Spectrogram function cutting off beginning and end of signal

Hi,

I have a 10 second audio clip (called originalsound) with a sampling frequency of 2,000 Hz (called downsampledrate) that I want to plot using spectrogram. When I do so with the following code:

figure; spectrogram(originalsound,N,[],[],downsampledrate,'yaxis')
    title('Original Sound')
    set(gca,'FontSize',16)
    yt = get(gca, 'YTick');
    set(gca, 'YTick',yt, 'YTickLabel',yt*1E+3) %change y-axis to Hz from kHz
    ylabel('Frequency (Hz)') %change y-axis label           

the plot generates without 0 or 10 labeled for time on the x-axis.

I tried fixing this by changing the xlim from 0-10, but then I get white bars on either side of the spectrogram.

If I am plotting a 10 second sound, why are there not 10 seconds shown in the spectrogram? Is there some sort of windowing effect going on?

Answer 1

[Paul]

Hi Danielle,

Create a fake signal using the the parameters in the question

downsampledrate = 2e3;
rng(100);
originalsound = rand(1,numel(0:(1/downsampledrate):10));

The quesiton didn't specify the window length, so I'll just pick one

N = 1024;

With this window length, the number of N-point segments in the signal is

nseg = (numel(originalsound)/N)

nseg = 19.5322

Of course there can only be an integer number of segments. The input data will be truncated.

nseg = floor(nseg)

nseg = 19

However, I believe with these parameters there will be 50% overlap of the segements, so there will really be double the number of segments

nseg = nseg*2;

Create the spectrogram

figure; spectrogram(originalsound,N,[],[],downsampledrate,'yaxis')
title('Original Sound')
set(gca,'FontSize',16)
yt = get(gca, 'YTick');
set(gca, 'YTick',yt, 'YTickLabel',yt*1E+3) %change y-axis to Hz from kHz
ylabel('Frequency (Hz)') %change y-axis label 
xlim([0 10])

We see the white bars. Use output arguments to get the actual time vector

[~,~,t] = spectrogram(originalsound,N,[],[],downsampledrate);
t([1 end])

ans = 1×2

0.2560 9.7280

According to spectrogram, "The time values in t correspond to the midpoint of each segment." The midpoint of the first segment is at N/2 and the midpoint of the last segment (with the assumed 50% overlap) is at nseg*N

[N/2 , nseg*N/2]/downsampledrate

ans = 1×2

0.2560 9.7280

We get a match.

Based on the documentation, I don't think you'll ever get the spectrogram to start/stop at the first/last time point. Where the actual time points fall will depend how spectrogram determines the length and locations of the segments based on the input arguments.