Professional Documents
Culture Documents
Abstract—Digital audio equalization is used to enhance the togheter. Indeed, due to the fact that the filters have a non-
listening experience introducing corrections in the auditory zero bandwidth, when the filters are connected together, their
frequency response. The equalizer performance depends on frequency responses will interfere each other [1], resulting in
its implementation which includes optimal filters with sharp undesirable distortions of the overall frequency response that
transition bandwidth development and a low computational will differ from the user’s gains setting. This problem has
complexity. Starting from a previous FIR implementation based
been deeply investigated in [1], [4], [5]. An iterative approach
on multirate systems and filterbanks theory, an optimized digital
audio equalizer is derived. The proposed approach employs an that involves the implementation of the filters using Remez
approximation of the FIR structure using IIR filters, in order to algorithm is presented in [6]. When the response of adjacent
improve the filterbanks structure developed to avoid interferences bands are added together, if the composite frequency response
between adjacent bands. The effectiveness of the optimized shows an unacceptable error deviations, a new filter with a new
implementation is shown comparing it with the previous ap- stopband cut-off frequency has to be designed. This procedure
proaches of the state of the art. The experimental results confirms is iterated until the deviation becomes acceptable resulting
that the presented solution has several advantages increasing in a too high computational complexity. A method based on
the equalization performance in terms of low computational the opposite filter concept has been proposed in [1]: a set of
complexity, low delay, and uniform frequency response. simple filters are adequately parametrized and serially inserted
into the equalizer bank, in order to reduce the interference
I. I NTRODUCTION effects. Also a frequency domain algorithm was presented in
Digital audio equalization is one of the most common [7]. The equalization consists of a complex multiplication of
operations in the acoustic field and it is used in order to the input spectrum with the frequency equalization function
enhance the listening experience by the correction of room that, when transformed in time domain, has all the properties of
and loudspeakers frequency response. It is based on the use a FIR filter. The computational complexity is very low but the
of equalization filters bank connected in serial, in parallel or algorithm efficiency is strictly bound to frequency resolution;
in a mixed connection form [1]. The equalization filters are large ripples at bands edges are easily observed.
developed in the frequency domain and they can boost or
A different solution could arise from the multirate tech-
attenuate specific part of the frequency spectrum. They are
niques applied to adaptive systems. In this context, the prob-
characterized by three parameters: the center frequency, the
lem of aliasing cancellation when an adaptive filtering is
gain and the quality factor, that gives the relation between the
included in a filterbank with perfect reconstruction is well-
filter bandwidth and the center frequency. The equalizer bank
known. Different approaches were presented such as cross
is called graphic equalizer when the center frequencies and
terms between subbands [8] or extra terms taking into ac-
the quality factor are fixed, while the gain is adjustable [1]. In
count adjacent bands [9]. Starting from these approaches, an
this context, different type of graphic equalizer can be found
innovative digital audio equalizer has been introduced in [10],
in the literature, that differ from filter development and final
[11]. Taking into consideration multirate systems and their
implementation.
property, the idea was to realize a linear phase FIR equalizer
Generally, a tree structured filter bank was used, as de- that overcame the well-known drawbacks. Therefore, a first
scribed in [2]: the analysis filter bank was built with equal optimized version of this algorithm has been proposed in
stages splitting the input signal in two subbands while the [12]: it is based on the use of IIR filters capable to reduce
synthesis filter bank had to recombine back the bands. The the required computational complexity preserving the audio
drawback of such a structure is the too high delay that quality [13], [14]. As for the FIR approach, the overall scheme
exponentially increases with the number of subbands. A similar of the algorithm derives from two analysis and two synthesis
approach has been described in [3]; a linear phase digital filter cosine modulated filter banks properly combined in order to
is implemented using frequency masking technique. This is one have the desired response when all bands are added together,
of the most computationally efficient techniques for the design exploiting multirate properties. However, the method proposed
