Sunteți pe pagina 1din 9

IEEE/ACM TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 23, NO.

5, MAY 2015 923

Design of Low Complexity Adjustable Filter Bank


for Personalized Hearing Aid Solutions
Ying Wei, Member, IEEE, and Yinfeng Wang

Abstract—The emerging demand for personalized hearing aids


requires the filter bank of a hearing aid system to be capable of de-
composing the sound waves in accordance with the characteristic
of the patient’s hearing loss. In this paper, an efficient adjustable
filter bank is proposed to achieve this goal. By careful design, the
number of the subbands as well as the location of the subbands can
be easily adjusted by changing a 4-bit control signal. The proposed
filter bank has extremely low complexity due to the adoption of
fractional interpolation and the technique of symmetric and com-
plementary filters. Only one prototype filter is needed for each of Fig. 1. A schematic diagram for digital hearing aid.
the stages, the multiple passbands generation stage and masking
stage. We show, by means of examples, that the proposed filter
bank can meet different needs of hearing loss cases with accept- spectrum, the adjustment is fully programmable and is able to
able delay. suit patients’ comfort.
Index Terms—Adjustable, filter bank, fractional interpolation, Much study has been invested into the design of digital filter
hearing aids. banks for selective amplification. Most of the current studies
focus on fixed (cannot be reconfigured) filter banks. Uniform
filter banks are the first and most widely used filter banks in
I. INTRODUCTION practice. Over the past decade, researchers have done much

T HE auditory system is a very sensitive and complex net- work to reduce the complexity of uniform filter banks. Lattice
work. Diseases, drugs, noise, trauma and aging may have wave digital filter bank (LWDFB) was employed for hearing
resulted in varying degrees of hearing loss, which makes hearing aids [4]. LWDFB has lower complexity than the FIR filter
impairments one of the most common sensory disturbances in bank and is not sensitive to the coefficients. Then, a DFT filter
the world. The most effective way to compensate hearing loss bank with a multi-dimensional logarithmic number system
is to employ a hearing aid system which is an integration of (MDLNS) was reported to obtain reduced complexity [5].
voice amplification, noise reduction, feedback suppression, au- Later, some well-known simple methods for critically sampled
tomatic program switching, environmental adaptation, and etc. filter banks were extended to the over-sampled case [6]. Its
The basic function of a hearing aid system is to amplify sounds efficiency comes from the flexibility to generate multiple pro-
selectively and then transfer the processed signal to the ear [1]. totype filters by one method. The complexity of hearing aid
A schematic diagram for digital hearing aids is shown in Fig. 1. system was further reduced by using a joint stereo filter bank
After the analog sound signal is transformed into the digital to satisfy the requirements of both the audio coding and the
signal by an A/D converter, the digital signal is divided into sub- hearing aid application [7].
band signals within different frequency bands by a filter bank Dividing the frequency range uniformly is straightforward
[2][3]. Each subband has its own amplification coefficient. The yet does not consider the unique characteristic of human
amplified subband signals are then synthesized and fed into the hearing. Therefore, non-uniform filter banks that mimic the
D/A converter. Filterbank-based algorithms permit an easy ad- resolution characteristic of human hearing have gained the
justment of speech amplification. Within the considered speech attention of hearing-aid researchers. A tree-structured filter
bank based on all-pass compensatory filters and elliptic min-
imal Q-factor (EMQF) filters was used as the analysis filter
Manuscript received February 09, 2014; revised October 08, 2014; accepted
February 13, 2015. Date of publication March 06, 2015; date of current version bank in [8]. An 8-band filter bank based on frequency response
April 09, 2015. This work was supported in part by the Shandong Province Sci- masking (FRM) was proposed for hearing compensation in [9].
ence and Technology Development Plan (No. 2013GGX10103), in part by the
Both the designs lower the complexity at the cost of delay. A
National Natural Science Foundation (No. 61201372), in part by the Promotive
Research Foundation for Excellent Young and Middle-Aged Scientists of Shan- critical band-like spaced filter bank was used in [10]. It can
dong Province (No. BS2013DX042), and in part by the Taishan Scholar Foun- obtain satisfactory hearing compensation, yet the irregularity
dation Project (No. 1170082963013). The associate editor coordinating the re-
of the subbands increases the difficulty of the design and
view of this manuscript and approving it for publication was Prof. Søren Jensen.
The authors are with the School of Information Science and Engineering, implementation. A 1/3 octave filter bank was realized to cover
Shandong University, Jinan 250010, China (e-mail: eleweiy@sdu.edu.cn; the hearing frequency range in [11]. It was based on IIR struc-
673645473@qq.com).
ture thus could not provide linear phase response. In general,
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. non-uniform filter banks have better performance in hearing
Digital Object Identifier 10.1109/TASLP.2015.2409774 compensation compared to the uniform filter banks. However,

