IEEE TRANSACTIONS ON BROADCASTING, VOL. 47, NO.

2, JUNE 2001 103

Soft Selection Combining for Terrestrial Digital

Audio Broadcasting in the FM Band
J. Nicholas Laneman, Student Member, IEEE and Carl-Erik W. Sundberg, Fellow, IEEE

Abstract—Methods of adaptive soft combining and channel

decoding are developed to combat the effects of multipath fading
and nonuniform interference channels, with particular applica-
tion to digital reception in hybrid in-band on-channel (HIBOC)
digital audio broadcast (DAB) systems in the FM band. These
systems transmit near CD quality digital audio and analog FM
simultaneously within the same license band, requiring the digital
audio to be protected with powerful channel codes and sophis-
ticated decoding algorithms to provide broad coverage under a
variety of fading and potentially severe interference conditions
created by first adjacent FM stations. In an example HIBOC DAB
system, digital transmissions are DQPSK/OFDM modulated in
two sidebands of the analog FM host signal, and a complementary
punctured pair convolutional (CPPC) inner coding scheme allows
for higher diversity benefit than code combining when both
sidebands are interference-free as well as full recovery of the
audio information when one of the sidebands is severely corrupted
by first adjacent interference. For the intermediate cases in
which one of the sidebands is partially useful, we demonstrate via
simulations that an unmodified receiver designed for a Gaussian
channel and corresponding to equal-gain combining performs Fig. 1. Frequency allocation at 100 MHz for analog FM host and DAB signals.
ineffectively for moderate to high interference levels. Motivated +
Also shown are two cases of a first-adjacent FM interferer at 100.2 MHz: 0 dB
by a clear need for more sophisticated receivers, we examine soft +
and 19 dB above the DAB signal.
combiners derived from the maximum-likelihood principle and
provide simulated performance bounds for the case in which per-
fect channel parameter estimates are available. We then discuss broadcasters require an intermediate solution in which the
more practical methods for performing adaptive soft combining analog and digital signals are transmitted simultaneously within
and channel decoding, focusing in particular on an appealing the same license band. Such systems are called hybrid, in-band
soft selection combining technique, based on successive erasures on-channel (HIBOC) DAB systems, and similar trends are
and Viterbi decoding, that requires only coarse estimates of the progressing in the AM band. For a thorough overview of the
channel parameters. An outer code used for error concealment
may be further utilized to perform the selection function. The per-
emerging all-digital and HIBOC DAB systems, see [1]–[7].
formance of this soft selection combining scheme under a variety Providing broad coverage for digital transmissions in a
of interference scenarios is also evaluated via simulation. Further HIBOC FM system is quite challenging due to a harsh channel
improvements may be obtained with a list Viterbi decoder. environment exhibiting multipath fading as well as FM inter-
Index Terms—Digital audio broadcasting (DAB), diversity ference induced by system constraints. Specifically, to prevent
combining, fading channel, interference mitigation, selection significant distortion in conventional analog FM receivers, the
combining, successive erasure decoding. DAB signal in currently prevailing standards is transmitted in
two sidebands of the analog FM host signal, as shown in Fig. 1
I. INTRODUCTION at 25 dB below the analog FM host in order to remain within
the FCC broadcast mask. From the point of view of the DAB

D IGITAL AUDIO BROADCAST (DAB) in the FM radio

band promises near CD-quality audio, data services, and
more robust coverage than existing analog FM transmission.
signal, first-adjacent analog FM interference, also shown in
Fig. 1, can be the most devastating of the numerous channel
impairments, because current FCC regulations allow this
However, before all-digital broadcasting can be made a reality, interference to be up to 19 dB above the DAB signal. Indeed,
even though first-adjacent analog FM interference is typically
Manuscript received May 10, 2000. present in only one digital sideband, if at all, this interference
J. N. Laneman was with the Multimedia Communications Research Labora- can lead to catastrophic results in improperly designed digital
tory, Bell Labs, Lucent Technologies, Murray Hill, NJ 07974 USA and is now receivers.
with the Research Laboratory of Electronics, Massachusetts Institute of Tech-
nology, Cambridge, MA 02139 USA. This work focuses on designing HIBOC DAB receivers to
C.-E. W. Sundberg was with the Multimedia Communications Research Lab- combat the effects of multipath fading and, more importantly,
oratory, Bell Labs, Lucent Technologies, Murray Hill, NJ 07974 USA and is analog FM interference. Although we restrict our attention to
now with the Media Signal Processing Research Department, Agere Systems,
Murray Hill, NJ 07974 USA. this application in the FM band and a particular singlestream
Publisher Item Identifier S 0018-9316(01)08403-7. transmission system under consideration, we emphasize that
0018–9316/01$10.00 © 2001 IEEE
104 IEEE TRANSACTIONS ON BROADCASTING, VOL. 47, NO. 2, JUNE 2001

Fig. 2. Block diagram for the digital portion of a hybrid, in-band on-channel Fig. 3. Partitioned receiver for adaptive signal combining and channel
(HIBOC) digital audio broadcast (DAB) system. decoding.

