离线版-端点检测代码重写
根据对双门限的理解和修改，重写了一个离线版端点检测。
function [st,en]=VAD(x, fs)
x=double(x);
x=x/max(abs(x));
framelen= floor(fs*40/1000);%
frameinc= floor(fs*10/1000);%
y=enframe(x,framelen,frameinc);
%计算短时间能量
amp=sum(abs(y),2);
%开始端点检测
tmp1=enframe(x(1:length(x)-1),framelen,frameinc);
tmp2=enframe(x(2:length(x)),framelen,frameinc);
signs=(tmp1.*tmp2)<0; %
diffs=(tmp1-tmp2)>0.01; zcr=sum(signs.*diffs,2); % zcr=[zcr;zcr(end)]; zcr_yu=0.2*mean(zcr);%
yuzhi=0.2*mean(amp);% minspeech=10; count=0;% start=[]; tail=[]; N=length(amp); flag=0;% biaozhi2=0; biaozhi3=0; %%%%%%%%%%%% for n=1:N if amp(n)<yuzhi || (zcr(n)<zcr_yu) continue; end kaitou=n; break; end for n=N:-1:1 if amp(n)<yuzhi || (zcr(n)<zcr_yu) continue; end jiewei=n; break; end noise=[amp(1:kaitou-1);amp(jiewei+1:end)]; noise_mean=mean(noise); noise_var = std(noise); speech_mean=mean(amp(kaitou:jiewei));%noise_mean yuzhi1= 0.3*speech_mean; % yuzhi2= max(0.3*speech_mean , (noise_mean + noise_var)*1.3);% noise_zcr=[zcr(1:kaitou-1);zcr(jiewei+1:end)]; noise_zcr_mean = mean(noise_zcr); noise_zcr_std = std(noise_zcr); speech_zcr_mean=mean(zcr(kaitou:jiewei)); zcr_yu1 = 0.3*speech_zcr_mean; zcr_yu2 = max(0.3*speech_zcr_mean , (noise_zcr_mean+noise_zcr_std)*0.3); %%%%%%%%%%%%%% st = []; en = []; bstart_state = 0; bend_state = 0; segment = 0; unvoice = 0; voice_min_len = 7;% 最短语音长度70ms unvoice_min_len = 5;%结束段最小持续50ms st_candicate = 0; en_candicate = 0; for i = 1:N if( (amp(i) >= yuzhi2 && zcr(i) >= zcr_yu1) && ~bstart_state ) %find start bstart_state = 1; % if(~st_candicate) st_candicate = i; end segment = segment + 1; elseif( (amp(i) >= yuzhi2 || zcr(i) >= zcr_yu1) && bstart_state )% if(unvoice >= unvoice_min_len) % st = [st; st_candicate]; en_candicate = en_candicate + unvoice_min_len - 1; en = [en; en_candicate];
bstart_state = 0; bend_state = 1; unvoice = 0; segment = 0; st_candicate = 0; else % unvoice = 0; % bend_state = 0; segment = segment + 1; end elseif( (amp(i) < yuzhi2 && zcr(i) < zcr_yu1) && bstart_state ) % if segment >= voice_min_len % unvoice = unvoice + 1; % if ~bend_state % en_candicate = i; end bend_state = 1; % else % bstart_state = 0; % segment = 0; bend_state = 0; unvoice = 0; st_candicate = 0; %prepare_for_start = 0; en_candicate = 0; end elseif((amp(i) >= yuzhi2 || zcr(i) >= zcr_yu1) && ~bstart_state) % if ~ st_candicate st_candicate = i; end else%
st_candicate = 0; continue; end end if(unvoice >= unvoice_min_len )%
st = [st; st_candicate]; en = [en; (en_candicate + unvoice_min_len -1)]; segment = 0; end if( segment >= voice_min_len) %
st = [st; st_candicate]; en = [en; N]; end figure(1);
subplot(3,1,1) plot(x); %原始语音波形 axis([1,length(x),-1,1]) ylabel('speech'); xlabel('样本点'); for k=1:length(st) line([st(k)*frameinc,st(k)*frameinc],[-1,1],'linestyle',':','color','blue','LineWidth',2); line([en(k)*frameinc,en(k)*frameinc],[-1,1],'linestyle',':','color','red','LineWidth',2); end subplot(3,1,2) plot(amp);%原始语音能量 axis([1,length(amp),0,max(amp)]) ylabel('energy'); xlabel('帧数'); line([1,N],[yuzhi1,yuzhi1],'color','yellow','LineWidth',2);%由语音能量 line([1,N],[yuzhi2,yuzhi2],'color','red','LineWidth',2);%由噪声平均能量和语音能量比较而得 for k=1:length(st) line([st(k),st(k)],[min(amp),max(amp)],'linestyle',':','color','blue','LineWidth',2); line([en(k),en(k)],[min(amp),max(amp)],'linestyle',':','color','red','LineWidth',2); end subplot(3,1,3) plot(zcr);%原始语音过零率 axis([1,length(zcr),0,max(zcr)]) ylabel('zcr'); xlabel('帧数'); line([1,N],[zcr_yu1,zcr_yu1],'color','yellow','LineWidth',2);%由语音能量 line([1,N],[zcr_yu2,zcr_yu2],'color','red','LineWidth',2);%语音加噪音 for k=1:length(st) line([st(k),st(k)],[min(zcr),max(zcr)],'linestyle',':','color','blue','LineWidth',2); line([en(k),en(k)],[min(zcr),max(zcr)],'linestyle',':','color','red','LineWidth',2); end end
用语音识别的强制对齐自动切分连续语音汉字，工程量和资源巨大，此处做了个简易版的汉字自动切分，效果还可以。