ech
T
P essScience
Compu e s, Ma e ials & Con inua
DOI:10.32604/cmc.2021.017568
A icle
Noise Reduc ion in Indus y Based on Vi ual Ins umen a ion
Radek Ma inek1, Rene Ja os1,JanBa os
1, Lukas Danys1, Aleksand a Kawala-S e niuk2,
Jan Nedoma3,*, Zdenek Machacek1and Ji i Kozio ek1
1Facul y o Elec ical Enginee ing and Compu e Science, Depa men o Cybe ne ics and Biomedical Enginee ing,
VSB–Technical Uni e si y o Os a a, 708 00, Os a a-Po uba, Czechia
2Facul y o Elec ical Enginee ing, Au oma ic Con ol and In o ma ics, Opole Uni e si y o Technology, Opole, Poland
3Facul y o Elec ical Enginee ing and Compu e Science, Depa men o Telecommunica ions, VSB–Technical
Uni e si y o Os a a, 708 00, Os a a-Po uba, Czechia
*Co esponding Au ho : Jan Nedoma. Email: jan.nedoma@ sb.cz
Recei ed: 03 Feb ua y 2021; Accep ed: 13 Ap il 2021
Abs ac : This pape discusses he educ ion o backg ound noise in an indus-
ial en i onmen o ex end human-machine-in e ac ion. In he Indus y 4.0
e a, he mass de elopmen o oice con ol (speech ecogni ion) in a ious
indus ial applica ions is possible, especially as ela ed o augmen ed eali y
(such as hands- ee con ol ia oice commands). As Indus y 4.0 elies hea ily
on adio equency echnologies, some b ie insigh in o his p oblem is p o-
ided, including he In e ne o hings (IoT) and 5G deploymen . This s udy
was ca ied ou in coope a ion wi h he indus ial pa ne B ose CZ spol.
s. .o., whe e sound eco dings we e made o p oduce a da ase . The expe i-
men al en i onmen comp ised h ee wo kplaces wi h backg ound noise abo e
100 dB, consis ing o a lase /magne ic welde and a p ess. A i ual de ice
was de eloped om a gi en da ase in o de o es selec ed commands om a
comme cial speech ecognize om Mic oso . We es ed a hyb id algo i hm
o noise educ ion and i s impac on oice command ecogni ion e iciency.
Using i ual de ices, he s udy was ca ied ou on la ge speake s wi h 20
pa icipan s(10 men and 10 women). The expe imen s included a la ge numbe
o epe i ions (100 imes o each command unde di e en noise condi ions).
S a is ical esul s con i med he e iciency o he es ed algo i hms. Lase
welding en i onmen e iciency was 27% be o e applied il e ing, 76% using
he leas mean squa e (LMS) algo i hm, and 79% using LMS +independen
componen analysis (ICA). Magne ic welding en i onmen e iciency was 24%
be o e applied il e ing, 70% wi h LMS, and 75% wi h LMS +ICA. P ess
wo kplace en i onmen e iciency showed no success be o e applied il e ing,
was 52% wi h LMS, and was 54% wi h LMS +ICA.
Keywo ds: 5G; hyb id algo i hms; signal p ocessing; speech ecogni ion
This wo k is licensed unde a C ea i e Commons A ibu ion 4.0 In e na ional License,
which pe mi s un es ic ed use, dis ibu ion, and ep oduc ion in any medium, p o ided
he o iginal wo k is p ope ly ci ed.
1074 CMC, 2021, ol.69, no.1
1 In oduc ion
Spoken wo d is s ill one o he mos na u al ways o di ec ly ans e in o ma ion be ween
people [1,2]. Di ec oice in e ac ion wi h compu e s and machine y is slowly gaining in impo -
ance, as he indus y is shi ing owa ds he Indus y 4.0 concep [3]. Voice communica ion
sys ems a e inc easingly in eg a ed in o bo h indus y and p i a e li e due o hei signi ican bene-
i s. Mos applica ions a e cu en ly limi ed o a small se o asks pe o med by speci ic machines
using p ede e mined commands ha hey can ecognize [4]. These sys ems a e pa icula ly use ul
when he ope a o mus do se e al hings a once.
In sma homes, con ol akes place be ween he de ice and he local ga eway by powe
line communica ion (PLC), T ansmission Con ol P o ocol (TCP), o Message Queue Teleme y
T anspo (MQTT) p o ocol, which enables con ol h ough sui able clien s (e.g., a sma phone o
Amazon Alexa) [5].
A digi al oice sys em was also designed o use wi h he in e ne o hings (IoT) o con ol
and simula e he p ocess o an assis ed obo ic load as pa o be e human- obo in e ac ion
(HMI) [6].
He e al. [7] designed a well-known A duino boa d wi h mo ion senso s and an audio ecei e
o con ol a obo ic ca using a cloud se e and IoT echnology, using p ese oice commands
in eg a ed wi h Google Voice API.
Indus ial applica ions o en ely on adio- equency echnologies, such as Wi-Fi, IoT ech-
nologies (e.g., SigFox o LoRa), o b oadband cellula ne wo ks (4G and 5G). In pa icula ,
5G is o en conside ed as a ool o a i icial in elligence (AI), Indus y 4.0, and IoT. New
communica ion s anda ds a e esilien , and designed wi h sma senso s o de ices and machine
communica ion in mind. I s la ency is much less han ha o 4G, and i is conside ably as e .
Reliabili y is eaching le els o wi ed connec ions, bu his is limi ed by he as scope o manu-
ac u ing plan s and Indus y 4.0. Howe e , 5G opens new ields o applica ion in indus y whe e
con en ional Wi-Fi ell sho in he pas . IoT and Indus y 4.0 senso s a e e e expanding–i
is es ima ed ha in 2022, almos 70 billion de ices will be connec ed o IoT ne wo ks. Cellula
ne wo ks o en o e unpa alleled co e age and scalabili y wi h obus and eliable connec ion.
Manu ac u e s a e expanding hei plan s by employing sma senso s o ack ca go o employees,
o o ga he manu ac u ing da a. The po en ial alue can uel he ise o au oma ed ac o ies.
Au oma ed ision quali y checks, augmen ed eali y cons uc ion, p edic i e main enance, sys em-
wide eal- ime p ocessing con ol, and au oma ed guided ehicles a e he u u e o Indus y 4.0.
While low-powe , wide-a ea solu ions a e su icien o some simple connec ed de ices, he opposi e
is ue in manu ac u ing, whe e machines a e da a-in ensi e and in close p oximi y. The powe
o mode n so wa e-de ined ne wo ks and he scalabili y suppo ed in 5G ne wo ks o e a mo e
agile and e icien model based on so wa e a he han adi ional ha dwa e solu ions. Vi ual
ne wo ks (o ne wo k slicing) and subne s adjus ed o speci ic needs a e possible wi h 5G. The
3 d Gene a ion Pa ne ship P ojec (3GPP) is wo king on a 17 h elease o mobile ne wo ks
(expanding 4G and 5G), which was o be comple ed in 2022–2023, bu may a i e la e due o he
ongoing COVID pandemic. Va ious eams a e ocusing on indus ial deploymen o 5G ne wo ks,
including indus ial IoT enhancemen , IoT o e non- e es ial ne wo ks, dynamic powe sa ing,
Na owband IoT enhancemen , p oximi y-based se ices, and 5G-based loca ion se ices. While
he p esen ed pape ocuses on di ec da a ga he ing, i can be conside ed a wo k in p og ess.
A combina ion o senso s such as Raspbe y Pi, came as, and o he inpu s ( oice, ex , and
isual) we e used o acili a e labo a o y communica ion [8]. While s udying he implemen a ion
CMC, 2021, ol.69, no.1 1075
o oice con ol o ope a ional and echnical unc ions in he i ualiza ion o a p oduc ion line,
Kennedy e al. [9] in es iga ed a passi e a ack called a inge p in ing a ack and ound ha i
may be possible o co ec ly de i e up o 33.8% o oice commands jus by ea esd opping on
enc yp ed a ic.