2329-9290 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
924 IEEE/ACM TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 23, NO. 5, MAY 2015

inserted in between two coefficients. The number of the pass-


bands of the magnitude response of , , is deter-
mined by the interpolation factor , as shown in equation (1),

(1)

The bandwidth of the th passband, , is determined by the


ratio of the decimation factor to the interpolation factor ,
Fig. 2. Frequency responses of fractional interpolated filter and masking filters.
which is shown in equation (2),
and
the complexity of non-uniform filter banks is generally much (2)
higher than that of the uniform filter banks.
With the development of hearing aid technology, new de- where is the bandwidth of the prototype filter . The
mands to hearing aid systems appear. One of them is “flexi- center frequencies of the passbands, , can be calculated by
bility,” which requires the core of the digital hearing system, equation (3),
the filter bank, to be adjustable according to the characteristic
of the patient’s hearing loss. Little study has been done on the (3)
design of adjustable (reconfigurable) filter banks for hearing
aids. A programmable spectrum cut-up permitted to adjust fil- From equations (1) and (3), we can see that by changing the
ters’ bands relatively to patient’s pathology in [12]. However, interpolation factor , the number and the positions of the pass-
the realization of each subband was not discussed. Recently a bands of can be changed. Furthermore, from equa-
three-channel variable filter bank was proposed [13][14]. The tion (2), we can see that by changing the interpolation factor
variable filter bank has simple structure due to the use of IIR as well as the decimation factor , the bandwidths of the
structure and non-linear transformation yet it cannot provide the passbands of can be changed. The above particular-
linear phase property. In [15], a filter bank with adjustable sub- ities of the fractional interpolated filters form the basis of the
bands was proposed. All the subbands were first generated and configurability of the proposed filter bank. Additionally, more
then some of them were selected and some of them were aban- passbands can be generated if the operation of complementa-
doned. This design can achieve satisfactory performance yet the tion is adopted and if the operation of subtraction is introduced
complexity is comparatively high. In [16], a reconfigurable filter between the fractional interpolated filters. It should be noted that
bank which employed the frequency response masking tech- the decimation process will cause degeneration of the passband
nique was proposed. The complexity of the filter bank was low and stopband ripples compared to the original filter. Therefore
while the throughput is too long to be used in practice. Though the prototype filter should be designed with smaller passband
the adjustable filter banks meet the new trends in digital hearing and stopband ripples than expected in order to make the final
aids, the efficiency and effectiveness of such kind of algorithms filter meet the specifications.
needs to be further improved. The variable passbands generated by fractional interpolation
In this paper, a linear phase reconfigurable digital FIR filter are then extracted by the masking filters. To generate uniform
bank with small complexity and acceptable delay is proposed. subbands, according to equation (2) and (3), the center frequen-
The proposed filter bank is based on fractional interpolation, cies of the passbands except for the first and the last ones should
which allows us to build the system using small number of pro- be in consistent with the cutoff frequencies of the masking fil-
totype filters. The whole frequency range is divided into three ters, which means that by masking a whole passband is divided
regions, each of which has three different subband decompo- into two parts and each part is extracted to be an individual sub-
sition schemes. By changing the control signal, different band band. It is straightforward to assign different masking filters for
decomposition schemes are chosen without changing the struc- different fractional interpolated filters. However, the complexity
ture of the filter bank. of the system will be huge. In order to reduce the complexity, the
The paper is organized as follows. In Section II, the funda- masking filters had better be reusable. Based on equation (3), as
mental ideas of the design are presented. In Section III, the struc- long as the interpolation factor is determined, the centers of the
ture of the proposed filter bank is described and the implemen- passbands are fixed. This fact suggests that the passbands gen-
tation issue is also discussed. Then in Section IV, design exam- erated by different filters with the same interpolation factor may
ples are given and the experimental results are analyzed. Finally, be extracted using the same masking filter that covers a certain
conclusion is drawn in Section V. frequency range. This idea is illustrated in Fig. 2. Two fractional
interpolated filters and are shown in
the figure. Since both of their interpolation factors are , the
II. THE FUNDAMENTAL IDEAS OF THE DESIGN center frequencies of their passbands are the same. Except for
the first and the last passbands, other passbands are divided into
The proposed filter bank is based on fractional interpolation two parts by two adjacent masking filters. The two fractional in-
[17]. For a fractional interpolated filter , every th terpolated filters can share the same set of masking filters which
coefficients of the prototype filter are grouped together together cover the whole frequency range. The number of the
discarding the in between coefficients and then ( ) zeros are masking filters is .
WEI AND WANG: DESIGN OF LOW COMPLEXITY ADJUSTABLE FILTER BANK 925

