Mike Brookes 26-Jan-06

Jan 2006 4

History of DAB

• 1986 – DAB
consortium formed
Digital Audio Broadcasting – France, Germany,
Netherlands, UK
– Eureka 147
development project
Mike Brookes • 1990 – First trial
broadcasts
SC5 - b
• 1993 – Public demonstration system in UK
• 1995 – Network broadcasts in UK
• 1997 – World DAB forum formed

Jan 2006 2 Jan 2006 5

Digital Audio Broadcasting Problems with AM and FM

• DAB Broadcasting • Multipath fading
– OFDM, SFN, Transmission frames – Reflections from aircraft, vehicles, buildings
– very large variations in signal strength over distances of ~1 m
– UK ensembles, System Parameters
• Source Coding – MP2
• Channel Coding
– Convolution, Puncturing, Freq & Time interleaving
• Receiver front end
Car @ 120 mph, f0 = 225 MHz
• Channel decoding
• Synchronization • Interference
– from equipment, vehicles and other radio stations

Jan 2006 3 Jan 2006 6

Main References OFDM

1. ETSI. “Radio Broadcasting Systems; Digital Audio Broadcasting (DAB) to
mobile, portable and fixed receivers”. EN 300 401, European • Orthogonal Frequency Division Multiplexing
Telecommunications Standards Institute, April 2000.
2. W. Hoeg and Thomas Lauterbach. “Digital Audio Broadcasting: Principles
• 1536 carriers at 1 kHz spacing
and Applications of Digital Radio”. John Wiley, 2003.
– symbol length: 0.246 + 1.0 =1.246 ms
3. C. Gandy. “DAB: an introduction to the Eureka DAB System and a guide
to how it works”. Technical Report WHP-061, British Broadcasting Corp, – 2 bits per carrier per symbol (DQPSK)
June 2003.
4. M. Bolle, D. Clawin, K. Gieske, F. Hofmann, T. Mlasko, M.J. Ruf, and G. 1 kHz
Spreitz. “The receiver engine chip-set for digital audio broadcasting”. In 3 kHz
Intl Symp on Signals Systems and Electronics, pages 338–342, October 7 kHz
1998. Cyclic Prefix
5. K. Taura, M. Tsujishita, M. Takeda, H. Kato, M. Ishida, and Y. Ishida. “A Sum
digital audio broadcasting (DAB) receiver”. IEEE Trans Consumer
Electronics, 42(3):322–327, August 1996.
0.246 ms 1 ms

SC5b: Digital Audio Broadcasting 1

Mike Brookes 26-Jan-06

Jan 2006 7 Jan 2006 10

Cyclic Prefix Fill-in transmitter

• Convolution with channel impulse response • Can have a low-power fill-in transmitter to
= sum of delayed, scaled copies
Prev Symbol Next Symbol
solve a local reception problem
• Add delay to synchronize with main Tx

Discrete Fourier Transform Interval

• If channel impulse response < 0.246 ms:

– No inter-symbol interference
– DFT gives input spectrum of symbol multiplied by
channel response:
• frequency dependent amplitude and phase shift
   

Jan 2006 8 Jan 2006 11

Frequency Domain Orthogonality UK DAB Multiplexes/Ensembles

• Taking DFT of 1 ms D
BBC: 12B (225.648 MHz)
segment is equal to C
12
1. convolving spectrum with B

sinc (FT of 1 ms window) A

2. sampling at multiples of Digital1: 11D (222.064MHz)
D
1 kHz
C
• Component frequencies are 11
B
orthogonal and do not
interfere. A • Each DAB multiplex: 1.536 MHz bandwidth
• Doppler spread damages D – 0.176 MHz gap between multiplexes
orthogonality: 10
C • Four multiplexes per 7 MHz TV channel
2
⎛ vf T ⎞ – 2 National ensembles
ICI power ≈ 0.5⎜ 0 ⎟ P0 = −14 dB at 190 km/h
⎝ c ⎠ – Regional + Local (3 for London)

Jan 2006 9 Jan 2006 12

Single Frequency Network Spectral Efficiency

• All transmitters send an identical signal • Existing FM transmissions
• Interference-free if delay + multipath < 0.246 – Each transmitter has a bandwidth of 0.2 MHz
ms relative to nearest transmitter – Nearby transmitters must be 0.4 MHz apart
• Optimal spacing ≈ c × 0.246 ms = 74 km – 2.2 MHz needed for a network covering entire
country
 
  • DAB
 – 1.5 MHz for 10 services covering entire
 country using a single frequency network
– 15 times more efficient !
• Transmitters further than ≈ 1.2 c × 0.246 ms do more harm than good