Objec i es o his wo k a e as ollows:
•Con ol assu ance o ope a ing and echnical unc ions in he p oduc ion line (p oduc ion
line on/o , a m ac i a ion, bel on, lase welde on/o , magne ic welde on/o , p ess
on/o );
•Pa icula command ecogni ion assu ance o he con ol o ope a ional and echnical
unc ions in he p oduc ion line;
•P o ision o a da a connec ion be ween speech ecogni ion echnologies;
•Addi i e noise supp ession in speech signals using he leas mean squa e (LMS) algo i hm
and he independen componen analysis (ICA).
•Ensu ing he highes possible e iciency in oice command ecogni ion in a eal en i onmen
wi h addi i e noise.
2 Rela ed Wo k
Speech signal p ocessing is a p omising esea ch a ea. Au oma ic speech ecogni ion, syn-
he ic speech, and na u al language p ocessing will ha e a signi ican impac in business and
indus y [10–12].
The mos impo an p oblems a e ela ed o au oma ed o semi-au oma ed equipmen con ol
(e.g., hea ing, cooling, ligh ing, en ila ion, and ai condi ioning). Am u ha (MATLAB imple-
men a ion) is one o he mos success ul ools used o spoken wo d iden i ica ion in indus ial
en i onmen s, wi h success a es o up o 90% [13], and Kamda and Kango a e o en used o
sma home appliances [14–17].
Voice in e ac ion is he mos na u al o m o human in e pe sonal communica ion. Di ec
oice commands make i easy o con ol sma de ices wi hou ime-consuming aining. Howe e ,
such a sys em con ollable om mul iple loca ions, equi es a sma a ay o mic ophones and
speake s connec ed o a cen alized p ocesso uni [18].
Au oma ic Speech Recogni ion (ASR) can be di ided in o h ee basic g oups [19,20]:
•Isola ed wo d ecogni ion sys ems ( oice commands a e used sepa a ely, such as in banking
o ai po elephone se ices);
•Small sys ems o applica ion commands and con ols;
•La ge sys ems o con inuous speech applica ions.
As ega ds ASR, sys ems applied di ec ly in indus y a e a mix o he second and hi d
g oups, which employ g amma ically limi ed commands o adminis a ion and con ol pu -
poses [19]. ASR sys ems can also be classi ied by oice in e ac ion in o wo ca ego ies [18]:
•Speci ic con ol applica ions, which c ea e he essence o sma homes ( oice con ol o
ope a ional and echnical unc ions and de ices);
•Gene al oice applica ions, which can be used in all ASR sys ems.
Obaid e al. [21] showed he sys em’s b oad applicabili y no only in indus y bu o pe sonal
use. Ou p oposed sys em consis s o oice ecogni ion and wi eless sys ems, implemen ed wi h
LabVIEW so wa e and ZigBee modules, espec i ely. The sys em’s g ea es ad an age is ha i
1076 CMC, 2021, ol.69, no.1
mus be ained only once. The equi ed ope a ions a e pe o med based on he da a ecei ed and
s o ed in he wi eless ecei e , which is connec ed di ec ly o he de ice.
A simila sys em, designed by Thaku e al. [22], can be used as a s and-alone po able uni
o wi elessly con ol ligh s, ans, ai condi ione s, ele isions, secu i y came as, elec onic doo s,
compu e sys ems, and audio isual equipmen [18].
Boeing is inco po a ing ASR in he new X-32 S ike Figh e ai c a , making i easie o he
pilo o con ol he ai c a and ocus on highe -p io i y aspec s o a mission [10,23].
I is possible o egula e mul iple ac o s o ASR sys ems, mainly speech a iabili y, which is
gene ally o limi ed use. The lexibili y o he language can be limi ed by a sui able g amma ical
design. The abili y o accu a ely ecognize cap u ed speech depends p ima ily on he size o he
dic iona y and he signal- o-noise a io (SNR). Thus, ecogni ion can be imp o ed by educing
he ocabula y and by imp o ing he SNR. Vocabula y es ic ions in Voice In elligence sys ems
a e based on he speci ic g amma . Reducing ocabula y, such as by sho ening indi idual com-
mands, can signi ican ly imp o e ecogni ion [24,25]. The quali y o cap u ed speech also a ec s
ecogni ion accu acy [26].
Real- ime esponse is ano he equi emen . Th ee aspec s a ec sys em pe o mance [27]:
•Recogni ion speed;
•Memo y equi emen s;
•Recogni ion accu acy.
I is challenging o combine all h ee aspec s, as hey end o con lic wi h each o he ; e.g., i
is ela i ely easy o imp o e ecogni ion speed while educing memo y a he expense o accu acy
o ecogni ion.
ASR sys ems can also be ound in indus ial applica ions such as obo ics, whe e oday’s
powe ul, inexpensi e mic op ocesso s and ad anced algo i hms con ol comme cial applica ions
in he a eas o compu e in e ac ion, da a en y, speech- o- ex con e sion, elephony, and oice
au hen ica ion. Robus ecogni ion sys ems o con ol and na iga ion a e cu en ly a ailable in
pe sonal compu e s [22].
ASR sys ems a e widely used in o he ields, such as wheelchai managemen [28], de ense
and a ia ion [29], and elecommunica ions.
The IoT pla o m [30] wi hin a cybe -physical sys em [31] can be unde s ood as a combi-
na ion o physical [32], ne wo k [33], and compu a ional p ocesses [34,35], and is impo an in
simul aneous oice ecogni ion.
Speech con ains in o ma ion usually ob ained by p ocessing a speech signal cap u ed by a
mic ophone h ough sampling, quan iza ion, coding [36], pa ame e iza ion, p ep ocessing, segmen-
a ion, cen e ing, p e-emphasis, and window weigh ing [37,38].
Fu he s eps include speech ecogni ion, whose implemen a ion may include he ollowing:
•A s a is ical app oach o con inuous speech ecogni ion using pe cep ual linea p edic ion
(PLP) o speech [39–42], such as:
◦Audio- o- isual con e sion in MPEG-4 [43];
◦Acous ic elemen modeling and ex ac ion [44];
◦Speech de ec o s [45] o aining o hyb id neu al ne wo ks o acous ic modeling in
au oma ic speech ecogni ion [46].
CMC, 2021, ol.69, no.1 1077
•RASTA (RelA i e SpecT Al) me hod [36]
•Mel- equency ceps al analysis (MFCC), such as:
◦Reduc ion o he pa hological sys em o oice quali y e alua ion dimensions [47];
◦De ec ion o clinical dep ession in adolescen s [48];
◦Sma wheelchai speech ecogni ion [49];
◦Speech ecogni ion using spoken wo d signals [50].
•Hidden Ma ko models (HMMs) [51]
•A i icial neu al ne wo ks (ANNs) [52], such as:
◦Au oma ic speech ecogni ion (ASR) o speech he apy and phased pa ien s [53];
◦Rapid adap a ion o neu al ne wo ks based on speech ecogni ion codes [54];
◦A combina ion o he unc ions o he HMM/MLP hyb id sys em and HMM/GMM
speech ecogni ion sys em [55];
◦Hyb id sys ems o con inuous speech ecogni ion HMM, MLP, and SVM [56].
•Supp ession o addi i e noise using single-o mul i-channel me hods [57], such as:
◦Speech enhancemen using spec al sub ac ion algo i hms [58];
◦Complex adap i e signal p ocessing me hods [59,60];
◦Model speech enhancemen [61,62];
◦Imp o ed emo al o addi i e noise by spec al sub ac ion [63];
◦Reduc ion o speech signal noise by wa ele ans o m wi h adjus ed uni e sal
h eshold alue [64] o supp ession o speech signals by wa ele ans o m [65].
