Gravitational Search Algorithm for IIR Filter-Based

Audio Equalization
Giovanni Pepe1 , Leonardo Gabrielli1 , Stefano Squartini1 , Luca Cattani2 , and Carlo Tripodi2
1
Department of Information Engineering, Università Politecnica delle Marche, Ancona, Italy
2
ASK Industries S.p.A., Reggio Emilia, Italy
[email protected], {l.gabrielli, s.squartini}@univpm.it, {CattaniL, TripodiC}@askgroup.it

Abstract—In this paper we present an evolutionary algorithm allow to partition the problem in sub-bands [3]. These are
for the design of stable IIR filters for binaural audio equalization. composed of several peaking filter, one per octave, or one per
The filters are arranged as a cascade of second-order sections third-octave, implemented as second-order sections (SOS’s).
(SOS’s) and the gravitational search algorithm (GSA) is used.
This process seeks for optimal coefficients based on a fitness One approach for automatic equalization was proposed in [8],
function, possibly leading to unstable filters. To avoid this, where two optimization methods, the Direct Search Method
we propose two alternative methods. Experiments have been (DSM) and the Rosenbrock Method were applied to the
performed taking an in-car listening environment as the use case, research of optimal gains for a graphic equalizer in order to
characterized by multiple loudspeakers, thus, multiple impulse obtain a given target curve. One issue with a graphic equalizer
responses (IR). This technique has been compared with a previous
heuristic method, achieving superior results. is the limited degrees of freedom it has. Other strategies can
Index Terms—Gravitational search algorithm, Audio equaliza- be envisioned for the design of the coefficients, where the
tion, IIR filters, Automotive filters are not designed by a closed-form equation, but have
coefficients derived from iterative heuristic processes looking
Audio equalization is a process aimed at increasing the
for solutions that are optimal according to some defined
sound quality in a listening environment [1]. Typical listening
constraints.
scenarios include indoor facilities, such as halls, venues or
Several algorithms have been proposed for filter design
rooms, or small environments such as vehicle cockpits. These
based on novel paradigms. Neural networks have been pro-
two categories present different issues. The in-car scenario
posed for the filter design task. In [9] a simple neural network
features small volumes, very short reverberation with no
is trained to correct target gains for a graphic equalizer in
separation between direct wave and reflections and peculiar
order to reduce the error due to interference of adjacent filters.
materials, such as car window glass, plastic, fabric, etc. [2].
In [10] a neural network is devised to design an IIR filter.
A large number of techniques for audio equalization have Wang et al. [11] proposes a two step optimization Frequency-
been proposed, based on linear filtering [1], [3]. Room re- Response Masking (FRM) technique based on the design of
sponse equalization, generally asks for the inversion of an a FRM filter optimizing the subfilters, further optimized by
impulse response. While this can be easily done for a single decomposing it into several linear neural networks. Agrawal
location, it is more difficult for multiple listening points, et al. [12] used Particle Swarm Optimization (PSO) with
or even for the binaural case, since the impulse responses fractional derivative to design a stable IIR filter, while Foresi et
perceived at the typical ear spacing of 18 cm can be different al. [13] implemented PSO with fractional derivative constraints
[4], [5]. to design a quasi-linear phase IIR filter for Audio Crossover
In the years, some nonlinear techniques have been used, Systems, achieving the desired filter with a flat magnitude
including machine learning-based ones. An early work from response. In [14] the authors achieved an IIR filter using the
Chang et al. [6] proposed a Time Delay Neural Network Artificial Immune Algorithm and compared the results with
(TDNN) to process the signal sample-wise to inverse-model Genetic Algorithm (GA), the Touring Ant Colony Optimiza-
the room acoustics. In [7], the authors used an end-to-end tion (TACO) and Tabu Search (TS). In [15] the authors use
Convolutional Neural Network (CNN). These approaches may Gravitational Search Algorithm (GSA) to model an IIR filter
have a very high computational cost, since the neural network and a nonlinear rational filter, then they compare the technique
is required to directly process the signal. with PSO and GA. In this case, the algorithms provide filter
Other approaches may be envisioned, based on efficient coefficients as outputs.
techniques such as linear filtering, but taking advantage of
machine learning techniques for the design of the filters A. Scope of the work
coefficients. The design of filters for room equalization is Motivated by the works above, in [16], we employed the
a well-known topic and a plethora of approaches have been GSA algorithm for the design of FIR filters for audio equal-
proposed (see, e.g., [1]). While FIR filters offer the advantage ization in a binaural scenario, with results superior to the state
of providing linear-phase, IIR can be less expensive. Some of the art techniques. In this work we extend the application
researchers have opted for IIR graphical equalizers which of GSA to the design of stable IIR filters for equalization.

