Frequency Domain Characterization of Signals

Yao Wang Polytechnic University, Brooklyn, NY11201 http: //eeweb.poly.edu/~yao

Signal Representation
What is a signal Time-domain description
Waveform representation Periodic vs. non-periodic signals

Frequency-domain description
Periodic signals Sinusoidal signals Fourier series for periodic signals Fourier transform for non-periodic signals Concepts of frequency, bandwidth, filtering Numerical calculation: FFT, spectrogram Demo: real sounds and their spectrogram (from DSP First)

Yao Wang, 2006

EE3414: Signal Characterization

What is a signal
A variable (or multiple variables) that changes in time
Speech or audio signal: A sound amplitude that varies in time Temperature readings at different hours of a day Stock price changes over days Etc

More generally, a signal may vary in 2-D space and/or time

A picture: the color varies in a 2-D space A video sequence: the color varies in 2-D space and in time

Continuous vs. Discrete

The value can vary continuously or take from a discrete set The time and space can also be continuous or discrete We will look at continuous-time signal only in this lecture

Yao Wang, 2006

EE3414: Signal Characterization

Waveform Representation
Waveform representation
Plot of the variable value (sound amplitude, temperature reading, stock price) vs. time Mathematical representation: s(t)

Yao Wang, 2006

EE3414: Signal Characterization

Sample Speech Waveform

0.08 0.06 0.04 0.02 0 -0.02 -0.04 -0.06 -0.08 -0.1 0.08 0.06 0.04 0.02 0 -0.02 -0.04 -0.06 -0.08 -0.1 2000

2000

4000

6000

8000

10000

12000

14000

16000

2200

2400

2600

2800

3000

Entire waveform [y,fs]=wavread('morning.wav'); sound(y,fs); figure; plot(y); x=y(10000:25000);plot(x);

Blown-up of a section. figure; plot(x); axis([2000,3000,-0.1,0.08]);

Signal within each short time interval is periodic Period depends on the vowel being spoken
5

Yao Wang, 2006

EE3414: Signal Characterization

Sample Music Waveform

Entire waveform

Blown-up of a section

[y,fs]=wavread(sc01_L.wav'); v=axis; sound(y,fs); axis([1.1e4,1.2e4,-.2,.2]) figure; plot(y); Music typically has more periodic structure than speech Structure depends on the note being played
Yao Wang, 2006 EE3414: Signal Characterization 6

Sinusoidal Signals
2.5

s (t ) = A cos(2f 0t + ) f 0 : frequency (cycles/second) T0 = 1 / f 0 : period A : Amplitude : Phase (time shift)

2 1.5 1 0.5 0 -0.5 -1 -1.5 -2 -2.5 -1.5 -1 -0.5 0 0.5 1 1.5

Sinusoidal signals are important because they can be used to synthesize any signal
An arbitrary signal can be expressed as a sum of many sinusoidal signals with different frequencies, amplitudes and phases

Music notes are essentially sinusoids at different frequencies

Yao Wang, 2006

EE3414: Signal Characterization

What is frequency of an arbitrary signal?

Sinusoidal signals have a distinct (unique) frequency An arbitrary signal does not have a unique frequency, but can be decomposed into many sinusoidal signals with different frequencies, each with different magnitude and phase The spectrum of a signal refers to the plot of the magnitudes and phases of different frequency components The bandwidth of a signal is the spread of the frequency components with significant energy existing in a signal Fourier series and Fourier transform are ways to find spectrums for periodic and aperiodic signals, respectively

Yao Wang, 2006

EE3414: Signal Characterization

Approximation of Periodic Signals by Sum of 2 Sinusoids sinusoids: 1 and 3d harmonics

st

0.5

s (t ) = Ak cos(2kf 0t )
k =0

-0.5

-1 -1.5
View note for matlab code

-1

-0.5

0.5

1.5

4 sinusoids: 1,3,5,7 harmonics

With many more sinusoids with appropriate magnitude, we will get the square wave exactly
Yao Wang, 2006 EE3414: Signal Characterization 9

Line Spectrum of Square Wave

Magnitude Spectrum for Square Wave 1.4

1.2
4 Ak = k 0

k = 1,3,5,... k = 0,2,4,...

