Spectrogram: HOW TO CONTROL ZERO-PADDING for individual fft's of an ULTRA LONG sequence

Hello,

I have a long audio sequence x at sampling rate fs, and would like to perform a spectrogram on it; (fs = 44000, and this is a several second recording).

I want this spectrogram

• to be built from half-overlapping individual audio snips of say 3000 audio data points
• I want to have the individual FFT of say length nFFT= 2^16 , i.e. with a very substantial amount of zero padding (for very fine interpolation purpose).

So in this case I want a specific control on the zero-padding, i.e. a zero-padding that is NOT governed by the criterium nFFT > size(x).

Can this be done using matlab spectrogram ?

Answer 1

[Chunru]

"doc spectrogram" for details.

[x, fs] = audioread('RockGuitar-16-44p1-stereo-72secs.wav');
whos

  Name            Size               Bytes  Class     Attributes

  fs              1x1                    8  double              
  x         3195904x2             51134464  double              

% zeros will be padded automatically
%                   win   overlap   nfft
spectrogram(x(:,1), 3000, 1500,     2^16, fs);

Answer 2

[Fabia Bayer]

Hello,

if I understand your question correctly, you want segments of length n_w = 3000, each overlapping by 1500 samples, and then zero-padded with 2^16-3000 zeros for your FFT.

This should be technically possible using

n_w = 3000;
n_fft = 2^16

window = ones(1, n_w) %no window - add here whichever window of length n_w, i.e. hamming, hanning etc you prefer
spectrogram(x, window, [], n_fft)

However, you may want to think about whether this solution actually computes what you are looking for. A high amount of zero-padding increases the frequency resolution of your result, but it also dismorphs yor result. A signal that is zero almost everywhere, as the one you specify, will reflect this in its FFT. So it may be worth thinking again whether adding this enormous zero padding truly is superior to increasing the frequency resolution by taking longer (potentially more overlapping) segments from your signals.

Answer 3

[Bjorn Gustavsson]

Yes, in principle, zero-padding is used by spectrogram if nFFT is larger than the window-length.

HTH