LAC 2009 - Paper: Application of Linux Audio in Hearing Aid Research

2009·1.5 MB·English

by  Giso Grimm, Tobias Herzke, Volker Hohmann

Application of Linux Audio in Hearing Aid Research
Giso GRIMM Tobias HERZKE Volker HOHMANN
Medizinische Physik Ho¨rTech gGmbH Ho¨rTech gGmbH
Universtita¨t Oldenburg Marie-Curie-Str. 2 Marie-Curie-Str. 2
26111 Oldenburg, 26129 Oldenburg 26129 Oldenburg
Germany, Germany Germany
[email protected] [email protected] [email protected]
Abstract still limited processing power and a low battery
capacity make it necessary to avoid every un-
Development of algorithms for digital hearing aid
processing includes many steps from the first algo- neededprocessingcycle. Thus,hearingaidalgo-
rithmic idea and implementation up to field tests rithms are substantially tested in as many and
with hearing impaired patients. Each of these as realistic situations as possible. A solution for
steps has its own requirements towards the devel- field testing is the implementation of a hearing
opment environment. However, a common platform aid in a programmable and portable signal pro-
throughout the whole development process is desir-
cessing system, e.g., the DASi [1]. The draw-
able. This paper gives an overview of the applica-
back of such DSP-based systems is the need of
tion of Linux Audio in the hearing aid algorithm
implementing the typically complex algorithms
development process. The performance of portable
in low level languages, which requires large ef-
hardwareintermsofdelay,batteryruntimeandpro-
cessing power is investigated. fort for re-configuration and modification. This
is different if standard hardware and standard
Keywords operating systems can be used for algorithm
Hearing aids, algorithm development, low delay, development and evaluation. Recent develop-
feedback cancellation ments in Linux Audio suggest that the Linux
platform might be well suited for hearing aid
1 Introduction
research. A key feature is the easy access to
Theintroductionofthefirstcommercialhearing low delay real-time audio processing on stan-
aid with digital signal processing in 1996 sub- dard hardware.
stantially changed the methods used in hearing In the first part of this paper, an overview
aid research. Whereas typical analogue hearing of algorithms for digital hearing aids is given.
aids have signal processing blocks for frequency Evaluation methods are briefly described. In
shaping, dynamic compression and static notch the second part, the application of Linux Au-
filters for feedback cancellation, hearing aids dio for hearing aid research is demonstrated by
with digital signal processing allow the imple- the example of a portable field testing device.
mentation of algorithms which have no coun- Processor performance data and the analogue
terpart in the analogue domain. Another re- delay of a Linux based hearing aid is given in
cent development in hearing aid technology is the results section.
the availability of a low-delay wireless binaural
2 Overview of Algorithms for
link between the left and right ear in commer-
Digital Hearing Aids
cial hearing aids. This link opens up new cate-
goriesofsignalprocessing: Binauralsignalanal- Key algorithms of hearing aids – be it digi-
ysis and presentation of a binaurally enhanced tal or analogue – is frequency shaping and dy-
audio signal. namic compression, for correction of audibil-
Hearing aid algorithms have to be beneficial ity and loudness recruitment. Loudness re-
forthehearingimpairedinanobjectivelyacces- cruitment denotes the reduced dynamic range
sible way. The most important benefit is an im- between hearing threshold and uncomfortable
provement of speech intelligibility in noisy envi- loudnessleveloftenfoundinsensorineuralhear-
ronments. Alsolessobviousimprovements,such ing loss, the most common type of hearing loss.
as reduction of listening effort, increase of lis- For an increase of signal-to-noise ratio (SNR)
tening comfort, improvement of spatial percep- and thus a better understanding in adverse lis-
tion, and artifact reduction are important. The tening conditions, hearing aids often use mul-
tiple microphones for directional filtering. The filtering and feedback cancellation schemes, a
easiest approach are delay-and-sum beamform- real-time version of the algorithms under test is
ers with a fixed beam, which can reach up to necessary, to account for head and jaw move-
6dBgaininSNRforon-axissounds. Moreelab- ments. Although most of the evaluation meth-
orated are adaptive beamformers [2] and side- ods can only be applied in laboratories, an eval-
lope cancellers. Single channel noise reduction uation of the algorithms under test in real-life
[3] and binaural de-reverberation [4] are algo- conditions or even in the field are desirable.
rithms which can increase the listening comfort However, the more realistic the test conditions,
andreducethelisteningeffort. However,speech the less reproducible are the results.