978-9-0827-9705-3 496 EUSIPCO 2020

Authorized licensed use limited to: Raytheon Technologies. Downloaded on September 03,2021 at 11:15:35 UTC from IEEE Xplore. Restrictions apply.
 Some considerations must be underlined: unlike some of where randj is a uniformly random number in the range [0, 1],
the works mentioned above, here we want to remove any con-  is a small value and Rij [n] is the euclidean distance between
straint on the SOS design and let the evolutionary algorithm the agents i and j, Rij [n] = ||Xi [n], Xj [n]||2 , Abest is a set
directly find optimal coefficients for several cascaded SOS’s. of the best A agents.
As a consequence, with this approach, the coefficients are free
to take any value, possibly leading to unstable impulse re- vi [n + 1] = randi · vi [n] + ai [n] (5)
sponses. To overcome this issue, we introduce two alternative Xi [n + 1] = Xi [n] + vi [n + 1] (6)
constraints to avoid instabilities, one based on pole reflection
and one on triangular stability. where randi is uniformly distributed random number in range
The equalization scenario we consider is that of a room, [0, 1].
specifically a car cabin. Generally, the environment is com- G[n] is the gravitational constant that varies over time. In
posed of S loudspeakers, each with an IIR filter associated. our case, G[n] linearly decrease, defining an initial and a final
M listening points are placed in the environment that is, thus, value (G0 and Gf respectively).
characterized by S ×M impulse responses (IR). The design of n
such filters is conducted by GSA. This process is performed G[n] = G0 − (G0 − Gf ) · (7)
N
off-line, due to the high computational cost. For this reason,
where N is the number of iterations.
the method is meant for pre-tuning. Time varying effects,
including the movement of the listener head and position, are II. D IRECT S EARCH M ETHOD
not taken into account.
The direct search method (DSM) is a heuristic method that
For a broader comparison, we also take the DSM in consid-
uses random variation Γ of the parameters c close to the initial
eration applying to our use case. Differently from [8] we use
values, where −γ ≤ Γ ≤ γ [8], [18]. If the new value of the
the DSM to directly design the filters coefficients, thus, the
parameter achieves a better fitness function, it is kept and used
two stability-preserving methods proposed for GSA are also
as the initial value for the next iteration, otherwise it is updated
applied to DSM.
using a random Γ value:
The paper outline follows: in Section I and Section II the
GSA and DSM, respectively, are described. In Section III ĉi = ci (1 + Γ) (8)
the proposed IIR filter design is exposed. In Section IV
experiments details are given, while in Section V the results where ĉi is the i-th parameter.
are reported. Finally, Section VI concludes the paper. Compared to [8], where the DSM is used to design IIR
filters optmizing gain level, filter quality factor and center
I. G RAVITATIONAL S EARCH A LGORITHM frequency (thus, all SOS’s are designed using closed-form
The gravitational search algorithm (GSA) is based on agents equations), we use this method to design each coefficient
subject to the Newton gravitational law [17]. The agent A of the SOS’s. Another difference concerns the initialization:
attracts every other agent with a force. The i-th agent is defined our experiments are performed by randomly initializing the
by the position Xi = [x1i , · · · , xdi , · · · , xD
i ], where D is the coefficients and the gains, while in [8] parameters have been
dimension of the search space. The mass Mi of i-th agent at choosen a priori at the initialization.
n-th iteration is defined as:
qi [n] III. IIR FILTER DESIGN
Mi [n] = A (1)
P Audio equalization is performed using IIR filters composed
qj [n] by a cascade of N Biquad filters and defined as [19]:
j=1
N
where qi [n] is defined as: Y bs,k,0 + bs,k,1 z −1 + bs,k,2 z −2
Hs (z) = gs (9)
f iti [n] − worst[n] 1 + as,k,1 z −1 + as,k,2 z −2
qi [n] = (2) k=1
best[n] − worst[n]
where Hs (z) is the s-th IIR filter composed by N SOS’s.
where f iti [n] is the fitness value of the i-th agent, best[n] and With the DSM, for each SOS, the algorithms designs the
worst[n] are the best and worst fitness value of all agents, filter coefficients (numerator and denominator) and the gain
respectively. gs . Similarly, for the GSA, each agent Ai , determines the
Like the law of gravity, the force of the i-th agent Fi is coefficients and gain of a SOS.
calculated considering a set Abest of heavier masses: The fitness function for the algorithms is the average
Mj [n]·Mi [n]
Mean Square Error (M SE) between the equalized magnitude
P
Fi [n] = randj · G[n] · Rij [n]+ · (Xj [n] − Xi [n]) (3)
j∈Abest , response Hm (ω) and the desired magnitude response Hdes (ω),
j6=i
averaged among all microphones:
Finally, acceleration, velocity and position of i-th agents
ωh
(ai [n], vi [n], Xi [n], respectively) are calculated.
(|Hm (ω)| − |Hdes (ω)|)2
P
M
ai [n] = Fi [n]
=
P
randj ·
Mj
G[n] Rij [n]+ · (Xj [n] − Xi [n]) (4) 1 ω=ωl
X
Mi [n] M SE = (10)
j∈Abest ,
j6=i
M m=1 ωh − ωl