SC5b: Digital Audio Broadcasting 2

Mike Brookes 26-Jan-06

Jan 2006 13 Jan 2006 16

Frame organization Source Coding

• Based on MP2 (MPEG 1 Layer 2)
Audio
– Simpler than MP3 but less good
MP2 frames • Masking – Psycho-acoustic model
(24 ms) – loud sounds make quieter sounds inaudible at nearby
frequencies and times
DAB frames • Sub-band Processing
(96 ms) – Input @ 48 kHz sample rate
– Divide into 32 subbands of 750 Hz @ 1.5 kHz
• 36 samples/subband in each 24 ms CIF frame
• Only low 27 subbands are used (0 to 20.25 kHz)
Synchronization and header CIF: Common Interleaved Frame
(24 ms of data)

Jan 2006 14 Jan 2006 17

DAB Transmission Frame Masking

• CIF: (2.3+55.3) kbits/24 ms = 2.4 Mbps total • Normal hearing threshold is A

– FIC: 96 kbps for multiplex config and service names • Threshold is changed to B because of tone C
– MSC: 2.304 Mbps for audio + data • Higher quantization noise allowed in bands near tone C
• Services: Radio 2 = 256 kbps, Radio 7 = 155 kbps • Band D can be completely eliminated
• Only need to decode the wanted portion of the MSC • Threshold calculated from FFT spectrum + Psycho-Acoustic model

Jan 2006 15 Jan 2006 18

System Parameters Subband Processing

• Centre frequency, f0 ≈ 220 MHz (Band III) 1 ↓32 encode decode ↑32 1 +
– Wavelength > 1 m ⇒ diffraction around objects 2 ↓32 encode decode ↑32 2
demultiplex

– Lower frequencies are full up

multiplex

 
3 ↓32 encode decode ↑32 3

• Total bandwidth = 1.537 MHz

– Needs to be > 1.5 MHz for fading to be frequency selective
↓32 ↑32
– < 1.6 MHz to fit four into a 7 MHz TV channel 32 encode decode 32
Bandpass Bandpass
• Cyclic Prefix = 0.246 ms filters filters

– Needs to be > (transmitter spacing)/1.2c to allow SFN • Sample Rates: Input @ 48 kHz, Subbands @ 1.5 kHz
– Wasteful if long compared to useful symbol length
– Total number of samples stays the same
• Carrier Spacing = 1 kHz ⇔ Useful symbol length = 1 ms
– Symbol length < 0.4/(Doppler spread) ≈ 0.4c/(f0v) = 10 ms • Noise and speech spectra are roughly flat within a
– Symbol length < cyclic prefix for efficiency sideband
• Transmission frame = 76 symbols = 96 ms • All bandpass filters are 750 Hz wide
– Long for efficiency, short for ease of synchronization – efficent to implement

SC5b: Digital Audio Broadcasting 3

Mike Brookes 26-Jan-06

Jan 2006 19 Jan 2006 22

Encoder Block Diagram Channel Coding

• Energy Dispersal randomizes carrier phases
• Convolution Code adds protection
• Time interleaving combats burst errors
• Freq interleaving combats freq selective fading
• Inverse DFT converts phases into a waveform

Use FFT to calculate a masking level for each subband

Jan 2006 20 Jan 2006 23

Subband Coding Energy Dispersal

• Scale Factor calculated for every 8 ms
– Scale factor = max absolute signal value
• If carrier phase changes linearly with frequency
– Samples are divided by scale factor before quantization then IDFT gives a single impulse
– 3 scale factors per 24 ms - quantized to 6 bits each – Bad news for the transmitter
– omit scale factors 2 and/or 3 according to how similar they are Input
• need 2 bits to say what the choice is. • Solution: Reset
Output
S-Reg
• Bit Allocation determined for entire 24 ms – XOR data bits with a Clock
pseudo-random
– Choose bits per sample for each subband: sequence
• < 2.25 kHz: 16 choices: 0, 1.7, 3, 4, …, 14, 15, 16
– Generator polynomial:
• < 8.25 kHz: 16 choices: 0, 1.7, 2.3, 3, 3.3, 4, 5, …, 12, 13, 16 P(X)=X9+X5+1
• < 17.25 kHz: 8 choices: 0, 1.7, 2.3, 3, 3.3, 4, 5, 16
– Reset shift register at
• < 20.25 kHz: 4 choices: 0, 1.7, 2.3, 16
start of each 24 ms
– n bits gives SNR of 6n+1.6 dB frame
– Subbands with 0 bits need no scale factors (save up to 18 bits) 9-bit shift register + two XNOR gates