intelligibility can not be improved significantly
2.2 Feedback howling and low delay
by this class of algorithms [5]. Hearing aids
constraints
also introduce artifacts, which can reduce au-
dio quality or even speech intelligibility. One of The amplification of hearing aids is limited by
the largest problems in hearing aids is the feed- the feedback from hearing aid receivers to the
back howling, caused by leakage of the receiver microphones. The criterion for feedback howl-
(speaker) signal back to the microphones, and ingtooccurisfulfillediftheamplificationofthe
a typically large amplification. Feedback arti- hearing aid is larger then the attenuation of the
factscanbereducedbythreecategoriesofalgo- feedback path – mainly the damping of the ear
rithms: Adaptivefeedbackcancellationschemes mould, and if the round trip phase, i.e., hear-
[6] estimate the feedback signal in an adaptive ingaidandfeedbackpath,isanintegermultiple
filter and remove the estimated feedback sig- of 2π. In Figure 1 the amplitude and phase re-
nal from the microphone input. Binaural infor- sponseofahearingaidfeedbackpathisplotted.
mation can be utilised to increase stability of Frequenciesatwhichthecriterionisfulfilledare
hearing aids [7], and feedback problems can be marked by a circle. The feedback criterion is
avoidedbyopeningtheloopincriticalfrequency time dependent because of changes in the feed-
ranges with non-linear operations in the hear- back path, e.g., ear mould leakage by jaw move-
ing aid, e.g., frequency transposition or phase ments, reflections by phone receivers and room
distortion [8]. For a more detailed overview of acoustics, and by the time variant amplification
algorithms see [9] and [10]. of the hearing aid, caused by dynamic compres-
sion and signal enhancement. The number of
2.1 Evaluation methods frequencies at which the criterion is fulfilled –
andthusfeedbackcanpossiblyoccur–increases
For the evaluation of hearing aid algorithms
withthedelay(seephaseresponsesfordelaysof
both objective evaluation methods and sub-
1 ms and 5 ms in Figure 1).
jective methods are necessary. The improve-
ment of SNR is a canonical measure for algo-
3 Implementation of a Linux-based
rithms which aim to remove or attenuate noise,
field test hearing aid
e.g., directional microphones and noise reduc-
tion schemes. The Speech Intelligibility Index
For subjective evaluation of advanced hearing
(SII) is commonly used to predict speech intel-
aid algorithms with hearing impaired subjects,
ligibility in non-fluctuating noise [11]. Exten-
a portable field test hearing aid based on a Net-
sions of the SII include the addition of binaural
book computer with a dedicated USB audio de-
processing [12]. Quality measures range from
vice to connect hearing aid shells has been im-
technical measures like spectral distortion [13]
plemented. This hardware runs UbuntuStudio
to methods based on audio perception models
with a real-time patched Linux 2.6.24 kernel,
[14].
generic ALSA driver for USB sound cards, and
To measure speech intelligibility subjec- the JACK audio connection kit. The H¨ortech
tively,typicallythespeechrecognitionthreshold Master Hearing Aid (MHA, [16]) connects to
(SRT) is measured by adaptively changing the JACK and hosts the hearing aid algorithms.
levelofspeechortheSNRofspeechinnoiseun- The MHA can be fitted to an individual hear-
til a certain percentage of the presented speech inglossusingaTCP/IPconfigurationinterface.
signal can be understood [15]. Audio quality End user control (i.e., program switching, vol-
and other subjective attributes can be assessed ume control) and data inspection and logging
by quality scaling or paired comparison. is also possible via the network connection. A
For the subjective evaluation of directional picture of the system is shown in Figure 2.
B ceivers (speakers) and provide sufficient output
d 20
0 level for subjects with a large hearing loss.
2
+ 0
G 3.1 CPU and battery performance
L−20
O 0.5 1 2 4 8
During the development of the field test hear-
ms) pi ing aid, two different Netbook computers have
1 been tested and compared regarding their pro-
e ( 0
s cessing performance and battery runtime. The
a
h−pi
p 0.5 1 2 4 8 dataareshowninFigure3andFigure4. Asoft-
frequency / kHz
s) pi ware hearing aid with four different programs
m
5 wasrunningduringthetest. Thesoftwarehear-
e ( 0 ing aid was running at a JACK sampling rate
s
ha−pi of 32 kHz, and with 1 ms blocks. Internally, the
p 0.5 1 2 4 8
frequency / kHz signalwasresampledto16kHz. Theprocessing
containedleveladaptationandfrequencyequal-
Figure 1: Amplitude (top panel) and phase re- isation, fouralternativetypesofsignalenhance-
sponse (middle and bottom panel) of a hearing ment (directional filtering, two versions of sin-
aid feedback path. Circles mark the frequencies gle channel noise reduction and binaural coher-
at which the feedback criterion is fulfilled. The ence based noise reduction, see [17] for details
phase response is given for a hearing aid with on the algorithms). The signal enhancement
1 ms delay (middle panel) and with 5 ms delay wasfollowedbyamulti-banddynamiccompres-
(bottom panel). sion scheme, which was fitted to a sample hear-
ing loss (typical high-frequency hearing loss).
