Box 实现基于连续概率密度隐含马尔科夫模型的汉语语音识别系统。在</div><div class="t m0 x9 h2 y14 ff1 fs0 fc0 sc0 ls0 ws0">实时录音的情况下 , 利用该语音识别系统 , 不同的人对 20 条 2～8 个字的语音命令进行识别 , 准确率可达到 95 % , 识别时间</div><div class="t m0 x9 h9 y15 ff2 fs3 fc0 sc0 ls0 ws0">115～3 s , 实现了小词汇量连续语音的非特定人的实时识别。</div><div class="t m0 x17 h2 y16 ff4 fs0 fc0 sc0 ls0 ws0">关键词 : 语音识别 ; MATLAB ; 连续概率密度隐含马尔科夫模型</div><div class="t m0 x18 h4 y17 ff2 fs3 fc0 sc0 ls0 ws0">(</div><div class="t m0 x19 h4 y16 ff2 fs3 fc0 sc0 ls0 ws0">CDHMM</div><div class="t m0 x1a h4 y17 ff2 fs3 fc0 sc0 ls0 ws0">)</div><div class="t m0 x1b ha y18 ff3 fs8 fc0 sc0 ls0 ws0">Speech Recognition System Based on HMM in MATLAB Environment</div><div class="t m0 x1c hb y19 ff2 fs1 fc0 sc0 ls0 ws0">Guo Shengquan , Lian Xiaofeng</div><div class="t m0 x1d h4 y1a ff2 fs3 fc0 sc0 ls0 ws0">(</div><div class="t m0 x1e h2 y1b ff2 fs3 fc0 sc0 ls0 ws0">Dept. of Automatic Control , North China Institute of Technology , Taiyuan　030051 , China</div><div class="t m0 x1f h4 y1a ff2 fs3 fc0 sc0 ls0 ws0">)</div><div class="t m0 x17 h9 y1c ff3 fs3 fc0 sc0 ls0 ws0">Abstract : Using Voice - Box a mandarin speech recognition system is realized based on CDHMM in MATLAB environment. Under the cir2</div><div class="t m0 x9 h9 y1d ff2 fs3 fc0 sc0 ls0 ws0">cumstance of the real - time recording , different people to say the connected words which include 2～8 words in 20 speech utterances , the accura2</div><div class="t m0 x9 h9 y1e ff2 fs3 fc0 sc0 ls0 ws0">cy can reach 95 % , and the recognition time is between 115～3 seconds in this system.</div><div class="t m0 x20 h5 y1f ff3 fs3 fc0 sc0 ls0 ws0">Key words : speech recognition ; MATLAB ; Continuo’s Density Hidden Markov Model</div><div class="t m0 x21 h2 y20 ff1 fs0 fc0 sc0 ls0 ws0">图 1 　语音输入与端点检测界面</div><div class="t m0 xe hc y21 ff3 fs9 fc0 sc0 ls0 ws0">0 　引言</div><div class="t m0 x22 hd y22 ff1 fsb fc0 sc0 ls0 ws0">目前 , 用于语音识别的方法有基于动态时轴归整</div><div class="t m0 x23 hb y23 ff2 fs1 fc0 sc0 ls0 ws0">(</div><div class="t m0 x24 hb y24 ff2 fs1 fc0 sc0 ls0 ws0">DTW</div><div class="t m0 x25 hb y23 ff2 fs1 fc0 sc0 ls0 ws0">)</div><div class="t m0 x17 hd y24 ff1 fsb fc0 sc0 ls0 ws0">技术的模板匹配法、基于概率统计的 HMM 法和</div><div class="t m0 xe hd y25 ff1 fsb fc0 sc0 ls0 ws0">基于人工神经网络</div><div class="t m0 x26 hb y26 ff2 fs1 fc0 sc0 ls0 ws0">(</div><div class="t m0 x27 hb y25 ff2 fs1 fc0 sc0 ls0 ws0">ANN</div><div class="t m0 x28 hb y26 ff2 fs1 fc0 sc0 ls0 ws0">)</div><div class="t m0 x29 hd y25 ff1 fsb fc0 sc0 ls0 ws0">的最优搜索法。