ELEN E4810: Digital Signal Processing

Topic 10: The Fast Fourier Transform

1. Calculation of the DFT 2. The Fast Fourier Transform algorithm 3. Short-Time Fourier Transform

Dan Ellis

2013-11-27

1. Calculation of the DFT

! !

Filter design so far has been oriented to time-domain processing - cheaper! But: frequency-domain processing makes some problems very simple:
DFT
!

x[n]

X[k]

Fourier domain Y[k] IDFT processing

y[n]

use all of x[n], or use short-time windows

Need an efficient way to calculate DFT

2013-11-27 2

Dan Ellis

The DFT
!

Recall the DFT:

kn X[k ] = " x[n]WN n =0
!

N !1

( WN = e

!j

2" N

discrete transform of discrete sequence

1 1 WN 2 WN . . .
(N 1) WN

WN@2/N WNr has only N distinct values Structure opportunities for efficiency

!
X [N 1] X [0] X [1] X [2] . . .

1 1 1 = . . . 1

Matrix form:
1 2 WN 4 WN . . .
2(N 1) WN

.. .

(N 1) WN 2(N 1) WN . . . (N 1)2 WN

x[N 1]

x[0] x[1] x[2] . . .

Dan Ellis

2013-11-27

Computational Complexity
X[k ] = "
!

N !1 n =0

kn x[n]WN

N complex multiplies + N-1 complex adds per point (k) ! N points (k = 0.. N-1)
!

cpx mult: (a+jb)(c+jd) = ac - bd + j(ad + bc) = 4 real mults + 2 real adds cpx add = 2 real adds

N points: 4N2 real mults, 4N2-2N real adds

2013-11-27 4

Dan Ellis

Goertzels NAlgorithm !1 k! ! Now: X [k ] = " x [! ]W N

!=0 kN = WN

"! x[!]

! k ( N ! !) WN

looks like a convolution

X [k ] = yk [ N ] i.e. where yk [n] = xe [n] ! hk [n]

yk[n] yk[-1] = 0 yk[N] = X[k]
2013-11-27

x[n] 0 ! n < N xe[n] = { 0 n = N WN-kn n " 0 hk[n] = { 0 n < 0

xe[n] + xe[N] = 0 W -k z-1 N

Dan Ellis

Goertzels Algorithm
!

Separate filters for each X[k]

!

can calculate for just a few values of k

! !

No large buffer, no coefficient table Same complexity for full X[k] (4N2 mults, 4N2 - 2N adds)
!

H (z)

k !1 for last step 1 1 ! WN z = = ! k !1 "k z !1 + z !2 2 real mults 1 ! WN z 1 ! 2 cos 2N per step

2013-11-27 6

but: can halve multiplies by making the denominator real: evaluate only

Dan Ellis

2. Fast Fourier Transform FFT

!

Reduce complexity of DFT from O(N2) to O(NlogN)

!

grows more slowly with larger N

Works by decomposing large DFT into several stages of smaller DFTs Often provided as a highly optimized library

Dan Ellis

2013-11-27

Decimation in Time (DIT) FFT

!

Can rearrange DFT formula in 2 halves:

N "1 n =0 N "1 2

nk X [k ] = # x [n] ! WN
k = 0.. N-1

Arrange terms in pairs... Group terms from each pair

= =

m =0 N "1 2 m =0

# ( x[2m ]

2 mk ! WN

+ x [2m + 1] ! WN
N "1 2

(2 m +1) k

mk k mk x 2 m ! W + W x 2 m + 1 ! W ] N # [ ] N N # [
2

X0[<k>N/2] N/2 pt DFT of x for even n

2013-11-27

X1[<k>N/2] N/2 pt DFT of x for odd n

8

m =0

Dan Ellis

Decimation in Time (DIT) FFT

DFTN {x [n]} = DFTN {x0 [n]}
2

x[n] for even n

x[n] for odd n

k + WN DFTN 2

{x1 [n]}

We can evaluate an N-pt DFT as two N/2-pt DFTs (plus a few mults/adds) But if DFTN{} ~ O(N2) then DFTN/2{} ~ O((N/2)2) = 1/4 O(N2) Total computation ~ 2 ! 1/4 O(N2) = 1/2 the computation (+) of direct DFT

Dan Ellis

2013-11-27

One-Stage DIT Flowgraph

X [k ] = X0 k
Even x[0] points x[2] from x[4] x[n] x[6] Odd x[1] points x[3] from x[5] x[n] x[7]

[ ]
N 2

k + WN X1

[ ]
k
N 2

DFTN
2

X0[0] X0[1] X0[2] X0[3] X1[0] X1[1] X1[2] X1[3]

twiddle factors: always apply to odd-terms output NOT mirror-image

0 WN 1 WN 2 WN 3 WN 4 WN 5 WN 6 WN 7 WN

X[0] X[1] X[2] X[3] X[4] Same as X[5] X[0..3] except for X[6] factors on X[7] X1[]

DFTN
2

Classic FFT structure

Dan Ellis 2013-11-27

terms

10

Multiple DIT Stages

!