The compressor, coherence filter and one single
channel noise reduction are based on an overlap
add FFT method [18].
70
IRQ
jack
60 MHA
algo
50
%
d in 40
a
o
U l 30
P
C
20
10
0
Figure 2: Portable field test hearing aid. The algo1algo2algo3algo4 algo1algo2algo3algo4 algo1algo2algo3algo4
Asus Eee PC 701 Acer Asp ire One Desktop P4
host PC is an Asus Eee PC 701 running Ubun-
tuStudio 7.10 and the Master Hearing Aid [16]. Figure 3: Performance of a field test hearing
A dedicated USB audio device is used for con- aid system on different hardwares. In addition
necting the hearing aid shells to the computer. to the algorithm CPU time, the CPU time used
by the jackd sound server and the sound card
interrupt handler was measured.
Thesoundcard1 providesfourinputchannels
and two output channels, and is USB powered.
Hearing aid shells (silver headsets in Figure 2) 3.2 Acoustic Delay of PC-based
with microphones can be directly connected via hearing aids
cables. With a low output impedance of 7.6 Ω
The acoustic delay in hearing aids should be as
it is able to drive low-impedance hearing aid re-
small as possible. A maximum acceptable delay
of 20 ms was found for hearing aids with closed
1The sound card was developed for application in a
fitting(theearcanalisclosedbytheearmould)
field test hearing aid by OFFIS e.V., Oldenburg, Ger-
many. [19]. For hearing aids with open fittings, i.e., an
3h30’ thesoftwarerequiresanti-aliasingfiltersaswell,
which will introduce an additional group delay.
3h15’ The delay of the hardware anti-aliasing filters
and signal handling τ has been estimated by
e 3h00’ sc
m subtracting number of periods times the period
nti
y ru 2h45’ size from the total delay.3 The delay τsc con-
er tainsthedelayoftheanti-aliasingfilterofboth,
batt 2h30’ AD- and DA-converters.
2h15’
Device f /kHz P τ /ms τ /ms
s sc t
2h00’ Echo 32 32 2.81 4.81
algo1algo2algo3algo4 algo1algo2algo3algo4 Layla 3G 44.1 32 2.04 4.22
Asus Eee PC 701 Acer Aspire One 44.1 64 2.04 4.94
48 32 1.88 3.88
Figure 4: Runtime of the hearing aid prototype
64 32 1.27 2.77
with a fully charged battery and the lid closed.
88.2 32 0.92 2.01
96 32 0.84 1.84
100 32 0.81 1.77
earmouldwithaventoronlyaslimtubewitha RME 32 64 3.34 7.34
dome,thedelayneedtobebelowapproximately HDSP9652 + 44.1 64 2.68 5.58
6 ms to avoid disturbances [20]. The delay of Behringer 48 64 2.52 5.19
typical audio signal processing algorithms with Ultramatch 64 64 2.16 4.16
block processing is determined by three sources SRC2496 64 128 2.14 6.14
ofdelay: (i)Bythegroupdelayofthealgorithm 88.2 128 1.8 4.71
itself, which can be frequency dependent and 96 128 1.73 4.40
time variant, (ii) by block processing, which is RME 32 64 2.13 6.12
typically two times the period size, and (iii) by HDSP9652 + 44.1 64 1.61 4.51
the anti-aliasing filters and signal handling in Behringer 48 64 1.46 4.12
theaudiointerfaceandAD-andDA-converters. Ultragain
The total delay of a hearing aid with no algo- ADA8000
rithmic group delay (i.e., identity processing), RME 32 64 2.09 6.09
was measured using JACK. The total delay τ HDSP9632+ 44.1 64 1.52 4.42
t
isthedelayfromtheelectricalinputtotheelec- ADI8QS 48 64 1.4 4.06
trical output. To allow an easier measurement, 64 128 1.55 5.55
instead of opening the loop at the electric in- 88.2 128 1.12 4.02
and output, a direct analogue connection from 96 128 1.03 3.7
output to input was made and the loop was OFFIS 16 16 6.81 9.81
opened digitally, i.e., the delay from a JACK USB SC-4/2 32 32 4.38 7.38
client output to a JACK client input via the 44.1 64 4.08 8.44
hardware was measured. For a validation of the 48 64 3.96 7.96
results, all measurements have been repeated, 96 128 1.97 5.97
drop-outs have been monitored and the total
harmonic distortion (THD) has been measured Table 1: Total delay τt from electrical input
to ensure a non-distorted signal output.2 The to electrical output, including delay caused by
data are shown in Table 1. The shortest de- block processing and the delay τsc caused by
lay of 1.77 ms is reached with the Echo Layla digital transmission and anti-aliasing filters, for
3G at 100 kHz sampling rate and 32 samples several professional audio devices clocked at the
period size. However, hearing aids are typi- sampling rate fs and using a JACK period size
cally processed at low sampling rates between P.