•Mul i-channel me hods, including:
◦Leas medium quad a u e (LMS) algo i hm [66,67];
◦Recu si e leas squa es (RLS) [68,69];
◦Independen componen analysis (ICA) [70,71];
◦P incipal componen analysis (PCA) [72,73] o beam o ming (BF) o speech acqui-
si ion in noisy en i onmen s [74] o linea ly cons ained adap i e beam o ming [75]
wi h a obus algo i hm [76].
2.1 Classi ica ion o Speech Signal P ocessing Me hods
Algo i hms a e applied o imp o e he quali y o speech signals be o e p ocessing hem in
speech ecogni ion applica ions. These algo i hms inc ease he in elligibili y o speech signals and
supp ess in e e ence while minimizing he loss o use ul in o ma ion. They can be ca ego ized as
adap i e o non-adap i e me hods.
2.1.1 Adap i e Me hods
Adap i e me hods use a lea ning sys em ha changes coe icien s based on he wo king en i-
onmen . They ely on con inuous adjus men s o con ol pa ame e s in luenced by luc ua ions
o en i onmen o inpu and auxilia y signals. The basic elemen is eedback, which is used o
adjus he pa ame e s o he il e . These me hods use a speech- o-noise signal as an inpu . The
noise signal is used as a e e ence, which is sub ac ed o il e he speech signal. The e a e wo
ca ego ies o adap i e me hods [66,77]:
•Linea il e s a e de i ed om a linea ime-in a ian sys em o one o which he p inciple
o supe posi ion applies. These include he Kalman il e , LMS, RLS, and he adap i e
linea neu on (ADALINE).

1078 CMC, 2021, ol.69, no.1
•Nonlinea il e s a e no subjec o he p inciple o supe posi ion. They include he adap-
i e neu o- uzzy in e ence sys em (ANFIS), mul i-laye neu al ne wo ks, and e olu iona y
algo i hms.
2.1.2 Non-Adap i e Me hods
Non-adap i e me hods do no apply a lea ning sys em, and hence equi e no e e ence signal
con aining only noise. A speech signal wi h noise is su icien . These me hods can be ca ego ized
as ollows:
•Mul i-channel me hods pe o m sensing using mul iple mic ophones, whe e he p ima y one
acqui es he noisy speech signal and he o he s pick up only in e e ence. Two o mo e
channels may sense noisy speech signals in di e en places. Me hods include ICA, PCA,
singula alue decomposi ion (SVD), and pe iodic componen analysis (µCA) [58,77].
•Single-channel me hods equi e only one channel, wi h inpu consis ing o a speech signal
con amina ed by in e e ence. In e e ence supp ession is based on he cha ac e is ics o he
use ul signal and he in e e ence. These sys ems a e simple and less cos ly han mul i-
channel me hods. They assume ha he use ul signal (speech) and backg ound in e e ence
ha e di e en cha ac e is ics. They use he calcula ion o he equency spec um om
sub-segmen s o he signal. Thei e ec i eness is usually limi ed because o non-s a iona y
in e e ence. Me hods include equency selec i e il e s o he ini e impulse esponse (FIR)
and in ini e impulse esponse (IIR) ype, me hods based on Wiene il e ing heo y, spec al
sub ac ion using he as Fou ie ans o m (FFT), wa ele ans o m (WT), and empi ical
modal decomposi ion (EMD) [77].
2.2 Compa ison o Speech Signal P ocessing Me hods
Tab. 1 shows he ad an ages and disad an ages o basic speech signal p ocessing me hods.
In 2010, Bo isaga e al. [78] es ed he adap i e LMS and RLS algo i hms o eal- ime
speech signal p ocessing. Bo h achie ed much highe accu acy in MATLAB simula ions han a
ixed il e designed by con en ional me hods. In addi ion, LMS has a simple s uc u e and is easy
o implemen . I s main disad an age is slowe con e gence, bu i equi es much less memo y han
RLS.
Wang e al. [79] in oduced a me hod in 2011 based on spec al eading using a mul i-
channel LMS algo i hm. They pe o med ecogni ion expe imen s on a dis o ed speech signal
simula ed by con olu ion o mul i-channel impulse esponses wi h pu e speech. The me hod’s e o
was 22.4% less han ha o con en ional ceps al mean no maliza ion. When imp o ed using
beam o ming, he e o was 24.5% less han ha o con en ional ceps al mean no maliza ion
wi h beam o ming. The es was ocused on analysis o indi idual wo ds wi h a du a ion o abou
0.6 s.
In 2008, Cole e al. [63] applied di e en wid hs o he Hanning window and FFT signal
con e sion o he equency domain o pe o m spec al eadings using spec al sub ac ion, i.e.,
sub ac ing he noise spec um om he spec um o a speech signal con amina ed wi h addi i e
noise, assuming no co ela ion be ween he signals. The signal was u he modi ied o o m blocks
called mic o-segmen s. A e con e sion o he equency domain, he in e e ence componen was
emo ed by spec al sub ac ion and he signal was con e ed using an in e se FFT in he ime
domain. Tes ing used a speech signal wi h digi ally added acuum cleane noise. Based on he
SNR calcula ion, he bes esul was ob ained using a Hanning window wi h a wid h o 256
poin s. Howe e , he me hod can be conside ed inapp op ia e, as i is necessa y o moni o he
CMC, 2021, ol.69, no.1 1079
amoun o inpu in o ma ion. The me hod’s e ec i eness depends on he de e mina ion o he
noise spec um, which is di icul in eal condi ions and unsui able in a e y noisy en i onmen .
Table 1: Ad an ages and disad an ages o abo e me hods o in e e ence emo al
Me hod Ad an ages Disad an ages
LMS Simplici y Lowe speed
RLS Op imiza ion speed Compu a ional complexi y
Spec al
sub ac ion (FFT)
Simplici y Necessi y o noise spec um
sepa a ion; unsui able o la ge
in e e ence
DWT In o ma ion in ime and equency
domain due o he di ision o
componen s; easie o il e ing
non-s a iona y signals
Challenging o calcula e s a iona y
signals in he equency domain
ICA Noise de ec ion accu acy Compu a ional complexi y;
Gaussian- ype noises canno be
in oduced
PCA So ing componen s acco ding o
hei ene gy
Compu a ional complexi y;
Gaussian- ype noises canno be
in oduced
µCA Robus o noise and il e ing Need o signal pe iodici y;
Gaussian- ype noises canno be
in oduced
In 2009, Miho e al. [65] employed he WT o educe speech signals con amina ed by
in e e ence. Files om a es da abase con aining 720 male oice eco dings we e sampled a
25 kHz. Noise was added o he speech signal wi h SNRs o 0, 5, 10, and 15 dB. Due o
i s compu a ional complexi y, sym3 and highe (Symle wa ele ) we e unusable o in e e ence
educ ion in eal- ime. The bes p ope ies we e shown by db3 and db5 (Daubechies wa ele s),
wi h a maximum SNR imp o emen o 14 dB.
Agga wal e al. [64] used he DWT algo i hm in 2011 o educe in e e ence, applying bo h
so and ha d h esholding. Analysis was ca ied ou on a speech signal con amina ed wi h noise
a SNR le els o 0, 5, 10, and 15 dB. The so h esholding me hod p o ided be e esul s a all
measu ed le els o SNR inpu , and he maximum pe o mance imp o emen was 35.16 dB. The
ha d h eshold eached a maximum imp o emen o 21.71 dB.
In 2003, Visse e al. [70] analyzed he e iciency o he ICA me hod in au omobiles. D i ing
a 40 km/h, he d i e spoke a sequence o numbe s while he passenge spoke on a mobile phone
and he adio and hea e we e u ned on. S e eo mic ophones on ei he side o he ea iew
mi o (15 cm apa ) we e used o eco ding, and eco ded da a we e sampled a 8 kHz. The
SNR o he mix u e eco ded by he mic ophone on he d i e ’s side anged om 2 o 5 dB.
