892

IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 60, NO. 12, DECEMBER 2013

Hardware-Efficient Implementation of Half-Band IIR
Filter for Interpolation and Decimation
I. H. H. Jørgensen, P. Pracný, and E. Bruun, Senior Member, IEEE

Abstract—This brief deals with a simple heuristic method for
the hardware optimization of a half-band infinite-impulse response (IIR) filter. The optimization method that is proposed here
is intended for a quick design selection at the system level, without
the need for computationally intensive calculations and simulations. The aim is to arrive at a design with low hardware complexity that is measured in terms of the number of adders. In the
approach that is presented here, the filter specification is treated
with some flexibility at the topmost system level. The half-band
filter is implemented as a parallel connection of two all-pass filter
cells. The filter is designed by first fixing the most sensitive filter
coefficient to a convenient value that can be quantized by using
only a few adders. Subsequently, the overdesign margin is used
to coarsely quantize the remaining filter coefficients and thereby
minimize hardware demands. The complexity of the resulting IIR
filter is evaluated by counting all the adders in the filter, i.e.,
the adders for both the filter coefficients and the filter cells. The
result of the method is compared with state-of-the-art works where
the filter is designed by using the fixed filter specification and
advanced algorithms to minimize the hardware that is used to
implement filter coefficients.

Fig. 1. Example of a multistage interpolation filter with four stages, performing a sample-rate increase of 64.

Fig. 2. IIR filter using a parallel connection of two all-pass cells H0 (z −2 )
and H1 (z −2 ), which is used as the first stage of the filter in Fig. 1. X(z −2 )
and Y (z −1 ) are the input and output signals, respectively. The filter runs at
2fs, in .

Index Terms—Decimation, digital filter, half-band filter, hearing
aid, infinite-impulse response (IIR) filter, interpolation, interpolation filter, low power, low voltage, sigma–delta Σ−Δ modulator.

I. I NTRODUCTION

O

VERSAMPLED digital-to-analog (D/A) converters, such
as a Σ−Δ modulator, are widely used in systems in
which the signal bandwidth can be limited to a fraction of the
system’s operating frequency [1]. This greatly helps in relaxing
the circuit requirements, e.g., the matching of components, at
the expense of a high-frequency operation. Due to the oversampling nature of Σ−Δ modulators, an interpolation filter is
needed prior to the modulator. To reduce the hardware demands
and power consumption, the state-of-the-art D/A converter
designs implement the interpolation filter as a multistage filter
[1]. An example of an interpolation filter that increases the input
signal sampling frequency fs, in 64 times and that consists of
four stages can be seen in Fig. 1.
With the first stage of the interpolation filter implemented as
a half-band FIR filter, this stage becomes the most hardware
demanding of all filter stages [1], [2]. To reduce the hardware

Manuscript received July 10, 2013; accepted October 9, 2013. Date of
publication November 1, 2013; date of current version December 24, 2013.
This brief was recommended by Associate Editor Y. Miyanaga.
The authors are with the Department of Electrical Engineering, Technical University of Denmark, 2800 Kongens Lyngby, Denmark (e-mail:
ihhj@elektro.dtu.dk; pp@elektro.dtu.dk; eb@elektro.dtu.dk).
Color versions of one or more of the figures in this brief are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TCSII.2013.2285975

Fig. 3. Second-order all-pass filter cell that is running at (a) 2fs, in and
(b) fs, in .

demands and power consumption, a half-band IIR filter can
be used instead if the requirement for phase linearity is not
strict [3]. For this purpose, the design optimization using a
polyphase IIR structure of all-pass filter cells for sample-rate
conversion was originally proposed in [4] (see Figs. 2 and 3). A
thorough description is given in [3], and a genetic algorithm
was used to design such a filter in [5]. Various options for
the all-pass IIR filter cells (with transfer functions H0 (z −2 )
and H1 (z −2 ) in Fig. 2) and their sensitivity to coefficient
quantization can be found in [6] and [7].
The filter that is used in this brief to implement the interpolation is shown in Fig. 2. Each of the all-pass filters, i.e.,
H0 (z −2 ) and H1 (z −2 ), in Fig. 1 are constructed using one or
more second-order all-pass filters, effectively resulting in a halfband IIR filter [3], i.e.,
H(z) = H0 (z −2 ) + z −1 H1 (z −2 )