Each line corresponds to one harmonic frequency. The line magnitude (height) indicates the contribution of that frequency to the signal. The line magnitude drops exponentially, which is not very fast. The very sharp transition in square waves calls for very high frequency sinusoids to synthesize.

Amplitude

0.8

0.6

0.4

0.2

0 1 3 5 7 9 k,fk=f0*k 11 13 15

Yao Wang, 2006

EE3414: Signal Characterization

10

Period Signal
Period T: The minimum interval on which a signal repeats
Sketch on board

Fundamental frequency: f0 =1/T Harmonic frequencies: kf0

Yao Wang, 2006

EE3414: Signal Characterization

11

Approximation of Periodic Signals by Sinusoids

Any periodic signal can be approximated by a sum of many sinusoids at harmonic frequencies of the signal (kf0 ) with appropriate amplitude and phase. The more harmonic components are added, the more accurate the approximation becomes. Instead of using sinusoidal signals, mathematically, we can use the complex exponential functions with both positive and negative harmonic frequencies

Yao Wang, 2006

EE3414: Signal Characterization

12

Complex Exponential Signals

Complex number:
A = A exp( j ) = A cos + j A sin = Re + j Im

Complex exponential signal

s (t ) = A exp( j 2f 0t ) = A cos(2f 0t + ) + j A sin( 2f 0t + )

Euler formula
exp( jt ) + exp( jt ) = 2 cos(t ) exp( jt ) exp( jt ) = j 2 sin(t )

Yao Wang, 2006

EE3414: Signal Characterization

13

Fourier Series Representation of Periodic Signals

Fourier Series Synthesis (inverse transform) : s (t ) = A0 + Ak cos(2kf 0t + k ) (single sided, for real signal only)
k =1

k =

exp( j 2kf 0t ) (double sided, for both real and complex)

Fourier series analysis (forward transform) : S k= 1 T0

T0

s (t ) exp( j 2kf 0t ) dt ; k = 0,1,2,...

S k is in general a complex number

For real signals, Sk=S*-k |Sk|=|S-k| (Symmetric spectrum)

Yao Wang, 2006 EE3414: Signal Characterization 14

Fourier Series Representation of Square Wave

Applying the Fourier series analysis formula to the square wave, we get
2 S k = jk 0 k = 1,3,5,... k = 0,2,4,...

Do the derivation on the board

Yao Wang, 2006

EE3414: Signal Characterization

15

Line Spectrum of Square Wave

Magnitude Spectrum for Square Wave 1.4

1.2

4 Ak = k 0

k = 1,3,5,... k = 0,2,4,...

Amplitude

0.8

0.6

0.4

0.2

0 1 3 5 7 9 k,fk=f0*k 11 13 15

Only the positive frequency side is drawn on the left (single sided spectrum), with twice the magnitude of the double sided spectrum.

Yao Wang, 2006

EE3414: Signal Characterization

16

Fourier Transform for Non-Periodic Signals

Aperiodic signal T 0= f 0 = 0 uncountable number of harmonics integral instead of sum

Fourier synthesis (inverse transform) : s (t ) =

S ( f ) exp( j 2ft )df

Fourier analysis (forward transform) : S(f) =

s(t ) exp( j 2ft )dt

For real signals, |S(f)| =|S(-f)| (Symmetric magnitude spectrum)

Yao Wang, 2006 EE3414: Signal Characterization 17

Pulse Function: Time Domain

A Rectangular Pulse Function 1.2

0.8 s(t)

0.6

0.4

0.2

0 -1

-0.8

-0.6

-0.4

-0.2

0 t

0.2

0.4

0.6

0.8

Derive Fourier transform on the board

Yao Wang, 2006 EE3414: Signal Characterization

1 T / 2 < t < T / 2 s (t ) = otherwise 0

18

Pulse Function: Spectrum

Magnitude Spectrum of Rectangular Pulse 1.2 1 0.8 0.6 |S(f)| 0.4 0.2 0 -0.2 -0.4 -10

The peaks of the FT magnitude drops slowly. This is because the pulse function has sharp transition, which contributes to very high frequency in the signal.

