Lab 2

Sampling
0. Preface
Now that you have learned MATLAB to a sufficient and necessary level, in this lab you will use it to explore sampling. Your lab report should answer all questions in all sections.

1. Are you down with DFT? How about the FFT?

The DFT X[k] of a finite-length sequence x[n] of length N , defined over

0 n N 1, is given by:
N 1

2 X[k] = x[n]e j kn / N , 0 k N 1 . n= 0

(1)

The sequence x[n] is created by sampling a continuous signal x a (t) at a sampling rate of Fs = 1/Ts: x[n] = x a (nTs ) , for some range of n . The significance of the DFT is more apparent written in the following way:
N 1

X[k] = x a (nTs )e
n= 0

j 2

kFs nTs N

, 0 k N 1 .

(2)

It can be seen in equation (2) that the samples of x[n] are being correlated with N samples of N sampled complex sinusoids having different angular frequencies. The value of X[k] is the (complex) amplitude of the k th frequency bin; the kF corresponding frequency of this bin in radians per second is given by k = 2 s , N when the sampling rate is in Hz. For instance X[0] is the zero-frequency, or D.C., term; X[1] is the frequency at 1 = 2

Fs Hz, and so on. For real signals X[k] is N

only meaningful for 0 k N /2 + 1for N even. All other k signify negative

frequencies; their values for can be inferred from the positive ones.

Copyright 2006 Bob L. Sturm

ECE158 Lab2: Sampling

You will now perform your first DFT using the fast Fourier transform (FFT)which is an implementation of the DFT. 1.1 Sample a sinusoidal signal with amplitude 1 and frequency of 128 Hz at a sampling rate of 2,048 Hz indexed by 0 n N 1, where N = 32 . That means sample the following: (3) x[n] = sin(2128n /2048) , at the points n=[0:31]. Plot it using stem and include your code and plot in
1.2 Now find the DFT of the sequence created in 1.1 using the fft MATLAB

your report.

routine (just call fft with the sequence as an argument). Plot the magnitude spectrum of the result, abs(X), versus the corresponding frequency (in Hertz and not k). Plot only the frequencies between 0 and the Nyquist frequency (Fs/2). Your plot should look like Figure 1. That means use stem instead of
plot.

Include your code.

Figure 1: The magnitude spectrum of a real 128 Hz sinusoid sampled at 2048 Hz over 0 n 31 1.3 What is the phase in pi radians at the frequency 128 Hz? Do this using the
angle

routine on the relevant value of X[k]. Why does this answer make

sense? Copyright 2006 Bob L. Sturm

ECE158 Lab2: Sampling

1.4 Now sample a sinusoid at 220 Hz at the same sampling rate as in 1.1. Create a 32-sample sequence and plot it versus n using stem. Include the plot in your report, but not your code. 1.5 Draw a picture of what you hypothesize will be the magnitude DFT of this signal. Include in your lab this hypothesized picture. Explain why you drew what you drew. 1.6 As in 1.2 find the DFT of this signal and plot the magnitude of the result versus frequency. Include this plot in your report. How does this plot compare to that in 1.2? Was your hypothesized plot, created in 1.5, correct? 1.7 Verify that you get the exact same time-domain signal from this spectrum by taking an inverse DFT using the ifft routine. Plot the resulting time-domain sequence, which should be completely real, just like in 1.4. If it is not exactly the same, then you have done something wrong. What the heck is going on? The input is a pure sinusoid at 220 Hz, but the DFT of this new signal, evaluated in 1.6, shows that all frequencies possible with 32 uniformly spaced samples are present. Why? There are two reasons for this, both of which are related. First, since the only frequencies that can be represented by a DFT of length-32 and sampling rate of 2,048 Hz are integer multiples of f k =

2048 = 64 Hz , a sinusoid with a frequency 32

of 220 Hz is between two frequency bins. Second, looking at the differences between the time-domain signal created in 1.1 and that created in 1.4, one can see in the latter there is not an integer number of periods contained in the 32 samples. Thus a discontinuity is created when this sequence is abruptly cut off. In short this discontinuity is responsible for the extra frequencies when the DFT periodizes the signal to both ends of infinity. This effect, when a frequency falls in a crack, is colorfully called spectral leakage (Figure 2).