our results are useful in broader contexts when channels ex- subcarriers over 400 kHz bandwidth, with 80 subcarriers per
hibit nonuniform interference. We develop signal combining re- digital sideband, and transmitted symbols are differentially
ceivers consisting of an adaptive weighting module followed by quadrature phase-shift-keyed (DQPSK) in frequency, with
channel decoding via the Viterbi algorithm. Our receivers per- one pilot tone in each sideband serving as a phase reference.
form interference mitigation in contrast to interference cancella- Interleaving and placement of the code bits into one of ten
tion such as, e.g., explicit analog demodulation and subtraction frequency subbands in the dual sideband rate
of the interfering signals. Even when such cancelers are em- code is optimized so that the more important bits, in terms of
ployed, receivers may utilize our interference mitigation tech- contributing to the free distance of the individual rate
niques to combat the residual interference at the output of the codes on each sideband, are placed in the inner frequency
canceler. bands, farther away from the potential first-adjacent analog FM
An outline of this paper is as follows. In Section II, we interference carrier frequencies.
describe at a high level one singlestream HIBOC DAB system, In the following two sections, we develop adaptive com-
leaving details of the particular channel codes as well as bining and channel decoding methods for receivers partitioned
channel modeling and simulation issues to Appendices A as shown in Fig. 3. The received OFDM-demodulated subcar-
and B. We introduce a partitioned receiver consisting of an rier signals are differentially demodulated in frequency,
adaptive weighting module followed by a Viterbi-based signal and the differential demodulator outputs are multiplied by
combiner and decoder. Sections III and IV go on to design adaptive combiner weight sequences . Following dein-
several combiner weighting schemes and evaluate their relative terleaving, the Viterbi algorithm performs signal combining
performance via simulation. Finally, Section V ends with some and decoding. This partitioning of the receiver is convenient
discussion of the results and concluding remarks. because design of the adaptive combiner and decoder reduces
to the design of the adaptive combiner weights; these weights
II. SYSTEM MODEL appropriately modify the received signals so that an unmodified
Fig. 2 shows the structure of the digital portion of the HIBOC Viterbi algorithm for the Gaussian channel can be utilized for
DAB system considered in this paper. Audio signals are com- signal combining and decoding. Moreover, multiple replicas of
pressed by the Lucent Perceptual Audio Coder (PAC) [8] and this partitioned receiver can be utilized in multistream HIBOC
encoded with a cyclic-redundancy check (CRC) error detecting DAB systems [7].
block code to provide a flag mechanism for error mitigation in Throughout this paper, we evaluate the performance of our
the PAC decoder. As Fig. 2 indicates, we focus in this paper on combining methods for the singlestream HIBOC DAB system
the channel coding and modulation subsystem of the HIBOC over the “Strong-Echo” multipath fading channel model, a
DAB system. three-ray model with 0 dB echoes at 10 and 20 s, and vehicle
Appendix A outlines the various channel distortions and pro- speed of 88.5 km/hr. This channel model is described in more
vides details of the coding and modulation subsystem for one detail in Appendix B. We measure the performance of the
possible singlestream HIBOC DAB system. We develop a con- channel coding and modulation system in terms of the bit-error
venient discrete-time, baseband equivalent multicarrier channel rate at the output of the Viterbi algorithm as a function of
model of the form the average received subcarrier signal-to-noise ratio (SNR)
per symbol, or equivalently for our low-rate situation with
(1) the fading appropriately normalized, the ratio . Since
audio quality is the ultimate measure of system performance,
for and , for simulation is a reasonable target bit-error rate for the coding
purposes. Here is the transmitted QPSK symbol of energy and modulation system, because the breakdown regime for the
, and with variance , with variance , and audio coder occurs near for the AWGN channel
with variance are the fading coefficient, first-adjacent analog [9], [10], [11].
FM interference, and additive noise, respectively, in subcarrier Useful performance bounds for this system are shown
at sample time . in Fig. 4. We take as our lower bound on bit-error rate the
The system described in Appendix A employs a combined performance of the dual sideband rate CPPC code
rate complementary punctured-pair convolutional with fading, additive noise and no interference. It is clear
(CPPC) inner coding scheme [10], in which the CRC-encoded that first-adjacent analog FM interference in one sideband
data is encoded by two complementary rate , memory degrades performance, but our combining schemes are capable
punctured convolutional codes and transmitted over of suppressing the interference and approaching this curve. As
the respective sidebands. The modulation method is orthogonal an upper bound on the bit-error rate, we take the performance
frequency-division multiplexing (OFDM) in a total of of either of the single sideband rate CPPC codes with
LANEMAN AND SUNDBERG: SOFT SELECTION COMBINING FOR TERRESTRIAL DIGITAL AUDIO BROADCASTING IN THE FM BAND 105

symbol of the th subcarrier, namely, , which requires

as a phase reference, where in the lower sideband, and
in the upper sideband. For simplicity of exposition,
we suppress the time index in the following development.
We may appropriately rotate our perspective so that the
transmitted pilot tone takes the value , and write the
adjacent transmitted symbols for a particular phase difference
in vector form as

(3)

where, for QPSK, , . We also write

the fading in vector form as

(4)

and the interference plus noise in vector form as

(5)
Fig. 4. Performance bounds for combining methods on the “Strong-Echo”
multipath fading channel. The lower bound on bit-error rate corresponds to the
dual sideband R = 4=10 CPPC code with fading and no interference, while Letting , the received vector for a particular
the upper bound corresponds to the single sideband R = 4=5 CPPC code with value of is
fading, modeling severe interference in one sideband.
(6)
fading and no interference. This curve corresponds to situations
When the fading is Rayleigh and perfectly corre-
with a very strong first-adjacent analog FM interferer, which
lated in two adjacent subcarriers, is a complex-valued,
the receiver knows exists and so essentially ignores (erases) an
circularly-symmetric Gaussian random vector with mean and
entire sideband. Any acceptable adaptive combining method
covariance matrix
should do at least as well as this upper bound.

III. SOFT COMBINING

This section develops several continuous-valued combiner In practice, the fading varies between adjacent subcarriers
weights, leading to receivers that we collectively refer to as soft due to the multipath propagation inherent in the channel, but
combiners. We first show that a maximum-likelihood (ML) re- our model is reasonable if the OFDM symbol time is large
ceiver, which is optimal in terms of minimizing sequence error relative to the delay spread of the channel. Moreover, taking
probability, performs soft combining via the Viterbi algorithm this variation into account requires more detailed estimates
applied to appropriately weighted outputs of a conventional dif- of the fading channel statistics, which may be impractical
ferential demodulator, thus motivating our partition of the re- to acquire. We also conservatively model the interference
ceiver as in Fig. 3. Because the ML combiner weights rely on sequences as mutually independent, circularly-symmetric
accurate estimates of the various channel parameters, we also Gaussian white-noise sequences. Consequently, the vector is
consider several other soft combiners employing the Viterbi al- also a complex-valued, circularly-symmetric Gaussian random
gorithm and alternate weights that may be more attractive for vector with mean and covariance
implementation purposes. Finally, we compare the performance
of the various soft combiners via simulations.