-8

-6

-4

-2

0 f

10

1 T / 2 < t < T / 2 s (t ) = 0 otherwise

Yao Wang, 2006

S( f ) = T

sin(Tf ) = T sinc(Tf ) Tf
19

EE3414: Signal Characterization

Exponential Decay: Time Domain

1 0.9 0.8 0.7 0.6 s(t) 0.5 0.4 0.3 0.2 0.1 0 0 0.5 1 1.5 2 2.5 t 3 3.5 4 4.5 5 s(t)=exp(- t), t>0); =1

t >0 exp(t ) s (t ) = 0 otherwise

Yao Wang, 2006

S( f ) =

1 1 ; S( f ) = + j 2f 2 + 4 2 f 2
20

EE3414: Signal Characterization

Exponential Decay: Spectrum

1 0.9 0.8 0.7 0.6 |S(f)| 0.5 0.4 0.3 0.2 0.1 0 -10 -8 -6 -4 -2 0 f 2 4 6 8 10 S(f)=1/( +j 2 f), =1

The FT magnitude drops much faster than for the pulse function. This is because the exponential decay function does not has sharp transition.

t >0 exp(t ) s (t ) = 0 otherwise

Yao Wang, 2006

S( f ) =

1 1 ; S( f ) = + j 2f 2 + 4 2 f 2
21

EE3414: Signal Characterization

(Effective) Bandwidth
1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 -10

fmin (fma): lowest (highest) frequency where the FT magnitude is above a threshold Bandwidth: B=fmax-fmin
The threshold is often chosen with respect to the peak magnitude, expressed in dB dB=10 log10(ratio) 10 dB below peak = 1/10 of the peak value 3 dB below=1/2 of the peak
22

B

-8 -6 -4 -2 0 2 4 6 8 10

fmin
Yao Wang, 2006

fmax
EE3414: Signal Characterization

More on Bandwidth
Bandwidth of a signal is a critical feature when dealing with the transmission of this signal A communication channel usually operates only at certain frequency range (called channel bandwidth)
The signal will be severely attenuated if it contains frequencies outside the range of the channel bandwidth To carry a signal in a channel, the signal needed to be modulated from its baseband to the channel bandwidth Multiple narrowband signals may be multiplexed to use a single wideband channel

Yao Wang, 2006

EE3414: Signal Characterization

23

How to Observe Frequency Content from Waveforms?

A constant -> only zero frequency component (DC compoent) A sinusoid -> Contain only a single frequency component Periodic signals -> Contain the fundamental frequency and harmonics -> Line spectrum Slowly varying -> contain low frequency only Fast varying -> contain very high frequency Sharp transition -> contain from low to high frequency Music: contain both slowly varying and fast varying components, wide bandwidth Highest frequency estimation?
Find the shortest interval between peak and valleys

Go through examples on the board

Yao Wang, 2006

EE3414: Signal Characterization

24

Estimation of Maximum Frequency

Blown-Up of the Signal 0.04

S(t)
0.03

0.02

0.01

-0.01

-0.02 2450

2460

2470

2480

2490

2500

2510

2520

2530

2540

2550

Time index Yao Wang, 2006 EE3414: Signal Characterization 25

Numerical Calculation of FT
The original signal is digitized, and then a Fast Fourier Transform (FFT) algorithm is applied, which yields samples of the FT at equally spaced intervals. For a signal that is very long, e.g. a speech signal or a music piece, spectrogram is used.
Fourier transforms over successive overlapping short intervals

Yao Wang, 2006

EE3414: Signal Characterization

26

Spectrogram
0.08 0.06 0.04 0.02 0 -0.02 -0.04 -0.06 -0.08 -0.1 2000

S(t)

FFT

FFT

FFT

FFT
2600

FFT FFT
2800 3000

2200

2400

t
Yao Wang, 2006 EE3414: Signal Characterization 27

Sample Speech Waveform

(click to hear the sound)
0.08 0.06 0.04 0.02 0 -0.02 -0.04 -0.06 -0.08 -0.1 0.08 0.06 0.04 0.02 0 -0.02 -0.04 -0.06 -0.08 -0.1 2000

T
2200 2400 2600 2800 3000

2000

4000

6000

8000

10000

12000

14000

16000

Entire waveform

Blown-up of a section.

