[Emotion recognition] Speech emotion recognition based on matlab K nearest neighbor classification algorithm [Include Matlab source code 542]

1. Introduction

The K-Nearest Neighbor (KNN) classification algorithm is a theoretically mature method and one of the simplest machine learning algorithms. The idea of this method is: if most of the k most similar (ie, the nearest neighbors in the feature space) samples of a sample in the feature space belong to a certain category, the sample also belongs to this category.
1 Definition.
If most of the k most similar (ie nearest neighbors in the feature space) samples of a sample in the feature space belong to a certain category, then the sample also belongs to this category, which is inferred by your "neighbors" Out of your category.
2 Distance formula
The distance between two samples can be calculated by the following formula, also called Euclidean distance
Insert picture description here

3 Steps of the KNN algorithm
(1) Calculate the distance between each point in the known category data set and the current point;
(2) Select the K points with the smallest distance from the current point;
(3) Calculate the statistics of each category in the first K points The frequency of the sample;
(4) Return the category with the highest frequency of the previous K points as the predicted classification of the current point.

4 KNN principle
Insert picture description here

5 KNN advantages and disadvantages

6 KNN performance problems
NN performance problems are also one of the shortcomings of KNN. Using KNN, it is easy to construct a model, but when classifying samples to be classified, in order to obtain K nearest neighbors, a brute force search method must be used to scan all training samples and calculate the distance between them and the samples to be classified. The system overhead is very high. Big.

Second, the source code

clc
clear all;
close all;

%% 载入各情感的特征向量矩阵
load A_fear.mat;
load F_happiness.mat;
load N_neutral.mat;
load T_sadness.mat;
load W_anger.mat;
NumberOfTrain=size(fearVec,2)/2; %一半测试用，一半训练用
trainVector=[fearVec(:,1:NumberOfTrain),hapVec(:,1:NumberOfTrain),neutralVec(:,1:NumberOfTrain),sadnessVec(:,1:NumberOfTrain),angerVec(:,1:NumberOfTrain)]; % 构建训练样本集
testVector=[fearVec(:,(NumberOfTrain+1):size(fearVec,2)),hapVec(:,(NumberOfTrain+1):size(hapVec,2)),neutralVec(:,(NumberOfTrain+1):size(neutralVec,2)),sadnessVec(:,(NumberOfTrain+1):size(sadnessVec,2)),angerVec(:,(NumberOfTrain+1):size(angerVec,2))]; % 构建测试样本集
k=9; %k 最近邻
distanceMatrix=zeros(size(trainVector,2),size(testVector,2)); % 每一列表示某个测试语音与所有训练集样本的距离
%% 计算每个测试样本和训练样本集各样本的距离
for i=1:size(testVector,2)
    for j=1:size(trainVector,2)
        distanceMatrix(j,i)=norm(testVector(:,i)-trainVector(:,j)); %计算欧氏距离
    end
end
%% 统计分类结果 （根据相应的特征向量在数组trainVector或testVector中所处的位置来辨别类型）
totalTestNumber=size(fearVec,2)-NumberOfTrain;
emtionCounter=zeros(1,5);
n1=NumberOfTrain;
n2=n1+NumberOfTrain;
n3=n2+NumberOfTrain;
n4=n3+NumberOfTrain;
n5=n4+NumberOfTrain;
p1=size(fearVec,2)-NumberOfTrain;
p2=p1+size(hapVec,2)-NumberOfTrain;
p3=p2+size(neutralVec,2)-NumberOfTrain;
p4=p3+size(sadnessVec,2)-NumberOfTrain;
p5=p4+size(angerVec,2)-NumberOfTrain;
if(n5~=size(trainVector,2)||p5~=size(testVector,2))
    disp('data error')
    return;
end

for i=1:size(distanceMatrix,2)
    flag=zeros(1,5);
    [sortVec,index]=sort(distanceMatrix(:,i));
    % 统计K个近邻中各类别的数量
    for j=1:k
        if(n1>=index(j)&&index(j)>=1)
            flag(1)=flag(1)+1;
        elseif(n2>=index(j)&&index(j)>n1)
            flag(2)=flag(2)+1;
        elseif(n3>=index(j)&&index(j)>n2)
            flag(3)=flag(3)+1;
        elseif(n4>=index(j)&&index(j)>n3)
            flag(4)=flag(4)+1;
        else
            flag(5)=flag(5)+1;
        end
    end
    [~,index1]=sort(flag);
    % 如果K个近邻中数量最多的类别与该样本实际的类别一致，则认为算法识别正确，相应counter加一。
    if((p1>=i&&i>=1)&&index1(5)==1)
        emtionCounter(index1(5))=emtionCounter(index1(5))+1;
       
    elseif((p2>=i&&i>p1)&&index1(5)==2)
        emtionCounter(index1(5))=emtionCounter(index1(5))+1;
        
    elseif((p3>=i&&i>p2)&&index1(5)==3)
        emtionCounter(index1(5))=emtionCounter(index1(5))+1;
        
    elseif((p4>=i&&i>p3)&&index1(5)==4)
        emtionCounter(index1(5))=emtionCounter(index1(5))+1;
       
    elseif((p5>=i&&i>p4)&&index1(5)==5)
        emtionCounter(index1(5))=emtionCounter(index1(5))+1;
       
    end

end
function feature=featvector(filename)
[y,fs]=wavread(filename); 
L=length(y);
ys=y;
for i=1:(length(y)-1)
    if (abs(y(i))<1e-3)  %  剔除较小值，计算短时能量时使用  %
        ys(i)=ys(i+1);
        L=L-1;
    end
end 
y1=ys(1:L);
s=enframe(y,hamming(256),128); %  分帧加窗  %
s1=enframe(y1,hamming(256),128); 
[nframe,framesize]=size(s);  
[nframe1,framesize1]=size(s1);
E=zeros(1,nframe1);  
Z=zeros(1,nframe);
F=zeros(1,nframe);
for i=1:nframe
    Z(i)=sum(abs(sign(s(i,framesize:2)-s(i,framesize-1:1))))/2;  %  过零率  %
end
for i=1:nframe1
    E(i)=sum(s1(i,:).*s1(i,:)); %  短时能量  %
end
s1=x2-x1;s2=x3-x4;
E_Reg_coff=s1/s2;
x=0;
for i=1:nframe1
    t=E(i)-(mean(E)-s1/s2*x4/nframe1)-s1/s2*i;
    x=x+t^2/nframe1;
end
E_Sqr_Err=x;
feature(1:7,1)=[max(E);min(E);mean(E);var(E);E_shimmer;E_Reg_coff;E_Sqr_Err];%  短时能量相关特征  %

%  能量比  %
feature(8,1)=Eratio;

end

Three, running results

Insert picture description here

Four, remarks