聲音識(shí)別_聲音分類_算法實(shí)現(xiàn)

1、模型框架生態(tài)環(huán)境聲音信號(hào)預(yù)處理訓(xùn)練樣本測(cè)試樣本特征提取特征提取高斯混合模型判決規(guī)則投票識(shí)別結(jié)果 訓(xùn)練識(shí)別前端處理圖圖1基基于于GMM的的生生態(tài)態(tài)環(huán)環(huán)境境聲聲音音識(shí)識(shí)別別系系統(tǒng)統(tǒng) 設(shè)計(jì)模塊 預(yù)處理 使用cooledit轉(zhuǎn)換格式/聲道，或用Matlab處理 特征提取 訓(xùn)練數(shù)據(jù)的特征提取 測(cè)試數(shù)據(jù)的特征提取 不同的特征提取可以分開或合在一起設(shè)計(jì) 分類算法設(shè)計(jì) GMM，HMM，SVM，KNN,ANN. 測(cè)試和決策特征提取%讀取聲音文件 wav= wavread(data.filenamesi_files,1 307712); eg. 200個(gè)文件 i_files= 1:200 307712是提取的樣本

2、數(shù)（限制讀取的長(zhǎng)度，約28s）% 要提取的MFCC系數(shù)個(gè)數(shù)num_ceps_coeffs = 20;c.fs = 11025; %采樣頻率% 設(shè)置每幀大?。ò瑯颖緮?shù)）c.seg_size = 1024; c.hop_size = 512; % c.seg_size-交疊部分=c.hop_size 特征提?。∕FCC）% 幀數(shù)計(jì)算num_segments = floor(length(wav)-c.seg_size)/c.hop_size)+1; % 初始化功率譜矩陣P = zeros(c.seg_size/2+1,num_segments); % 設(shè)置窗函數(shù)c.w = 0.5*(1-cos(

3、2*pi*(0:c.seg_size-1)/(c.seg_size-1);%漢寧窗函數(shù) % 逐幀做FFT for i_p = 1:num_segments, idx = (1:c.seg_size)+(i_p-1)*c.hop_size; x = abs(fft(wav(idx).*c.w)/sum(c.w)*2).2; P(:,i_p) = x(1:end/2+1);%end c.num_filt = 36; % Mel頻帶數(shù)f = linspace(0,c.fs/2,c.seg_size/2+1);%初始平均劃分fmel = log(1+f/700)*1127.01048; %1127.0

4、1048=2595/log10 ,Matlab中l(wèi)og=lnmel_idx = linspace(0,mel(end),c.num_filt+2);%初始平均劃分mel（38個(gè)點(diǎn)）f_idx = zeros(c.num_filt+2,1);for i=1:c.num_filt+2,% f_idx(i)存的是mel中與mel_idx(i)最近的元素的地址 tmp f_idx(i) = min(abs(mel - mel_idx(i);%近似的平均劃分 endfreqs = f(f_idx);h = 2./(freqs(3:c.num_filt+2)-freqs(1:c.num_filt);%三角

5、的高度c.mel_filter = zeros(c.num_filt,c.seg_size/2+1);for i=1:c.num_filt, c.mel_filter(i,:) =(f freqs(i) & f freqs(i+1) & f freqs(i+2).* . h(i).*(freqs(i+2)-f)/(freqs(i+2)-freqs(i+1);end特征提取M = zeros(c.num_filt,num_segments); %初始化 for i_m = 1:num_segments, M(:,i_m) = c.mel_filter*P(:,i_m);% 通過三角濾波器end%

6、 做對(duì)數(shù)變換M(M1)=1; M = 10*log10(M);%DCT函數(shù)c.DCT = 1/sqrt(c.num_filt/2) * .cos(0:num_ceps_coeffs-1)*(0.5:c.num_filt)*pi/c.num_filt);c.DCT(1,:) = c.DCT(1,:)*sqrt(2)/2;%離散余弦變換mfcc= c.DCT * M;分類算法設(shè)計(jì)（GMM示例）ncentres = 16;% 高斯分量個(gè)數(shù)input_dim = 16; %特征維數(shù)% 設(shè)置混合模型mix = gmm(input_dim, ncentres, diag);% 特征數(shù)據(jù)輸入siz=600;features = zeros(siz,input_dim); for k=1:siz for j=1:input_dim features(k,j)=data.feat.mfcc(i_files,j,k); endend% 初始化模型參數(shù)mix = gmminit(mix, features, options);option