1549-7747 © 2013 IEEE

=

K1
K0  

a0, k + z −2
a1, k + z −2
−1
+
z
.
−2
1 + a0, k z
1 + a1, k z −2

k=1

k=1

(1)

The approach that is chosen to design and minimize the IIR Fig. the number of coefficients and adders is lower. the order of the parallel all-pass filters must be determined so that the filter specifications are met after the filter coefficients are quantized and implemented using only the shifters and adders/subtractors. Changing a0. 3(b). IIR filter from Fig. as shown in Fig. F ILTER -D ESIGN M ETHOD D ISCUSSION A transfer function of a half-band polyphase IIR filter as a parallel connection of two all-pass filters is shown in Fig. ωs . 5. The filter coefficients are listed in the Step 1 column in Table I. 1 results in the dashed (green) curve and the solid (black) curve. running at half the sampling rate [3]. which is shown as the dotted (red) curve. II. where ωp and ωs are the normalized passband and stopband cutoff frequencies. it can be seen that the . The optimization methods in the state-of-the-art works is discussed in Section II. The transfer function is most sensitive to the change of the coefficient corresponding to the pole that is closest to the unity circle in the z-domain. it is an advantage to use as few adders as possible. Results are summarized in Section IV. In [7]. which results in filter designs with hardware complexities on par with the designs that are obtained using much more computationally intensive methods. 1 corresponds to the pole that is farthest from the unity circle. such as in [6]–[8]. respectively. IIR filter transfer function sensitivity to coefficient changes. e. therefore. Thus. The smallest coefficient a1. 3 by 1% while leaving the other coefficients unchanged results in the dash–dotted (blue) transfer function. as the all-pass sections are constructed using H0 (z −2 ) and H1 (z −2 ). This allows the coefficients to be implemented using only the shifters and adders/subtractors. the complexity of the filter in the state-of-the-art works. in . and dp . ωp = 1 − ωs . In Section III. First. and the resulting diagram for the second-order all-pass filter is shown in Fig. The adders represent the majority of the hardware that is needed to construct the filter. 6. 4. reducing the number of adders is a good system-level optimization to achieve both high hardware and power efficiency in the application-specific integratedcircuit implementation. the power consumption is of extreme importance. a step-by-step method is proposed. therefore. 3(a). Due to the symmetric properties of a half-band filter. hearing aids [2]. but for some battery-powered applications. Fig. Example designs are included. avoiding multipliers. Given ωp . offering the opportunity for hardware-efficient implementation.g. 6. To illustrate this effect. 2. Here. 5. 6 shows a transfer function of a half-band IIR filter using a parallel connection of two allpass cells. A diagram for one of the second-order all-pass filters can be implemented by only using two adders. where filters H0 (z −2 ) and H1 (z −2 ) are operated at fs. This half-band filter holds two advantages. Second.: IMPLEMENTATION OF HALF-BAND IIR FILTER FOR INTERPOLATION AND DECIMATION 893 Fig. filter in the state-of-the-art works is the following [6]. The largest of the coefficients corresponds to the pole that is closest to the unity circle. 2 and a1. Since shifts can be simply implemented by rewiring the individual bits of a binary word. K0 and K1 are the number of the second-order allpass filters that are used in each branch of the filter in Fig. the stateof-the-art works focus on filter coefficients and their implementation as a sum of integer powers of two. it is shown that the sensitivity of the coefficients increases as the position of the pole corresponding to a coefficient approaches the unity circle in the z-domain. Fig. Applying a 1% change to a0. respectively. respectively. and dp ∼ d2s /2 [4]. the area of an adder is insignificant. they can be realized. avoiding multipliers.. In a modern CMOS process. such as hearing aids. In Fig. and dp and ds are the passband ripple and the stopband ripple. To reduce the hardware demands in the IIR filter.JØRGENSEN et al. 4. 3 corresponds to the pole that is closest to the unity circle. as shown in Fig. The largest coefficient a0. Simplified transfer function of a general half-band filter. A genetic algorithm is used to find the combination of the quantized coefficients. is judged by counting the number of adders that is needed to implement the coefficients. 1. resulting in the smallest number of adders [6]. This makes this filter suited for the implementation in which both area and current consumption are a primary consideration. they do not incur any additional hardware.