Fig. 3. The structure of the proposed filter bank.

III. THE STRUCTURE OF THE PROPOSED FILTER BANK


Based on the fundamental principles discussed in Section II,
we propose a novel adjustable filter bank for hearing aids. The
whole frequency range is uniformly divided into three regions,
the low frequency region (0, ), the middle frequency region
( , ) and the high frequency region ( , ). For each
region, there are three band-schemes to be selected,
Scheme 1: 1 subbands with bandwidth ,
Scheme 2: 2 subbands with bandwidth ,
Scheme 3: 4 subbands with bandwidth .
Based on the specifications, the structure of the proposed filter
bank is illustrated in Fig. 3. There are two functional blocks
for the proposed filter bank, the multiple passbands generation
block and the masking block. Since the order of the cascade
does not affect the final results, to reduce the delay, we put the
masking block in front of the multiple passbands generation
block. Proper delays should be added to the branches to balance
the group delay, which are not shown in the figure.

A. The Masking Block


The masking block is to extract the desired passbands. The
function of the subfilter is to produce the masking filters
which divide the whole frequency range into 3 uniform regions
naturally. The cutoff frequency of is . The “o” port of
the subfilter provides the output of the original filter while
the “h” port provides the output of the highpass filter symmetric
with at , denoted as . This highpass filter can be
easily produced from the original filter, which will be discussed
later. The -transform transformations of the three masking fil-
ters, denoted as , ,2,3, are shown in the following
equations,
Fig. 4. The generation of the subbands.
(4)
(5) B. The Multiple Passbands Generation Block
(6) and its fractional interpolated versions consist of a
block whose function is to generate multiple passbands. The
The outputs of , and are kept in the storages “o” ports of the interpolated filters provide the original outputs
and then sent to the multiple passbands generation block one by while the “c” ports provide the outputs of the complementary
one in accordance with the 2-bit region-selection signal of filters. Since the number of the masking filters is the same
switch . The low frequency region is selected if . with the interpolation factor, equals to 3. The generation of
The middle frequency region is selected if . The high the subbands is shown in Fig. 4. The outputs of this block are
frequency region is selected if . Switches , and denoted as , which means the th subband for
together are responsible for selecting the scheme to be used scheme in region (region 1: , region 2:
in that region. and region 3: ).
926 IEEE/ACM TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 23, NO. 5, MAY 2015

For scheme 1, the masking filters already produce three TABLE I


uniform subbands with bandwidth , therefore no addi- SUBBANDS GENERATION
tional passband generation is needed. For scheme 2, half of
the passbands is generated by and the other half
of the passbands is generated by the complementary filter of
. These passbands are then extracted by the masking
filters, as shown in Fig. 4(b) and (c). Let us suppose the cut-off
frequency of and are and , respectively.
To generate uniform passbands with bandwidth (half of
the bandwidth of ), in accordance with equation (2) and
Fig. 4(b), we have,

(7)

For scheme 3, , and together are


used to generate the desired passbands with bandwidth
(fourth of the bandwidth of ), as shown in Fig. 4(d) to (h).
In order to facilitate the following description, let us denote,