</div><div class="t m0 xe hc y27 ff3 fs9 fc0 sc0 ls0 ws0">1 　系统过程设计</div><div class="t m0 xe he y28 ff2 fs1 fc0 sc0 ls0 ws0">111 　实时录音和端点检测控件</div><div class="t m0 x2a hf y29 ff2 fsc fc0 sc0 ls0 ws0">[1 ]</div><div class="t m0 x22 hd y2a ff2 fs1 fc0 sc0 ls0 ws0">ActiveX 控件是 Microsoft 公司制订的一种软件接口标</div><div class="t m0 xe hd y2b ff1 fsb fc0 sc0 ls0 ws0">准 , 在 MATLAB 中 , 通常是将 ActiveX 控件嵌入到 figure</div><div class="t m0 xe hd y2c ff1 fsb fc0 sc0 ls0 ws0">窗口中 , 以 GUI 程序的方式使用。通过 ActiveX 录音控</div><div class="t m0 xe hd y2d ff1 fsb fc0 sc0 ls0 ws0">件可实现与 MATLAB 主程序的双向交互。当录音控件录</div><div class="t m0 xe hd y2e ff1 fsb fc0 sc0 ls0 ws0">音完成并进行端点检测后 , 向主程序发送一个事件 , 通</div><div class="t m0 xe hd y2f ff1 fsb fc0 sc0 ls0 ws0">知主程序读取语音数据及短时参数信息等. 端点检测算</div><div class="t m0 xe hd y30 ff1 fsb fc0 sc0 ls0 ws0">法主要采用短时平均能量和短时平均过零率作为判决的</div><div class="t m0 xe hd y31 ff1 fsb fc0 sc0 ls0 ws0">主要特征。</div><div class="t m0 x2b h10 y32 ff6 fsd fc0 sc0 ls0 ws0">e</div><div class="t m0 x13 h10 y33 ff6 fsd fc0 sc0 ls0 ws0">(</div><div class="t m0 x2c h10 y32 ff6 fsd fc0 sc0 ls0 ws0">i</div><div class="t m0 x20 h10 y33 ff6 fsd fc0 sc0 ls0 ws0">)</div><div class="t m0 x2d hd y32 ff2 fs1 fc0 sc0 ls0 ws0">= ∑</div><div class="t m0 x2e h11 y34 ff6 fse fc0 sc0 ls0 ws0">N</div><div class="t m0 x2f h12 y35 ff6 fse fc0 sc0 ls0 ws0">n = 1</div><div class="t m0 x30 hb y32 ff2 fs1 fc0 sc0 ls0 ws0">logx</div><div class="t m0 x31 hf y36 ff2 fsc fc0 sc0 ls0 ws0">2</div><div class="t m0 xf h13 y37 ff6 fs10 fc0 sc0 ls0 ws0">i</div><div class="t m0 x10 h10 y33 ff6 fsd fc0 sc0 ls0 ws0">(</div><div class="t m0 xb h10 y32 ff6 fsd fc0 sc0 ls0 ws0">n</div><div class="t m0 x32 h10 y33 ff6 fsd fc0 sc0 ls0 ws0">)</div><div class="t m0 x33 hd y32 ff1 fsb fc0 sc0 ls0 ws0">　计算第 i 帧的短时能量;</div><div class="t m0 x22 hb y38 ff2 fs1 fc0 sc0 ls0 ws0">zcr</div><div class="t m0 x2c h10 y39 ff6 fsd fc0 sc0 ls0 ws0">(</div><div class="t m0 x34 h10 y38 ff6 fsd fc0 sc0 ls0 ws0">i</div><div class="t m0 x35 h10 y39 ff6 fsd fc0 sc0 ls0 ws0">)</div><div class="t m0 x36 hd y38 ff2 fs1 fc0 sc0 ls0 ws0">= ∑</div><div class="t m0 x37 h12 y3a ff6 fse fc0 sc0 ls0 ws0">N - 1</div><div class="t m0 x37 h12 y3b ff6 fse fc0 sc0 ls0 ws0">n = 1</div><div class="t m0 x38 hb y38 ff2 fs1 fc0 sc0 ls0 ws0">| x</div><div class="t m0 x39 h13 y3c ff6 fs10 fc0 sc0 ls0 ws0">i</div><div class="t m0 x3a h10 y39 ff6 fsd fc0 sc0 ls0 ws0">(</div><div class="t m0 x27 h10 y38 ff6 fsd fc0 sc0 ls0 