A. Maximum-Likelihood Combining
Furthermore, we assume that and are independent.
With two-symbol (conventional) differential demodulation at Under these assumptions, given a particular transmitted se-
the receiver, the optimal soft combiner based on the maximum- quence is also a complex Gaussian random vector, having
likelihood principle uses the combiner weights mean and covariance

(2)
The likelihood function is therefore
as we now explain using a formulation similar to [14].
To show that (2) corresponds to the appropriate weighting for (7)
the branch metrics of the Viterbi algorithm, consider uncoded
DQPSK transmissions over the OFDM channel specified by (1). where “ ” denotes conjugate transpose. It is straightforward
Without loss of generality, we examine demodulation of the th to show that is independent of , and that the
106 IEEE TRANSACTIONS ON BROADCASTING, VOL. 47, NO. 2, JUNE 2001

maximum-likelihood rule is equivalent to choosing the phase

index to maximize

(8)

We have purposely written (8) so that the ML combiner weight

(2) is evident. Clearly, the weight plays no role in an uncoded
system; however, the result (8) indicates that the ML soft com-
biner may be partitioned as in Fig. 3. The combiner weights
are crucial for reducing the impact of nonuniform interference
in a coded system, because the Viterbi algorithm chooses the
largest path metric consisting of a sum of branch metrics, each
in the form of the metric in (8). Quite intuitively, the denomi-
nator of the ML combiner weights measure the nonsignal en-
ergy, involving energy in the interference and noise as well as
in the cross-terms involving signal, interference, and noise. The
ML combiner weights attenuate subcarriers with higher levels
of interference, thereby reducing their impact on the decision Fig. 5. Performance of maximum-likelihood (ML) combining on the
making about sequences. “Strong-Echo” multipath fading channel. Successively lower solid curves
Of course, implementation of the ML combiner weights (2) + + + 0 0
correspond to interference-to-signal ratios of 10, 5, 0, 5, and 10 dB,
respectively. The dashed curves correspond to the performance bounds from
requires accurate estimates of the fading variance , the noise Fig. 4.
energy , and the interference energies . Furthermore, the
random sequences involved are rarely long-term stationary, so
that some form of adaptation on the weights is also necessary.

B. Equal-Gain Combining and Inverse-Energy Combining

In the absence of analog interference, the combiner weights
are no longer necessary, or equivalently, the combiner uses the
weights

(9)

for all and . Though suboptimal in general, this equal-gain

combiner may be appealing when the interference levels are suf-
ficiently low and accurate estimates of the channel parameters
are difficult to obtain. Moreover, this approach serves as a useful
baseline for comparison with other methods.
As a third alternative, an inverse-energy combiner weights the
data by the inverse product of the total energies in the adjacent
received sequences and via the weight
Fig. 6. Performance of equal-gain combining on the “Strong-Echo”
multipath fading channel. Successively lower solid curves correspond to
(10) + + + 0 0
interference-to-signal ratios of 10, 5, 0, 5, and 10 dB, respectively.
The dashed curves correspond to the performance limits from Fig. 4.
Clearly, we see that excess energy caused by the interference
results in a smaller weight; however, simulations are required for low to moderate interference levels, only ML combining re-
in order to determine the relative effectiveness of the various mains useful for very high interference levels.
schemes. We emphasize that the simulations generating Figs. 5 and 7
use error-free estimates of the channel parameters , , and
C. Performance Comparison in the weight equations; consequently, these results should
We examine the bit-error rate performance for the various be viewed as lower bounds on bit-error rates for comparing
schemes in several interference scenarios, parameterized by the the various methods. In a practical system, noisy estimates are
(peak) interference-to-signal ratio at the first-adjacent analog derived from the received signal and pilot tones, and the ef-
FM interferer carrier frequency. The results for ML, equal-gain, fects of estimation error on the performance of the system must
and inverse-energy soft combining are shown in Figs. 5–7, re- be taken into account. Ad hoc methods for approximating the
spectively. These results indicate that equal-gain combining is ML combiner weights (2) through time- and frequency-domain
not effective for moderate to high interference levels, clearly smoothing of parameter estimates are described in [15], with
demonstrating the need for suppressing the first-adjacent inter- application to the case of moderate interference levels after im-
ference. While inverse-energy combining shows some promise perfect first-adjacent interference cancellation. Estimating the
LANEMAN AND SUNDBERG: SOFT SELECTION COMBINING FOR TERRESTRIAL DIGITAL AUDIO BROADCASTING IN THE FM BAND 107

Fig. 8. Conceptual block diagram of a soft selection combining receiver for

CPPC codes.

Fig. 7. Performance of inverse-energy combining on the “Strong-Echo”

