Designing Anaylsis and Synthesis Filterbanks

in GNU Radio
Thomas W. Rondeau Timothy J. OShea
University of Pennsylvania & Rondeau Research The Hume Center
Shelburne, VT Virginia Tech
[email protected] Arlington, VA
[email protected]

Index TermsGNU Radio, software radio, SDR, dy- sign tools. Use of these, however, is not necessarily
namic spectrum access, DSA, filters, polyphase filter- straight-forward.
banks, analysis filterbanks, synthesis filterbanks, dsp In this paper, we discuss the principles of using
Abstract the filterbank blocks, the FIR filter design tools, and
This paper provides a look at the polyphase analy- how to build the necessary prototype filters required
sis and synthesis filterbanks and filter design tools in to effectively use polyphase filterbanks in GNU Ra-
GNU Radio. The filterbanks are two powerful blocks
dio. Design decisions for each step are explored to
for performing signal detection, collection, and analy-
sis operations that are of increasing interest to the cog- enable further use of these tools and techniques.
nitive radio (CR) and dynamic spectrum access (DSA) As an extension to the filter design process, we will
communities. We look at how to design and use the also look into the process of building and using per-
filterbanks to channelize, synthesize, and reconstruct fect reconstruction filters in GNU Radio. A recon-
signals. We use a few simple examples for explanatory struction filter allows us to channelize spectrum with
purposes but which are readily available to experiment an analysis filterbank and then recombine any set of
with and learn from. The processes, descriptions, and
channels with a synthesis filterbank with no loss of
tools used in this paper are designed to allow others to
adapt filtering and filterbank tools of GNU Radio for
signals split between channels. By combining the
other, more complex purposes and applications. analysis and synthesis filterbanks like this, we can
create arbitrary bandwidth channelizers.
A companion webpage1 is available to download
I. I NTRODUCTION the flowgraphs and filters discussed in this paper.
Cognitive radio (CR) and dynamic spectrum ac-
cess (DSA) problems generally involve discovery and II. A SSUMPTIONS
then use or reuse of spectrum. After the identification As we cover the topics of polyphase analysis and
of available spectrum, the DSA system will seek to synthesis filterbanks in GNU Radio, we recognize
transmit and receive in it. These spectrum opportuni- that they are fairly advanced topics in signal process-
ties, however, may be close to other users, and so we ing. Using these tools requires quite a bit of knowl-
risk receiving or creating adjacent channel interfer- edge to appropriately set the parameters for an appli-
ence with other users. Even in the discovery process, cation. Here, we will give a very basic understanding
we desire efficient structures for isolating and testing of the theory and properties of the filterbank tools,
areas of spectrum. Polyphase analysis and synthesis but we will mostly leave the real theoretical details to
filterbanks are useful tools for these problems. And other papers referenced through the text that discuss
while there is a nice body of literature about how the these concepts in-depth. Instead, we want to provide
filterbanks themselves work, the design of the filter context about the use of these filterbanks in GNU Ra-
taps is still not as widely understood. dio to help enable their use elsewhere. As such, we
GNU Radio as a software radio platform provides a assume a basic literacy of the GNU Radio2 , such as
number of tools to build, use, and analyze polyphase the what it is, what it does, and the basic principles of
filterbanks for use in DSA systems. We have the prin- 1
http://www.trondeau.com/examples/2014/1/
ciple tools necessary to construct analysis and syn- 23/pfb-channelizers-and-synthesizers.html
2
thesis channelizers as well as a large suite of filter de- http://gnuradio.org
a flowgraph and a GNU Radio block. We also assume which means that each filter is processing a down-
a basic familiarity here with digital filtering [1]. sampled version of the signal. Because of the down-
When discussing specific GNU Radio blocks, we sampling of the signal, we design the prototype filter
mention the full C++ namespace of the block to make at the full signal sample rate. In a channelizing filter-
it easy to find the entry in the GNU Radio Manual3 . bank, the sample rate for the design process is the in-
We further assume that the radio front-end has suf- put rate, and for a synthesis filterbank, we work with