of sharp low-pass, high-pass, band-pass, and band-stop filters in [12] shows some disadvantages: it is not possible to ex-
with arbitrary passband but the introduced delay dramatically ploit the lattice structure behaviour considering the polyphase
increases with stricter constraints. Apart from too high delay components in the filtering procedure, in order to have an
introduced in the filtering operation, the main disadvantages efficient filter bank structure. In this context, a more feasible
of these approaches are the appearance of extra ripples in the structure for the filter coefficients development is presented,
overall frequency response, when adjacent bands are added in order to fully exploit the filter bank structure, reducing the
Signal Processing
Speech, Music, and Audio Processing 410
8th International Symposium on Image and Signal Processing and Analysis (ISPA 2013) September 4-6, 2013, Trieste, Italy
Signal Processing
Speech, Music, and Audio Processing 411
8th International Symposium on Image and Signal Processing and Analysis (ISPA 2013) September 4-6, 2013, Trieste, Italy
Signal Processing
Speech, Music, and Audio Processing 412
8th International Symposium on Image and Signal Processing and Analysis (ISPA 2013) September 4-6, 2013, Trieste, Italy
DI ∆t Order DI ∆t Order
the proposed approach and the approaches of [11] and [12], M = 16 0.9985 1.192s 2048 1.0049 0.343s 50
taking into consideration also the delay introduced by the final M = 32 0.9985 2.413s 2048 1.0026 0.457s 30
equalizer. Considering the method proposed in [11], for a L-
taps FIR filter prototype, there are exactly L multiplications M = 64 0.9985 4.950s 2048 1.0009 0.677s 22
and L − 1 additions, giving a complexity of 2L − 1 for
each equalizer band. Taking into consideration the method TABLE IV. D ISTORTION I NDEX FOR UNIFORM M BAND EQUALIZER
WITH A FIR PROTOTYPE OF LENGTH 4096 AND A IIR PROTOTYPE : ∆t
proposed in [12], the IIR prototype length is bounded to the
REPRESENTS THE TIME INTERVAL IN SECOND TO FILTER 4096 SAMPLES .
number of IIR lattice structure used and to the number of
subbands as shown in Table II. This facts allows to have a FIR of [11] IIR proposed approach
good resolution and a sharp transition of the IIR filter, to
DI ∆t Order DI ∆t Order
the detriment of the computational complexity. Considering
the proposed approach, if Ns is the number of fourth order M = 32 0.99978 5.043s 4096 1.0018 0.657s 46
sections for a N length IIR filter, the computational cost M = 64 0.99978 10.025s 4096 1.0034 0.832s 28
is evaluated taking into account Eq.(17) where each section
needs 5 multiplications and 4 additions both for the numerator
and denominator. Furthermore, it is needed to accumulate the M = 8, M = 16, M = 32, and M = 64: incremental gains
output of each second order section in order to obtain the from 1 to 8 have been repeated every 8 bands. Figs. 4ii-7ii
output sample and also N − 1 additions have to be considered. shows a detail of the equalizer curve for each value of M ,
More in details, taking into consideration the example since it is very difficult to distinguish the difference between
reported in Fig. 3, a L = 4096 taps FIR filter prototype for the filter considering all the spectrum as reported in Figs. 4i-7i.
M = 64 subbands equalizer requests a total amount of 8191 It is evident that for the FIR prototype, increasing the number
operations per output bin. If we consider the IIR filter of order of subbands, the length of 1024 is no more sufficient in order to
N = 28 with Ns = 7 fourth order sections, the total amount have a sharp transition band. Therefore, increasing the number
of computational complexity will be equal to 265 operations. of subband, longer FIR prototype are requested. Considering
To obtain good performance considering the approach of [12], the IIR prototype, it has good performance also with a small
a IIR filter with Nl = 5 lattice sections had to be considered number of samples, and it preserve its behaviour increasing
with a computational complexity of 1923 samples. the number of the subbands. Differently from [11] and [12],
there is a computational saving considering the proposed IIR
III. E XPERIMENTAL RESULTS prototype maintaining a good performance level. Taking as
term of comparison the process proposed in [19], it is evident
The validation of the algorithm has been performed through
that the proposed approach is capable to avoid the ripple
a comparison between the FIR Equalizer implemented as
around the transition band while [19] shows an increase of the
described in Section II-A, the proposed approach and the
ripple amount, increasing the number of subband. Furthermore,
method proposed in [19]. This approach has been chosen since
the IIR filter prototype preserve a good phase linearity in the
it tries to solve the problem of frequency response interference,
passband resulting from the least squares approximation of the
considering each band as composed by an high number of
complete FIR filter frequency response, as shown in Fig.(8).