multipath fading channel. Successively lower solid curves correspond to . The higher the value of , the more sub-
+ + + 0 0
interference-to-signal ratios of 10, 5, 0, 5, and 10 dB, respectively. bands that have been selected. For example, if the interferer lies
The dashed curves correspond to the performance limits from Fig. 4.
in the upper sideband, corresponds to selecting the
subbands in the lower sideband as well as the subbands G and
inverse-energy combiner weight (10) appears to be more prac- G in the upper sideband, and erasing subbands G , G , and
tical; for example, simulations of an exponential-memory en- G in the upper sideband. Note however that there is no rate
ergy estimator indicate that we can achieve the performance normalization for these punctured codes: the transmitted energy
curves shown in Fig. 7. per bit remains the same, but the receiver actually ignores some
of this energy in highly corrupted subcarriers.
IV. SOFT SELECTION COMBINING: SOFT COMBINING AFTER Fig. 8 shows a conceptual block diagram of a soft selection
ERASURES combining receiver. When an error is detected in the audio block
In the previous section, we observed that effective soft by the outer CRC code, a flag signal is generated to initiate
combining methods require detailed estimates of the channel the soft selection combining algorithm. If the pilot tones on
parameters for each of the OFDM subcarriers. In this section, either side of host signal indicate the presence of an adjacent
we consider selection combining methods that simply ignore channel interference signal, e.g., in the upper DAB sideband,
or erase OFDM subcarriers containing too much interference. the bits corresponding to the tones in the outermost subband
Specifically, we set of the OFDM frequency subcarriers are erased, and a new de-
coding attempt is performed as if a rate CPPC code
or (11) was used rather than a rate code for the full band. If
the second CRC is satisfied, the block of audio bits is passed on
depending on the energy in the interference . These to the audio decoder. If the second CRC fails, further erasures
methods require less accuracy in estimates of the channel and alternative decodings are performed. In principle, succes-
parameters, and therefore may be more effective than the soft sive erasure decoding with up to 10 attempts can be done for
combining methods in practice. More generally, soft combiner all blocks. However, the most efficient way of operating the de-
weights may be applied to the selected subcarriers, but a coder appears to be in a learning fashion, in which pilot tone
thorough investigation of such hybrid approaches is beyond the measurements indicate the presence and strength of the first ad-
scope of this work. jacent interferer, and successive erasures generate several likely
We introduced a crude form of selection combining when alternatives. If none of these alternative decodings satisfy the
we discussed the performance limits in Section II, namely, the CRC, the soft selection combining receiver generates a flag to
scheme yielding the worst case performance in Fig. 4. For this trigger the error mitigation unit in the audio decoder. We should
case, when the interference level is exceedingly high in one side- point out that these successive erasure techniques are applicable
band, all of the combiner weights for that sideband are set to when other coding schemes, e.g., code combining with identical
zero. When interference levels are moderate, refined selection codes in each sideband, or other modulation formats, e.g., single
combining should prove more effective, because some of the carrier modulation in each subband, are used. Furthermore, this
inner subcarriers with less interference can be utilized in the de- approach may be readily combined with List Viterbi Decoding
coding process. Algorithms (LVAs) [11]–[13] to further improve performance
As one of many possible schemes, we consider selection com- at the cost of additional complexity. Finally, successive erasures
bining of the CPPC code subbands in Fig. 19. This approach of the outer subbands of both sidebands can also be done in the
can be equivalently viewed as a further puncturing of the CPPC unlikely event of experiencing (moderate) interference in both
code, and consequently we denote these schemes by , sidebands.
108 IEEE TRANSACTIONS ON BROADCASTING, VOL. 47, NO. 2, JUNE 2001

Fig. 11. Performance of soft selection combining on the “Strong-Echo”

Fig. 9. Performance of soft selection combining on the “Strong-Echo”
multipath fading channel with no first-adjacent FM interference. The +
multipath fading channel with 0 dB first-adjacent analog FM interference.
successively lower curves correspond to punctured code rates of R = ,=45 The successively lower (at high SNR) solid curves correspond to punctured
code rates of R = 4 10
= , 4/9, 4/8, 4/7, and 4/6, respectively. The dashed curves
4/6, 4/7, 4/8, 4/9, and 4/10, respectively.
correspond to the performance limits from Fig. 4.

Fig. 12. Performance of soft selection combining on the “Strong-Echo”

Fig. 10. Performance of soft selection combining on the “Strong-Echo” 0
multipath fading channel with 5 dB first-adjacent analog FM interference.
+
multipath fading channel with 5 dB first-adjacent analog FM interference. The successively lower (at high SNR) solid curves correspond to punctured
The successively lower (at high SNR) solid curves correspond to punctured code rates of R = 4 10
= , 4/9, 4/8, 4/7, and 4/6, respectively. The dashed curves
code rates of R=4 8= , 4/7 and 4/6, respectively. The dashed curves correspond correspond to the performance limits from Fig. 4.
to the performance limits from Fig. 4.

code rate performs better than ignoring an entire side-

Fig. 9 shows the performance for selection combining of all band for interference levels up to roughly 0 dB, while the selec-
six punctured code rates in fading and additive noise. These tion combiner of punctured code rate breaks down for
results constitute lower bounds for selection combining in the interference levels of roughly dB. These results suggest that
presence of interference. Figs. 10–13 show the performance re- soft selection combining as described above could achieve close
sults for selection combining of several punctured code rates to the minimum of the bit-error rates of the curves in Figs. 10–13
for fixed interference-to-signal ratios, and Figs. 14–16 show the at each SNR.
performance results for selection combining of fixed punctured For the soft selection combining algorithms to be successful,
code rates for several interference-to-signal ratios. We see from both the location of a possible first-adjacent interferer as
these results that increasing the effective code rate through fur- well as a rough estimates of the interference and noise power
ther puncturing degrades the performance when no interference levels may be obtained from training and pilot data, and the
is present, but makes the system more robust in the presence of appropriate punctured code rate can be selected. Finally, we
interference. For example, the selection combiner of punctured note that the results of this and the previous section suggest
LANEMAN AND SUNDBERG: SOFT SELECTION COMBINING FOR TERRESTRIAL DIGITAL AUDIO BROADCASTING IN THE FM BAND 109

Fig. 13. Performance of soft selection combining on the “Strong-Echo” Fig. 15. Performance for soft selection combining of punctured code rate
0
multipath fading channel with 10 dB first-adjacent analog FM interference. R = 4=7 on the “Strong-Echo” multipath fading channel with first-adjacent
The successively lower (at moderate SNR) solid curves correspond to punctured analog FM interference. The successively lower solid curves correspond to
code rates of R =4 6= , 4/7, 4/8, 4/9, and 4/10, respectively. The dashed curves interference-to-signal ratios of +10, +5, +0, 05, and 010 dB, respectively.
correspond to the performance limits from Fig. 4. The dashed curves correspond to the performance limits for this rate from
Fig. 9.