ficient selectivity and dynamic range in its out-of- the output sample rate. This value turns out to be the
band filtering and digitization to allow the desired sig- highest sample rate the system will ever operate at.
nal to be negligibly distorted by hardware. Given this Analysis, or channelizer, filterbanks bring in a wide
condition, we can do the rest of our signal analysis, spectrum and split it into equally-space, equal band-
recovery, and isolation using software filters. width channels. Think of splitting the US FM broad-
With the highly dynamic nature of software filters, cast band from 88 to 108 MHz into the one hundred
we expect these rapidly reconfigurable and efficient 200 kHz FM stations using a single filterbank. Each
channelizer techniques to be widely applicable and an station is 200 kHz wide and separated by 200 kHz for
appealing option in the development of next genera- each channel. Instead of filtering and moving from
tion cognitive and dynamic spectrum radio systems. its RF channel to baseband for each channel, we pro-
cess all one hundred channels at the same time in the
channelizer.
III. P OLYPHASE F ILTERBANKS : A NALYSIS AND The channelization process uses the concept of
S YNTHESIS F ILTERS aliasing. While we generally work hard to avoid alias-
We begin with the basic structures of the analysis ing, in the channelizing filterbank, we are aliasing in
and synthesis filterbanks. The one is basically the in- a very controlled way such that we can pull out the
verse of the other. Both involve first the design and individual channels. We know from sampling theory
then the partition of a prototype filter, where the parti- that when we down-sample a signal by a factor of M
tioning occurs over a bank of smaller filters, hence the we also create M Nyquist zones that are aliased onto
name filterbank. There is then the filtering of the sig- baseband. We typically avoid the effects of aliasing
nal through the bank of filters and combining the filter by removing energy from the parts of the spectrum
output such that they separate or combine the signals that will alias by using a low pass or band pass filter.
into the different channels, practically done using a If we filter out enough of the signal, the aliased zones
discrete Fourier transform. Details of the filterbank will not fold in and distort our desired signal.
structures can be found in [2]. Here, we focus on the In the channelization process, we want those
development and understanding of the prototype fil- aliases. Recall that the filterbanks use the same fil-
ter. ter with a different phase. When filtering, each of the
One of the fundamental concepts required when aliased zones is processed by each arm of the filter-
understanding what is going on in the channelizer fil- bank, which has its own phase. At this point, the out-
terbank is the ability to analyze the filters in terms of put of each filterbanks arms is a set of aliases from
phase. In Figure 1, we have created a prototype filter the different channels. We can then despin by apply-
and partitioned it among four filters (M = 4). The ing a correction factor for filter m of exp(j2km/M )
partitioned filters are simply taking every M = 4 tap for a given aliased zone k. The correction factor is re-
starting at tap m where m is the channel number. We lated to the phase difference between each of M filter
can think of this as each of the channels is the same arms. If we despin with this function for a given k
filter with a phase delay of m 2
M.
over all arms of the filterbank and sum them together,
A consequence of the partitioning process is that we constructively add the alias from that channel
we are down-sampling the filter. Each partitioned fil- while destructively canceling all other aliases from
ter is now at 1/M the rate of the original filter. This the other channels. If we think about that correction
clues us in to both the nature of the operations being factor and summing all arms together over every arm
performed as well as the design methods for making and every channel, we make the leap to see that this is
the prototype filter. In the application of the filter- actually a discrete Fourier transform (DFT) operation
bank, we successively insert samples into the filters, that we can efficiently implement using fast Fourier
transform (FFT) algorithms. This topic is given much
3
http://gnuradio.org/doc/doxygen greater coverage with figures explaining the aliasing
Fig. 1. Prototype filter partitioned into four filterbanks. Each filterbank filter is a phase-delayed version of the same filter.