If decomposing one DFTN into two smaller DFTN/2s speeds things up ... Why not further divide into DFTN/4s ? k i.e. X [k ] = X0 k N + WN X1 k N

[ ] [ ] make: X [k ] = X [ k ] + W X [ k ]
0!k<N
2 2

00

0 ! k < N/2 N/4-pt DFT of even points N/4-pt DFT of odd points from even subset in even subset of x[n]

N 4

N 2

01

N 4

Similarly, X1 [k ] = X10 k
2013-11-27

[ ]
N 4

+ WNk X11 k
2 11

[ ]
N 4

Dan Ellis

Two-Stage DIT Flowgraph

different from before

x[0] x[4] x[2] x[6] x[1] x[5] x[3] x[7]

DFTN X 00
4

0 WN /2

DFTN X01
4

X0[0] X0[1] X0[2] X0[3]

same as before
0 WN 1 WN 2 WN 3 WN 4 WN 5 WN 6 WN 7 WN

DFTN X 10
4

DFTN X11
4

3 WN /2 X1[0] 0 WN/2 X1[1] X1[2] X1[3] 3 WN /2

X[0] X[1] X[2] X[3] X[4] X[5] X[6] X[7]

Dan Ellis

2013-11-27

12

Multi-stage DIT FFT

!

Can keep doing this until we get down to 2-pt DFTs:

butterfly element X[0] = x[0] + x[1] X[1] = x[0] - x[1]

DFT2

" N = 2M-pt DFT reduces to M stages of twiddle factors & summation (O(N2) part vanishes) " real mults < M4N , real adds < 2!M2N " complexity ~ O(NM) = O(Nlog2N)
Dan Ellis 2013-11-27 13

1 = W20 -1 = W21

FFT Implementation Details

!

Basic butterfly (at any stage):

XX0[r] XX1[r] WNr WNr+N/2 XX[r] 2 cpx mults XX[r+N/2] 2 " ( r+ N )
r+ N WN 2

=e

!j !j

N 2 "r N !j 2 "N / 2 N

Can simplify:
XX0[r] XX1[r] XX[r]

=e #e r = !WN

just one cpx mult! WNr -1 XX[r+N/2] i.e. SUB rather than ADD
2013-11-27 14

Dan Ellis

bit-reversed indexing

8-pt DIT FFT Flowgraph

000 100 010 110 001 101 011 111
! ! !

x[0] x[4] x[2] x[6] x[1] x[5] x[3] x[7]

W4 W4

W8 2 W8 3 W8

X[0] X[1] X[2] X[3] X[4] X[5] X[6] X[7]

-1s absorbed into summation nodes WN0 disappears in-place algorithm: sequential stages
2013-11-27 15

Dan Ellis

FFT for Other Values of N

! !

Having N = 2M meant we could divide each stage into 2 halves = radix-2 FFT Same approach works for:
! ! !

N = 3M radix-3 N = 4M radix-4 - more optimized radix-2 etc...

Composite N = abcd " mixed radix (different N/r point FFTs at each stage)
!

.. or just zero-pad to make N = 2M

2013-11-27 16

Dan Ellis

Inverse FFT
!

1 Recall IDFT: x[n] = N Thus: Nx [n] = "

* N !1 k =0

N !1

"

only differences from forward DFT

! nk X [k ]WN

k =0

Forward DFT of x#[n] = X*[k]|k=n i.e. time sequence made from spectrum
! nk * X [k ]WN

) = "X
k =0
Re -1 Im

N !1

nk [k ]WN

Hence, use FFT to calculate IFFT:

N 1 *
pure real flowgraph

x [n] =

1 N

X [k ]
*

nk WN

k =0

Re{X[k]} Im{X[k]}
2013-11-27

DFT

Re Im

1/N

-1/N

Re{x[n]} Im{x[n]}
17

Dan Ellis

DFT of Real Sequences

! ! !

