Borrar filtros
Borrar filtros

How to generate audio from colored MFCC in the form of image?

9 visualizaciones (últimos 30 días)
In above fig, 41 images are tiled. Each image is generated mfcc. Its size is 28x28x3. It is rgb image. I have to find inverse of mfcc to generate sound. I have used inverse mfcc function from mathworks website. but it is applicable to 2-D matrix of mfcc. It can not be used with rgb image. I am unable to find audio signal from coloured mfcc.
please provide guidance.

Respuesta aceptada

Sarthak el 15 de Mayo de 2023
Hi Shilpa,
MFCC goes through multiple computations like STFT, Mel Projection, etc. Assuming you are mentioning the invMFCC function available on FileExchange, it should reverse all such computations and introduce necessary approximations. If it is only applicable for double 2-D matrices, you can use the im2double function. However, if the function doesnt behave as expected, you may need to write your own inverse functions for MFCC or leverage other third-party libraries.
Attaching links to the mentioned functions for your reference
  2 comentarios
Shilpa Sonawane
Shilpa Sonawane el 15 de Mayo de 2023
Thank you so much. I will use im2double function before invmfccs. Thanks a lot.
Shilpa Sonawane
Shilpa Sonawane el 15 de Mayo de 2023
In my project, mfccs are stored in RGB image format. It is 3-D matrix. I used im2double function before calling invmfcc. I tried invmfccs which is available on Mathworks website.I am facing so many errors.

Iniciar sesión para comentar.

Más respuestas (1)

Brian Hemmat
Brian Hemmat el 15 de Mayo de 2023
Editada: Brian Hemmat el 15 de Mayo de 2023
Inverting MFCC will require knowledge of the algorithm and parameters used to extract the MFCC. Note that perfectly reconstructing the audio is by definition impossible since you are discarding a lot of information during the feature extraction. It's just an interesting toy to understand what information is actually preserved in MFCC.
Here is an example of extracting MFCC using Audio Toolbox functionality and then attempting to reconstruct.
[audioIn,fs] = audioread("Counting-16-44p1-mono-15secs.wav");
audioIn = resample(audioIn,16e3,fs);
fs = 16e3;
audioIn = audioIn./max(abs(audioIn));
%% Extract MFCC
% This is roughly equivalent to how audioFeatureExtractor works with
% default settings.
win = hann(round(fs*0.025),"periodic");
overlap = round(fs*0.01);
fftlength = 2048;
numBands = 40;
% Design auditory filter bank.
filterBank = designAuditoryFilterBank(fs, ...
FrequencyScale="mel", ...
NumBands=numBands, ...
FFTLength=fftlength, ...
FrequencyRange=[0,4000], ...
% Compute STFT of speech signal.
X = stft(audioIn, ...
Window=win, ...
OverlapLength=overlap, ...
FFTLength=fftlength, ...
% Compute Mel auditory spectrogram
B = filterBank*abs(X);
% Compute cepstral coefficients
coeffs = cepstralCoefficients(B);
%% Visualize MFCC
imh = imagesc(normalize(coeffs'));
%% Inverse MFCC
% Inverse Cepstral Coefficients
B_reconstruct = 10.^(idct(coeffs',size(filterBank,1),Type=2));
% Inverse Mel Spectrum
% Scale the band per time step energy.
B_reconstruct = permute(B_reconstruct,[1,3,2]);
bands = filterBank.*B_reconstruct;
% Sum the bands at each time step so that you have a single spectrum per
% time step.
X = squeeze(sum(bands,1));
% Reconstruct signal from magnitude spectrum
audioOut = stftmag2sig(X,fftlength,fs,FrequencyRange="onesided", ...
% Clean up the edges
audioOut([1:numel(win),end-numel(win):end]) = 0;
audioOut = audioOut./max(abs(audioOut));
%% Listen to and plot the reconstruction
plot(audioIn,'bo'),hold on
plot(audioOut,'r*'),hold off
title("Original audio")
title("Reconstructed audio")

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by