The ecogni ion success a e was 46.9% be o e applying ICA sepa a ion. A e using ICA, he
success a e inc eased o 72.8%. The bes esul s we e achie ed by a combina ion o he ICA and
WT me hods, whe e he ecogni ion success a e was 79.6%. In he same yea , Visse e al. [71]
1080 CMC, 2021, ol.69, no.1
examined he e ec i eness o he ICA in a oom (3 ×4×6 m) wi h wo di ec ional mic ophones
placed 10 cm apa . Speake s placed in he ou co ne s o he oom gene a ed spa ially dis ibu ed
noise. Two o he speake s we e placed 30 cm om he mic ophones. The i s speake ansmi ed
a sequence o numbe s, and he second ansmi ed in e e ence consis ing o p e eco ded wo ds.
The SNRs o he mix u es eco ded by he mic ophone we e in he ange o 5, 0, 5, and 10 dB.
The ecogni ion success a e was up o 49.34%, and his was imp o ed o 84.89% h ough he
ICA me hod.
In 2010, Kandpal e al. [80] used he PCA algo i hm o bo h speech ecogni ion and speech
sepa a ion. A eco ding o se en oices, which was 2 s long, wi h a sampling equency o 8 kHz,
was used o analysis. Based on he co ela ion coe icien , hey e alua ed he ou pu o he PCA
me hod agains hese se en oices and concluded ha he p obabili y o a ma ch be ween he
PCA ou pu and he oices was a ound 0.8.
In 2001, Saul e al. [81] de eloped he µCA me hod o speech ecogni ion. The algo i hm
had ou phases. Fi s , hey used he eigen alue me hod o combine and ampli y weak pe iodic
signals. They used a Hilbe ans o m o adjus he phase changes ac oss he channels. They used
e ec i e sinus seizu es o measu e he pe iodici y. They pe o med a hie a chical analysis o he
in o ma ion h ough di e en equency bands. The expe imen was pe o med on syn he ic da a
a a sampling equency o 8 kHz. They showed ha he µCA me hod enabled ex ac ion o he
equi ed signal segmen om di e en pa s o he equency spec um, and ha he me hod is
also e ec i e on signals wi h an SNR inpu o 20 dB. They men ioned ha he µCA me hod is
qui e esis an o noise and il e ing.
3 Applied Ma hema ical Me hods
Based on he abo e s udies ha use ad anced signal p ocessing me hods o speech il e ing,
he ICA me hod combined wi h an adap i e LMS algo i hm was selec ed o in e e ence supp es-
sion. A ho ough s udy o he li e a u e indica es ha hese me hods p o ide p omising esul s in
a ious applica ions. We desc ibe he selec ed me hods below.
3.1 Independen Componen Analysis
The independen componen (ICA) me hod is a possible solu ion o he “cock ail-pa y
p oblem,” as i can de ec hidden ac o s ha a e he bases o g oups o andom a iables, mea-
su emen s, o signals. I is a mul i-channel me hod, whe e wo o mo e signals a e con e ed o i s
inpu . The ICA is o en used o analysis o a highly a iable da a om a la ge sample da abase.
The a iables a e conside ed as linea mix u es o some unknown hidden a iables, wi h no known
mixing sys em. Hidden a iables a e conside ed o be non-Gaussian and independen , so hey a e
called independen componen s o he obse ed da a. Also called sou ces o ac o s, hey can
be ound by ICA. Be o e applying his me hod, da a p ep ocessing is necessa y using cen e ing
(c ea ing a ec o wi h ze o mean alue) and whi ening (c ea ing unco ela ed da a wi h uni
a iance). Eq. (1) ep esen s he measu ed signals using mic ophones, whe e he ma ix ep esen s
he mixing ma ix Amix (e.g., en i onmen and dis ance o he mic ophone om he sou ce signals)
and 
s ep esen s he con ained sou ce signals. The ICA me hod enables implemen a ion o Eq. (2),
o which i needs o es ima e a ma ix W ha is he in e se o Amix [70–72].

x=Amix ·
s,(1)

s=W·
x.(2)
CMC, 2021, ol.69, no.1 1081
An algo i hm de i ed om ICA, called Fas ICA, is o en used o sol e such p oblems. I
has ou s eps. A andom ec o 
xis c ea ed, and ku osis is calcula ed using Eq. (3),whe e
w
is he ec o o weigh s gand is he de i a i e o he non-quad a ic unc ion G. The da a a e
s anda dized, and a scala p oduc is calcula ed be ween he new ec o 
wand i s coun e pa
om he p e ious i e a ion. These s eps a e epea ed un il he scala p oduc eaches a alue
smalle han he selec ed con e gence c i e ion, o he maximum numbe o i e a ions is eached.
When wo king wi h Fas ICA, i is necessa y o selec he con e gence c i e ion, maximum numbe
o i e a ions, and numbe o ou pu componen s, which is gi en by he numbe o sou ce signals
we a e ying o es ima e. In he case o speech p ocessing, a leas wo componen s a e used,
whe e one should con ain he speech i sel , and he o he only noise [70–72].

w+=E
xg 
wT
x−E{g(
wT
x)}
w.(3)
3.2 Leas Mean Squa es Fil e
The LMS algo i hm is cu en ly one o he mos widely used adap i e algo i hms. I s main
s eng h lies in i s ma hema ical simplici y. Adap i e algo i hms a e in gene al used in unknown
en i onmen s because hey can adjus hei coe icien s based on a ying ci cums ances. They a e
based on a g adien sea ch algo i hm, o maximum g adien me hod. The dependence o he
s anda d de ia ion o he ou pu e o signal o he adap i e FIR il e on he il e coe icien s is
a quad a ic cu e wi h one global minimum. The basis o he adap i e algo i hm is he calcula ion
o he e o unc ion 
e(n)using Eq. (4),whe e
d(n)is he equi ed ou pu and 
y(n)is he eal
ou pu . The ou pu o each i e a ion o he LMS algo i hm is de ined acco ding o Eq. (5) and i s
modi ica ion, Eq. (6). Fil e ecu sion (adjus men o il e weigh s) is gi en acco ding o Eq. (7),
whe e µis he s ep size o he adap i e il e (which g ea ly a ec s he con e gence a e), 
w(n)
is he ec o o il e coe icien s, and 
x(n)is he inpu ec o . These s eps a e epea ed in each
i e a ion un il con e gence is achie ed. Ano he impo an pa ame e o he LMS algo i hm is he
o de o he il e M, which has a signi ican e ec on he compu a ional complexi y [67–69,78,82].

e(n)=
d(n)−
y(n),(4)

y(n)=
w(n)
x(n),(5)

y(n)=
wT(n)
x(n),(6)

w(n+1)=
w(n)+2µ
e(n)
x(n).(7)
These algo i hms equi e ewe demanding ma hema ical ope a ions han RLS algo i hms.
Fu he mo e, hey a e one o de less in complexi y, and a e he e o e as e . The main disad-
an age o LMS is i s lowe pe o mance in ime- a ying en i onmen s and lowe con e gence
speed [67,69,70,78].
4 Expe imen s
Fi e expe imen s we e conduc ed in labo a o y o eal condi ions o e i y he abo e ech-
nologies. Fi e scena ios we e e alua ed by so wa e-based simula ions. The in e e ence models
we e combined wi h audio eco dings o indi idual commands o es speech p ocessing me hods
in di e en condi ions.