If x[n] is pure-real, DFT wastes mults Real x[n] " Conj. symm. X[k] = X*[-k] Given two real sequences, x[n] and w[n] call y[n] = jw[n] , v[n] = x[n] + y[n] N-pt DFT V[k] = X[k] + Y[k] X[k] -Y[k] but: V[k]+V*[-k] = X[k]+X*[-k]+Y[k]+Y*[-k] X[k]=1/2(V[k]+V*[-k]) , W[k]=-j/2(V[k]-V*[-k]) i.e. compute DFTs of two N-pt real sequences with a single N-pt DFT
2013-11-27 18

Dan Ellis

3. Short-Time Fourier Transform (STFT)

! ! !

Fourier Transform (e.g. DTFT) gives spectrum of an entire sequence: How to see a time-varying spectrum? e.g. slow AM of a sinusoid carrier:
2 1 0 -1 -2 0 200 400 600 800 1000

# 2"n & x [n] = %1 ! cos ( cos ) 0 n $ N '

x[n]

n
19

Dan Ellis

2013-11-27

Fourier Transform of AM Sine

!

X[k] Spectrum of 400 whole sequence 200 indicates modulation 0 0 0.02 indirectly... Nsin2kn ... as N cancellation between -Nsin2(k-1)n closely2 N tuned -Nsin2(k+1)n sines 2 N
1 0.5 0 -0.5 -1 1 0.5 0 -0.5 -1 1 0.5 0 -0.5 -1

600

N N/2

2cAcB = cA+B +cA-B

0.04

0.06

0.08

k/(N/2)

128

256

384

512

640

768

896

Dan Ellis

2013-11-27

20

Fourier Transform of AM Sine

!

Sometimes wed rather separate modulation and carrier: x[n] = A[n]cos0n

!

A[n]

A[n] varies on a different (slower) timescale

One approach:
! ! !

chop x[n] into short sub-sequences .. .. where slow modulator is ~ constant DFT spectrum of pieces " show variation
2013-11-27 21

Dan Ellis

FT of Short Segments
!

Break up x[n] into successive, shorter chunks of length NFT, then DFT each:
2 1 0 -1

x[n]

NFT -2 = N/8 0 128 256 384 512 640 768 896 1024 = N x0[n] x1[n] x2[n] x3[n] x4[n] x5[n] x6[n] x7[n]
100 50 0

X0[k] X1[k] X2[k] X3[k] X4[k] X5[k] X6[k] X7[k]

0

64

Shows amplitude modulation of 0 energy

Dan Ellis

k0 = 0 NFT 2
2013-11-27

22

The Spectrogram
!

Plot successive DFTs in time-frequency:

X0[k] X1[k] X2[k] X3[k] X4[k] X5[k] X6[k] X7[k] k k
120 100 80 60 40 5 0 0 128 256 384 512 640 768 896 1024 20 0

Xi[k] X[k,n]
15

10

This image is called the Spectrogram

2013-11-27 23

time hopsize (between successive frames) = 128 points

Dan Ellis

Short-Time Fourier Transform

! !

Spectrogram = STFT magnitude plotted on time-frequency plane STFT is (DFT form): X [k , n0 ] =

N FT "1 n =0

$ x[n0 + n] ! w[n] ! e
window

2 #kn " jN FT

frequency time index index

!

NFT points of x starting at n0

DFT kernel

intensity as a function of time & frequency

2013-11-27 24

Dan Ellis

STFT Window Shape

!

w[n] provides time localization of STFT

!

e.g. rectangular selects x[n], n0 ! n < n0+NW

w[n]

DTFT form of STFT

But: resulting spectrum has same problems as windowing for FIR design: j! X e , n0 = DTFT{x [n0 + n] " w[n]}

&$ % e

j#n 0

X e

( )W (e
j#

j (! $# )

)d#

spectrum of short-time window is convolved with (twisted) parent spectrum

Dan Ellis 2013-11-27 25

STFT Window Shape

!

e.g. if x[n] is a pure sinusoid,

X(ej)

W(ej)

blurring (mainlobe) + ghosting (sidelobes)

W(ej)

Hence, use tapered window for w[n]

w[n]

e.g. Hamming

w[n] = n ) 0.54 + 0.46 cos(2! 2 M +1

-10 -5 0 5 10

sidelobes < -40 dB

26

Dan Ellis

2013-11-27

STFT Window Length

!

Length of w[n] sets temporal resolution

0.2 0.1 0 -0.1

x[n]

wS [n]

0.2 0.1 0

x[n]

wL[n]

short window measures only local properties

!
1 0.5

200

400

600

800

1000

-0.1

Window length " 1/(Mainlobe width)

wS[n]
N1 pts
0 -100 -50 0 20

longer window averages spectral character

WS(ej)

200

400

600

800

1000

10

n
100

zero at 4 N1
-0.5 0

wL[n]
N2 pts

50

0 - 20

1 0.5 0 10

WL(ej)
zero at 4 N2

0.5

shorter window " more blurred spectrum

n
-100 -50 0 50 100

more time detail

0 -

less frequency detail

27

-0.5

0.5

Dan Ellis

2013-11-27

STFT Window Length

!