Signal within each short time interval is periodic. The period T is called pitch. The pitch depends on the vowel being spoken, changes in time. T~70 samples in this ex. f0=1/T is the fundamental frequency (also known as formant frequency). f0=1/70fs=315 Hz. k*f0 (k=integers) are the harmonic frequencies.
Yao Wang, 2006 EE3414: Signal Characterization 28

Sample Speech Spectrogram

Power Spectrum -15 -20 fs=22,050Hz Power Spectrum Magnitude (dB) -25 -30 -35 -40 -45 -50 -55 -60 Frequency 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 GOOD MOR NING 1 0.9 Spectrogram

f0
0 2000 4000

0 0 1000 2000 3000 4000 Time 5000 6000 7000

f0

6000 Frequency

8000

10000

12000

figure; psd(x,256,fs);
Signal power drops sharply at about 4KHz

figure; specgram(x,256,fs);
Line spectra at multiple of f0, maximum frequency about 4 KHz What determines the maximum freq?

Yao Wang, 2006

EE3414: Signal Characterization

29

Another Sample Speech Waveform

Entire waveform

Blown-up of a section.

In the course of a December tour in Yorkshire

Yao Wang, 2006

EE3414: Signal Characterization

30

Speech Spectrogram

figure; psd(x,256,fs);
Signal power drops sharply at about 4KHz

figure; specgram(x,256,fs);
Line spectra at multiple of f0, maximum frequency about 4 KHz

Yao Wang, 2006

EE3414: Signal Characterization

31

Sample Music Waveform

Entire waveform

Blown-up of a section

[y,fs]=wavread(sc01_L.wav'); v=axis; sound(y,fs); axis([1.1e4,1.2e4,-.2,.2]) figure; plot(y); Music typically has more periodic structure than speech Structure depends on the note being played
Yao Wang, 2006 EE3414: Signal Characterization 32

Sample Music Spectrogram

figure; psd(y,256,fs);
Signal power drops gradually in the entire frequency range

figure; specgram(y,256,fs);

Line spectra are more stationary, Frequencies above 4 KHz, more than 20KHz in this ex.

Yao Wang, 2006

EE3414: Signal Characterization

33

Summary of Characteristics of Speech & Music

Typical speech and music waveforms are semi-periodic
The fundamental period is called pitch period The inverse of the pitch period is the fundamental frequency (f0)

Spectral content
Within each short segment, a speech or music signal can be decomposed into a pure sinusoidal component with frequency f0, and additional harmonic components with frequencies that are multiples of f0. The maximum frequency is usually several multiples of the fundamental frequency Speech has a frequency span up to 4 KHz Audio has a much wider spectrum, up to 22KHz

Yao Wang, 2006

EE3414: Signal Characterization

34

Demo
Demo in DSP First, Chapter 3, Sounds and Spectrograms
Look at the waveform and spectrogram of sample signals, while listening to the actual sound Simple sounds Real sounds

Yao Wang, 2006

EE3414: Signal Characterization

35

Advantage of Frequency Domain Representation

Clearly shows the frequency composition of the signal One can change the magnitude of any frequency component arbitrarily by a filtering operation
Lowpass -> smoothing, noise removal Highpass -> edge/transition detection High emphasis -> edge enhancement

One can also shift the central frequency by modulation

A core technique for communication, which uses modulation to multiplex many signals into a single composite signal, to be carried over the same physical medium.

Yao Wang, 2006

EE3414: Signal Characterization

36

Typical Filters
Lowpass -> smoothing, noise removal Highpass -> edge/transition detection Bandpass -> Retain only a certain frequency range
Low-pass H(f) High-pass H(f) Band-pass H(f)

Yao Wang, 2006

EE3414: Signal Characterization

37

Low Pass Filtering (Remove high freq, make signal smoother)

Magnitude Spectrum of Rectangular Pulse 1.2 1 0.8 0.6 |S(f)| 0.4 0.2 0 -0.2 -0.4 -10

Filtering is done by a simple multiplification: Y(f)= X(f) H(f) H(f) is designed to magnify or reduce the magnitude (and possibly change phase) of the original signal at different frequencies. A pulse signal after low pass filtering (left) will have rounded corners.