Fig. 14. Performance for soft selection combining of punctured code rate
R = 4=6 on the “Strong-Echo” multipath fading channel with first-adjacent Fig. 16. Performance for soft selection combining of punctured code rate
analog FM interference. The successively lower solid curves correspond to
interference-to-signal ratios of +10, +5, +0, 05, and 010 dB, respectively.
R = 4=8 on the “Strong-Echo” multipath fading channel with first-adjacent
analog FM interference. The successively lower solid curves correspond to
The dashed curves correspond to the performance limits for this rate from interference-to-signal ratios of +5, +0, 05, and 010 dB, respectively. The
Fig. 9. dashed curves correspond to the performance limits for this rate from Fig. 9.

that a hybrid approach would be quite appealing. Using fairly

code combining is obtained. In many cases, one of sidebands is
coarse channel estimates, we can select subbands in which
exposed to various levels of first adjacent analog FM interfer-
first-adjacent analog interference is manageable, and we
ence, the level depending upon, among other things, the loca-
can use soft combiner weights in these selected subbands to
tion of the receiver in the coverage area. In the extreme case,
improve upon the equal-gain results in this section.
one of the sidebands is exposed to such high interference levels
that the information therein is completely jammed, and if the re-
V. DISCUSSION AND CONCLUSIONS ceiver does not account for the interference and instead employs
HIBOC DAB systems in the FM band with complementary equal-gain combining, the combined full bandwidth channel
punctured pair convolutional (CPPC) channel codes transmit code becomes overwhelmed by the interference. If the receiver
identical source information in two separate sidebands of the identifies the presence of this strong interferer via pilot tone
analog FM host signal, so that, when both sidebands are free measurements and falls back to the half bandwidth code, much
from interference, an overall code that is more powerful than less frequency diversity gain is achieved. For the intermediate
110 IEEE TRANSACTIONS ON BROADCASTING, VOL. 47, NO. 2, JUNE 2001

subcarriers, has been chosen for HIBOC DAB systems in this

environment for several reasons. OFDM is often chosen for
time dispersive channels to avoid complex adaptive equalizers,
but in this particular application, it provides a convenient
framework for tailoring the receiver to handle the nonuniform
adjacent channel interference, which is a potentially devastating
channel distortion.
We assume a discrete Fourier transform (DFT) based OFDM
scheme consisting of subcarriers in a bandwidth of
400 kHz centered about the carrier frequency of the analog
FM host. This choice results in an OFDM subcarrier bandwidth
Fig. 17. Block diagram for the channel coding and modulation subsystem of of 781.25 Hz, or equivalently, an OFDM symbol time of
the singlestream HIBOC DAB system. ms. Since each sideband of the analog FM host has roughly
70 kHz available, we may safely utilize 90 OFDM subcarriers
cases in which the corrupted sideband is partially useful, adap- for digital transmission; in fact, we utilize only 80 OFDM sub-
tive combining techniques are required, yielding overall system carriers in each sideband in the present simulated system.
results in between the two extremes. We have outlined several For convenience, we utilize a discrete-time, baseband equiv-
such techniques and illustrated their performance by means of alent multicarrier channel to model the effects from the input
simulation. We have developed a constructive method based on of the OFDM modulator to the output of the OFDM demod-
selection of useful fractions of a corrupted sideband, and our ulator. In this model, received OFDM-demodulated subcarrier
simulations have shown promising results requiring only coarse sequences are written as
estimates of the interference environment. (12)
Additional simulations with other delay profiles and vehicle
speeds are necessary for a more complete evaluation of these Here, indexes OFDM subcarriers (fre-
algorithms. Further work in several areas may lead to improved quency), with corresponding to roughly kHz
overall system performance; for example, multiple symbol dif- and corresponding to roughly kHz, and
ferential detection [16] or even coherent modulation should in indexes OFDM symbols (time). In the sidebands, corre-
general yield better performance at a price of additional com- sponding to indices for the lower sideband
plexity. Our results above are given for systems without a first and for the upper sideband, the digital
adjacent interference canceler, but they can be interpreted as re- data in (12) are QPSK sequences with energy , differ-
sults with a nonideal canceler with residual interference. Well entially modulated across OFDM subcarriers. To accomplish
functioning first adjacent interference cancelers should improve the differential modulation, pilot sequences with energy are
coverage. inserted at subcarriers indexed by and . Finally,
for , the inputs .
APPENDIX A The next three subsections describe the models used to gen-
SINGLESTREAM HIBOC FM SYSTEM DETAILS erate the fading , interference , and noise se-
quences in (12).
Shown in Fig. 17, the channel coding and modulation
1) Fading Generation: The fading coefficients in (12)
subsystem of one singlestream HIBOC DAB system consists
correspond to the short-time discrete Fourier transform of the
of several modules in tandem: a CPPC encoder/decoder;
time-varying multipath fading channel response, and are gener-
a time- and frequency-interleaver/deinterleaver; a
ated according to
differential modulator/ demodulator; and an OFDM
modulator/demodulator. The adaptive combiner weighting
module, the focus of this paper, serves to compensate for the (13)
effects of fading and analog FM interference. In this appendix
we outline the various channel distortions and restrictions, where
and develop a convenient discrete-time, baseband equivalent is a set of mutually independent, complex-
multicarrier channel model for simulation purposes. We then valued, circularly-symmetric Gaussian random
highlight several important components of the coding and sequences;
modulation subsystem that have been designed with these is a set of fixed multipath delays in seconds;
channel impairments in mind. is a normalization for the OFDM symbol period
in seconds;
A. Discrete-Time, Baseband-Equivalent Channel Model is a normalization for the number of OFDM sub-
Transmissions in the FM band, whether analog or digital, are carriers in an OFDM frame.
subject to such distortions as receiver thermal noise, multipath Appendix B addresses the issue of accurate simulation of
propagation and signal fading, and adjacent channel interfer- .
ence from other broadcasts. Orthogonal frequency-division 2) Interference Generation: The interference in
multiplexing (OFDM), with DQPSK modulation across (12) is the frequency content of the first-adjacent analog FM
LANEMAN AND SUNDBERG: SOFT SELECTION COMBINING FOR TERRESTRIAL DIGITAL AUDIO BROADCASTING IN THE FM BAND 111

interference. To simulate this interferer, we FM modulate a

narrowband Gaussian sequence1 according to

(14)

where the gain determines the power of the FM interference,

kHz is the sampling frequency or system bandwidth,
MHz is the interferer carrier frequency, and
kHz is the maximum frequency deviation.2 Fig. 18. CPPC code encoder. The outputs of a R = 1=3 “mother”
convolutional code are divided among the two sidebands through
Then is formed via the short-time discrete Fourier trans- complementary puncturing operations of period four. A “1” in the puncturing
form of , i.e., matrix indicates that the particular code bit is present in the output.

(15)

for in the digital sideband of interest; otherwise, .