(8)
(9)
(10)
(11)
Furthermore, to avoid aliasing, the stopband edge of , ,
The magnitude response of provides the subbands
should satisfy
and after masking. To obtain
uniform subbands, the following equations should be held,
(20)
(12)
Using and to replace , we have,
(13)
(21)
Set and substitute into the equations (7),
(12) and (13), we have,
In the multiple passbands generation block, switches ,
(14) and are controlled by parameters and . For switch ,
the upper branch is connected when and the lower
(15)
branch is connected when . For switch , it is open
(16) when and it is close when (we will see later
that is the symbol which represents a set of switches). The
Now we know that the cut-off frequencies of the prototype
state of switch is opposite to the state of and controlled by
filters and are and , respectively. In ad-
the complement of . The control signals for the switches ,
dition, we need to have some understanding of the transition
and together with the corresponding inputs and outputs
bandwidths of them. To ensure that the masking filters can ex-
are shown in Table I. For example, when the control signal
tract the desired subband clearly, the stopband edge of ,
is 1010, scheme 2 is used in region 3. Since there
, should be smaller than the left passband edge of the second
are three regions, three rounds are needed to get all the outputs
passband of , , as shown in Fig. 4(f),
in accordance with the structure. The outputs of each round are
(17) stored and wait for being outputted.

Suppose the transition bandwidths of and are and C. Working Procedure


, respectively. Using the variables already known to replace To make the algorithm more easily to be understood, the
and , we have, working procedure of the proposed reconfigurable filterbank is
shown in Fig. 5.
(18)
D. Implementation
which results in the relationship between and , The proposed filter bank can be implemented efficiently. The
prototype filters and are both linear phase FIR filters
(19) with odd lengths. Suppose coefficients of and are
WEI AND WANG: DESIGN OF LOW COMPLEXITY ADJUSTABLE FILTER BANK 927

Fig. 6. Implementation of and .

lation was obtained by replacing each delay element of the orig-


inal filter by delay elements. An implementation example is
shown in Fig. 7. In this example, the length of the prototype
filter is assumed to be 7. The switches are also shown in
the figure.

IV. EXPERIMENT RESULT AND ANALYSIS


Let us illustrate the idea of the proposed filter bank by an ex-
ample. The cutoff frequency of is and in accordance
with equation (14), the cutoff frequency of is . From
equation (19) we know that the sum of the transition bandwidths
of the two prototype filters is at most . In order to reduce the
complexity, we employ . Set , we get
. Based on the cutoff frequencies and the transi-
tion bandwidths, the passband and stopband edges of and
can be calculated, which are ( , ) and ( ,
), respectively. The desired passband ripple is 0.005 and
the desired stopband attenuation is dB.
The performance of the filterbank was evaluated for two dif-
ferent hearing loss cases represented by the audiograms. An au-
diogram presents a pure tone curve that illustrates the hearing
thresholds which are the softest sounds one can hear. A person
with impaired hearing tends to have a low sensitivity towards
certain frequencies where hearing thresholds are higher than the
normal speech intensities. The main task of the hearing aid is
to selectively amplify the audio sounds such that the processed
sound equalizes or “matches” one’s audiogram [18]. Since the
more subbands are used, the more flexible compensation can
Fig. 5. Working procedure.
be provided in that region, we use the scheme with more sub-
bands (narrower subbands) in the region where the slope of the
[ , , , ] and [ , , , ], respectively, where threshold curve is greater. The gain of each subband is then op-
and are the orders of the filters, we have, timized to reduce the maximum matching error (The recruit-
ment-phenomenon is not considered here).
Example 1 Audiogram for Mild Hearing Loss in the High Fre-
(22)
quencies: People with such kind of hearing loss cannot hear
and s’s, z’s, th’s, v’s, and other soft, high frequency consonants. The
audiogram is shown in Fig. 8(a). The right ear hearing thresh-
(23) olds (represented by ‘O’) were considered to be compensated.
The slope of the threshold curve is greater in the middle fre-
In the masking block, the original filter and its symmetric quency range, thus more subbands are put in that region. Since
filter at can be integrated together, as shown in Fig. 6. the hearing is fine in low frequency range, wider subbands are
can be produced by changing the sign of the coefficients of used in low frequency region. The decomposition plan for re-
alternatively. Therefore the multipliers can be shared. gion 1, 2 and 3 are scheme 1, scheme 3 and scheme 2, respec-
In the multiple passbands generation block, the implementa- tively. Altogether 7 subbands are generated. The magnitude re-
tion of the fractional interpolated filters and their complemen- sponse of the filter bank is shown in Fig. 8(b) and the matching
tary filters need to be considered. In order to share the mul- result is shown in Fig. 8(c).
tipliers, transposed form was used. The decimation was real- We can see that the proposed filterbank provides satisfactory
ized by selecting every th output of multipliers. The interpo- matching between the audiogram and the magnitude response
928 IEEE/ACM TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 23, NO. 5, MAY 2015

Fig. 7. Example of the implementation of the fractional interpolated filters (suppose the coefficients of is [ , , , , , , ]).