in fred harris book on the subject [2]. TABLE I

The synthesis filter behaves the same way in re- C HANNELIZER PARAMETERS
verse. It uses an FFT to spin the data properly for the
aliases to be separated and then filtered through the Parameter Value
filterbanks. The output samples are then successively Type FIR
taken from the filterbanks. Style Low Pass
Later, we will modify these structures to allow us Window Hann Window
to handle signals where the sampling rate is twice the Sample Rate 20 MHz
channel bandwidth. The changes necessary to these Filter Gain 1
structures is important and complicated. The work End of Pass Band 125 kHz
in [3], [4] explains the required changes. Start of Stop band 225 kHz
Stop Band Atten. 60 dB
IV. P ROTOTYPE F ILTER D ESIGN IN GNU R ADIO
In analysis and synthesis filterbanks, each channel these values using a normalized sample rate or 1 and
is 1/M th the total bandwidth of the received or trans- make the other parameters relative to this.
mitted, respectively, bandwidth. We either break the The other parameter we set for this type of filter is
signal apart in the channelizer or put them together in the window type. In this section and in Figure 1, we
the synthesizer. We do the same with the prototype use a Hann window because its parameters are well
filter. By partitioning the filter, we split the original understood [1]. In latter parts of the paper, we found
filter into M pieces. So building the prototype means it easier to build perfect reconstruction filters using
building a filter capable of filtering a single channel the Blackman-harris window [5].
at the full bandwidth (i.e., the sample rate we receive Continuing with our broadcast FM example, we
or transmit on) of all M channels. This is important know that the full spectrum bandwidth is 20 MHz and
to understand that in a polyphase filterbank, the pro- each channel is 200 kHz. We make a lowpass filter
totype filter is designed at the total bandwidth of all based on this. However, we will not design this filter
M channels. to have the full bandwidth of 200 kHz. The issue is
GNU Radio has tools for building filters that we that in the process of channelizing, we will alias bands
will use to construct the prototype filter, including a together at the channel bandwidth, so filtering beyond
graphical program called gr filter design. In this ap- the channel boundaries ends up aliasing in pieces of
plication, we set the type of filter we want, which will the neighboring channels. Instead, we will select pa-
be a windowed low pass FIR filter. We set the sample rameters that will filter out the edges of our band with
rate, the end of the pass band, and the start of the stop a transition period that ends roughly with the channel.
band along with the stop band attenuation (in dB). The final parameters we select for this experiment are
The pass band and stop band parameters are speci- shown in Table I.
fied in units relative to the sample rate. Using the real We made the start of the stop band slightly beyond
sample rate of the full channel in samples per second, the channel in order to give us a filter with a smaller
we can also specify the pass band and stop band pa- number of taps and knowing that any aliasing with
rameters in Hertz. An alternative approach is to set this will be negligible to the FM quality. Even still,
the resulting filter consists of 1091 taps, which seems
quite long. However, since we separate this filter
among a filterbank of 100 channels, each channel will
actually consist of only 11 taps (i.e., d1091/100e).
Computationally, these become very light-weight fil-
ters, especially if we can parallelize the computations.
A bigger concern may be memory usage, depend-
ing on the platform on which the filterbank is imple-
mented.
The synthesis filter design issues are the same as
the analysis filter. We can easily take in 100 channels
of frequency modulated audio and synthesize them
together to broadcast 100 channels of FM.
Fig. 2. GNU Radio flowgraph for channelizing the FM broadcast
band and pulling out a single channel for FM demod.
V. U SING A P OLYPHASE C HANNELIZER
In the next section, we will look more heavily into
analysis and synthesis filterbanks work well. But for
using the synthesis filterbanks. Here, though, we con-
cases becoming more relevant in developing ideas of
tinue with the channelizer example from the proto-
DSA and cognitive radio, we may not have channels
type filter we built in Table I.
of equal spacing or bandwidth. For these cases, we
The GNU Radio flowgraph is shown in Fig-
need to modify the standard filterbank structures to
ure 2. The signal is taken from a USRP4 ,
allow for arbitrary bandwidth signals. Again, much
sampled at 20 Msps, and fed to the channel-
of this work on the design of the filterbanks and the
izer (gr::filter::pfb channelizer ccf ) after some gain
prototype filters is already published. We require few
adjustment (gr::analog::agc2 cc). In this flow-
modifications on the filterbank designs, some modifi-
graph, we take a single channel out for FM de-
cations on how we partition the filters, and very spe-
modulation and simply discard the rest. We use
cific design requirements of the prototype filter. Pa-
the Busports5 concept in the GNU Radio Com-
pers [3], [4] provide a good overview of the chan-
panion to handle the large number of connec-
nelizer and synthesizer changes required to support
tions. The FM demodulation is done with a
the reconstruction and [6] provides details on the re-
WBFM receive block (wfm rcv Python block in
quired changes to handling the filter partitioning in
gr::analog), and this is followed up by a resampler
the updated synthesis filterbank.
(gr::filter::pfb arb resampler ccf ) to get the sample
rates correct for the audio sink (gr::audio::sink). A First, the changes in the analysis and synthesis fil-
multiplier (gr::blocks::multiply const cc) is in there terbank comes from the need to be able to completely
for volume control. stitch together two channels already split up by a
Setting the receiver to a center frequency of channelizer filterbank without any loss. That means
97.9 MHz, we channelize the entire FM broadcast that, unlike before, we must design the filters such
spectrum. The output of a slightly modified flow- that the roll-off of the filter hits 6 dB just at the
graph from Figure 2 where we plot the output of two channel edge. Previously when channelizing, we de-
channels is shown in Figure 3. The input spectrum is signed the filter to make sure that it was within the
shown along with channels 5 and 10. We only show bandwidth of the channel to prevent aliasing. Here,
the two channels here for space even though all 100 we have a requirement that when combining adjacent
channels are available. channels, the sum of the signals at the edges will re-
combine with the correct energy, which would intro-
duce adjacent channel aliasing that we need to avoid.
VI. D ESIGNING R ECONSTRUCTION F ILTERS
To overcome the aliasing problem, we redesign the
For the FM example, and other signals that have filterbanks to produce channels of the same band-
equal channel bandwidths and spacing, the standard width but now at twice the sample rate. With this, we
4
http://ettus.com
move the adjacent channels far outside of our transi-
5
http://gnuradio.org/redmine/projects/ tion width of the filter, and so the signal remaining in
gnuradio/wiki/Busports the adjacent channels is specifically designed to hold
Fig. 3. Channels 5 (98.9 MHz) and 10 (99.9 MHz) displayed out of all 100 channels of the input FM band spectrum centered at 97.9
MHz.