the value slots (16. Fig. 7. even for a person with normal hearing. The all-pass cell in Fig.1 kHz) = 0.e. Due to the symmetry of the half-band filter. it can be reused for other designs. To fulfill the Nyquist criterion. a passband ripple of 0. e. Slot (16. in = 22. For example. 1 is changed by 1%. with the x of the 16 bits set to +/−1. a stopband ripple of ds = −58 dB has been selected. The 16-bit values are sorted into slots. Filter 1 is used to explain the simple heuristic step-by-step design method that results in hardware complexity comparable to the state-of-theart designs that are obtained using more complicated design techniques.4535. 2 with the all-pass sections of Fig.g. performing decimation. A detailed study of coefficient sensitivity and the use of all-pass filter cells that are more resistant to coefficient quantization can be found in [7].1 kHz/2 = 22. For the present design for hearing aids. and input sampling frequency fs. out = 44. Step 2) A list of 16-bit-long words with all the possible combinations of signed bits set to one is created using simple MATLAB scripts. This allows the first filter stage to be implemented as an IIR filter [2].05 kHz. Therefore. 2) are investigated. the audio quality requirements are relaxed compared with high fidelity. x) consists of all possible 16-bit values. 1) does not contain any . The analytic design method in [4] gives the transfer function in Fig. A detailed discussion of the interpolation filter specifications for hearing aids is given in [2]. the all-pass filter cells also need adders for the implementation. Thus.5 dB is chosen for the whole filter chain in Fig. H EURISTIC F ILTER -D ESIGN M ETHOD W ITH E XAMPLES Three design examples are presented. ωs = 1 − ωp = 0.1 dB is allocated to the first half-band IIR filter. the filter cell that is shown in Fig.1 dB. in = 44. 7 (the red dotted curve) and the coefficients in the Step 1 column in Table I. 3 should be designed for ωp = 0. In the first step.1 kHz. Hearing aids normally have a bandwidth of BW = 10 kHz [2]. 1 is designed as a parallel connection of two all-pass filter cells. and 4. VOL. A passband ripple of 1 dB is audible. NO. In this example. However. The filters are designed for interpolation. IIR filter transfer function plot in Steps 1. and dp = 0. Filter 2 and Filter 3 are used to show a comparison of the proposed method with the current state-of-the-art designs of [6] and [8] where numerical optimization methods are used. and ds = −58 dB. and the normalized transition band is ωt = ωp − ωs = 0. the filter is designed for an attenuation of 80 dB.4535. the first stage of the multistage filter in Fig. 1. the largest is the most sensitive and is likely to incur the largest number of adders when quantized. performing a sample rate increase by a factor of two (see Fig. 2). These slots contain values with the most coarse quantization (i. the output sampling frequency is fs. but because of the duality.1 dB is used for the high-fidelity audio application [2]. slot (16. 60. the same coefficients can be used for filters. 12. when evaluating the filter complexity. The methods that are described in [5]–[8] deal with a minimization of the number of adders that is needed for the coefficients. Filter 1: In applications such as hearing aids. It is the most sensitive to quantization [7] and is likely to incur the largest number of adders of all the coefficients. 3. 1) and (16. and it is optimized with respect to the hardware demands. The interpolation factor is two.894 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS. For this reason.5465. an overdesign is applied: instead of a stopband attenuation of 58 dB. A word length of 16 bits is sufficient as it is well above the coefficient sensitivity. 3 only requires two adders. III. the input sampling frequency in this example is half of the standard high-fidelity audio sampling frequency fs. of all the coefficients. specifying the number of bits that is set to one. 3 is used for the design. and a ripple of 0. the normalized passband cutoff frequency is ωp = (2π × 2 × 10 kHz)/(2π × 44. therefore. the total number of adders that is used for the coefficients and for the filter cells should be calculated.05 kHz.. Once this list is available. leaving a margin of approximately 20 dB for the quantization of the filter coefficients. A stopband ripple in the range of 55–60 dB is sufficient for a hearing aid since a speech input signal of normal intensity is about 60 dB above the threshold of hearing. can be implemented using the smallest number of adders). Performing similar investigations on other filters of any order will also reveal that. DECEMBER 2013 TABLE I IIR F ILTER C OEFFICIENT VALUES IIR filter transfer function shows the most substantial change when coefficient a0. For the present design. 3 is changed by 1% and the smallest change when coefficient a1. dp = 0..093. Step 1) The half-band IIR filter in Fig. 3 in this case) corresponds to the pole that is closest to the unity circle in the z-domain. Step 3) The largest of all the coefficients (a0.