TABLE II TABLE III


COMPARISON OF THE FILTERBANKS FOR EXAMPLE 1 COMPARISON OF THE FILTERBANKS FOR EXAMPLE 2

of the filterbank. The comparisons of the maximum matching TABLE IV


COMPARISON OF COMPLEXITY WITH OTHER TYPES OF FILTERBANKS
error and the delay to that of the fixed uniform filterbank (direct
design), the fixed non-uniform filterbank in [9] and the recon-
figurable filterbank in [16] are shown in Table II. To make a
reasonable comparison, different types of filterbanks were de-
signed with the same attenuation and the same number of sub-
bands. It is shown in Table II that the maximum matching error
of the proposed filterbank is reduced by about 12% compared
with that of the uniform filterbank and reduced by about 44%
compared with that of the non-uniform filterbank in [9]. The
maximum matching error is increased by about 14% compared bank is shown in Fig. 9(b) and the matching result is shown in
with that of the reconfigurable filterbank in [16], however, the Fig. 9(c). The comparisons of the maximum matching error and
delay is reduced by about 58%. In fact the filterbank in [16] for the delay to that of the uniform filterbank (direct design), the
this example cannot be used in a real application because its fixed non-uniform filterbank in [9] and the reconfigurable filter-
delay is too large while the proposed filterbank can be used due bank in [16] are shown in Table III. It can be seen that the max-
to its acceptable delay. imum matching error of the proposed filterbank is reduced by
Example 2 Audiogram for Mild Hearing Loss in All Fre- about 69% compared with that of the uniform filterbank, about
quencies: Conversational speech can be understood at distances 49% compared with that of the non-uniform filterbank in [9] and
of 1-2 meters for a person with hearing loss of the severity about 31% compared to the reconfigurable filterbank in [16].
shown in the audiogram of Fig. 9(a). This communication dis- The specifications of the filter bank leads to a of length
tance is significantly lower than the up to 12 meter distance at 81 and a of length 69. Since both are linear phase filters,
which a person with normal hearing can understand conversa- 76 different coefficients are needed for generating total 21 dif-
tional speech. The right ear hearing thresholds (represented by ferent subbands. The complexity comparison with other types of
‘O’) were considered to be compensated. The hearing threshold adjustable filter banks (with the same specifications) is shown in
curve has bigger fluctuation in the low frequency range than in Table IV. Compared to the adjustable IIR filter bank proposed
the middle and high frequency ranges. Therefore we consider in [13], the proposed filter bank has slightly smaller complexity
putting more subbands in the low frequencies. The decompo- as well as more subbands and a linear phase property. Further-
sition plan for region 1, 2 and 3 are scheme 3, scheme 2 and more, the proposed filter bank achieves more than 54.1% sav-
scheme 1, respectively. The magnitude response of the filter ings in terms of the number of multipliers per band than the
WEI AND WANG: DESIGN OF LOW COMPLEXITY ADJUSTABLE FILTER BANK 929

Fig. 9. Example 2 (a) audiogram for mild hearing loss in all frequencies,
(b) The magnitude response of the filter bank, (c) matching result.