1088 CMC, 2021, ol.69, no.1
Figu e 9: P ess machine in e e ence spec og ams: “Homing” command
5 Discussion
The p oposed concep o in e e ence educ ion can be applied in o he indus ial a eas. The
so-called acous ic/mechanical analysis o p oduc ion, which is pa o he p edic i e main enance
concep , seems o be a p omising app oach. In coope a ion wi h ou indus ial pa ne , B ose
CZ spol. s. .o., some pilo expe imen s we e ca ied ou on lase and magne ic welde s, ocusing
on an acous ic analysis o welding quali y (Fig. 10). The expe imen s ocused on acous ic and
mechanical analysis o a speci ic ool, as seen in Figs. 11 and 12. To ca y ou a eliable acous ic
analysis equi es he comple e elimina ion o backg ound noise, since i signi ican ly in luences he
esul s. The p esen ed ea ly designs seem like an op imal adeo be ween cos s and esul s, and
will be he subjec o u he esea ch.
The p esen ed sys em can be used in o he en i onmen s. Wo ke s such as cons uc ion pe -
sonnel, designe s, a is s, police, and i e igh e s can le e age he powe o di ec oice commands
in en i onmen s obs uc ed by noise. The p esen ed algo i hms a e ully ans e able and can be
deployed o o he uses. Noise educ ion can be used in au omobiles ( o il e ehicula noise),
cons uc ion ( o il e backg ound noise), o e en in medicine ( o il e li e signs o a mo he
and/o e us). The sys em has he ad an age ha i does no obs uc he wo ke in any way. O he
sys ems equi e di ec con ac wi h he employee’s body. These solu ions a e o en ob usi e and
could po en ially in luence wo ke s’ capabili ies, o e en hei sa e y. A wi eless and unob usi e
app oach mi iga es hese p oblems and inc eases employees’ com o .
Howe e , some deploymen a eas migh equi e e ec i e ad anced signal p ocessing me hods.
Apa om he es ed LMS and ICA combina ion, adap i e me hods include no malized LMS
(NLMS), RLS, QR-decomposi ion-based RLS (QR-RLS), and as ans e sal il e ing (FTF).
The esul ing signal can also be enhanced by pos -p ocessing echniques such as he wa ele
ans o m (WT), empi ical mode decomposi ion (EMD), and ensembled EMD (EEMD). Ou

CMC, 2021, ol.69, no.1 1089
u u e esea ch will seek he combina ion o he mos sui able algo i hms, and will also ocus on
ad anced AI echniques.
Figu e 10: Acous ic analysis o welding p ocess (de ec ion o ejec ed p oduc s)
Figu e 11: Long- e m p oduc ion line moni o ing: cu ing machine wi hin p ess machine (use o
a i icial in elligence o p edic i e main enance)
1090 CMC, 2021, ol.69, no.1
Figu e 12: Collage om pilo measu emen s
Da a in sma ac o ies end o di e g ea ly om egula IoT a ic, since hey o en
con ain highe amoun s o da a ans e ed o e sho e pe iods o ime. Manu ac u ing lines and
machine y wo k a he bes quali y- ime a io and mus main ain he highes e ec i i y possible.
The machine y o en con ains a as a ay o di e en commands, which epo impo an da a o
in luence p ecise manu ac u ing p ocesses. I is he e o e impo an o ha e a solu ion ha o e s
highe ansmi speeds [88] and he lowes possible la ency [89], o which 5G is slowly su acing
as a candida e echnology. The ne wo k is obus , ope a es in a licensed spec um, has low la ency,
and is pa ially ailo ed o indus ial deploymen . Machine- o-machine (M2M) communica ion is
a c i ical pa o Indus y 4.0, and is necessa y o main ain coo dina ion be ween a ious de ices
and componen s. Coo dina ion can be main ained no only wi hin a single plan bu ac oss plan s.
Synch onized machines o e a signi ican ad an age in p ecise manu ac u ing p ocesses, such as
o au omo i e applica ions. Many eams a e ocused on equi emen s and challenges o wi eless
echnologies in Indus y 4.0. Mos machines a e cu en ly wi ed. Howe e , wi es can obs uc
machine mo emen s, which can lead o mal unc ion o ejec manu ac u ing. The indus y is
slowly shi ing owa d wi eless echnologies and he e a e ce ain necessa y equi emen s needed
o seamless ansi ion. Va ghese e al. [90] ocused on hese challenges, ocusing on design c i e ia
o la ency, longe i y, and eliabili y. The eam benchma ked bo h WiFi and 5G in e ms o la ency
and eliabili y pa ame e s. Based on hei in o ma ion, a single wi eless s anda d will no add ess
all o he s ic equi emen s o Indus y 4.0; howe e , i oo ea ly o ejec some echnologies,
since many a e s ill unde going e ision. O donez-Lucena e al. [91] analyzed he newes 3GPP
Release 16 speci ica ion o 5G and iden i ied a numbe o deploymen op ions ele an o non-
public ne wo ks. Thei wo k included a easibili y analysis co e ing echnical, egula o y, and
business aspec s. They also discussed business models and egula o y aspec s.
CMC, 2021, ol.69, no.1 1091
Based on his in o ma ion, he p esen ed speech ecogni ion sys em could be in eg a ed as a
pa o a 5G Indus y 4.0 au oma ed ac o y. The ea ly concep is shown in Fig. 13.Thesys em
could employ an a ay o wi eless mic ophones o assis in speech ecogni ion. Wo ke s would
ca y hei own e e ence mic ophones as a sou ce o oice commands, in luenced by backg ound
noise. A ays deployed on machine y could be used as a sou ce o noise. The ga he ed da a would
be au oma ically e alua ed on a local se e based on a ious quali a i e pa ame e s. Due o he
low la ency o 5G ne wo ks, e alua ion and in e ac ion could be seamless. The machine y could
he e o e be ope a ed by eliable and well- ecognized oice commands, inc easing wo k low and
sa e y. The local non-public ne wo k (NPN) can pa se ele an da a ga he ed om machines and
e alua e ad anced unc ions on auxilia y se e s. The esul s could be dis ibu ed h ough public
land mobile ne wo ks (PLMN) o o he manu ac u ing plan s.
Figu e 13: Concep o sma ac o y wi h 5G compa ible oice e alua ion pla o m
6 Conclusions
We p esen ed inno a i e me hods o speech signal p ocessing o oice con ol o a p oduc ion
line in Indus y 4.0. A comme cially a ailable Windows ecognize was used in o de o ecognize
speci ic commands. The sys em was based on a comme cially a ailable sound ca d and LabVIEW
p og amming en i onmen . The analyzed da a we e ga he ed di ec ly on he p oduc ion line,
making i possible o analyze a lase welde , magne ic welde , and p ess machine.
The linea adap i e il e LMS and he ICA me hod we e chosen o en i onmen al noise
il e ing. A da ase o 100 epe i ions o each command was used o e alua e he designed sys em.
A o al o eigh commands we e es ed in combina ion wi h h ee ypes o in e e ence. The
a e age ecogni ion success be o e and a e il e ing was up o 49% highe in case o he LMS
algo i hm, and up o 52.3% o he hyb id il e ing scena ios.
The o e all esul s showed ha he hyb id me hod had a 5% ad an age o e a con en ional
LMS algo i hm. Howe e , due o he compu a ional complexi y o he ICA me hod, i is sig-
ni ican ly be e o implemen he LMS algo i hm, which is much simple and o e s simila
esul s. As he pe o mance and p ice o a ailable echnology change apidly, many mo e powe ul
algo i hms migh su ace in he coming yea s.
1092 CMC, 2021, ol.69, no.1
Acknowledgemen : This wo k was suppo ed by he Eu opean Regional De elopmen Fund in
Resea ch Pla o m ocused on Indus y 4.0 and Robo ics in Os a a p ojec CZ.02.1.01/0.0/0.0/17_-
049/0008425 wi hin he Ope a ional P og amme Resea ch, De elopmen and Educa ion, and in
pa by he Minis y o Educa ion o he Czech Republic unde P ojec SP2021/32 and SP2021/45.
Funding S a emen : This wo k was suppo ed by he Eu opean Regional De elopmen Fund in
Resea ch Pla o m ocused on Indus y 4.0 and Robo ics in Os a a p ojec CZ.02.1.01/0.0/0.0/17_-
049/0008425 wi hin he Ope a ional P og amme Resea ch, De elopmen and Educa ion, P ojec
Nos. SP2021/32 and SP2021/45.