497

Authorized licensed use limited to: Raytheon Technologies. Downloaded on September 03,2021 at 11:15:35 UTC from IEEE Xplore. Restrictions apply.
 (a) (b) (c)

Fig. 1: Frequency responses of first and second microphone: (a) using DSM with 4 SOS’s; (b) using GSA with 4 SOS’s; (c)
using GSA with 32 SOS’s. Green lines are the non-equalized frequency responses, the blue lines are the equalized frequency
responses

In designing coefficients with DSM and GSA, it is possible ran on a Linux-based Intel i7 server, with 32GB RAM and a
to incur in unstable SOS’s. For this reason, we propose two GTX Titan GPU. The DSM and GSA have been implemented
methods to avoid unstable filters: the pole reflection technique using tensors to allow the processing to run on the GPU.
[19] and a technique based on triangular stability [20]. For the In our experiments, the IIR filters have been designed by
first we calculate the poles of each second order filter at each cascading N = 4, 8, 16 or 32 SOS’s. In the GSA case, the
iteration. If the pole is outside the unit circle, pole reflection number of agents have been varied from 10 to 50 agents, while
is applied: the number of best agents Abest was set from 2 to 10 agents. G
1 has been varied setting the initial and final value: 1.0 · 10−4 ≤
pi = ∗ if |pi | ≥ 1 for i = 1, 2 (11) G0 ≤ 0.1 and 0.1 ≤ Gf ≤ 10.0. On DSM, γ has been set to
pi · (1 + ε)
0.05. The number of iterations is equal to 1000 for both GSA
where ε is a small value to avoid conditional stability. Finally both for DSM.
the coefficients are calculated and updated: For the initialization phase, agent and parameter values have
as,k,1 = −(p1 + p2 ) as,k,2 = p1 · p2 (12) been randomly set with an uniform distribution between -1.0
and 1.0.
With the second method, we check whether the denominator
coefficients are inside the stability triangle, i.e.: V. R ESULTS
(
|as,k,2 | < 1 Experiments are compared in terms of M SE, as in Eq.
(13) (10), and the average standard deviation σ, calculated as:
|as,k,1 | < 1 + as,k,2 .
M
If this condition is not met, coefficients are updated according 1 X
to |as,k,2 | = 1 −  and |as,k,1 | = 1 + as,k,2 − . σ= σm (14)
M m=1
IV. E XPERIMENTS
where σm is the standard deviation of m-th microphone:
Experiments were performed using binaural impulse re- q Pωh
1 2
sponses (M = 2) from an Alfa Romeo Giulia cockpit. σm = ωh −ω l +1 ω=ωl (10 · log10 |Hm (ω)| − D) (15)
The impulse responses have been measured using a Kemar
ωh
mannequin type 45BA placed on the driver’s seat and Audition 1 X
D= (10 · log10 |Hm (ω)|). (16)
3.0 with Aurora plug-in in order to generate a Sine Sweep ωh − ωl + 1 ω=ω
l
and to measure the impulse response [21], using a Roland
Octa-Capture as audio interface. Impulse responses have been Moreover, to assess the perceptual impact of the equaliza-
measured with a sampling rate of 28.8 kHz, later oversampled tion, we weight the M SE with the normalized A-curve ψA (ω)
to 48 kHz. Since the original sampling rate was 28.8 kHz, the [22], as follows:
upper frequency bound ωh was 14.4 kHz, while the lower ωh
ψa (ω)(|Hm (ω)| − |Hdes (ω)|)2
P
frequency bound ωl is equal to 20 Hz. The car is fitted with M
1 X ω=ωl
S = 7 loudspeakers, positioned as follows: two on the front M SE A =
M m=1 ωh − ωl
doors, two on the back doors, one the center part of the car,
(17)
one near the driver headrest and a subwoofer within the trunk.
Results are presented in Table I. Very low M SE are
GSA is performed using Python as programming language
obtained by the GSA using both pole reflection and triangular
and the Tensorflow1 library as a framework for the experi-
stability. However, the solutions obtained by the triangular
ments. DSM was implemented with Matlab. All experiments
stability constraint with 16 and 32 SOS’s are trivial. Indeed,
1 https://www.tensorflow.org/ the solutions found by the algorithm in these two cases applies