Fig. 8. Example 1 (a) audiogram for mild hearing loss in the high frequen-
cies,(b) The magnitude response of the filter bank, (c) matching result. and visual integration. The delay of the proposed filter bank can
be calculated by equation (24),

non-uniform filter bank proposed in [9] and achieves more than (24)
15.6% savings in terms of the number of multipliers per band
than the reconfigurable filter bank proposed in [16].
where is the delay of masking and is the delay of multiple
passbands generation for region . For a linear phase fractional
V. DELAY ANALYSIS interpolated filter , its group delay can be obtained
using equation (25),
In hearing aid applications, it is necessary to study the delays
of the decomposition filter banks. It was reported in [19] that de-
(25)
lays longer than 20 ms may cause interference between speech
930 IEEE/ACM TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 23, NO. 5, MAY 2015

TABLE V VI. CONCLUSIONS


DELAYS OF DIFFERENT CASES
A flexible filter bank based on fractional interpolation is pro-
posed. The fractional interpolation provides the reconfigura-
bility of the number of subbands and the locations of the sub-
bands. Multiple passbands are generated after fractional inter-
polation and the employment of the technology of symmetric
filters and complementary filters. The passbands are extracted
by the masking filters which are obtained by the same prototype
filter. Totally two prototype filters are needed for the whole filter
bank which makes the complexity extremely low. A 4-bit signal
is employed to control the selection of the schemes. We show,
by examples, that the proposed filter bank can meet different
needs of hearing loss cases with acceptable delay.