the part of the current channel that spills over into the
adjacent channel, but unaliased. From here, we can
see that synthesizing these two channels, using the
same prototype filter, will combine the signals such
that the edges of both channels add together to create
no loss between them.
Following from the channelizer now producing
channels at twice the sample rate, we must also
Fig. 4. Diagram of the reconstruction filter to split a signal into
change the synthesis filterbanks to perform in a simi- many pieces and reassemble them again.
lar way. This structure takes in the channels at twice
the sample rate and produces a signal that combines
the channels now at twice the output sample rate. width of our prototype filter equal to the bandwidth
That is, synthesize N channels of 2fs produces a sig- of the channel as the tool defines this as the end of
nal at 2fs N . the pass band. We then have the window, attenuation,
To show how the reconstruction works, we will and transition band as degrees of freedom to tweak
use an impulse as the input to the reconstruc- our filter.
tion filter, which is made up of a channelizer In our analysis here, we will look at six channels
(gr::filter::pfb channelizer ccf ) and synthesis filter- with a sample rate of 2 kHz per channel. We lose
bank (gr::filter::pfb synthesizer ccf ) connected di- no generality in the approach, but these few num-
rectly to each other as shown in Figure 4. The pro- ber of channels allows use to plot and visualize the
totype filters for both filterbanks are the same except process much easier than a large number of channels.
for a gain factor of M/2 (M is the number of chan- Given this, our filter design parameters are shown in
nels) of the synthesis taps so that the output signal is Table II.
at the same level as the input signals. Passing an impulse through this structure will re-
GNU Radios firdes filter design program lets us sult in an impulse out at the receiver at twice the sam-
fairly easily design the reconstruction filters we need. ple rate. While this is an arbitrary example with a
The firdes tool has a low pass 2 function to design clean, known signal, the important consequence of
a low pass filter with the gain, bandwidth, transition this structure and the prototype filter design is that
width, window function, and out-of-band attenuation the resulting signal has been segmented and put back
as design parameters. To hit the required 6 dB roll-off together with almost no change to the signal. The
at the channels edge, we just need to set the band- output, shown in Figure 5, shows that the signal is at
 TABLE II
I MPULSE C HANNELIZER PARAMETERS