Copyright 2006 Bob L. Sturm

ECE158 Lab2: Sampling

Figure 2: A Leicester County worker listens for spectral leakage Now we are going to perform a trick. If the frequencies of the DFT are integer multiples of f k =

kFs , one way to make the spacing of the bins closer together is N

to increase N . So instead of taking 32 samples of the sinusoid in equation (3), we can generate 512 samples, which gives a frequency spacing of 4 Hz. Theoretically we could then get rid of spectral leakage for our sinusoids of 220 Hz because it falls right in bin number 56. Lets prove this now. 1.8 Create a 512-length sequence of the sinusoid in equation (3) at a frequency of 220 Hz, amplitude of 1, and a sampling rate of 2,048 Hz. As in 1.2 find the DFT of this signal and plot the magnitude of the result versus frequency. Include this plot in your report. Problem solved? Well what if we only had the 32-sample version created in 1.4 and there was no way to extend it or get more data? We will return to this problem in a later lab called windowing. Through learning the FFT implementation of the DFT you have taken your first steps to understanding not only the frequency domain representation of signals, but also the concept of sampling. When you evaluate some continuous function at uniformly spaced samples, you are in effect sampling that signal. To gain a Copyright 2006 Bob L. Sturm

ECE158 Lab2: Sampling

much clearer understanding of this process, lets move to the next portion of this lab.

2. Sampling
Much more than walking through CostcoTM on food promotion day, sampling is essential to any digital system. Sampling, and specifically the Shannon-Nyquist sampling theorem, makes it possible to represent a continuous signal exactly with a finite number of observations of it. Did I say exactly? I meant, exactly, meaning no difference, no error, nothing, nada, exactamundo. Lets see how! 2.1 Sample the sinusoid x(t) = cos(2ft) having frequencies of f=100, 200, 300, 400, 500, 600, 700, 800, 900, 1000 Hz, at a sampling rate of 1 kHz. Plot using stem each signal over 0 n N 1, where N=128. Use subplot to make it look fancy! Include your code. 2.2 What do you notice about these plots? For instance, can you distinguish between the sinusoids that have frequencies of 400 and 600 Hz? What is going on when f=1000 Hz? 2.3 For each of the sinusoids sampled in 2.1, evaluate its DFT and plot the normalized magnitude spectrum with stem . Normalize the magnitude spectrum by dividing it with the largest magnitude. Include your figure and code. 2.4 What do you notice about these plots? For instance, can you distinguish between the sinusoids that have frequencies of 400 and 600 Hz? At what frequency is most of the energy located for the sampled sinusoid have f=700? Hopefully by now you have come to realize that the sampling rate has dire consequences for the output if you are unsure as to the frequency content of the continuous signal being sampled. As above, a 700 Hz sinusoid produces the same output as a 300 Hz sinusoid when both are sampled at 1000 Hz. And God forbid if you try to sample a signal having a frequency that is the same as the sampling rateall you get in that case is a constant! Copyright 2006 Bob L. Sturm

ECE158 Lab2: Sampling

2.5 Retrieve from the class website, and load, the speech sound file speech_female.wav. This signal is sampled at a rate of 44.1 kHz. You are going to make a really pretty picture of its distribution of energy in time and frequency over the first 1.4 seconds. Over this time the woman says, in an English accent: To administer medicine to a. You can listen to it by using
sound .

(Make sure you pass the correct sampling rate, otherwise you will

hear a zombie with an English accent.) Run the following program and include the resulting picture in your report. Comment the program, line-byline, to describe what it is doing.
clear all sndfile = 'speech_female.wav'; [x,Fs] = wavread(sndfile); N = 512; [S,F,T] = spectrogram(x(1:Fs*1.4),N,3*N/4,N*4,Fs); f = figure('Position',[500 300 700 500],'MenuBar','none', ... 'Units','Normalized'); set(f,'PaperPosition',[0.25 1.5 8 5]); axes('FontSize',14); imagesc(T,F./1000,20*log10(abs(S))); axis xy; set(gca,'YTick',[0:2000:Fs/2]./1000, ... 'YTickLabel',[0:2000:Fs/2]./1000); ylabel('Frequency (kHz)'); xlabel('Time (s)');

You might have heard about the Short-Time Fourier Transform (STFT). The figure you just created is the log magnitude of the STFT, and is called a spectrogram or sonogram. Essentially it displays the distribution of energy in a signal as a function of time and frequency. 2.6 Using the spectrogram you created in 2.5, in what frequency band is most of the energy located? Knowing the Nyquist-Shannon Sampling Theorem, at what minimum rate would you sample this signal such that it could be reconstructed wellmeaning without significant aliasing. 2.7 Looking at the distribution of energy across frequencies in your spectrogram, and knowing the woman says To administer medicine to a, which sounds of which words correspond to the energy between 4 and 18 kHz? Copyright 2006 Bob L. Sturm