For example, in all of our simulations, the first-adjacent inter- Fig. 19. Optimal code bit placement into subbands of OFDM subcarriers. The
ferer is assumed to lie in the upper sideband, so (15) applies for subscripts indicate the index of the code bit in the puncturing period. For our
, and for . system, there are 80 subcarriers per sideband, each with two bits per symbol,
and five code bits per sideband; thus, the groups of subcarriers for each code bit
We treat the interference in each OFDM subcarrier as it if were contain 32 subcarriers.
additive Gaussian noise of variance .
3) Noise Generation: The receiver thermal noise is assumed
to be white Gaussian noise with power spectral density . determined through exhaustive search in [10]. An attractive fea-
Because the DFT in the receiver is a unitary transformation, ture of these CPPC codes is that the partitioned bits can be re-
the noise sequences in (12) are mutually independent, combined and decoded using an appropriately modified Viterbi
complex-valued, circularly-symmetric Gaussian white-noise algorithm for the mother convolutional code. Though we only
processes, each with variance . consider equal error protection codes, these codes can be con-
figured for unequal error protection as well. A more compre-
B. Complementary Punctured-Pair Convolutional (CPPC) hensive treatment of CPPC codes and puncturing can be found
Codes in [10] and the references therein.
Anticipating the nonuniform first-adjacent FM interference
The channel effects modeled in the previous section can be apparent in Fig. 1, the authors in [10] also optimized the place-
quite devastating to uncoded transmissions. For example, if ment of the code bits into subbands of OFDM subcarriers. Intu-
the first-adjacent analog FM interference is very strong, the itively, their result places the most important code bits, in terms
entire upper sideband is completely jammed. Consequently, of contributing to the free distance of each rate code,
information bits are protected from channel errors by means on the innermost OFDM subcarriers. The optimal bit placement
of complementary punctured-pair convolutional (CPPC) codes is shown in Fig. 19. The particular memory CPPC codes
[10]. These codes essentially repeat the digital information in used in this paper are merely used as examples. For further de-
both sidebands to provide diversity, but send different coded tails on optimum CPPC codes under a variety of conditions, see
streams to perform better than code combining—sending the [10].
same coded streams over the two sidebands—when the two
sidebands are clean.
C. Time- and Frequency-Interleaving
For the proposed system, the overall channel coding rate is
, allowing for in each sideband. For memory For the CPPC code of a given constraint length to be most
codes, the dual-band code has free distance , effective over a fading channel, sufficient interleaving must
and the single-band code has free distance . (Code com- be introduced in order to make the channel appear memo-
bining for the dual-band case has free distance .) Fig. 18 ryless. Furthermore, because the fading is usually time- and
shows how the CPPC coded streams are generated from the frequency-selective in the FM band, it is helpful to interleave
memory , “mother” convolutional encoder, as in both time and frequency. Following the approach in [15],
well as the puncturing operations with period four, which were we interleave within each code bit subband established by the
optimal bit placement of the previous section. In our case, the
subbands consist of 32 bit streams corresponding to 16 OFDM
1The sequence m[n] is a Gaussian white-noise sequence with variance 1/9 subcarriers. We specify the minimum time separation
appropriately filtered to bandwidth of 15 kHz. between two consecutive code bits, and choose the overall
2Normally, f = 75 kHz; however, to simplify the computational complexity
depth of the interleaver to be some multiple of . Then
of simulating the spectrum of a commercial analog FM system, we have used
a mono message signal and increased f to achieve a symmetric spectrum with the interleaver can be interpreted as filling a matrix of size
0
slope of 0.35 dB/kHz for frequencies above the carrier. -by-32.
112 IEEE TRANSACTIONS ON BROADCASTING, VOL. 47, NO. 2, JUNE 2001

TABLE I
“STRONG ECHO” CHANNEL MULTIPATH INTENSITY PROFILE. OTHER
CHANNEL PARAMETERS INCLUDE: CARRIER FREQUENCY 100 MHz, AND
MOBILE VELOCITY 88.5 km/hr

Fig. 20. Differential demodulation across OFDM subcarriers. Also shown is

[]
the adaptive weighting operation, multiplication by w n . Due to the placement
= 80
of the pilot tones at subcarriers k and k = 431 , for the lower sideband,
= 0 1 . . . 79 = + 1
i.e., k ; ; ; ,l k , while for the upper sideband, i.e., k=
432 433 . . . 511 = 0 1
; ; ; ,l k .

TABLE II
Consider a block of code bits of size indexed by EIA “URBAN FAST” MULTIPATH INTENSITY PROFILE. OTHER CHANNEL
within a given partition. The interleaver places PARAMETERS INCLUDE: CARRIER FREQUENCY 94.1 MHz, AND
MOBILE VELOCITY 60 km/hr
code bit at position in the matrix, where

(16)

(17)