Parameter Value
Sample Rate 12 kHz
Filter Gain 1 or 3
End of Pass Band 1000 Hz
Start of Stop band 400 Hz
Stop Band Atten. 80 dB
Window Blackman-harris [5] Fig. 6. Diagram of the reconstruction filter to split a signal into
channelizer many pieces and reassemble into a set of four and two channels.
synthesizer

Fig. 7. Output of the reconstruction filter where an impulse is

segmented into two signals of four and two channels.
Fig. 5. Output of the reconstruction filter. Each channel filter is
displayed in the top figure while the reconstructed signal is shown
in the bottom figure.
this style, which means that we can now more eas-
ily channelize signals of different bandwidths and at
twice the original sampling rate with very minor pass different frequencies. One restriction comes into the
band distortion due to the filters. design process, though. To work at twice the sample
Next, we show a more complex example that indi- rate, both the channelizer and synthesizer must have
cates better how this concept can be used. We stick an even number of filters or channels. This may mean
with the impulse as the input signal to give a clean using one more channel than required for some appli-
PSD of how the reconstruction produces perfect sig- cations.
nals at the output. Instead of synthesizing all six chan-
nels back together, we use two synthesis filterbanks to
combine the first four and last two channels. The de- A. Channel Mappings
sign is shown in Figure 6 and the results in Figure 7. The GNU Radio channelizer and synthesis filter-
Not shown in these figures is the close-up of the re- banks have the ability to control the location of the
constructed pass band; however, the two, four, and output channels using a channel map, or permuta-
full six reconstructed channel examples all have the tion, through the set channel map function call. This
same maximum ripple of 5 104 dB. While the rip- mechanism provides a mapping of where each input
ple is dependent on the original filter, the results do channel will end up at the output of the filterbank. For
not change with the number of channels, making the the channelizer, the default mapping is that channel
technique scalable to large numbers of channels. 0 is the 0 Hz channel, or the middle of the incoming
In this example, the more general use of the recon- frequency band. Channel numbers increase positively
struction filter becomes more obvious. Any signal can in the spectrum and wrap around to the negative half
be broken up and reconstructed with a set of filters of of the spectrum at M/2. For an even number of chan-
nels, the M/2 channel spans the positive and negative TABLE III
edges of the spectrum. FM C HANNELIZER PARAMETERS
The channel map function takes a vector of num-
bers that is specified by an input channel value that Parameter Value
will map to the output channel as the index of the ar- Sample Rate 1 MHz
ray. This setup gives us flexibility in where and how Filter Gain 1
to direct the channel positions, including being able End of Pass Band 50 kHz
to direct the same channel to multiple outputs. Start of Stop band 20 kHz
The synthesis filterbank does the same: the index Stop Band Atten. 80 dB
of the array is the output channel number and the Window Blackman-harris [5]
value at that index is the input channel. Figure 8
TABLE IV
demonstrates this. Remember that the output of the
FM S YNTHESIZER PARAMETERS
oversampling synthesis filterbank is at two times the
sampling rate, which gives us twice the number of Parameter Value
channels. By default, the channel mapping is from
Sample Rate 600 kHz
channel 0 to M for M input channels. All of the in-
Filter Gain 5
coming channels from the channelizer are mapped di-
End of Pass Band 50 kHz
rectly to the corresponding output index value. That
Start of Stop band 20 kHz
is why in the above graphs we saw the impulse cov-
Stop Band Atten. 80 dB
ering the positive half of the spectrum; the other half
Window Blackman-harris [5]
of the channels produced by the synthesis filterbank
had no input and therefore produce 0 on the output.
In Figure 8, we are using the mapping sequence [10, to adjust the volume of the audio as well as resample
11, 0, 1, 2, 3]. The channels are mapped in order of it to match the required rate of the audio subsystem,
their output from the channelizer, but we have used just as we did above with the FM channelizer. And
this mapping technique to shift the entire spectrum the synthesizer filterbanks channel map to set the re-
down by two channels. sulting signal at DC is [10, 11, 0, 1, 2, 3].
At different places in the flowgraph, we use fre-
B. Reconstructing FM quency sinks to visualize the PSD of the signal. Fig-
In this final example, we channelize part of the FM ure 10 shows the output of these instrumentation
broadcast band into ten 100 kHz channels, select six tools, which shows the recovered signal with the stan-
of these to recombine, and demodulate the FM to au- dard FM in the middle as well as the HD radio side-
dio. The GNU Radio flowgraph is shown in Figure 9 bands. On the input spectrum, we added the channel
that uses a USRP B2006 to pull in a 1 MHz signal boundaries of the ten channels to show how it was
centered at 101.3 MHz. There is a strong FM sta- split up in the channelizer. In this case, we only re-
tion at 101.5 MHz. We channelize the signal into quire five of the channels to reconstruct the signal, but
ten channels using the parameters in Table III. The we use six to meet the requirement that the filterbanks
first six of these channels, covering the positive half must work with an even number of channels.
of the spectrum, go into the synthesis filterbank. We
terminate the other four channels for this experiment. VII. C ONCLUSION
The synthesizer parameters are shown in Table IV and
From the details of this paper, we should under-
are nearly the same as for the channelizer. Instead,
stand the basics of how to design a polyphase filter-
though, this filter is designed at 600 ksps because we
bank prototype filter for analysis and synthesis filter-
are only combining six 100 kHz channels together.
banks. This information gives us the basic knowl-
The gain parameter is the half of the total number
edge of how to build filters with GNU Radio and use
of channels in the channelizer, which normalizes the
them in the GNU Radio polyphase filterbank blocks.
output. For convenience, we decimate the output by
We also discussed the added challenges of design-
two using a very inexpensive half-band decimating
ing filterbank filters that allow perfect reconstruction,
filter. The rest of the flowgraph performs the FM de-
which we can used to break apart and then recombine
modulation and some signal processing to allow us
channels as we need to in such a way that any signals
6
http://ettus.com on channel boundaries are reconstructed perfectly.
Fig. 8. Channel mapping in the synthesis filterbank using [10, 11, 0, 1, 2, 3] as a method of moving the spectrum down two channels.