498

Authorized licensed use limited to: Raytheon Technologies. Downloaded on September 03,2021 at 11:15:35 UTC from IEEE Xplore. Restrictions apply.
 Pole reflection Triangular stability
Number of
Method M SE M SE A σ M SE M SE A σ
SOS’s
4 0.15 0.11 2.27 0.16 0.12 2.34
8 0.17 0.13 2.29 0.16 0.12 2.45
GSA 16 0.20 0.15 2.62 0.25 0.18 2.86
32 0.09 0.04 1.06 0.28 0.22 0.63
4 0.22 0.16 2.72 0.25 0.19 2.82
8 0.23 0.17 2.72 0.24 0.18 2.94
DSM 16 0.26 0.20 3.00 0.31 0.24 3.11
32 0.30 0.19 3.20 0.25 0.23 2.96

TABLE I: Best results of GSA and DSM with 4, 8, 16, 32

SOS’s IIR filters using pole reflection and stability constraint.
The experiments indicated in italic provide a trivial solution
to the problem, as detailed in Section V. Fig. 2: Frequency response of the experiment with GSA,
triangular stability, 32 SOS. The equalized curve is almost flat,
Method Filter (Order) M SE M SE A σ
No EQ - 2.14 1.23 3.57 however, the solution to the equalization problem is trivial, as
SD FIR (640) 0.98 0.72 7.13 all designed filters but one have extremely low gains.
PD FIR (512) 0.80 0.52 5.6
GSA FIR (768) 0.13 0.10 2.07
GSA IIR (32 SOS’s) 0.09 0.04 1.06

TABLE II: Best results from different methods, from [16]. 700

Number of pole reflections

600

500
a flat equalization to one of the loudspeakers and silences 400
all other loudspeakers by selecting extremely low gains in
300
the filters transfer function. The frequency response of all
200
filters but one stays always below -90 dB, thus the overall
100
magnitude frequency response, plotted in Figure 2, is mainly
0
given by the energy contribution of a single loudspeaker. These 0 100 200 300 400 500 600 700 800 900 1000
solutions, although valid for the optimization algorithm, are Iterations

not of interest for a practical application, thus these two results Fig. 3: Overall number of pole reflection occurrences
have been discarded. The best result not presenting such an
issue, is the 32 SOS experiment using pole reflection, which
achieves a 0.09 M SE. Its frequency response is shown in
Figure 1 where it is compared to the 4 SOS pole reflection significantly with respect to the non-equalized case, but the
case and the best DSM case. As can be seen, the frequency GSA method outperforms the other ones. In particular the
responses achieved with the 4-SOS’s GSA IIR presents a proposed IIR GSA design slightly improves over the previous
deep notch below 700 Hz, several peaks and notches at low method based on FIR filters, at a much lower computational
frequencies and many notches at high frequency. The 32- cost. Indeed, considering the cost of a SOS in terms of sums
SOS’s GSA IIR is able to reduce these, resulting in a quite flat and products, the cascade of 7 filters composed by 32 SOS’s
response over 1 kHz, and in the low range only a large bell requires 2016 floating point operations per sample, while 7
around 200 Hz is present. Regarding the DSM (Figure 1(c)), FIR filters with 768 taps requires 10745 operations. On the
its frequency responses present more notches and peaks than other hand, the PD method, as implemented in [16], employs
GSA, especially at high frequency. only one FIR filter for all seven loudspeakers. Since the best
Figure 1 shows the frequency responses of GSA and DSM. results is for the FIR length of 512 taps, only 1023 operations
The frequency responses achieved with the 4-SOS’s GSA IIR are required. Its use may be considered in low-computational
presents a deep notch below 700 Hz, several peaks and notches cost settings, however, the GSA method allows to achieve
at low frequencies and many notches at high frequency. The superior results even with 4 SOS’s, i.e. 252 floating point
32-SOS’s GSA IIR is able to reduce these, resulting in a quite operations.
flat response over 1 kHz, and in the low range only a large bell For what concerns the two stability constraints, after each
around 200 Hz is present. Regarding the DSM (Figure 1(a)), iteration, the SOS’s are checked for instability and, if it is the
its frequency responses present more notches and peaks than case, either pole reflection or triangular stability is enforced,
GSA, especially at high frequency. depending on the chosen method. In Figure 3 we show the
number of occurrences of pole reflection for all agents at
A. Considerations a given iteration. At initialization, a large number of pole
We first compare the GSA and DSM IIR design methods reflections occur, due to initial random values having high
with results from [16], based on the following FIR design probability of instability. As the iterations increase, there is a
methods: prototype design (PD) [2], stochastic descent (SD) decrease of pole reflections. This means that the algorithm is
[23] and GSA. Table II clearly shows that all methods improve able to update the coefficients to provide stable filters.

