myvoicemark.m

一个关于端点检测的程序,大家学习用,很有用的哦~

%%%  voice - unvoice speech part decision 
%%%  2004-3jingjing:端点检测程序使,适应于多种采样率

function [T]=myvoicemark(x,Fs)


FrameLength=round(Fs*0.0025);   %%2.5ms, Fs=16000时为40 samples
energy_threshold=0.01;
zero_threshold=3.5;

N=length(x);

%% cross zero rate

Len=round(20*Fs/16000);%过零率分析单元

for n=1:N-Len
   cross=0;
   if(x(n)>0) 
      flag=0;
   else 
      flag=1;
   end
   for m=n+1:n+Len                 %求出短时（Len＝20）的过零率
      if(flag==0&x(m)<=0)  
         cross=cross+1;
         flag=1;
      elseif(flag==1&x(m)>0)
         cross=cross+1;
         flag=0;
      end
   end
   crslt(n)=cross;                 %共求出N-Len帧短时过零率，存于crslt中 
end

%% cross rate smoothing

C=fft(crslt);			   %过零率波形的谱,形状不平滑
N2=length(C);                      %N2=N-Len  

C1=zeros(1,N2);
lp=floor(N2/200);

%去掉过零率波形中的高频，即fft谱中间置零
C1(1:lp)=C(1:lp);                  %将C的前面一段复制到C1中
C1(N2-lp:N2)=C(N2-lp:N2);          %将C的后面一段复制到C1中  

C2=ifft(C1);                       %C1的中间一大段都为0，应该是为了平滑
C2=real(C2);

%% mark the voice part of speech by cross-zero-rate

C3=zeros(1,N2);

for n=1:N2
   if(C2(n)<zero_threshold)
      C3(n)=1;                     %C3中保存的为判断是否超过零率门限的标志，超过置0，未超过置1
   end
end
    

%%energy

for n=1:N-2*FrameLength            %共有N-2*FrameLength帧
   E(n)=0;
   for m=n:n+2*FrameLength         %帧长2*FrameLength+1 
      E(n)=E(n)+(x(m).^2);
   end
end

%% energy smoothing

E1=fft(E);
N1=length(E1) ;                     %N1=N-3*FrameLength

lp2=floor(N1/200);

E2=zeros(1,N1);
E2(1:lp2)=E1(1:lp2);
E2(N1-lp2:N1)=E1(N1-lp2:N1);
E3=ifft(E2);                       %平滑算法同过零率的平滑

%% energy normalization 
E3=real(E3);
E3=E3./max(E3);


%% mark the voice part of speech by energy
%% be sure N1<N2

E4=zeros(1,N2);

for n=1:N1
   if(E3(n)>energy_threshold)
      E4(n)=1;                     %E4中保存的为判断是否超能量门限的标志，超过置1，未超过置0（对前N1个元素）
   end
end

v=E4&C3;   %% both zero cross rate and energy decision   

if(v(1)==1)%v(1)=1则起始点为浊音点
   flag=1;                          %v(1)同时超能量门限和不超过零率门限则flag置1，否则置0
else
   flag=0;
end
cross2=0;%cross2记录v总共的过零次数

for n=2:N2
   if(flag==1&v(n)==0)
      flag=0;
      cross2=cross2+1;
      pt(cross2)=n;
   end
   if(flag==0&v(n)==1)
      flag=1;
      cross2=cross2+1;
      pt(cross2)=n;
   end
end                              %pt保存的是cross2分别为1、2……最开始的n值


T=[];

if(v(1)==1&(mod(cross2,2)==0))   %% High, even 
   b=1;
   e=pt(1);                      %e为cross2等于1时的序号  
   if((e-b)>8*FrameLength)
      T=[T [b,e]'];
   end
   pt(cross2+1)=N2;              %% add a psudo point 加个伪点：序号为N2时等于cross2+1（因为是偶数）
   for n=2:2:cross2
      b=pt(n);                   %b为cross2等于n时的序号 
      e=pt(n+1);                 %b为cross2等于n+1时的序号
      if((e-b)>8*FrameLength)
         T=[T [b,e]'];
      end
   end
   
elseif(v(1)==1&(mod(cross2,2)==1))  %% High, odd
   b=1;
   e=pt(1);
   if((e-b)>8*FrameLength)
      T=[T [b,e]'];
   end
   for n=2:2:cross2-1
      b=pt(n);
      e=pt(n+1);
      if((e-b)>8*FrameLength)
         T=[T [b,e]'];
      end
   end
   
elseif(v(1)==0&(mod(cross2,2)==0))  %% Low, even
   for n=1:2:cross2-1
      b=pt(n);%起点
      e=pt(n+1);%终点
      if((e-b)>8*FrameLength)%8*FrameLength在Fs=16000时即为20ms帧长
            T=[T [b,e]'];
      end
   end
elseif(v(1)==0&(mod(cross2,2)==1))  %% Low, odd
   pt(cross2+1)=N2;
   for n=1:2:cross2
      b=pt(n);
      e=pt(n+1);
      if((e-b)>8*FrameLength)
         T=[T [b,e]'];
      end
   end
end

%判断两组分割点的间隔
[M,N]=size(T);
flag=0;%表示最后一列未被存储过
if N>1
    T_temp=[];
    for n=1:N-1
        if T(1,n+1)-T(2,n)>8*FrameLength
            T_temp=[T_temp,T(:,n)];
            flag=0;
        else
            T_temp=[T_temp,[T(1,n),T(2,n+1)]'];
            flag=1;
        end
    end
    if flag==0
        T_temp=[T_temp,T(:,n+1)];
    end
    T=T_temp;
end
%T                    %每一列应该是标注语音的始末点