ws0">n</div><div class="t m0 x3b h10 y39 ff6 fsd fc0 sc0 ls0 ws0">)</div><div class="t m0 x3c hb y38 ff2 fs1 fc0 sc0 ls0 ws0">- x</div><div class="t m0 x3d h13 y3c ff6 fs10 fc0 sc0 ls0 ws0">i</div><div class="t m0 xd h10 y39 ff6 fsd fc0 sc0 ls0 ws0">(</div><div class="t m0 x3e hb y38 ff6 fsd fc0 sc0 ls0 ws0">n + 1</div><div class="t m0 x3f h10 y39 ff6 fsd fc0 sc0 ls0 ws0">)</div><div class="t m0 x40 hd y38 ff2 fs1 fc0 sc0 ls0 ws0">| 　计算第 i 帧的</div><div class="t m0 xe hd y3d ff1 fsb fc0 sc0 ls0 ws0">过零率</div><div class="t m0 x22 hd y3e ff1 fsb fc0 sc0 ls0 ws0">本系统采用 8 kHz 采样频率、16 bit 、单声道的</div><div class="t m0 xe hd y3f ff2 fs1 fc0 sc0 ls0 ws0">PCM 录音格式 , 帧长 N 为 30 ms , 帧移 M 为 10 ms ,</div><div class="t m0 x41 h10 y40 ff6 fsd fc0 sc0 ls0 ws0">x</div><div class="t m0 x42 h10 y41 ff6 fsd fc0 sc0 ls0 ws0">(</div><div class="t m0 x43 h10 y40 ff6 fsd fc0 sc0 ls0 ws0">n</div><div class="t m0 x44 h10 y41 ff6 fsd fc0 sc0 ls0 ws0">)</div><div class="t m0 x45 hd y40 ff1 fsb fc0 sc0 ls0 ws0">为输入的语音信号 , x 为第 i 帧的第 n 个样本。</div><div class="t m0 x46 hd y42 ff1 fsb fc0 sc0 ls0 ws0">系统界面如图 1 所示。</div><div class="t m0 x46 he y43 ff2 fs1 fc0 sc0 ls0 ws0">112 　语音信号特征参数提取</div><div class="t m0 x47 hf y44 ff6 fs10 fc0 sc0 ls0 ws0">[1 - 2]</div><div class="t m0 x48 hd y45 ff1 fsb fc0 sc0 ls0 ws0">　　语音信号特征参数的提取是语音识别的一个重要</div><div class="t m0 x46 hd y46 ff1 fsb fc0 sc0 ls0 ws0">环节。目前常用的方法是基于人的发音器官建立声道模</div><div class="t m0 x46 hd y47 ff1 fsb fc0 sc0 ls0 ws0">型和基于听觉器官建立听觉模型。基于听觉模型得到的</div><div class="t m0 x46 hd y48 ff2 fs1 fc0 sc0 ls0 ws0">MEL 倒 谱 系 数</div><div class="t m0 x16 hb y49 ff2 fs1 fc0 sc0 ls0 ws0">(</div><div class="t m0 x49 hb y48 ff2 fs1 fc0 sc0 ls0 ws0">Mel - Frequency Cepstrum Coefficients ,</div><div class="t m0 x46 hb y4a ff2 fs1 fc0 sc0 ls0 ws0">MFCC</div><div class="t m0 x4a hb y4b ff2 fs1 fc0 sc0 ls0 ws0">)</div><div class="t m0 x4b hd y4a ff1 fsb fc0 sc0 ls0 ws0">比基于声道模型得到的 LPC 倒谱系数更符合人</div><div class="t m0 x46 hd y4c ff1 fsb fc0 sc0 ls0 ws0">耳的听觉特性 , 在有信道噪声和频谱失真的情况下 , 能</div><div class="t m0 x46 hd y4d ff1 fsb fc0 sc0 ls0 ws0">产生更高的识别精度。</div><div class="t m0 x4c hd y4e ff2 fs1 fc0 sc0 ls0 ws0">MFCC 建立在 Fourier 频谱分析基础上 , 首先利用人</div><div class="t m0 x46 hd y4f ff1 fsb fc0 sc0 ls0 ws0">耳的感知特性 , 在语音的频谱范围内设置若干个带通滤</div><div class="t m0 x46 hd y50 ff1 fsb fc0 sc0 ls0 ws0">波器 , 每个滤波器具有三角形或正弦形滤波特性 , 然</div></div><div class="pi" data-data='{"ctm":[1.592040,0.000000,0.000000,1.592040,0.000000,0.000000]}'></div></div> </body> </html>