[Voice recognition] based on matlab VQ specific person isolated words speech recognition [including Matlab source code 536]

1. Introduction

VQ (Vector Quantization) is a commonly used compression technology. This article mainly reviews:

1) VQ principle

2) VQ-based speaker recognition (SR, speaker recognition) technology
〇. Classification problem

Speaker recognition is actually a classification problem:
Insert picture description here
Speaker recognition technology mainly includes these methods:

Template matching methods
These methods are relatively mature, and the main principles are: feature extraction, template training, and matching. Typical ones are: dynamic time warping DTW, vector quantization VQ, etc.

DTW uses the idea of dynamic programming, but it also has shortcomings: 1) Too much reliance on VAD technology; 2) It does not make full use of the timing dynamic characteristics of speech, so it is easy to understand that it is replaced by HMM.

The VQ algorithm is a method of data compression. Codebook resume and codeword search are two basic problems. Codebook resume is to train a better codebook from a large number of signal samples. Codeword search is to find a codeword that best matches the input. The method is simple and suitable for small System and sound with obvious differences are more suitable.

Classification methods based on statistical models.
The nature of this type of method is still a pattern recognition system. It needs to extract features, then train the classifier, and finally classify decision-making. Typical framework:
Insert picture description here
Commonly used models are: GMM, HMM, SVM, ANN, DNN or various Joint model, etc.

GMM basic framework:
Insert picture description here
similar to the GMM-UBM (Universal background model) algorithm. The difference from GMM is that it trains a large GMM on the overall sample of the L class, instead of training a GMM model for each class like GMM. In the case of SVM, MFCC is used as a feature, and each frame is used as a sample. VAD can be used to delete invalid audio segments and directly train the classification. In recent years, there have also been methods of using sparse expression:
Insert picture description here
2 Principle
Vector Quantization This technology is widely used in signal processing and data compression and other fields. In fact, there is a VQ step in multimedia compression formats such as JPEG and MPEG-4.

The name Vector Quantization sounds a bit mysterious, but in fact it is not so sophisticated. Everyone knows that analog signals are continuous values, and computers can only process discrete digital signals. When converting analog signals into digital signals, we can replace an interval with a certain value in the interval, for example, [ All values on 0, 1) become 0, all values on [1, 2) become 1, and so on. This is a VQ process. A more formal definition is: VQ is a process of encoding a point in a vector space with a finite subset of it.

A typical example is the encoding of images. In the simplest case, consider a grayscale image, 0 is black, 1 is white, and the value of each pixel is a real number on [0, 1]. Now to encode it as a 256-level grayscale image, one of the simplest methods is to map each pixel value x to an integer floor(x 255). Of course, the original data space does not necessarily have to be continuous. For example, if you want to compress this picture, each pixel uses only 4 bits (instead of the original 8 bits) to store. Therefore, you need to use [0, 15] for the integer values in the original [0, 255] interval. ] On the integer value to encode, a simple mapping scheme is x 15/255.

However, this mapping scheme is quite Naive. Although it can reduce the number of colors to achieve a compression effect, if the original colors are not evenly distributed, the quality of the resulting picture may not be very good. For example, if a 256-level grayscale image is completely composed of two colors, 0 and 13, then a completely black image will be obtained through the above mapping, because both colors are mapped to 0. A better approach is to combine clustering to select representative points.

The actual method is: treat each pixel as a piece of data, run K-means to get k centroids, and then use the pixel values of these centroids to replace the pixel values of all points in the corresponding cluster. For color pictures, the same method can be used. For example, in RGB three-color pictures, each pixel is regarded as a point in a 3-dimensional vector space.