1). ωp = 0. Assuming three adders per cell. (16. it does not provide a candidate for quantizing a0. this results in ωs = 0.: IMPLEMENTATION OF HALF-BAND IIR FILTER FOR INTERPOLATION AND DECIMATION TABLE II C OMPARISON OF THE D ESIGN M ETHODS AND H ARDWARE D EMANDS value between 0. The quantized coefficients are listed in the Step 4 column in Table I. respectively. 3).2 dB. From slot (16. as in Step 1) such that a0. The analytic method in [4] and the given steps result in four second-order all-pass cells (one cell per coefficient). example 3] and [6.9375. this gives a reduction in the number of adders by approximately 20% (see Table II) for this design. example 6] are ωp = 0. The margin that is gained by the filter performance being better than needed is used for a coarse quantization of the coefficients (see Table III) to fit the specification. quantizing them one by one in descending order. Step 4) In this step. 7 and the new filter coefficients in the Step 3 column in Table I. Details of the calculation are given in Table IV. Moreover. with two adders in each cell. the next most sensitive coefficient to be determined is a1. but the filter still fulfills the original specification of Step 1. Again. Step 4 is then repeated for the rest of the coefficients. A similar approach that uses the overdesign was presented in [9] where it was applied to FIR filters but without specifically exploiting the coefficient sensitivity and without expressing the coefficients as the sums of integer powers of two.. Due to the symmetric properties of the half-band filter. A comparison with the design in [8] can be found in Table IV and the filter coefficients of the design using the simple method are shown in Table V.9375.425 and ds = −65 dB. 2). example 9]. 4). suggesting that the power consumption for this filter will be lower compared with the other designs. Moreover. With the most sensitive coefficient fixed to a0. 3. This results in ten adders to implement the coefficients. 3 . 3 can be now implemented using only one adder (see Table I).093 to 0. It can be seen that the simple method presented here results in filter complexities that are comparable with what can be achieved using more complicated approaches. resulting in a further reduction in the number of adders (see Table II). the filter is designed using the half-band IIR filter in Fig. At the same time. Filter 3: A final example for the comparison is the lattice wave half-band filter design with specifications taken from [6. If the filter using the quantized coefficient does not fulfill the specification. with the all-pass sections of Fig. and the passband ripple is well within the specification of dp = 0. 3). etc. There were no better results found for this design by running the script. The filter specification for [6.575. the number of adders that is used for the filter with the same specification with the works in [6] and [8] can be estimated to 30 adders and 17 adders.2.0746 and gives the dashed (green) transfer function in Fig. The filter is now redesigned by varying ωt (again. 3 exactly corresponds to 0. there are six second-order all-pass cells (one cell per coefficient). The resulting transfer function fulfills the original specification of Step 1 and can be seen in Fig. Compared with a direct quantization of the coefficients. the value that is closest to a0.9375. a half-band filter with specifications taken from the works in [6] and [8] has been designed. This changes the specification of normalized transition band ωt from 0. A simple Matlab script has been run to check a large number of possible quantized coefficient 895 TABLE III IIR F ILTER C OEFFICIENT VALUES FOR F ILTER 2 TABLE IV C OMPARISON OF THE H ARDWARE D EMANDS AND S AMPLING R ATES IN F ILTERS 2 AND 3 combinations. Filter 2: For comparison purposes. i. the stopband attenuation was degraded. using the analytic method of [4].44 and ds = −46 dB.e. 7 (the solid (blue) curve). The designs .JØRGENSEN et al.56 and dp ∼ d2s /2 = 1. 3 is 0. 2). The search starts with choosing the closest possible smaller or larger values from slot (16. It is possible to relax the specification of the transition band to ωt = 0. notice that the sampling rate of the proposed filter is only fs .5 and 1. 2. the quantization is refined by examining the closest possible values that are available in slot (16. The resulting number of adders used for all coefficients is eight. This gives 16 adders in total for the entire filter using the simple heuristic method that is presented here. 3 = 0. This results in 22 adders in total. Due to the symmetric properties of the half-band filter. 3 . The result of this step is that coefficient a0. with two adders in each cell (see Fig. this results in ωs = 0. The search is continued until the filter fulfills the specification by choosing values from slots (16.1 × 10−4 dB. thus. the remaining stopband ripple margin is used for coarse quantization.