Jan 2006 21 Jan 2006 24

Bit Allocation Procedure Convolution Coding

• Aim: Maximize the minimum (over all subbands) • Four separate convolution codes with constraint
mask – to - quantization noise ratio length of 7
– If this ratio is > 1 then quantization noise inaudible – M bits → 4(M+6) bits
• Method • Mother code has
1. Initialize bit allocation to 0 for each subband rate ¼
2. Find the worst subband • Extra 6 bits from
emptying the shift
3. Give it an extra bit (or fraction of a bit) register
4. Go back to step 2. • Reset Shift-Reg
5. Stop when all available bits are used up every 24 ms

Output: x0,1 x1,1 x2,1 x3,1 x0,2 x1,2 … x3,M+5 x0,M+6 x1,M+6 x2,M+6 x3,M+6

SC5b: Digital Audio Broadcasting 4

Mike Brookes 26-Jan-06

Jan 2006 25 Jan 2006 28

Puncturing Frequency Interleaving

2 sync + 75 data symbols = 96 ms
• Not all 4M+6 bits are transmitted
• Predefined puncturing patterns. Examples:
data for one symbol = 3072 bits
– Rate 1/3 code
• 8 input bits → 32 mother → 24 transmitted
• Transmit: 9998 9998 9998 9998 9998 9998
1536 carriers (excluding 0 Hz)
– Rate 2/3 code –768 kHz 0 +768 kHz
• 8 input bits → 32 mother → 12 transmitted
• The first 1536 bits of the symbol are assigned to carriers in a
• Transmit: 9988 9888 9988 9888 9888 9888
pseudo random sequence (same for all symbols).
• Code for each service defined in FIC • The next 1536 bits use the same sequence.
– FIC itself always uses rate 1/3 code • Each carrier gets 2 bits (0 Hz carrier is not used)
• Prevents fading causing burst errors

Jan 2006 26 Jan 2006 29

Unequal Error Protection DQPSK Modulation

• Some audio code bits are much more critical • If xk, yk ∈ {0,1} are the bits that map onto yk
10 00
than others carrier k, then the complex amplitude for
symbol n of a transmission frame is: xk
– e.g. bit allocation, scale factors, samples, text
• Predefined unequal protection rates ak (n) = ak (n − 1) × (1 − 2 xk + (1 − 2 yk ) j ) / 2 11 01

• Example: 128kbps UEP level 3 • All carriers have constant |ak(n)|=1

• Called ¼π DQPSK because phase increment is an odd
multiple of ¼π
• Worst case discontinuity at symbol boundary is 1.71
(instead of 2 for plain DQPSK without the ¼π)

Jan 2006 27 Jan 2006 30

Time Interleaving Transmitter Output

6040 bits / 24 ms = complex DAC ~
~
real ×
cos(ωt)
0 8 4 12 2 10 6 14 1 9 5 13 3 11 7 15 0 8
Zero
Pad
IFFT
Cyclic
Prefix imag ~ sin(ωt)
+ ~
~
~
• Makes data robust to burst errors 2048 @ 2552 ×
~
1537 2048

• Delay each bit by between 0 and 15 CIF frames

2.048 MHz DAC
~
– Delay between 0 and 360 ms
• 1537 carriers padded with zeros to 2048 for efficient IFFT
– Imposes a coding delay o at least 360 ms
• cyclic prefix added to complex IFFT output
– Requires memory in the receiver
• Real/Imag parts modulate cos(ωt) and sin(ωt)
• Delays of adjacent bits differ by ≥ 4 × 24 ms • Bandpass filtered to remove sidelobes
– Adjacent bits are always in different transmission frames – –71 dB bandwidth = 1.94 MHz
• Not used for Fast Information Channel • DC carrier unused – difficult to control phase

SC5b: Digital Audio Broadcasting 5

Mike Brookes 26-Jan-06

Jan 2006 31 Jan 2006 34

DAB Receiver Receiver Demodulation

888 MHz 30.72 MHz
(fixed)

• Time-sync selects a 1 ms window for the FFT

jω offset t
• Freq-sync multiplies complex signal by e
220 MHz
1.139 GHz to correct for tuning errors and doppler shifts
919 MHz (Tunable) • CORDIC block calculates phase and amplitude for DQPSK
demodulation
Bandpass filters have
1.537 MHz bandwidth
Bosch D-FIRE design

Jan 2006 32 Jan 2006 35

Receiver Downsampling Channel Decoding

• Must reverse the channel coding:

• FFT (and later processing) need only process symbols

that contain the wanted service
• ADC sample freq of 24.6 MHz aliases 30.72 MHz to 6.144 MHz
• Complex band-pass filter passes 6.144 MHz ±768 kHz DAB channel • Time de-interleaving requires time delays of up to 15 CIF
• Downsample by 12 to give sample rate of 2.048 MHz with DAB frames (24 ms each) – needs memory
channel aliased down to DC
• Lowpass filter energy to detect Null symbol every 96 ms