Con lic s o In e es : The au ho s decla e ha hey ha e no con lic s o in e es o epo ega ding
he p esen s udy.
Re e ences
[1] S. D. Goldinge , “Wo ds and oices: Episodic aces in spoken wo d iden i ica ion and ecogni ion
memo y,” Jou nal o Expe imen al Psychology: Lea ning, Memo y, and Cogni ion, ol. 22, no. 5, pp. 1166–
1183, 1996.
[2] K. Hend ickson, J. Spinelli and E. Walke , “Cogni i e p ocesses unde lying spoken wo d ecogni ion
du ing so speech,” Cogni ion, ol. 198, no. 4, pp. 1–15, 2020.
[3] S. Ela a asi and G. Suseend an, “Au oma ic obo p ocessing using speech ecogni ion sys em,” in Da a
Managemen , Analy ics and Inno a ion,1
s ed., ol. 2. Singapo e: Sp inge , pp. 185–195, 2020.
[4] T. Zoughi, M. M. Homayounpou and M. Deypi , “Adap i e windows mul iple deep esidual ne wo ks
o speech ecogni ion,” Expe Sys ems wi h Applica ions, ol. 139, no. 10, pp. 1–16, 2020.
[5] R. Do ihal, A. Sopo i, A. Muku, N. Deochake and D. Va pe, “Sma homes using alexa and powe
line communica ion in IoT,” in In . Con . on Compu e Ne wo ks and Communica ion Technologies,1
s
ed., ol. 15. Singapo e: Sp inge , pp. 241–248, 2019.
[6] A. P. Naik and P. Ab aham, “A duino based oice con olled obo ic a m,” Robo ics, ol. 1, pp. 1–8,
2020.
[7] S. He, A. Zhang and M. Yan, “Voice and mo ion-based con ol sys em: P oo -o -concep imple-
men a ion on obo ics ia in e ne -o - hings echnologies,” in P oc. o he 2019 ACM Sou heas Con .,
Kennesaw, GA, USA, pp. 102–108, 2019.
[8] N. J. Knigh , S. Kanza, D. C uickshank, W. S. B ocklesby and J. G. F ey, “Talk2Lab: The sma lab
o he u u e,” IEEE In e ne o Things Jou nal, ol. 7, no. 9, pp. 8631–8640, 2020.
[9] S.Kennedy,H.Li,C.Wang,H.Liu,B.Wange al., “I can hea you alexa: Voice command inge -
p in ing on sma home speake s,” in IEEE Con . on Communica ions and Ne wo k Secu i y, Washing on,
DC, USA, pp. 232–240, 2019.
[10] J. Vajpai and A. Bo a, “Indus ial applica ions o au oma ic speech ecogni ion sys ems,” In e na ional
Jou nal o Enginee ing Resea ch and Applica ions, ol. 6, no. 3, pp. 88–95, 2016.
[11] K. Koumpis and K. Pa i , “Co po a e ac i i ies in speech ecogni ion and na u al language: Ano he
“new science”-based echnology,” In e na ional Jou nal o Inno a ion Managemen , ol. 3, no. 3, pp. 335–
366, 1999.
[12] J. Mocnej, T. Lojka and I. Zolo o á, “Using in o ma ion en opy in sma senso s o decen alized
da a acquisi ion a chi ec u e,” in IEEE 14 h In . Symp. on Applied Machine In elligence and In o ma ics,
Dan e s, IEEE, pp. 47–50, 2016.
[13] S. Am u ha, S. A a ind, A. Ma hew, S. Suga han, R. Rajas ee e al., “Voice con olled sma home,”
In e na ional Jou nal o Eme ging Technology and Ad anced Enginee ing, ol. 5, no. 1, pp. 272–275, 2015.
[14] R. Ma inek, J. Vanus, J. Nedoma, M. F id ich, J. F nda e al., “Voice communica ion in noisy
en i onmen s in a sma house using hyb id LMS +ICA algo i hm,” Senso s, ol. 20, no. 21,
pp. 1–24, 2020.
CMC, 2021, ol.69, no.1 1093
[15] H. Kamda , R. Ka ke a, A. Khanna, P. Kulka ni and S. Ag awal, “A e iew on home au oma ion
using oice ecogni ion,” In e na ional Resea ch Jou nal o Enginee ing and Technology, ol. 4, no. 10,
pp. 1795–1799, 2017.
[16] R. Kango, P. Moo e and J. Pu, “Ne wo ked sma home appliances-enabling eal ubiqui ous cul u e,”
in P oc. 3 d IEEE In . Wo kshop on Sys em-on-Chip o Real-Time Applica ions, Li e pool, UK, pp. 76–
80, 2002.
[17] I. Zolo o á, P. Papcun, E. Kajá i, M. Mišku and J. Mocnej, “Sma and cogni i e solu ions o ope -
a o 4.0: Labo a o y H-CPPS case s udies,” Compu e s & Indus ial Enginee ing, ol. 139, no. 105471,
pp. 1–15, 2020.
[18] I. V. McLoughlin and H. R. Sha i zadeh, “Speech ecogni ion o sma homes, Speech Recogni ion,”
Technologies and Applica ions, ol. 2008, pp. 477–494, 2008.
[19] L. R. Rabine , “Applica ions o oice p ocessing o elecommunica ions,” P oceedings o he IEEE,
ol. 82, no. 2, pp. 199–228, 1994.
[20] D. Yu and L. Deng, Au oma ic Speech Recogni ion,1
s ed., London, UK: Sp inge , pp. 1–321, 2016.
[21] T.Obaid,H.Rashed,A.A.ElNou ,M.Rehan,M.M.Salehe al., “ZigBee based oice con olled
wi eless sma home sys em,” In e na ional Jou nal o Wi eless & Mobile Ne wo ks, ol. 6, no. 1, pp. 47–
59, 2014.
[22] D. S. Thaku and A. Sha ma, “Voice ecogni ion wi eless home au oma ion sys em based on Zigbee,”
IOSR Jou nal o Elec onics and Communica ion Enginee ing, ol. 6, no. 1, pp. 65–75, 2013.
[23] D. W. Beeks and R. Collins, Speech Recogni ion and Syn hesis,2
nd ed., London, UK: CRC P ess, pp. 1–
316, 2001.
[24] H. Che alie , C. Ingold, C. Kunz, C. Moo e, C. Ro en e al., “La ge- ocabula y speech ecogni ion
in specialized domains,” in In . Con . on Acous ics, Speech, and Signal P ocessing,De oi ,MI,USA,
pp. 217–220, 1995.
[25] C. A. Kamm, C. Shamieh and S. Singhal, “Speech ecogni ion issues o di ec o y assis ance applica-
ions,” Speech Communica ion, ol. 17, no. 3–4, pp. 303–311, 1995.
[26] H. Sun, L. Shue and J. Chen, “In es iga ions in o he ela ionship be ween measu able speech quali y
and speech ecogni ion a e o elephony speech,” in IEEE In . Con . on Acous ics, Speech, and Signal
P ocessing, Mon eal, QC, Canada, pp. 865–868, 2004.
[27] M. K. Ra ishanka , “E icien algo i hms o speech ecogni ion,” Ph.D. disse a ion, Ca negie Mellon
Uni e si y, Pi sbu gh, 1996.
[28] A. Rogowski, “Indus ially o ien ed oice con ol sys em,” Robo ics and Compu e -In eg a ed Manu ac-
u ing, ol. 28, no. 3, pp. 303–315, 2012.
[29] C. Spi ze , U. Fe ell and T. Fe ell, Digi al A ionics Handbook,3
d ed., London, UK: CRC p ess,
pp. 1–848, 2017.
[30] V. Ma ik, P ˚
umysl 4.0-Výz a P o ˇ
Ceskou Republiku. P ague, Czechia: Managemen P ess, 2016.