16 and 32 kHz. A downsampling from 100 kHz
to the desired hearing aid sampling rate within
2The measurement of THD was necessary because 3Forperiodsizesbelow1mstheEchoLayla3Gdriver
some sound cards produced audible distortion at short seemstouseanadditionthirdperiodforbuffering,which
period sizes, but JACK did not report an error or xrun. adds to the delay.
4 Discussion is possible to develop closed software on the
Linux platform without infringing any licenses,
Recent developments in the Linux Audio world
and this does not place a greater burden on the
made it possible to develop and evaluate algo-
development company than using closed-source
rithms for digital hearing aids on the Linux op-
software. Instead, it opens a wealth of exist-
erating system. A portable field test hearing
ing third-party libraries that may help in get-
aidwithalowoveralldelaycanbeimplemented
ting the product to market fast. This requires
based on standard hardware and the Linux op-
careful consideration of what software compo-
erating system.
nents to use, how they are being used, knowl-
4.1 Hardware and performance edge of the relevant licenses, and compliance
The processing performance of miniature lap- with their terms. Sometimes this means, that a
top PCs is sufficient for many advanced hear- third-partylibrarycanbeused,butmayonlybe
ing aid algorithms. However, the runtime with- linked dynamically and shipped together with
out recharging the battery is below four hours, itssourcecode. Otherfreesoftwarecomponents
which is not sufficient for full-day testing. The may not be used at all, or different terms of
round-tripdelayofmosttestedsoundcardsisin licensing have to be negotiated with the copy-
the order of 5 ms at 44.1 kHz. The tested USB right holder. This process of considering soft-
soundcardhasadelayof8.4ms,however,thisis ware components and their licenses, however,
the only portable bus powered sound card with does not differ between closed and open soft-
a sufficient number of input channels, which is ware development platforms.
required for multi-microphone processing, e.g.,
5 Conclusions
beamforming.
The advantage of dedicated DSP-based field Basedontheresultsofthisstudyandtheobser-
testingenvironmentsistheefficientexecutionof vation of general aspects of Linux Audio it can
algorithmsandlow-delaycapabilities. However, be concluded that Linux Audio is applicable in
to achieve this benefit, low-level program code hearing aid research. Requirements of technical
has to be written, which is typically time con- kind,e.g.,processingperformanceandaudiode-
suming and does not offer access to high-level lay are fulfilled, and the availability of software
libraries. and supported hardware is sufficient for the re-
search work. The authors think that these fea-
4.2 Usability and Accessibility
tures make it superior to DSP-based hardware
A major advantage of using Linux Audio solutions.
againstDSPbasedhardwaresolutionsistheac-
cessibility of low-delay audio performance and 6 Acknowledgements
a development environment: Real-time patched
We thank the members of the Medical Physics
kernels are available from several audio-related
groupatOldenburgUniversityandB.Kollmeier
Linux distributions. Most standard hardware is
for continuous support. Special thanks to
supported by these mainstream kernels. Sound
Carsten Beth and Frank Poppen from OFFIS
cardsandtoothersourcesandsinksofaudioare
e.V.forthedevelopmentoftheUSBsoundcard.
easily accessible via ALSA and the Jack Audio
Work supported by grants from the European
Connection Kit. This is opposed by a restricted
Union FP6, Project 004171 HEARCOM.
access to hardware and development environ-
All trademarks mentioned in the text are
ments for DSP-based hardware solutions.
property of their respective owners.