The deinterleaver simply inverts this mapping between the ma-

trix and the code sequence. For all of the simulations in this
paper, , and for comparisons with [15]. This
depth translates into slightly over 300 ms when we multiply by
the OFDM symbol period. It is anticipated that longer inter-
leavers are required to deal with slow fading.
Gaussian random sequences; is a set of fixed multipath
D. Differential Encoding delays in seconds; is a normalization for the OFDM symbol
Following coding and interleaving, the transmitter modulates period in seconds; and is a normalization for the number
the OFDM subcarriers in QPSK format, and differentially en- of OFDM subcarriers in an OFDM frame. Tables I and II
codes across OFDM subcarriers, to provide some inherent ro- list multipath intensity profiles and other parameters for two
bustness to frequency-selective fading and oscillator phase drift. channels of interest to us.
The corresponding differential demodulator in the receiver is The carrier frequencies and mobile velocities, as in Tables I
shown in Fig. 20. For simplicity, we consider differential de- and II, determine the maximum Doppler rate, which in turn
modulation based on only two adjacent received symbols be- parameterizes Jakes’s model [17] for the temporal correlation
cause this case allows us to employ the Viterbi algorithm for of the individual multipath gains . We now discuss two
decoding the CPPC code. Multiple-symbol differential detec- methods for generating sequences with the appropriate correla-
tion methods [16] can offer coding gains of up to 3 dB over this tion structure. The first approach is based on an algorithm given
approach, but are considerably more complex. Also shown in in [17] for simulating a single multipath gain sequence. We have
Fig. 20 is the combiner weighting operation, multiplication by observed that this approach produces inconsistent simulation re-
. These weights essentially modify the branch metrics in sults for fading channels with nonuniform multipath intensity
the Viterbi decoder according to various criteria, as explored in profiles, such as the EIA “Urban Fast” channel of Table II. We
Sections III and IV. then discuss another method based on filtering white Gaussian
sequences to generate these multipath gain sequences, for which
APPENDIX B more consistent simulations results have been obtained.
MULTIPATH RAYLEIGH FADING CHANNEL SIMULATION
As we said in Appendix A, we utilize a baseband A. Modified Jakes’s Algorithm
equivalent, discrete-time multicarrier model for time- and Jakes [17] provides an algorithm for generating a single mul-
frequency-selective Rayleigh fading channels, in which the tipath gain sequence with the appropriate correlation, but it is
QPSK symbol at time in OFDM subcarrier is multiplied not clear from his discussion how the algorithm can be used to
by the fading coefficient generate multiple, independent sequences, each approximating
the temporal correlation of his model.
(18) Generating independent sequences by initializing the algo-
rithm with random initial conditions for each tap appears to
work reasonably well for the “Strong Echo” channel of Table I.
Here: is a set of multipath gains modeled as For fading channels with nonuniform multipath intensity, we
mutually independent, complex-valued, circularly-symmetric found that the channel simulator needed further modifications.
LANEMAN AND SUNDBERG: SOFT SELECTION COMBINING FOR TERRESTRIAL DIGITAL AUDIO BROADCASTING IN THE FM BAND 113

Fig. 21. Filtering method for generating the fading coefficients in (18).

TABLE III
COEFFICIENTS OF THE FILTERING METHOD SIMULATOR FOR THE EIA “URBAN
FAST” FADING CHANNEL
Fig. 22. Comparison of the Jakes’s Doppler spectrum and the filter designed
for the filtering method simulator for the EIA “Urban Fast” channel.

the relative gain between zero and the maximum Doppler fre-
quency is close to that of the Jakes’s spectrum.
Simulation results suggest that the filtering method is suit-
B. Filtering Method able for the multistream systems. We note, however, that results
Statistical examination of the multipath gain sequences gen- from the filtering method fading simulator may exhibit slightly
erated by the modified Jakes algorithm for the nonuniform mul- faster temporal variations (and thus potentially higher coding
tipath intensity profile above indicates that the sequences are gains) due to the components at frequencies higher than the
not independent, violating one of our assumptions about the maximum Doppler frequency; however, these gains have been
channel. To ensure independence, we consider another simula- slight (around 0.2 dB for the bit-error rates of interest) in the
tion approach whereby each multipath gain sequence is several cases we have examined.
generated by filtering a white, complex-valued, circularly sym-
metric Gaussian random sequence . Fig. 21 shows how the ACKNOWLEDGMENT
fading coefficients of (18) are generated from filtering mutually
This work benefited from discussions with H.-L. Lou, D.
independent, white Gaussian sequences .
Sinha, M. Zarrabizadeh and V. Weerackody.
To simulate a multipath gain sequence with the appropriate
temporal correlation, we design the real-valued filter
in Fig. 21 so that its autocorrelation function REFERENCES
approximates the temporal correlation of Jakes’s model, or, [1] J. A. C. Bingham, “In-band digital audio radio: An update on AT&T
equivalently, its magnitude squared matches the Amati PAC/DMT solution,” in Proc. International Symposium on Dig-
ital Audio Broadcasting, Toronto, Canada, Mar. 1994, pp. 270–277.
corresponding power spectrum. Because the Jakes’s spectrum [2] N. S. Jayant, E. Y. Chen, J. D. Johnston, S. R. Quackenbush, S. M.
is narrowband, requiring excessive computation for accurate Dorward, K. Thompson, R. L. Cupo, J.-D. Wang, C.-E. W. Sundberg,
modeling with finite impulse response (FIR) filters, we use and N. Seshadri, “The AT&T in-band adjacent channel system for dig-
ital audio broadcasting,” in Proc. International Symposium on Digital
low-order, infinite impulse response (IIR) filters with rational Audio Broadcasting, Toronto, Canada, Mar. 1994, Conference Record,
spectra, so that with pp. 254–267.
[3] Second International Symposium on Digital Audio Broadcasting, The
Sound of 2000, vol. 1–2, Toronto, Canada, Mar. 1994, Conference
Record.
[4] N. S. Jayant and E. Y. Chen, “Audio compression: Technology and appli-
cations,” AT&T Technical Journal, vol. 74, pp. 23–34, Mar./Apr. 1995.
[5] B. W. Kroeger and A. J. Vigil, “Improved IBOC DAB technology for
Table III lists the coefficients of a second order ( ) filter AM and FM broadcasting,” in NAB Show Conference Record, Las
Vegas, NV, Oct. 1996, pp. 1–10.
for simulating the EIA “Urban Fast” channel, and Fig. 22 com- [6] C.-E. W. Sundberg, “Digital audio broadcasting in the FM band,” in
pares the magnitude squared to the corresponding ISIE’97 Conference Record, July 1997, pp. SS37–SS41.
Jakes’s spectrum. This filter can be conveniently obtained (and [7] C.-E. W. Sundberg, D. Sinha, D. Mansour, M. Zarrabizadeh, and J. N.
Laneman, “Multistream hybrid inband on channel FM systems for dig-
appropriately modified for other Doppler rates) from the filter ital audio broadcasting,” IEEE Trans. Broadcast., vol. 45, no. 4, pp.
design tool in MATLAB by selecting a second order, Cheby- 410–417, Dec. 1999.
shev Type I filter. We specify the passband cutoff frequency so [8] D. Sinha, J. D. Johnston, S. M. Dorward, and S. R. Quackenbush, “The
perceptual audio coder (PAC),” in The Digital Signal Processing Hand-
that the peak of the magnitude response falls at the maximum book, V. K. Madisetti and D. B. Williams, Eds: CRC/IEEE Press, 1998,
Doppler frequency, and we specify the passband ripple so that p. 42-1–42-17.
114 IEEE TRANSACTIONS ON BROADCASTING, VOL. 47, NO. 2, JUNE 2001