raised-cosine filters, in order to avoid ripple between adjacent
bands. Moreover, a comparison with other simple filter bank To better underline the performances of the proposed
has been already performed in [10], [11]. Then, four cases have approach, an objective performance measurement parameter
been considered with a different number of subbands (i.e., M has been used. Taking into consideration the distortion transfer
= 16, 32, 64, 128): for each case, three FIR filter prototypes of function [16], calculated as follows,
length N = 1024, 2048, 4096 and different IIR filter prototypes
have been considered for the tests. 1 X
M
T (z) = Hi (z)Fi (z) (22)
Figs. 4i-7i show the equalizer frequency response for M i=1
Signal Processing
Speech, Music, and Audio Processing 413
8th International Symposium on Image and Signal Processing and Analysis (ISPA 2013) September 4-6, 2013, Trieste, Italy
10
Magnitude [dB] 8 5
Magnitude [dB]
6
4.5
4
2 a a
b 4 b
c c
0
0 0.1 0.2 0.3 0.4 0.5 0.22 0.23 0.24 0.25 0.26 0.27 0.28
Frequency [Normalized] Frequency [Normalized]
i ii
Fig. 4. Equalizer frequency responses for M = 8 subbands considering (a) 1024 FIR filter, (b) IIR of order N = 50, (c) method of [19]: (i) frequency
behaviour of the first 8 bands, (ii) detail of one band.
10
8 5
Magnitude [dB]
Magnitude [dB]
6
4.5
4
2 a a
b 4 b
c c
0
0 0.05 0.1 0.15 0.2 0.25 0.11 0.115 0.12 0.125 0.13 0.135 0.14
Frequency [Normalized] Frequency [Normalized]
i ii
Fig. 5. Equalizer frequency responses for M = 16 subbands considering (a) 1024 FIR filter, (b) IIR of order N = 34, (c) method of [19]: (i) frequency
behaviour of the first 8 bands, (ii) detail of one band.
8
5
Magnitude [dB]
Magnitude [dB]
4 4.5
2 a a
b 4 b
c c
0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.055 0.06 0.065 0.07
Frequency [Normalized] Frequency [Normalized]
i ii
Fig. 6. Equalizer frequency responses for M = 32 subbands considering (a) 2048 FIR filter, (b) IIR of order N = 30 ,(c) method of [19]: (i) frequency
behaviour of the first 8 bands, (ii) detail of one band.
where Hi and Fi are the frequency responses of the analysis different prototype lengths. Also a representation of the time
and synthesis filter banks, respectively, a distortion index has interval needed to filter 4096 samples is reported. It is evident
been defined allowing direct comparison between different that for all the considered prototype, good performance has
equalizer with different number of bands, i.e., been achieved, obtaining values very close to 1. In particular,
increasing the number of subbands, the IIR equalizer shows
max T (ejω ) + min T (ejω ) better performance in terms of distortion index, confirming the
DI = . (23)
2 validity of the proposed approach in comparison with the first
FIR implementation. Furthermore, as underlined in Section
This index takes into account the amplitude distortion of II-C, the computational complexity of the IIR implementation
each band: better performance is achieved when it is ap- is lower than FIR implementation as shown in the Table,
proximately 1. Tabs.(II)-(IV) show the results in terms of considering ∆t values.
distortion index for the FIR equalizer considering different
number of subbands and different prototype lengths and for
the IIR equalizer considering different number of subbands and
Signal Processing
Speech, Music, and Audio Processing 414
8th International Symposium on Image and Signal Processing and Analysis (ISPA 2013) September 4-6, 2013, Trieste, Italy
10
Magnitude [dB] 8 5
Magnitude [dB]
6
4.5
4
2 a a
b 4 b
c c
0
0 0.01 0.02 0.03 0.04 0.05 0.06 0.028 0.029 0.03 0.031 0.032 0.033 0.034 0.035
Frequency [Normalized] Frequency [Normalized]
i ii
Fig. 7. Equalizer frequency responses for M = 64 subbands considering (a) 4096 FIR filter, (b) IIR of order N = 28, (c) method of [19]: (i) frequency
behaviour of the first 8 bands, (ii) detail of one band.
IV. C ONCLUSION [4] R. A. Greiner and M. Schoessow, “Design aspects of graphic equaliz-
ers,” J. Audio Eng. Soc., vol. 31, no. 6, pp. 394–407, 1983.
In this work, a new approach to M -band uniform IIR [5] D. A. Bohn, “Constant-q graphic equalizers,” J. Audio Eng. Soc.,
equalizer has been presented starting from a previous FIR vol. 34, no. 9, pp. 611–626, 1986.
implementation based on multirate systems and filter banks [6] J. Henriquez, T. Riemer, and R. Trahan Jr., “A Phase Linear Audio
theory. The proposed approach employs IIR filters to improve Equalizer: Design and Implementation,” J. Audio Eng. Soc., vol. 38,
the filter banks structure developed to avoid ripple between no. 9, pp. 653–666, Sept. 1990.
adjacent bands. The solution presented here has several advan- [7] H. Schopp and H. Hetzel, “A Linear Phase 512 Band Graphic Equalizer
tages increasing the equalization performance in terms of low using the Fast Fourier Transform,” in Proc. 96th Audio Engineering
Society Convention, Amsterdam, The Netherlands, Feb. 1994.