Jan 2006 33 Jan 2006 36

Downsampling Viterbi Decoding

• Mother code is ¼ rate
– Punctured before transmission
• Receiver must unpuncture to
restore original mother code
–30 –6 0 2 6 18 24 30 42 (MHz)
but with “unknown” bits
• Constraint length of 7 ⇒ trellis has 26 = 64 states
• Analog filter removes images at 12 MHz spacing • Branch metric compares input 4 bit sequence with
• Digital sampling @ 24 MHz ⇒ 30 MHz aliased to 6 MHz “correct” value:
• Complex digital filter removes images at 2 MHz spacing – Cost for each bit = –1 if correct, +1 if wrong, 0 if punctured
– negative frequencies removed completely • Delay decisions for ≈32 bits
– use polyphase filter and combine with downsampling
• Re-encode and compare with input to estimate BER
• Downsample to 2 MHz ⇒ 6 MHz aliased to DC See Data Comms lecture 19

SC5b: Digital Audio Broadcasting 6

Mike Brookes 26-Jan-06

Jan 2006 37 Jan 2006 40

Soft Decisions Frame Synchronization

2 sync + 75 data symbols = 96 ms
• “hard decision” decoder uses branch metric of ±1
• Ideal Branch metric is log ( prob( z | x ) ) Null Symbol Null Symbol
– z is observed bit, x is “correct” bit
– adding and/or multiplying by a constant makes no difference | |2 ∫ – Null
Detect
Phase Lock Loop
(×2359296 )
÷512

• Can calculate ideal metric if you know the noise Delay

1 ms
10.417 Hz 24.576 MHz 48 kHz

characteristics:
– Flat Rayleigh fading with complex FFT output sn • Detect Null symbols by low energy in 1 ms
– Ideal branch metrics for the two QPSK bits are • (96 ms)–1 = 10.417 Hz × 2359296 = 24.576 MHz
( ) (
± ℜ sn sn*−1 and ± ℑ sn sn*−1 ) – Exact frequency multiplication is done using a phase lock loop
– 24.576 MHz is the ADC sample clock
– Use a 4-bit signed number to represent this
• 24.576 MHz ÷ 512 = 48 kHz audio sample clock
• Also finds approximate start of first symbol

Jan 2006 38 Jan 2006 41

Error Concealment Effect of Frequency Offsets

Δf
• Errors in MP2 bit allocation • FFT samples spectrum at
bits or high bits of scale multiples of 1 kHz
factors are catastrophic • If carrier frequencies have an
– CRC check words are included offset, Δf, then they will no
in the MP2 bit stream longer be orthogonal
– If these are wrong then sound – must measure Δf and
is muted FFT sample frequencies
compensate
– compensation can be combined with the FFT
• Divide Δf into integer and fractional multiple of 1 kHz
– Integer part ⇒ wrong carriers
– Fractional part ⇒ inter carrier interference

Jan 2006 39 Jan 2006 42

Synchronization Requirements Fine AFC

• The 48 kHz audio sample clocks must be identical in • Frequency error of Δf ⇒ additional phase shift between
transmitter and receiver (long term average) successive symbols of Δφ = 2πΔfT
– otherwise receiver will have too many/few samples – T is symbol period = 1.246 ms
• At the input to the FFT, the carrier frequencies must be • Phase shift for each carrier should be (¼+½k)π in the
integer multiples of 1 kHz (= sample freq/2048) absence of noise
– otherwise the carriers will not be orthogonal – Find deviation from nearest correct value
– carrier frequencies are altered by doppler shifts – Form energy-weighted average of phase error over all carriers
• The FFT processing window must be timed to make the – calculate Δf
most constructive use of multipaths – apply correction before or during FFT calculation by multiplying
– In practice the FFT window aims to start at the end of the cyclic input signal by by exp(–2πjΔft)
prefix of the strongest received signal • Only works to within a multiple of ¼π

SC5b: Digital Audio Broadcasting 7

Mike Brookes 26-Jan-06

Jan 2006 43

Coarse AFC
• Transmitted phases of phase
reference symbol carriers are
known:
– Subtract transmitted phases from Phase ref symbol
FFT output and do inverse FFT
– Try lots of values of Δf in the
range ±8 kHz or so
• Subtract phases due to Δf
• Result is impulse response of Ideal impulse response
channel
– Pick Δf that gives the highest peak
• Position of peak indicates where
to put the end of the cyclic prefix
Δf (kHz)

Jan 2006 44

Benefits of DAB
• CD quality
• Mobile reception
• Spectral Efficiency
• European Standardisation
• Data as well as Audio
• Lower transmitter power
• Receiver features
– easy tuning
– pause

SC5b: Digital Audio Broadcasting 8