Fig. 9. GNU Radio flowgraph for channelizing and reconstructing a US broadcast FM radio station.

Fig. 10. Output of the reconstruction of the FM signal. This figure shows the input spectrum, the six highlighted channels sent to the
synthesis filterbank, and the output of the filterbank and decimate-by-2 filter.
 The details of how the filterbanks work is the
subject of many other texts and papers referenced
throughout. GNU Radio provides one of the few areas
where the code, documentation, and examples exist
to allow us to study and understand all of the aspects,
features, and trade-offs of using these structures.
While this paper focused on the analysis and syn-
thesis filterbanks, the same filter design principles are
equally valid for any of the other polyphase filterbank
tools we have in GNU Radio. Included among these
are the arbitrary resampler and the clock timing syn-
chronization blocks. In both of these filterbanks, we
can set the quantization error by selecting the number
of filters we use, which is set to 32 by default. In the
terms used above, the number of channels is the num-
ber of filters in the filterbank. We then use the same
concepts of designing the prototype filter by under-
standing this simple shift in nomenclature.

R EFERENCES
[1] A. V. Oppenheim, R. W. Schafer, and J. R. Buck, Discrete-
time Signal Processing (2Nd Ed.). Upper Saddle River, NJ,
USA: Prentice-Hall, Inc., 1999.
[2] fred harris, Multirate Signal Processing For Communication
Systems. Upper Saddle River, NJ: Prentice Hall, 2004.
[3] E. Venosa, X. Chen, and fred harris, Polyphase analysis fil-
ter bank down-converts unequal channel bandwidths with ar-
bitrary center frequencies - design i, in SDR10-WinnComm,
2010.
[4] X. Chen, E. Venosa, and fred harris, Polyphase analysis fil-
ter bank up-converts unequal channel bandwidths with arbi-
trary center frequencies - design ii, in SDR10-WinnComm,
2010.
[5] f. j. harris, On the use of windows for harmonic analysis
with the discrete fourier transforms, in Proc. IEEE, vol. 66,
Jan. 1978.
[6] f. j. harris, C. Dick, X. Chen, and E. Venosa, Wideband 160-
channel polyphase filter bank cable tv channeliser, in IET
Signal Processing, 2010.