8th Int. 2007. 167–172.” IEEE Trans.” in Understanding Delta-Sigma Data Converters. K. 8. starting with the most sensitive coefficient and proceeding with the less sensitive coefficients. Devices Syst. 364–367. pp. IEEE APCCAS. P. 2005. Modern Satellite.896 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS.” in Proc. no. pp. which is measured in terms of the number of adders. the proposed filter offers the opportunity for low-power implementation in. no. NO. Sep. Conf. “FIR filter design methodology for hardware optimized implementation. Anzova. I. 55. 2006. Saramäki. G. “A systematic algorithm for designing multiplierless computationally efficient recursive decimators and interpolators. Dec. Papers. Khan. Constantinides. Saramäki. [5] J. Pracný. Anzova. Jan. Cable Broadcast. Renfors and T. [3] M. Saramäki. Taking the lower sampling rate into account. Nikolova. 2007. C ONCLUSION A simple optimization method is presented for a half-band IIR filter in order to obtain a low hardware complexity of the filter. “A systematic algorithm for the design of lattice wave digital filters with short-coefficient wordlength. Although the design presented here does use a slightly higher number of adders. [8] J. it has been obtained using a quick and easy design methodology.. 54. pp. 6. vol. Dec. Llimós. “Interpolation filter design for hearing-aid audio class-D output stage application. 4th Int. DECEMBER 2013 TABLE V IIR F ILTER C OEFFICIENT VALUES FOR F ILTER 3 coefficients. vol. 2005. 9. Ivanova. 130. vol. and E. A. [2] P. 225–235. 1983. “Recursive Nth-band digital filters— Part I: Design and properties. Saramäki.” in Proc.. 1987. [9] R. The optimization relies on the overdesign of the IIR filter at system level and the application of simple all-pass filter cells. and R. The proposed method results in hardware complexities on par with the stateof-the-art filter examples that are designed using more computationally intensive methods. NJ. pp.g. R EFERENCES fulfill the same specification. 299–308. 1. ISPA. Electron. “An algorithm for the design of multiplierless IIR filters as a parallel connection of two allpass filters.. 3. “Example modulator systems. Aug. G Circuits. 2009. Dec. Yli-Kaakinen. “Design and realization of efficient IIR digital filter structures based on sensitivity minimizations. Schreier and G.. Consum. USA: IEEE Press. hearing aid applications [2]. [4] R. The filter coefficients are quantized.” in Proc. Yli-Kaakinen and T. e. 12. J. Symp. no. Telecommun. vol. 2012. Z. Stoyanov. Qamar. Sep. [6] V. Hoboken. VOL. Circuits Syst. M.” IEEE Trans. Temes. 744–747.” IEEE Trans. pp. IV. Circuits Syst. The margin that is gained from the overdesign in the stopband attenuation is utilized to permit a coarse quantization of the [1] R. ch. Serv. pp. pp. I. Aug. C. Yli-Kaakinen and T. pp.” in Proc. Mehboob. “Digital signal processing schemes for efficient interpolation and decimation. ICECS. and T. S. Reg.. The complexity of the resulting IIR filter is evaluated by counting all the adders in the filter. Valenzuela and A. [7] G. 34. 60. A. Bruun. 1669–1673. 1838–1851. 24–39. .” IEE Proc. no. and V.