singleWord.rar_HMM_识别_词汇

clc; clear all; close all; fs=8000; Tw = 25; % analysis frame duration (ms) 帧长 Ts = 10; % analysis frame shift (ms) 帧移 alpha = 0.97; % preemphasis coefficient 语音增强 M = 20; % number of filterbank channels C = 12; % number of cepstral coefficients MFCC个数 L = 22; % cepstral sine lifter parameter LF = 300; % lower frequency limit (Hz) HF = 3700; % upper frequency limit (Hz) %训练用数据 [train_signals train_labels] = load_audio_from_folder('train');%读取语音文件、生成标签数据 %转MFCC参数 train_feature={}; for speech=train_signals [ MFCCs, FBEs, frames ] = mfcc( speech{1}, fs, Tw, Ts, alpha, @hamming, [LF HF], M, C+1, L ); train_feature(end+1,1)={MFCCs}; end; %测试用数据 [test_signals test_labels] = load_audio_from_folder('test');%读取语音文件、生成标签数据 %转MFCC参数 test_feature={}; for speech=test_signals [ MFCCs, FBEs, frames ] = mfcc( speech{1}, fs, Tw, Ts, alpha, @hamming, [LF HF], M, C+1, L ); test_feature(end+1,1)={MFCCs}; end; %% HMM模型建立和训练 d = C+1; %一个观察符号的维度 nstates = 5;%状态个数 nmix = 3; % 混合高斯分布个数 %训练apple的HMM pi0=[1 0 0 0 0]; %初始状态概率 trans0=[1/2 1/2 0 0 0 ; %转移概率 0 1/2 1/2 0 0 ; 0 0 1/2 1/2 0 ; 0 0 0 1/2 1/2 ; 0 0 0 0 1 ]; models={}; [unique_train_labels, ~, indices] = unique(train_labels);%重复去掉，只取唯一值 for i = 1:length(unique_train_labels) display(sprintf('modeling %s...', char(unique_train_labels(i)))) %发射概率初始值计算 stackedData = cell2mat(train_feature(indices == i)')';%训练用数据 mu = zeros(d, nmix);%均值 Sigma = zeros(d, d, nmix);%方差 for k = 1:nmix XX = stackedData + randn(size(stackedData)); mu(:, k) = colvec(mean(XX));%所有训练数据的均值 Sigma(:, :, k) = cov(XX);%所有训练数据的方差 end M = normalize(rand(nstates, nmix), 1);%混合高斯系数 emission = condMixGaussTiedCpdCreate(mu, Sigma, M);%发射状态描述结构体获得 %训练HMM %verbose：是否打印信息 %nRandomRestarts:随机训练几次 %maxiter：最大迭代次数 %nmix：混合高斯个数 %pi0：HMM初始状态概率 %trans0：HMM初始转移概率 %emission0：初始发射状态描述 %piPrior：防止除数为0的情况出现 a/b=a+piPrior/b+piPrior 变成这种形式 %transPrior：防止除数为0的情况出现 a/b=a+piPrior/b+piPrior 变成这种形式 models(end+1,1) = {hmmFit(train_feature(indices == i), nstates, 'mixGaussTied', 'verbose', false, ... 'nRandomRestarts', 3, 'maxiter', 50, 'nmix', nmix,... 'pi0',pi0,'trans0',trans0,'emission0',emission, ... 'piPrior',pi0,'transPrior',trans0.*10)}; end; %训练样本识别率计算 errorcount=0; for j=1:length(train_feature) p=zeros(length(unique_train_labels),1); for i = 1:length(unique_train_labels) p(i)=hmmLogprob(models{i}, train_feature(j));%计算概率值 end; [~, i]=max(p);%取最大概率的模型作为识别 %display(sprintf('"%s" is recognized as "%s"', train_labels{j},char(unique_train_labels(i)))) if ~strcmp(train_labels{j},char(unique_train_labels(i))) errorcount=errorcount+1;%错误累计 end; end; display(sprintf('train accuracy is %0.2f', (length(train_feature)-errorcount)*100/length(train_feature))); %% 测试样本识别率计算 errorcount=0; for j=1:length(test_feature) p=zeros(length(unique_train_labels),1); for i = 1:length(unique_train_labels) p(i)=hmmLogprob(models{i}, test_feature(j));%计算概率值 end; [~, i]=max(p);%取最大概率的模型作为识别 %display(sprintf('"%s" is recognized as "%s"', test_labels{j},char(unique_train_labels(i)))) if ~strcmp(test_labels{j},char(unique_train_labels(i))) errorcount=errorcount+1;%错误累计 end; end; display(sprintf('test accuracy is %0.2f', (length(test_labels)-errorcount)*100/length(test_labels)));