[Speech recognition] Isolated word speech recognition based on matlab dynamic time warping (DTW) [including Matlab source code 573]

1. Introduction

The birth of Dynamic Time Warping (DTW) has a certain history (proposed by Japanese scholar Itakura), and its purpose is relatively simple. It is a method to measure the similarity of two time series with different lengths. The application is also relatively wide, mainly in template matching, such as isolated word speech recognition (recognizing whether two speeches represent the same word), gesture recognition, data mining and information retrieval, etc.

1 Overview

In most disciplines, time series are a common representation of data. For time series processing, a common task is to compare the similarity of two series.
In the time series, the lengths of the two time series that need to be compared for similarity may not be the same. In the field of speech recognition, the speech speed of different people is different. Because the voice signal has considerable randomness, even if the same person makes the same tone at different times, it may not have the full length of time. Moreover, the pronunciation speed of different phonemes in the same word is also different. For example, some people will drag the sound of "A" very long, or pronounce "i" very short. In these complex situations, the distance (or similarity) between two time series cannot be effectively obtained using the traditional Euclidean distance.
2 Principle of DTW method

In the time series, the lengths of the two time series that need to be compared for similarity may not be the same. In the field of speech recognition, the speech speed of different people is different. Moreover, the pronunciation speed of different phonemes in the same word is also different. For example, some people will drag the sound of "A" very long, or pronounce "i" very short. In addition, different time series may only have a displacement on the time axis, that is, in the case of a reduction displacement, the two time series are the same. In these complex situations, the distance (or similarity) between two time series cannot be effectively obtained using the traditional Euclidean distance.

DTW calculates the similarity between the two time series by extending and shortening the time series:
Insert picture description here
as shown in the figure above, the upper and lower solid lines represent the two time series, and the dashed line between the time series represents the two time series. Similarities between. DTW uses the sum of the distances between all these similar points, called Warp Path Distance, to measure the similarity between two time series.

2 DTW calculation method:

Let the two time series to be calculated for similarity be X and Y, and the lengths are |X| and |Y|.
Warp Path The Warp Path has
the form of W=w1,w2,…,wK, where Max(|X|,|Y|)<=K<=|X|+|Y|.
The form of wk is (i, j), where i represents the i coordinate in X, and j represents the j coordinate in Y.
The reorganization path W must start with w1=(1,1) and end with wK=(|X|,|Y|) to ensure that each coordinate in X and Y appears in W.
In addition, the i and j of w(i,j) in W must increase monotonically to ensure that the dashed lines in Figure 1 do not intersect. The so-called monotonic increase refers to the
Insert picture description here

cost matrix (Cost Matrix) D, D( i, j) represents the normalized path distance between two time series of length i and j.

Second, the source code

function trimmed_X = my_vad(x)
%端点检测；输入为录入语音，输出为有用信号

Ini = 0.1;          %初始静默时间
Ts = 0.01;          %窗的时长
Tsh = 0.005;        %帧移时长
Fs = 16000;         %采样频率
counter1 = 0;       %以下四个参数用来寻找起始点和结束点
counter2 = 0;
counter3 = 0;
counter4 = 0;
ZCRCountf = 0;      %用于存储过零率检测结果
ZCRCountb = 0;     
ZTh = 40;           %过零阈值
w_sam = fix(Ts*Fs);                   %窗口长度
o_sam = fix(Tsh*Fs);                  %帧移长度
lengthX = length(x);
segs = fix((lengthX-w_sam)/o_sam)+1;  %分帧数
sil = fix((Ini-Ts)/Tsh)+1;            %静默时间帧数
win = hamming(w_sam);

Limit = o_sam*(segs-1)+1;             %最后一帧的起始位置

FrmIndex = 1:o_sam:Limit;             %每一帧的起始位置
ZCR_Vector = zeros(1,segs);           %记录每一帧的过零点数
                                     
%短时过零点
for t = 1:segs
    ZCRCounter = 0; 
    nextIndex = (t-1)*o_sam+1;
    for r = nextIndex+1:(nextIndex+w_sam-1)
        if (x(r) >= 0) && (x(r-1) >= 0)
         
        elseif (x(r) > 0) && (x(r-1) < 0)
         ZCRCounter = ZCRCounter + 1;
        elseif (x(r) < 0) && (x(r-1) < 0)
         
        elseif (x(r) < 0) && (x(r-1) > 0)
         ZCRCounter = ZCRCounter + 1;
        end
    end
    ZCR_Vector(t) = ZCRCounter;
end

%短时平均幅度
Erg_Vector = zeros(1,segs);
for u = 1:segs
    nextIndex = (u-1)*o_sam+1;
    Energy = x(nextIndex:nextIndex+w_sam-1).*win;
    Erg_Vector(u) = sum(abs(Energy));
end

IMN = mean(Erg_Vector(1:sil));  %静默能量均值（噪声均值）
IMX = max(Erg_Vector);          %短时平均幅度的最大值
I1 = 0.03 * (IMX-IMN) + IMN;    %I1，I2为初始能量阈值
I2 = 4 * IMN;
ITL = 100*min(I1,I2);            %能量阈值下限,前面系数根据实际情况更改得到合适结果
ITU = 10* ITL;                  %能量阈值上限
IZC = mean(ZCR_Vector(1:sil));  
stdev = std(ZCR_Vector(1:sil)); %静默阶段过零率标准差

IZCT = min(ZTh,IZC+2*stdev);    %过零率阈值
indexi = zeros(1,lengthX);      
indexj = indexi;               
indexk = indexi;
indexl = indexi;

%搜寻超过能量阈值上限的部分
for i = 1:length(Erg_Vector)
    if (Erg_Vector(i) > ITU)
        counter1 = counter1 + 1;
        indexi(counter1) = i;
    end
end
ITUs = indexi(1);        %第一个能量超过阈值上限的帧

%搜寻能量超过能量下限的部分
for j = ITUs:-1:1
    if (Erg_Vector(j) < ITL)
        counter2 = counter2 + 1;
        indexj(counter2) = j;
    end
end
start = indexj(1)+1;    %第一级判决起始帧

Erg_Vectorf = fliplr(Erg_Vector);%将能量矩阵关于中心左右对称，如果是一行向量相当于逆序 

%重复上面过程相当于找结束帧
for k = 1:length(Erg_Vectorf)
    if (Erg_Vectorf(k) > ITU)
        counter3 = counter3 + 1;
        indexk(counter3) = k;
    end
end
%初始化DTW判别矩阵
Scores1 = zeros(1,N);                
Scores2 = zeros(1,N);
Scores3 = zeros(1,N);


%加载模板数据
s1 = load('Vectors1.mat');
fMatrixall1 = struct2cell(s1);
s2 = load('Vectors2.mat');
fMatrixall2 = struct2cell(s2);
s3 = load('Vectors3.mat');
fMatrixall3 = struct2cell(s3);


%计算DTW
for i = 1:N
    fMatrix1 = fMatrixall1{
    
    i,1};
    fMatrix1 = CMN(fMatrix1);
    Scores1(i) = myDTW(fMatrix1,rMatrix);
end

for j = 1:N
    fMatrix2 = fMatrixall2{
    
    j,1};
    fMatrix2 = CMN(fMatrix2);
    Scores2(j) = myDTW(fMatrix2,rMatrix);
end

Three, running results

Insert picture description here

Four, remarks