Ideal lowpass filter

Spectrum of the pulse signal

-8 -6 -4 -2 0 f 2 4 6 8 10

Yao Wang, 2006

EE3414: Signal Characterization

38

The original pulse function and its low-passed versions

S(t)

1.2

0.8

0.6 original averaging over 11 samples filter=fir1(10,0.25)

0.4

0.2

-0.2 -1

-0.8

-0.6

-0.4

-0.2

0.2

0.4

0.6

0.8

Impulse Response of the Filters 0.3

0.25

h(t)
0.2 averaging over 11 samples fir1(10,0.25) 0.15

0.1

0.05

-0.05 1 2 3 4 5 6 7 8 9 10 11

Yao Wang, 2006

EE3414: Signal Characterization

40

Frequency Response of the Filters

0 Magnitude (dB)

Averaging

-20 -40 -60 -80

0.1

0.2 0.3 0.4 0.5 0.6 0.7 0.8 Normalized Angular Frequency ( rads/sample)

0.9

100

fir11(10,0.25)

Magnitude Phase (degrees) (dB)

0 0 -50 -100 -100 -200 0 -150 0 0.1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Normalized Angular Frequency ( rads/sample) 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Normalized Angular Frequency ( rads/sample) 0.9 0.9 1 1

Yao Wang, 2006

EE3414: Signal Characterization

41

High Pass Filtering (remove low freq, detect edges)

Magnitude Spectrum of Rectangular Pulse 1.2 1 0.8 0.6 |S(f)| 0.4 0.2 0 -0.2 -0.4 -10

Ideal high-pass filter

Spectrum of the pulse signal

-8 -6 -4 -2 0 f 2 4 6 8 10

Yao Wang, 2006

EE3414: Signal Characterization

42

The original pulse function and its high-passed version

S(t)
1

0.8 original high-pass filtered

0.6

0.4

0.2

-0.2

-0.4 -1

-0.8

-0.6

-0.4

-0.2

0.2

0.4

0.6

0.8

t
Yao Wang, 2006 EE3414: Signal Characterization 43

The High Pass Filter

0.5

h(t) fir1(10,0.5,high); Impulse response: Current sample neighboring samples

0.4 0.3 0.2 0.1 0 -0.1 -0.2 -0.3

10

11

Magnitude (dB)

Frequency response

0 -20 -40 -60 -80 -100 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Normalized Angular Frequency ( rads/sample) 0.9 1

Yao Wang, 2006

400

EE3414: Signal Characterization

44

Filtering in Temporal Domain (Convolution)

Convolution theorem X ( f )H ( f ) x(t ) * h(t )

x(t ) * h(t ) = x(t )h( )d

Interpretation of convolution operation

replacing each pixel by a weighted sum of its neighbors Low-pass: the weights sum = weighted average High-pass: the weighted sum = left neighbors right neighbors

Yao Wang, 2006

EE3414: Signal Characterization

45

Implementation of Filtering
Frequency Domain
FT -> Filtering by multiplication with H(f) -> Inverse FT

Time Domain
Convolution using a filter h(t) (inverse FT of H(f))

You should understand how to perform filtering in frequency domain, given a filter specified in frequency domain Should know the function of the filter given H(f) Computation of convolution is not required for this lecture Filter design is not required.
Yao Wang, 2006 EE3414: Signal Characterization 46

What Should You Know (I)

Sinusoid signals:
Can determine the period, frequency, magnitude and phase of a sinusoid signal from a given formula or plot

Fourier series for periodic signals

Understand the meaning of Fourier series representation Can calculate the Fourier series coefficients for simple signals (only require double sided) Can sketch the line spectrum from the Fourier series coefficients

Fourier transform for non-periodic signals

Understand the meaning of the inverse Fourier transform Can calculate the Fourier transform for simple signals Can sketch the spectrum Can determine the bandwidth of the signal from its spectrum Know how to interpret a spectrogram plot

Yao Wang, 2006

EE3414: Signal Characterization

47

What Should You Know (II)

Speech and music signals
Typical bandwidth for both Different patterns in the spectrogram Understand the connection between music notes and sinusoidal signals

Filtering concept
Know how to apply filtering in the frequency domain Can interpret the function of a filter based on its frequency response
Lowpass -> smoothing, noise removal Highpass -> edge detection, differentiator Bandpass -> retain certain frequency band, useful for demodulation

Yao Wang, 2006

EE3414: Signal Characterization

48

References
Oppenheim and Wilsky, Signals and Systems, Sec. 4.2-4.3 (Fourier series and Fourier transform) McClellan, Schafer and Yoder, DSP First, Sec. 2.2,2.3,2.5 (review of sinusoidal signals, complex number, complex exponentials)

Yao Wang, 2006

EE3414: Signal Characterization

49