499

Authorized licensed use limited to: Raytheon Technologies. Downloaded on September 03,2021 at 11:15:35 UTC from IEEE Xplore. Restrictions apply.
 VI. C ONCLUSION [5] ——, “New factors in room equalization using a fuzzy logic approach,”
in 111th Audio Engineering Society Convention, New York, USA, 2001.
In this paper we proposed the use of an evolutionary algo- [6] P. Chang, C. G. Lin, and B. Yeh, “Inverse filtering of a loudspeaker and
rithm, the GSA for IIR filter design in the audio equalization room acoustics using time-delay neural networks,” The Journal of the
scenario. In order to prevent filter instabilities we proposed Acoustical Society of America, vol. 95, no. 6, pp. 3400–3408, 1994.
[7] M. A. Martinez Ramirez and J. D. Reiss, “End-to-end equalization with
two methods: pole reflection and triangular stability. Similarly, convolutional neural networks,” in Proceedings of the 21st International
we compare GSA to a modified version of the DSM, previ- Conference on Digital Audio Effects (DAFx-18), Aveiro, Portugal, 2018,
Retrieved from: http://dafx2018.web.ua.pt.
ously used for computing the gains, here modified to directly [8] H. Behrends, A. von dem Knesebeck, W. Bradinal, P. Neumann, and
compute the filter coefficients. We pick the in-car equalization U. Zölzer, “Automatic equalization using parametric iir filters,” J.
scenario as a use case, where multiple loudspeakers are Audio Eng. Soc, vol. 59, no. 3, pp. 102–109, 2011. [Online]. Available:
http://www.aes.org/e-lib/browse.cfm?elib=15778
associated to one IIR filter each, and a binaural head is used [9] V. Välimäki and J. Rämö, “Neurally controlled graphic equalizer,”
to record the individual impulse responses. IEEE/ACM Transactions on Audio, Speech, and Language Processing,
The experiments show that the GSA with pole reflection vol. 27, no. 12, pp. 2140–2149, 2019.
[10] N. Allakhverdiyeva, “Application of neural network for digital recursive
technique is superior to state of the art methods, including filter design,” in 2016 IEEE 10th International Conference on Applica-
the GSA algorithm for FIR filters. With respect to the latter, tion of Information and Communication Technologies (AICT), Oct 2016,
a computational cost reduction is also achieved, which may pp. 1–4.
[11] X.-H. Wang, Y.-G. He, and T.-Z. Li, “Neural network algorithm for
prove noteworthy for embedded applications. designing FIR filters utilizing frequency-response masking technique,”
In this work, IIR filter stability was introduced as a con- Journal of Computer Science and Technology, vol. 24, no. 3, pp. 463–
straint. Future works may, instead, reformulate the problem 471, May 2009.
to include stability as a penalty term within the fitness func- [12] N. Agrawal, A. R. Kumar, and V. Bajaj, “A new design method for
stable IIR filters with nearly linear-phase response based on fractional
tion. Furthermore, additional constraints or penalties may be derivative and swarm intelligence,” IEEE Transactions on Emerging
required to avoid trivial solutions, such as those found in this Topics in Computational Intelligence, vol. 1, no. 6, pp. 464–477, Dec
work. Finally, since the proposed methods are heuristic, the 2017.
[13] F. Foresi, P. Vecchiotti, D. Zallocco, and S. Squartini, “Designing quasi-
hyperparameters search may be further enlarged by launching linear phase IIR filters for audio crossover systems by using swarm
a large number of batch experiments in order to seek for better intelligence,” in Audio Engineering Society Convention 144, May 2018.
performance. [Online]. Available: http://www.aes.org/e-lib/browse.cfm?elib=19509