computational complexity, low delay, and uniform frequency
[8] A. Gilloire and M. Vetterli, “Adaptive filtering in subbands with
response. The effectiveness of the optimized implementation critical sampling: analysis, experiments and application to acoustic echo
has been tested taking into comparison the previous FIR cancellation,” IEEE Trans. Signal Process., vol. 40, no. 8, pp. 1862–
approach and a technique of the literature that try to solve 1875, Aug. 1992.
the same problem. The obtained results in terms of desired [9] M. Petraglia and P. Batalheiro, “Non uniform Subband Adaptive Fil-
final frequency response and distortion index have confirmed tering with Critical Sampling,” in Proc. International Symposium on
the validity of the proposed approach, that shows a low compu- Circuits and Systems, vol. 56, no. 2, Island of Kos, Greece, May 2006,
pp. 565–575.
tational complexity and filtering delay especially considering a
[10] S. Cecchi, P. Peretti, L. Palestini, F. Piazza, F. Bettarelli, and A. Lattanzi,
great number of subband. Future works will be oriented to the “Real Time implementation of an Innovative Digital Audio Equalizer,”
development of the system on a real time platform, in order in Proc. 123rd Audio Engineering Society Convention, New York, NY,
to analyze the obtained performance. USA, Oct. 2007.
[11] S. Cecchi, L. Palestini, E. Moretti, and F. Piazza, “A New Approach
to Digital Audio Equalization,” in Proc. Workshop on Applications
R EFERENCES of Signal Processing to Audio and Acoustics, New Paltz, NY, USA,
[1] S. A. Azizi, “A new concept of interference compensation for parametric October 2007, pp. 62–65.
and graphic equalizer banks,” in Proc. 112nd Audio Engineering Society [12] M. Virgulti, S. Cecchi, A. Primavera, L. Romoli, E. Ciavattini,
Convention, Munich, Germany, May 2002. F. Bettarelli, and F. Piazza, “Optimized Implementation of an Innovative
[2] J. Vieira, “Digital Five Band Equalizer with Linear Phase,” in Proc. Digital Audio Equalizer,” in Proc. 133rd Audio Engineering Society
100th Audio Engineering Society Convention, Copenhagen, Denmark, Convention, San Francisco, CA, USA, Oct. 2012.
May 1996. [13] R. D. Koilpillai and P. P. Vaidyanathan, “Cosine-modulated fir filter
[3] Y. Lim, “A Digital Filter Bank for Digital Audio Systems,” IEEE Trans. banks satisfying perfect reconstruction,” Signal Processing, IEEE Trans-
Circuits Syst., vol. 33, no. 8, pp. 848–849, Aug. 1986. actions on, vol. 40, no. 4, pp. 770–783, apr 1992.
[14] S. G. Kim and C. D. Yoo, “Highly selective m-channel iir cosine-
modulated filter banks,” Electronics Letters, vol. 39, no. 20, pp. 1478–
1479, oct. 2003.
0
a [15] M. Petraglia, R. Alves, and P. Diniz, “New Structures for Adaptive
−20 b Filtering in Subbands with Critical Sampling,” IEEE Trans. Signal
Process., vol. 48, no. 12, pp. 3316–3327, Dec. 2000.
−40 [16] P. P. Vaidyanathan, Multirate Systems and Filterbanks. Prentice-Hall,
Phase [Deg]
Inc., 1993.
−60
[17] F. Cruz-Roldan, P. Amo-Lopez, S. Maldonado-Bascon, and S. S. Law-
−80 son, “An Efficient and Simple Method for Designing Prototype Filters
for Cosine-Modulated Pseudo-QMF Banks,” IEEE Signal Process. Lett.,
−100 vol. 9, no. 1, pp. 29–31, Jan. 2002.
[18] N. Wong and C. Lei, “Fir filter approximation by iir filters based
−120 on discrete-time vector fitting,” in Proc. International Symposium on
0 0.05 0.1 0.15
Frequency [Normalized] Circuits and Systems, May 2007.
[19] D. McGrath, J. Baird, and B. Jackson, “Raised cosine equalization
Fig. 8. Phase response for M = 8 subbands considering a FIR filter of 1024 utilizing log scale filter synthesis,” in Proc. 117th Audio Engineering
samples (a) and a IIR filter of order 50 (b). Society Convention, San Francisco, CA, USA, Oct 2004.
Signal Processing
Speech, Music, and Audio Processing 415