[31] Cybe -Physical Sys ems. The P olemy p ojec . 2021. [Online]. A ailable: h p://cybe physicalsys ems.o g/.
[32] B. Ma diana, H. Hazu a, S. Fauziyah, M. Zaha iah, A. Hanim e al., “Homes appliances con olled
using speech ecogni ion in wi eless ne wo k en i onmen ,” in In . Con . on Compu e Technology and
De elopmen , Ko a Kinabalu, Malaysia, pp. 285–288, 2009.
[33] Sma De ice. Techopedia. 2015. [Online]. A ailable: h ps://www. echopedia.com/de ini ion/31463/
sma -de ice.
[34] M. Schie e , “Sma home de ini ion and secu i y h ea s,” in Nin h In . Con . on IT Secu i y Inciden
Managemen & IT Fo ensics, Magdebu g, Ge many, pp. 114–118, 2015.
[35] O. Kyas, “How o sma home,” Tanggal Akses Te akhi , ol. 3, pp. 1–208, 2013.
[36] J. Psu ka, L. Mulle , J. Ma ousek and V. Rado a, Mlu íme s poˇ
cí aˇ
cem ˇ
Cesky,1
s ed., ol. 111. P ague,
Czechia: Academia, pp. 1–746, 2006.
[37] H. Sakoe and S. Chiba, “Dynamic p og amming algo i hm op imiza ion o spoken wo d ecogni ion,”
IEEE T ansac ions on Acous ics, Speech, and Signal P ocessing, ol. 26, no. 1, pp. 43–49, 1978.

1094 CMC, 2021, ol.69, no.1
[38] R. E. Bellman and S. E. D ey us, Applied Dynamic P og amming,1
s ed., P ince on, New Je sey, USA:
P ince on Uni e si y P ess, pp. 1–390, 2015.
[39] A. Kuma , M. Dua and T. Choudha y, “Con inuous hindi speech ecogni ion using monophone based
acous ic modeling,” In e na ional Jou nal o Compu e Applica ions, ol. 24, pp. 1–5, 2014.
[40] S. J. A o a and R. P. Singh, “Au oma ic speech ecogni ion: A e iew,” In e na ional Jou nal o Compu e
Applica ions, ol. 60, no. 9, pp. 34–44, 2012.
[41] S. K. Saksamud e, P. Sh ish imal and R. Deshmukh, “A e iew on di e en app oaches o speech
ecogni ion sys em,” In e na ional Jou nal o Compu e Applica ions, ol. 115, no. 22, pp. 23–28, 2015.
[42] H. He mansky, “Pe cep ual linea p edic i e (PLP) analysis o speech,” Jou nal o he Acous ical Socie y
o Ame ica, ol. 87, no. 4, pp. 1738–1752, 1990.
[43] L. Xie and Z.-Q. Liu, “A compa a i e s udy o audio ea u es o audio- o- isual con e sion in mpeg-4
complian acial anima ion,” in In . Con . on Machine Lea ning and Cybe ne ics, Dalian, China, pp. 4359–
4364, 2006.
[44] A. Ga g and P. Sha ma, “Su ey on acous ic modeling and ea u e ex ac ion o speech ecogni ion,”
in 3 d In . Con . on Compu ing o Sus ainable Global De elopmen (INDIACom), New Delhi, India,
pp. 2291–2295, 2016.
[45] J. Rajnoha and P. Pollak, “De ek o y ˇ
eˇ
co é ak i i y na bázi pe cep i ní keps ální analýzy,” in ˇ
Ceské
Vysoké uˇ
cení Technické P aze, Fakul a Elek o echnická P ague, Czechia, pp. 1–9, 2008.
[46] G. A. Saon and H. Sol au, “Me hod and sys em o join aining o hyb id neu al ne wo ks o
acous ic modeling in au oma ic speech ecogni ion,” U.S. Pa en No. 9,665,823, Washing on, DC: U.S.
Pa en and T adema k O ice, pp. 1–18, 2017.
[47] J. I. Godino-Llo en e, P. Gomez-Vilda and M. Blanco-Velasco, “Dimensionali y educ ion o a pa ho-
logical oice quali y assessmen sys em based on Gaussian mix u e models and sho - e m ceps al
pa ame e s,” IEEE T ansac ions on Biomedical Enginee ing, ol. 53, no. 10, pp. 1943–1953, 2006.
[48] L.-S. A. Low, N. C. Maddage, M. Lech, L. Sheebe and N. Allen, “Con en based clinical dep ession
de ec ion in adolescen s,” in 17 h Eu opean Signal P ocessing Con ., Glasgow, UK, pp. 2362–2366, 2009.
[49] N. T. Hai, N. Van Thuyen, T. T. Mai and V. Van Toi, “MFCC-DTW algo i hm o speech ecogni ion
in an in elligen wheelchai ,” in 5 h In . Con . on Biomedical Enginee ing in Vie nam, Ho Chi Minh Ci y,
Vie nam, pp. 417–421, 2015.
[50] C. I ichaicha eon, S. Suks i and T. Ying hawo nsuk, “Speech ecogni ion using MFCC,” in In . Con .
on Compu e G aphics, Simula ion and Modeling, Pa aya, Thailand, pp. 135–138, 2012.
[51] J. Va ak, “Možnos i hlaso ého o ládání bezpilo ních d on˘
u,” Bc. hesis, VSB-Technical Uni e si y o
Os a a, Czechia, 2017.
[52] M. Cu aja , E. Ga , I. G ech, O. Casha and J. Micalle , “Compa a i e s udy o au oma ic speech
ecogni ion echniques,” IET Signal P ocessing, ol. 7, no. 1, pp. 25–46, 2013.
[53] N. Jamal, S. Shan a, F. Mahmud and M. Sha’abani, “Au oma ic speech ecogni ion (ASR) based
app oach o speech he apy o aphasic pa ien s: A e iew,” in AIP Con . P oc., Bydgoszcz, Poland,
pp. 1–8, 2017.
[54] S. Xue, O. Abdel-Hamid, H. Jiang, L. Dai and Q. Liu, “Fas adap a ion o deep neu al ne wo k based
on disc iminan codes o speech ecogni ion,” IEEE/ACM T ansac ions on Audio, Speech, and Language
P ocessing, ol. 22, no. 12, pp. 1713–1725, 2014.
[55] P. Pujol, S. Pol, C. Nadeu, A. Hagen and H. Bou la d, “Compa ison and combina ion o ea u es in
a hyb id HMM/MLP and a HMM/GMM speech ecogni ion sys em,” IEEE T ansac ions on Speech and
Audio P ocessing, ol. 13, no. 1, pp. 14–22, 2004.
[56] E. Za ouk, Y. B. Ayed and F. Ga gou i, “Hyb id con inuous speech ecogni ion sys ems by HMM,
MLP and SVM: A compa a i e s udy,” In e na ional Jou nal o Speech Technology, ol. 17, no. 3,
pp. 223–233, 2014.
[57] A. Chaudha i and S. Dhonde, “A e iew on speech enhancemen echniques,” in In . Con . on Pe asi e
Compu ing, Pune, India, pp. 1–3, 2015.
CMC, 2021, ol.69, no.1 1095
[58] N. Upadhyay and A. Ka maka , “Speech enhancemen using spec al sub ac ion- ype algo i hms:
A compa ison and simula ion s udy,” P ocedia Compu e Science, ol. 54, no. 2, pp. 574–584, 2015.
[59] R. Ma inek, “The use o complex adap i e me hods o signal p ocessing o e ining he diagnos ic
quali y o he abdominal e al elec oca diog am,” Ph.D. disse a ion, VSB-Technical Uni e si y o
Os a a, Czechia, 2014.
[60] J. Jan, ˇ
Císlico á Fil ace, Analýza a Res au ace Signál˚
u,2
nd ed., B no, Czechia: Vu ium, pp. 1–427, 2002.