[9] D. Sinha and C.-E. W. Sundberg, “Unequal error protection (UEP) for Carl-Erik W. Sundberg (S’69–M’75–SM’81–F’90)
perceptual audio coders,” in Conf. Proc. ICASSP’99, Phoenix, AZ, April was born in Karlskrona, Sweden on July 7, 1943. He
1999, pp. 2423–2426. received the M.S.E.E. and Dr. Techn. degrees from
[10] B. Chen and C.-E. W. Sundberg, “Complementary punctured-pair con- the Lund Institute of Technology, University of Lund,
volutional codes for digital audio broadcasting,” IEEE Trans. Commun., Lund, Sweden, in 1966 and 1975, respectively.
vol. 48, no. 11, pp. 1829–1839, Nov. 2000. Currently he is a Distinguished Member of the
[11] , “List Viterbi algorithms for continuous transmission of concate- Technical Staff in the Media Signal Processing
nated cyclic redundancy check and convolutional codes,” IEEE Trans. Research Department at Agere Systems. He was
Commun., May 2001, to be published. with Bell Laboratories, Lucent Technologies,
[12] N. Seshadri and C.-E. W. Sundberg, “List Viterbi decoding algorithms Murray Hill, NJ from 1984 to 2000. Before 1976
with applications,” IEEE Trans. Commun., vol. 42, no. 2/3/4, pp. he held various teaching and research positions
311–323, Feb./Mar./Apr. 1994. at the University of Lund. During 1976, he was with the European Space
[13] C. Nill and C.-E. W. Sundberg, “List and soft symbol output Viterbi Research and Technology Centre (ESTEC), Noordwijk, The Netherlands, as
algorithms: Extensions and comparisons,” IEEE Trans. Commun., vol. an ESA Research Fellow. From 1977 to 1984 he was a Research Professor
43, no. 2/3/4, pp. 277–287, Feb./Mar./Apr. 1995. (Docent) in the Department of Telecommunication Theory, University of Lund,
[14] W. C. Dam and D. P. Taylor, “An adaptive maximum likelihood receiver Lund, Sweden. He has held positions as Consulting Scientist at LM Ericsson,
for correlated Rayleigh-fading channels,” IEEE Trans. Commun., vol. SAAB-SCANIA, Sweden, and at Bell Laboratories, Holmdel. His consulting
42, no. 9, pp. 2684–2692, Sept. 1994. company, SUNCOM, has been involved in studies of error control methods
[15] B. W. Kroeger and D. Cammarata, “Robust modem and coding tech- and modulation techniques for the Swedish Defense, a number of private
niques for FM hybrid IBOC DAB,” IEEE Trans. Broadcast., vol. 43, companies and international organizations. His research interests include
no. 4, pp. 412–419, Dec. 1997. source coding, channel coding, digital modulation methods, digital audio
[16] T. R. Giallorenzi and S. G. Wilson, “Noncoherent demodulation tech- broadcasting systems, fault-tolerant systems, digital mobile radio systems,
niques for Trellis coded M-DPSK signals,” IEEE Trans. Commun., vol. spread-spectrum systems, digital satellite communications systems, and optical
43, no. 8, pp. 2370–2380, Aug. 1995. communications. He has published over 90 journal papers and contributed
[17] W. C. Jakes, Microwave Mobile Communications. New York: IEEE over 130 conference papers. He has 62 patents, granted and pending. He is
Press, 1994. coauthor of Digital Phase Modulation, (New York: Plenum, 1986), Topics
in Coding Theory, (New York: Springer-Verlag, 1989) and Source-Matched
Digital Communications (New York: IEEE Press, 1996).
Dr. Sundberg has been a member of the IEEE European-African-Middle
J. Nicholas Laneman (S’93) was born in St. East Committee (EAMEC) of COMSOC from 1977 to 1984. He is a member
Charles, MO, USA. He received B.S. degrees in of COMSOC Communication Theory Committee and Data Communications
electrical engineering and in computer science Committee. He has also been a member of the Technical Program Committees
from Washington University, St. Louis, MO, in for the International Symposium on Information Theory, St. Jovite, Canada,
1995. He earned the S.M. degree in electrical October 1983, the International Conference on Communications, ICC’84, Am-
engineering in 1997 from the Massachusetts Institute sterdam, The Netherlands, May 1984, the 5th Tirrenia International Workshop
of Technology (MIT), Cambridge, MA, where he is on Digital Communications, Tirrenia, Italy, September 1991, the International
currently pursuing the Ph.D. degree. Telecommunications Symposium, ITS’94, Rio de Janeiro, Brazil, August 1994
Since 1995, he has been affiliated with the and for the 2000 International Zurich Seminar, Zurich, Switzerland, February
Department of Electrical Engineering and Computer 2000. He has organized and chaired sessions at a number of international
Science and the Research Laboratory of Electronics, meetings. He has been a member of the International Advisory Committee
MIT, where he has held a National Science Foundation Graduate Research for ICCS’88 to ICCS’98 (Singapore). He served as Guest Editor for the IEEE
Fellowship and served as both a Teaching and Research Assistant. During Journal on Selected Areas in Communications in 1988–1989. He is a member
1998 and 1999 we was also with Lucent Technologies, Bell Laboratories, of SER (Svenska Elektroingenjörers Riksförening) and the Swedish URSI
Murray Hill, NJ, both as a Member of the Technical Staff and as a Consultant, Committee (Svenska Nationalkommittén för Radiovetenskap). In 1986 he
developing robust source and channel coding methods for digital audio and his coauthor received the IEEE Vehicular Technology Society’s Paper of
broadcasting. He has 4 patents pending. His current research interests lie in the the Year Award and in 1989 he and his coauthors were awarded the Marconi
broad areas of communications and signal processing, with particular emphasis Premium Proc. IEE Best Paper Award. He is a fellow of the IEEE since 1990.
on resource-efficient wireless network algorithms and architectures. He is a He is listed in Marquis Who’s Who in America and Who’s Who in the World.
member of Eta Kappa Nu and Tau Beta Pi.