Second, the source code

17 / 18
% Demo script that generates all graphics in the report and demonstrates our results.
[s6 fs6] = wavread('s6.wav');
[s1 fs1] = wavread('s1.wav');
%Question 2
disp('> Question 2：画出原始语音波形');
t = 0:1/fs1:(length(s1) - 1)/fs1;
plot(t, s1), axis([0, (length(s1) - 1)/fs1 -0.4 0.5]);
title('原始语音s1的波形');
xlabel('时间/s');
ylabel('幅度')
pause 
close all
%Question 3 (linear)
disp('> Question 3: 画出线性谱');
M = 100;%当前帧数
N = 256;%帧长
frames = blockFrames(s1, fs1, M, N);%分帧
t = N / 2;
tm = length(s1) / fs1;
subplot(121);
imagesc([0 tm], [0 fs1/2], abs(frames(1:t, :)).^2), axis xy;
title('能量谱(M = 100, N = 256)');
xlabel('时间/s');
ylabel('频率/Hz');
colorbar;
%Question 3 (logarithmic)
disp('> Question 3: 画出对数谱');
subplot(122);
imagesc([0 tm], [0 fs1/2], 20 * log10(abs(frames(1:t, :)).^2)), axis xy;
title('对数能量谱(M = 100, N = 256)');
xlabel('时间/s');
ylabel('频率/Hz');
colorbar;
D=get(gcf,'Position');
set(gcf,'Position',round([D(1)*.5 D(2)*.5 D(3)*2 D(4)*1.3]))
pause
close all
%Question 4
disp('> Question 4: 画出不同帧长语谱图');
lN = [128 256 512];
u=220;
for i = 1:length(lN)
    N = lN(i);
    M = round(N / 3);
    frames = blockFrames(s1, fs1, M, N);
    t = N / 2;
    temp = size(frames);
    nbframes = temp(2);
    u=u+1;
    subplot(u)
    imagesc([0 tm], [0 fs1/2], 20 * log10(abs(frames(1:t, :)).^2)), axis xy;
    title(sprintf('能量对数谱(第 = %i帧, 帧长 = %i, 帧数 = %i)', M, N, nbframes));
    xlabel('时间/s');
    ylabel('频率/Hz');
    colorbar
end
D=get(gcf,'Position');
set(gcf,'Position',round([D(1)*.5 D(2)*.5 D(3)*1.5 D(4)*1.5]))
pause
close all
%Question 5
disp('> Question 5: Mel空间');
plot(linspace(0, (fs1/2), 129), (melfb(20, 256, fs1))');
title('Mel滤波');
xlabel('频率/Hz');
pause
close all
%Question 6
disp('> Question 6: 修正谱');
M = 100;
N = 256;
frames = blockFrames(s1, fs1, M, N);
n2 = 1 + floor(N / 2);
m = melfb(20, N, fs1);
z = m * abs(frames(1:n2, :)).^2;
t = N / 2;
tm = length(s1) / fs1;
subplot(121)
imagesc([0 tm], [0 fs1/2], abs(frames(1:n2, :)).^2), axis xy;
title('原始能量谱');
xlabel('时间/s');
ylabel('频率/Hz');
colorbar;
subplot(122)
imagesc([0 tm], [0 20], z), axis xy;
title('通过mel倒谱修正后的能量谱');
xlabel('时间/s');
ylabel('滤波器数目');
colorbar;
D=get(gcf,'Position');
set(gcf,'Position',[0 D(2) D(3)*2 D(4)])
pause
close all
%Question 7
disp('> Question 7: 2D plot of accustic vectors');
c1 = mfcc(s1, fs1);
c2 = mfcc(s2, fs2);
plot(c1(5, :), c1(6, :), 'or');
hold on;
plot(c2(5, :), c2(6, :), 'xb');
xlabel('5th Dimension');
ylabel('6th Dimension');
legend('说话人1', '说话人2');
title('2D plot of accoustic vectors');
pause
close all
%Question 8
disp('> Question 8: 画出已训练好的VQ码本')
d1 = vqlbg(c1,16);
d2 = vqlbg(c2,16);
plot(c1(5, :), c1(6, :), 'xr')
hold on

Three, running results

Insert picture description here

Four, remarks