[14] A. Kalinli and N. Karaboga, “Artificial immune algorithm for IIR filter
Other experiments will be carried out in other acoustic design,” Engineering Applications of Artificial Intelligence, vol. 18,
scenes (i.e. multi-point equalization) to improve the study no. 8, pp. 919 – 929, 2005.
of machine learning techniques for audio equalization and to [15] E. Rashedi, H. Nezamabadi-pour, and S. Saryazdi, “Filter modeling
using gravitational search algorithm,” Engineering Applications of Artifi-
approximate linear phase on output responses. Moreover, deep cial Intelligence, vol. 24, no. 1, pp. 117 – 122, 2011. [Online]. Available:
learning techniques will be investigated for the design of audio http://www.sciencedirect.com/science/article/pii/S0952197610001120
equalization filters (both FIR and IIR) and perceptual tests will [16] G. Pepe, L. Gabrielli, S. Squartini, and L. Cattani, “Evolutionary
tuning of filters coefficients for binaural audio equalization,” Applied
be performed. Acoustics, vol. 163, p. 107204, 2020. [Online]. Available: http:
//www.sciencedirect.com/science/article/pii/S0003682X19305377
ACKNOWLEDGMENT [17] E. Rashedi, H. Nezamabadi-pour, and S. Saryazdi, “Gsa: A
The authors would like to thank ASK industries S.p.A. gravitational search algorithm,” Information Sciences, vol. 179, no. 13,
pp. 2232 – 2248, 2009, special Section on High Order Fuzzy
for financial support and technical assistance. This project is Sets. [Online]. Available: http://www.sciencedirect.com/science/article/
founded under the Italian Ministry of Economic Development pii/S0020025509001200
(MiSE) Ministerial Decree 24/05/2017 (Agreements for Inno- [18] G. Ramos and J. J. López, “Filter design method for loudspeaker
equalization based on iir parametric filters,” J. Audio Eng. Soc,
vation). vol. 54, no. 12, pp. 1162–1178, 2006. [Online]. Available: http:
R EFERENCES //www.aes.org/e-lib/browse.cfm?elib=13893
[19] J. G. Proakis and D. K. Manolakis, Digital Signal Processing (4th
[1] S. Cecchi, A. Carini, and S. Spors, “Room response equalization - a Edition), 4th ed. Prentice Hall, 2006.
review,” Applied Sciences, vol. 8, no. 1, p. 16, Dec 2017. [20] A. Antoniou, Digital Signal Processing, 2nd ed. Mcgraw-Hill, 2016.
[2] S. Cecchi, L. Palestini, P. Peretti, F. Piazza, and A. Carini, “Multipoint [21] A. Farina, “Advancements in impulse response measurements by sine
equalization of digital car audio systems,” in 2009 Proceedings of 6th sweeps,” in Audio Engineering Society Convention 122, May 2007.
International Symposium on Image and Signal Processing and Analysis, [Online]. Available: http://www.aes.org/e-lib/browse.cfm?elib=14106
Sep. 2009, pp. 650 – 655. [22] D. M. Howard and J. Angus, Acoustics and psychoacoustics. Routledge,
[3] V. Välimäki and J. D. Reiss, “All about audio equalization: Solutions 2017.
and frontiers,” Applied Sciences, vol. 6, no. 5, 2016. [23] W. Putnam, D. Rocchesso, and J. Smith, “A numerical investigation
[4] S. Bharitkar and C. Kyriakakis, “A cluster centroid method for room of the invertibility of room transfer functions,” in Proceedings of 1995
response equalization at multiple locations,” in Proceedings of the 2001 Workshop on Applications of Signal Processing to Audio and Accoustics,
IEEE Workshop on the Applications of Signal Processing to Audio and Oct 1995, pp. 249–252.
Acoustics (Cat. No. 01TH8575). IEEE, 2001, pp. 55–58.

500

Authorized licensed use limited to: Raytheon Technologies. Downloaded on September 03,2021 at 11:15:35 UTC from IEEE Xplore. Restrictions apply.