4.3 Licensing
References
Application in industrial context often requires
[1] U. Rass and G. H. Steeger. A high perfor-
closed-source software development. Suggest-
mancepocket-sizesystemforevaluationsin
ingLinuxas adevelopmentplatform sometimes
acousticsignalprocessing.EURASIPJour-
meets with reservations from industry decision
nal of Applied Signal Processing, 3:163–
makers who fear that the open source licens-
168, 2001.
ing of the platform might somehow spread to
theirproduct. Thepasthasalsoseencourtsen- [2] L. J. Griffiths and C. W. Jim. An alterna-
tences against companies who failed to comply tive approach to linearly constrained adap-
with the conditions of the GPL software that tive beamforming. IEEE Trans. on Anten-
they distributed. Our point of view is that it nas Propagation, 30:27–34, 1982.
[3] Y. Ephraim and D. Malah. Speech en- [14] R.HuberandB.Kollmeier. Pemo-q.anew
hancement using a minimum mean-square method for obejctive audio assessment us-
errorshort-timespectralamplitudeestima- ing a model of auditory perception. IEEE
tor. IEEE Trans. ASSP, (32):1109–1121, Transactions on Audio, Speech and Lan-
1984. guage Processing, 14:1902–1911, 2006.
[4] Th. Wittkop and V. Hohmann. Strategy- [15] K. Wagener and T. Brand. Sentence intel-
selective noise reduction for binaural dig- ligibility in noise for listeners with normal
ital hearing aids. Speech Communication, hearing and hearing impairment: influence
39:111–138, 2003. ofmeasurementprocedureandmaskingpa-
rameters. International Journal of Audiol-
[5] M. Marzinzik. Noise reduction schemes for
ogy, 44(3):144–156, 2005.
digital hearing aids and their use for hear-
ing impaired. PhD thesis, University of [16] G. Grimm, T. Herzke, D. Berg, and
Oldenburg, 2000. V. Hohmann. The Master Hearing Aid – a
PC-based platform for algorithm develop-
[6] J.M.Kates. Feedbackcancellationinhear-
ment and evaluation. Acta Acustica united
ing aids: Results from a computer simula-
with Acustica, 92:618–628, 2006.
tion. IEEE Trans. Signal Process., 39:553–
562, 1991. [17] K. Eneman, H. Luts, J. Wouters,
M. Bu¨chler, N. Dillier, W. Dreschler,
[7] V. Hohmann, V. Hamacher, I. Holube,
M. Froehlich, G. Grimm, V. Hohmann,
B. Kollmeier, and T. Wittkop. Method for
R. Houben, A. Leijon, A. Lombard,
the operation of a hearing aid device or
D. Mauler, M. Moonen, H. Puder,
hearing device system as well as hearing
M. Schulte, A. Spriet, and M. Vormann.
aid device or hearing device system, March
Evaluation of signal enhancement al-
2006. US Patent 7,013,015.
gorithms for hearing instruments. In
[8] M. Schroeder. Improvement of acoustic- Proceedings of the 16th European Sig-
feedback stability by frequency shifting. nal Processing Conference, Lausanne,
Journal of the Acoustical Society of Amer- Switzerland, 2008.
ica, 36(9):1718–1724, 1964.
[18] J. B. Allen. Short term spectral analy-
[9] J. M. Kates. Applications of digital signal sis, synthesis, and modification by discrete
processing to audio and acoustics, chapter Fourier transform. IEEE Trans. Acoust.,
Signal processing for Hearing Aids. Kluwer Speech, Signal Processing, 25(3):235–238,
Academic Publishers, 1998. June 1977.
[10] H.Dillon. Hearing Aids. BoomerangPress, [19] M. A. Stone and B. C. Moore. Tolerable
Australia, 2001. hearing aid delays. i. estimation of limits
imposed by the auditory path alone using
[11] ANSI-S3.5. American national standard
simulated hearing losses. Ear and Hearing,
methodsforthecalculationofthespeechin-
20(3):182–192, 1999.
telligibility index. AmericanNationalStan-
dards Institute, New York, 1997. [20] M. A. Stone, B. C. Moore, K. Meisen-
bacher, and R. P. Derleth. Tolerable hear-
[12] R. Beutelmann and T. Brand. Predic-
ing aid delays. v. estimation of limits
tion of speech intelligibility in spatial noise
for open canal fittings. Ear and Hearing,
and reverberation for normal-hearing and
29(4):601–617, 2008.
hearing-impaired listeners. Journal of the
Acoustical Society of America, 120(1):331–
342, 2006.
[13] K. Eneman, A. Leijon, S. Doclo, A. Spriet,
M. Moonen, and J. Wouters. Advances
in Digital Speech Transmission, chapter
Auditory-Profile-Based Physical Evalua-
tion of Multi-Microphone Noise Reduction
Techniques in Hearing Instruments. John
Wiley and Sons, Ltd, 2008.