[61] P. Ha ding, “Model-based speech enhancemen ,” Ph.D. disse a ion, Uni e si y o Eas Anglia,
UK, 2013.
[62] P. C. Loizou, Speech Enhancemen : Theo y and P ac ice,2
nd ed., London, UK: CRC P ess, pp. 1–
711, 2013.
[63] C. Cole, M. Ka am and H. Aglan, “Inc easing addi i e noise emo al in speech p ocessing using
spec al sub ac ion,” in Fi h In . Con . on In o ma ion Technology: New Gene a ions, Las Vegas, NV,
USA, pp. 1146–1147, 2008.
[64] R. Agga wal, J. K. Singh, V. K. Gup a, S. Ra ho e, M. Tiwa i e al., “Noise educ ion o speech
signal using wa ele ans o m wi h modi ied uni e sal h eshold,” In e na ional Jou nal o Compu e
Applica ions, ol. 20, no. 5, pp. 14–19, 2011.
[65] S. G. Miho , R. M. I ano and A. N. Popo , “Denoising speech signals by wa ele ans o m,” Annual
Jou nal o Elec onics, ol. 1, no. 6, pp. 2–5, 2009.
[66] R. Ma inek, J. Zidek, P. Bilik, J. Manas, J. Kozio ek e al., “The Use o LMS and RLS adap i e
algo i hms o an adap i e con ol me hod o ac i e powe il e ,” Ene gy and Powe Enginee ing, ol.5,
no. 4, pp. 1126–1133, 2013.
[67] B. Fa hang-Bo oujeny, Adap i e Fil e s: Theo y and Applica ions,2
nd ed., Hoboken, New Je sey, USA:
John Wiley & Sons, pp. 1–800, 2013.
[68] S. V. Vaseghi, Ad anced Digi al Signal P ocessing and Noise Reduc ion,4
h ed., Hoboken, New Je sey,
USA: John Wiley & Sons, pp. 1–544, 2009.
[69] R. Ma inek and J. Zidek, “The eal implemen a ion o NLMS channel equalize in o he sys-
em o so wa e de ined adio,” Ad ances in Elec ical and Elec onic Enginee ing, ol. 10, no. 5,
pp. 330–336, 2012.
[70] E. Visse , M. O suka and T.-W. Lee, “A spa io- empo al speech enhancemen scheme o obus speech
ecogni ion in noisy en i onmen s,” Speech Communica ion, ol. 41, no. 2, pp. 393–407, 2003.
[71] E. Visse and T.-W. Lee, “Speech enhancemen using blind sou ce sepa a ion and wo-channel
ene gy based speake de ec ion,” in IEEE In . Con . on Acous ics, Speech, and Signal P ocessing, P oc.,
Hong Kong, China, pp. 884–887, 2003.
[72] A. Hy ä inen and E. Oja, “Independen componen analysis: Algo i hms and applica ions,” Neu al
Ne wo ks, ol. 13, no. 4, pp. 411–430, 2000.
[73] A. Cichocki and S. Ama i, Adap i e Blind Signal and Image P ocessing: Lea ning Algo i hms and
Applica ions,1
s ed., Hoboken, New Je sey, USA: John Wiley & Sons, pp. 1–586, 2002.
[74] S. Fische and K. U. Simme , “Beam o ming mic ophone a ays o speech acquisi ion in noisy
en i onmen s,” Speech Communica ion, ol. 20, no. 3, pp. 215–227, 1996.
[75] L. G i i hs and C. Jim, “An al e na i e app oach o linea ly cons ained adap i e beam o ming,” IEEE
T ansac ions on An ennas and P opaga ion, ol. 30, no. 1, pp. 27–34, 1982.
[76] Q. Zou, Z. L. Yu and Z. Lin, “A obus algo i hm o linea ly cons ained adap i e beam o ming,”
IEEE Signal P ocessing Le e s, ol. 11, no. 1, pp. 26–29, 2004.
[77] A. Rajani and S. Sounda ya, “A e iew on a ious speech enhancemen echniques,” In e na ional
Jou nal o Ad anced Resea ch in Compu e and Communica ion Enginee ing, ol. 5, no. 8, pp. 296–
301, 2016.
[78] K. R. Bo isaga and G. Kulka ni, “Simula ion and compa a i e analysis o LMS and RLS algo i hms
using eal ime speech inpu signal,” Global Jou nal o Resea ch in Enginee ing, ol. 10, no. 5, pp. 44–
47, 2010.
1096 CMC, 2021, ol.69, no.1
[79] L. Wang, N. Ki aoka and S. Nakagawa, “Dis an - alking speech ecogni ion based on spec al sub ac-
ion by mul i-channel LMS algo i hm,” IEICE T ansac ions on In o ma ion and Sys ems, ol. 94, no. 3,
pp. 659–667, 2011.
[80] N. Kandpal and B. M. Rao, “Implemen a ion o PCA & ICA o oice ecogni ion and sepa a ion o
speech,” in IEEE In . Con . on Ad anced Managemen Science, Chengdu, China, pp. 536–538, 2010.
[81] L. K. Saul and J. B. Allen, “Pe iodic componen analysis: An eigen alue me hod o ep esen ing
pe iodic s uc u e in speech,” in Ad ances in Neu al In o ma ion P ocessing Sys ems. Vancou e , B i ish
Columbia, Canada, pp. 807–813, 2001.
[82] K. Cengiz, “Comp ehensi e analysis on leas squa es la e a ion o indoo posi ioning sys ems,” IEEE
In e ne o Things Jou nal, ol. 8, no. 4, pp. 2842–2856, 2013.
[83] NT5 Compac 1/2 Ca dioid Condense Mic ophone. RODE Mic ophones. [Online]. A ailable:
h p://www. ode.com/mic ophones/n 5.
[84] M. A. Jamshed, M. U -Rehman, J. F nda, A. A. Al huwayb, A. Nauman e al., “Dual band and dual
di e si y ou -elemen MIMO dipole o 5G handse s,” Senso s, ol. 21, no. 3, pp. 1–13, 2021.
[85] C. Zei ni z, “Wa eIO: A soundca d in e ace o lab iew,” Ch is ian Zei ni z, 2017. [Online]. A ailable:
h ps://www.zei ni z.eu/scms/wa eio.
[86] J. Ba os, R. Ma inek, R. Ja os, L. Danys and L. Sous ek, “De elopmen o applica ion o con ol o
SMART pa king lo ,” IFAC-Pape sOnLine, ol. 52, no. 27, pp. 19–26, 2019.
[87] VB-CABLE Vi ual Audio De ice. VB-Audio So wa e, 2020. [Online]. A ailable: h ps:// b-audio.com/
Cable/.
[88] W. Ta nebe g, O. Hamsis, J. Hedlund, K. B unns öm, E. Fi zge ald e al., “Towa ds in elligen indus y
4.0 5G ne wo ks: A i s h oughpu and QoE measu emen campaign,” in In . Con . on So wa e,
Telecommunica ions and Compu e Ne wo ks, Spli , H a , C oa ia, pp. 1–6, 2020.
[89] J. Ga cia-Mo ales, M. C. Lucas-Es an and J. Gozal ez, “La ency-sensi i e 5G RAN slicing o
indus y 4.0,” IEEE Access, ol. 7, pp. 143139–143159, 2019.
[90] A. Va ghese and D. Tandu , “Wi eless equi emen s and challenges in indus y 4.0,” in In .Con . on
Con empo a y Compu ing and In o ma ics, Myso e, India, pp. 634–638, 2014.
[91] J. O donez-Lucena, J. Folguei a Cha a ia, L. M. Con e as and A. Pas o , “The use o 5G non-
public ne wo ks o suppo indus y 4.0 scena ios,” in IEEE Con . on S anda ds o Communica ions and
Ne wo king, G anada, Spain, pp. 1–7, 2019.