where is the order of the prototype filter and is the sampling REFERENCES
frequency. [1] A. R. Moller, Hearing: Anatomy, Physiology and Disorders of the Au-
In the masking stage, the signal passes through the masking ditory System, 2 ed. New York, NY, USA: Academic, Sep. 11, 2006.
[2] M. A. Hersh and M. A. Johnson et al., Assistive Technology for the
filters (the length of is 81), thus the delay of the masking, Hearing-Impaired, Deaf and Deaf-Blind. London, U.K.: Springer-
, is 2.5 ms under the sampling frequency 16 kHz. The delays Verlag, 2003.
, ,2,3, caused by passbands generation depend on the [3] A. M. Engebretson, “Benefits of digital hearing aids,” IEEE Eng. Med.
Biol. Mag., vol. 13, no. 2, pp. 238–248, Apr.–May 1994.
schemes used. Let us denote the delay using scheme as [4] M. T. Tan, J. S. Chang, and Y. C. Tong, “A novel low-voltage low-
. For the region in which scheme 1 is employed, no further power wave digital filter bank for an intelligent noise reduction digital
process is needed, thus we have . For the region in which hearing instrument,” in Proc. IEEE Int. Symp. Circuits Syst., Sydney,
Australia, May 06–09, 2001, vol. 2, pp. 681–684.
scheme 2 is employed, the output of the masking stage passes [5] H. Li, G. A. Jullien, V. S. Dimitrov, M. Ahmadi, and W. Miller, “A
through or its complement filter. In accordance with 2-digit multidimensional logarithmic number system filter bank for
equation (25), we have ms. For the region in a digital hearing aid architecture,” in Proc. IEEE Int. Symp. Circuits
Syst., 2002, pp. II-760–II-763.
which scheme 3 is employed, the longest path is the one that the [6] D. Hermann, E. Chau, R. D. Dony, and S. M. Areibi, “Window based
signal passes through , thus we have prototype filter design for highly oversampled filter banks in audio ap-
plications,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process.,
ms. For example, in an application, two regions use Scheme Honolulu, HI, USA, Apr. 15–20, 2007, pp. II-405–II-408.
1 and one region uses Scheme 2. The delay is [7] R. Dong, D. Hermann, R. Brennan, and E. Chau, “Joint filter bank
ms.The group delays of different cases are shown in structures for integrating audio coding into hearing aid applications,”
in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process., Apr.–Mar.
Table V. “Scheme ( )” means scheme is used times. 31–4, 2008, pp. 1533–1536.
From Table V we can see that, the delay of the case 10 when [8] R. Cassidy and J. O. Smith, “A tunable, nonsubsampled, nonuniform
scheme 3 is used three times is slightly longer than 20 ms. The filter bank for multi-band audition and level modification of audio sig-
nals,” in Proc. 38th Asilomar Conf. Signals, Syst., Comput., Nov. 7–10,
delay of the other cases is satisfactory. In general, the more 2004, vol. 2, pp. 2228–2232.
subbands are used, the larger delay we have. The delays of ex- [9] Y. Lian and Y. Wei, “A computationally efficient non-uniform FIR
ample 1 and example 2 are both 12.1 ms which achieves 58.3% digital filter bank for hearing aid,” IEEE Trans. Circuits Syst. I: Reg.
Papers, vol. 52, no. 12, pp. 2754–2762, Dec. 2005.
and 51.6% savings compared to that of the same examples in [10] K. S. Chong, B. H. Gwee, and J. S. Chang, “A 16-channel low-power
[16]. Compared with the work in [9], the delay is reduced by nonuniform spaced filter bank core for digital hearing aid,” IEEE Trans.
Circuits Syst., vol. 53, no. 9, pp. 853–857, Sep. 2006.
29.8%. [11] Y.-T. Kuo, T.-J. Lin, Y.-T. Li, and C.-W. Liu, “Design and imple-
To achieve a reconfigurable filterbank, interpolation is often mentation of low-power ANSI S1.11 filter bank for digital hearing
used as the passband width of a filter can be adjusted by aids,” IEEE Trans. Circuits Syst. I: Reg. Papers, vol. 57, no. 7, pp.
1684–1696, Jul. 2010.
changing its interpolation factor. In general, the more are the [12] A. B. Hamida, “An adjustable filter-bank based algorithm for hearing
passbands, the larger is the interpolation factor. However, large aid systems,” in Proc. Int. Conf. Ind. Electron., Control, Instrum., 1999,
interpolation factor will cause long delay. In this work, the vol. 3, pp. 1187–1192.
[13] T. B. Deng, “Three-channel variable filter-bank for digital hearing
delay reduction lies in two aspects. aids,” IET Signal Process., vol. 4, no. 2, pp. 181–196, Apr. 2010.
(1) The function of the masking block is not only masking [14] N. Ito and T.-L. Deng, “Variable-bandwidth filter-bank for low-power
but also dividing one band into two, as shown in Fig. 4. There- hearing aids,” in Proc. 3rd Int. Congr. Image Signal Process., 2010,
pp. 3207–3201.
fore, to generate the same number of subbands, the interpo- [15] Y. Wei and D. Liu, “A design of digital FIR filter banks with adjustable
lation factor of the passband generation block is smaller than subband distribution for hearing aids,” in Proc. 8th Int. Conf. Inf.,
Commun., Signal Process., Singapore, Dec. 13–16, 2011, pp. 361–364.
that of [9] and [16]. [16] Y. Wei and D. Liu, “A reconfigurable digital filterbank for hearing
(2) By careful design, the masking filters were shared for the aid systems with a variety of sound wave decomposition plans,” IEEE
three schemes. Since the masking block does not need to be Trans. Biomed. Eng., vol. 60, no. 6, pp. 1628–1635, Jun. 2013.
[17] R. Mahesh and A. P. Vinod, “Coefficient decimation approach for real-
reconfigurable, we use simple filters (without any interpolation) izing reconfigurable finite impulse response filters,” in Proc. IEEE Int.
to realize its function which reduces the delay. Symp. Circuits Syst., Seattle, WA, USA, May 2008, pp. 81–84.
The price of such a delay reduction is that the way of masking [18] Y. C. Lim, “A digital filter bank for digital audio systems,” IEEE Trans.
Circuits Syst. , vol. CAS-3, no. 8, pp. 848–849, Aug. 1986.
leads to uniform subbands in each region, which reduces the [19] J. Agnew, “An overview of digital signal processing in hearing instru-
flexibility of compensation. ments,” The Hearing Review, Jul. 1997.
WEI AND WANG: DESIGN OF LOW COMPLEXITY ADJUSTABLE FILTER BANK 931

Ying Wei (M’04) received the B. S and M. S. de- Yinfeng Wang received the B.S. degree from Shan-
grees from Xi’an Jiaotong University, Xi’an, China, dong University, Jinan, China, in 2010 He is currently
in 2000 and 2003, respectively. She received the working toward the M.S. degree in signal and infor-
Ph.D. degree from National University of Singapore mation processing at Shandong University. His re-
(NUS), Singapore, in 2008. Dr. Wei worked as search interests include speech processing and filter
research fellow with the Department of Electrical design.
Engineering of NUS from 2008 to 2009. She joined
Shandong University, China, in 2010 where she is
currently an Associate Professor at the School of
Information Science and Engineering. Her research
interests include digital filter design, VLSI imple-
mentation of high-speed digital systems, and biomedical signal processing. She
served in technical program committees, organizing committees, and session
chairs for some international conferences.

S-ar putea să vă placă și