Question

如何对声音文件提取MFCC特征和LPC特征

如何以MATLAB或C语言编程实现相关特征提取算法

Answer 1

程序如下MusicSource = wavread('1');
Music_source = MusicSource';
N = 256; % window length，N = 100 -- 1000;
Hamm = hamming(N); % create Hamming window
frame = input('请键入想要处理的帧位置 = ');
% origin is current frame
origin = Music_source(((frame - 1) * (N / 2) + 1):((frame - 1) * (N / 2) + N));
Frame = origin .* Hamm';

%
%Short Time Fourier Transform
%
[s1,f1,t1] = specgram(MusicSource,N,N/2,N);
[Xs1,Ys1] = size(s1);
for i = 1:Xs1
FTframe1(i) = s1(i,frame);
end

N1 = input('请键入预测器阶数 = '); % N1 is predictor's order
[coef,gain] = lpc(Frame,N1); % LPC analysis using Levinson-Durbin recursion
est_Frame = filter([0 -coef(2:end)],1,Frame); % estimate frame(LP)
FFT_est = fft(est_Frame);
err = Frame - est_Frame; % error
% FFT_err = fft(err);
subplot(2,1,1),plot(1:N,Frame,1:N,est_Frame,'-r');grid;title('原始语音帧vs.预测后语音帧')
subplot(2,1,2),plot(err);grid;title('误差');
pause

%subplot(2,1,2),plot(f',20*log(abs(FTframe2)));grid;title('短时谱')

%
% Gain solution using G^2 = Rn(0) - sum(ai*Rn(i)),i = 1,2,...,P
%
fLength(1 : 2 * N) = [origin,zeros(1,N)];
Xm = fft(fLength,2 * N);
X = Xm .* conj(Xm);
Y = fft(X , 2 * N);
Rk = Y(1 : N);
PART = sum(coef(2 : N1 + 1) .* Rk(1 : N1));
G = sqrt(sum(Frame.^2) - PART);

A = (FTframe1 - FFT_est(1 : length(f1'))) ./ FTframe1 ; % inverse filter A(Z)
subplot(2,1,1),plot(f1',20*log(abs(FTframe1)),f1',(20*log(abs(1 ./ A))),'-r');grid;title('短时谱');
subplot(2,1,2),plot(f1',(20*log(abs(G ./ A))));grid;title('LPC谱');
pause

%plot(abs(ifft(FTframe1 ./ (G ./ A))));grid;title('excited')
%plot(f1',20*log(abs(FFT_est(1 : length(f1')) .* A / G )));grid;
%pause

%
% find_pitch
%
temp = FTframe1 - FFT_est(1 : length(f1'));

% not move higher frequnce
pitch1 = log(abs(temp));
pLength = length(pitch1);
result1 = ifft(pitch1,N);

% move higher frequnce
pitch1((pLength - 32) : pLength) = 0;
result2 = ifft(pitch1,N);

% direct do real cepstrum with err
pitch = fftshift(rceps(err));
origin_pitch = fftshift(rceps(Frame));
subplot(211),plot(origin_pitch);grid;title('原始语音帧倒谱(直接调用函数)');
subplot(212),plot(pitch);grid;title('预测误差倒谱(直接调用函数)');
pause

%subplot(211),plot(1:length(result1),fftshift(real(result1)));grid;title('预测误差倒谱(根据定义编写，没有去除高频分量)');
%subplot(212),plot(1:length(result2),fftshift(real(result2)));grid;title('预测误差倒谱(根据定义编写，去除高频分量)');
coef