Can illustrate time-frequency tradeoff on the time-frequency plane:

250 200 150 100 50 0 1 0.5 00 100 200 300

disks show blurring due to window length; area of disk is constant ! Uncertainty principle:

ft ! k
n

Alternate tilings of time-freq:

half-length window " half as many DFT samples
2013-11-27 28

Dan Ellis

Spectrograms of Real Sounds

0.1 0

time-domain
10 0

freq / Hz

4000

3000 -10 2000 -20 -30 1000 -40 0 2.35 4000 2.4 2.45 2.5 2.55 2.6 -50

successive short DFTs

intensity / dB

time / s

freq / Hz

3000

2000

1000

individual t-f cells merge into continuous image

0 0.5 1 1.5 2 2.5

time / s

Dan Ellis

2013-11-27

29

Narrowband vs. Wideband

!

Effect of varying window length:

0.2 0 4000

Window = 256 pt Narrowband

freq / Hz

3000

2000

10 0

1000 -10 0 -20 -30 -40 -50 2000

Window = 48 pt Wideband

freq / Hz

4000

3000

level / dB

1000

1.4

1.6

1.8

2.2

2.4

2.6

time / s

Dan Ellis

2013-11-27

30

Spectrogram in Matlab
>> >> >> >> >>
8000
Frequency

[d,sr]=wavread(mpgr1_sx419.wav'); (hann) window length Nw=256; specgram(d,Nw,sr) actual sampling rate caxis([-80 0]) (to label time axis) colorbar
0 -20 -40 -60 0.5 1 1.5 Time 2 2.5 3 -80

dB

6000 4000 2000 0

Dan Ellis

2013-11-27

31

STFT as a Filterbank
!

Consider one row of STFT: X k [n0 ] = $ x [n0 + n] ! w[n] ! e

N "1
kn " j 2# N

just one freq.

n =0 " ( N "1) m =0

$ hk [m ] x[n0 " m ]
1

convolution with complex IR

Im{x[n]}

where hk [n] = w[!n] " e

!

kn j 2# N

-1 -60 -40 0 -20 0 -1

Each STFT row is output of a filter (subsampled by the STFT hop size)
2013-11-27 32

Re{x[n]}

Dan Ellis

STFT as a Filterbank
!

If hk [n] = w[(!)n] " e then H k e

kn j 2# N

( ) = W (e
j!
H2(ej)

#k (" ) j (! " 2N )

shift-in-

Each STFT row is the same bandpass response defined by W(ej), frequency-shifted to a given DFT bin:
H1(ej)

W(ej)

A bank of identical, frequency-shifted bandpass filters: filterbank

33

Dan Ellis

2013-11-27

STFT Analysis-Synthesis
!

IDFT of STFT frames can reconstruct (part of) original waveform e.g. if X [k , n0 ] = DFT{x [n0 + n] ! w[n]} then IDFT{X [k , n0 ]} = x [n0 + n] ! w[n] ^ [n]: Can shift by n0, combine, to get x
^[n] x
n0

x[n]w[n-n0] n

Could divide by w[n-n0] to recover x[n]...

2013-11-27 34

Dan Ellis

STFT Analysis-Synthesis
! !

Dividing by small values of w[n] is bad

x[n]w[n-rH]

Prefer to overlap windows:

^[n] x n

i.e. sample X[k,n0] at n0 = rH where H = N/2 (for example)

hopsize window length
!

Then x [n] = "r x [n]w[n ! rH ] = x [n] if # w[n ! rH ] = 1

"r
2013-11-27 35

Dan Ellis

STFT Analysis-Synthesis
Hann or Hamming windows w[n] + w[n-N/2] with 50% overlap w[n] w[n-N/2] sum to constant n) 0.54 + 0.46 cos(2 ! ( N )
!
1 0.8 0.6 0.4 0.2 0 0 20 40 60 80

+ 0.54 + 0.46 cos(2!

!

n" N 2 N

) = 1.08

Can modify individual frames of X[k,n] and then reconstruct

! !

complex, time-varying modifications tapered overlap makes things OK

2013-11-27 36

Dan Ellis

STFT Analysis-Synthesis
!

e.g. Noise reduction:

STFT of original speech Speech corrupted by white noise Energy threshold mask
k
120 100 80 60 40 20 100 200 300

r
37

Dan Ellis

2013-11-27