application of variations of the lms adaptive filter …

1Introduction

This paper describes a proposed method for optimaladjustment of parameters of variations of the LMS adaptivefilter in the application of suppression of the additive noisein the speech signal. The proposed method is applied foroptimal settings of the parameters of the adaptive noisecanceller with the Conventional LMS algorithm, with theSigned – Regressor LMS algorithm, with the Sign LMSalgorithm and with the Sign-Sign LMS algorithm. Selectedvariations of the LMS adaptive noise canceller is used invoice communication with the control system.

The Least Mean – Square (LMS) algorithm wasdeveloped by Widrow and Hoff in 1960. This algorithm is amember of the stochastic gradient algorithms [2].

The LMS algorithm is a linear adaptive filteringalgorithm, which consists of two basic processes:a) a filtering process, which involves

computing the output ( ) of linear filter in response toan input signal ( ) (1),

� y nx n

2Description of the adaptive filter with the LMS algorithmand the other variations of the LMS algorithms

2.1The conventional LMS algorithm

553

J Vaňuš, V Stýskala. .

ISSN 1330-3651

UDC/UDK 621.391.8:004.421

APPLICATION OF VARIATIONS OF THE LMS ADAPTIVE FILTERFOR VOICE COMMUNICATIONS WITH CONTROL SYSTEM

Jan Va uš, Vít zslav Stýskalaň ě

This paper describes a proposed method for optimal adjustment parameters of variations of the LMS adaptive filter in the application of suppression of theadditive noise from the speech signal. Selected variation of the LMS adaptive noise canceller is implemented on the TMS320C6713 DSK. This practicalrealization was used in a voice communication with a control system for controlling of operating – technical functions in buildings.

Keywords: variations of the LMS algorithm, adaptive noise canceller, speech signal processing, voice communication, adaptive filter

Original scientific paper

U ovom se radu opisuje predložena metoda za optimalno podešavanje parametara varijacija LMS adaptivnog filtera kod prigušenja aditivne buke iz govornogsignala. Izabrana varijacija LMS adaptivnog poništa a buke je implementirana na TMS320C6713 DSK. U praksi je ovo primijenjeno na govornu komunikacijus upravlja kim sustavom za pra enje operativno-tehni kih funkcija u zgradama.

čč ć č

Klju ne rije i:č ć varijacije LMS algoritma, adaptivni poništa buke, obrada govornog signala, govorna komunikacija, adaptivni filtarč

Izvorni znanstveni članak

Primjena varijacija LMS adaptivnog filtera na govornu komunikaciju sa sustavom upravljanja

Tehni ki vjesnikč 18, 4(2011) 553-560,

b) an adaptive process (3), which involves the automaticadjustment of the parameters ( +1) of the filter inaccordance with the estimation error ( ).

w ne n


,)1()()(1

0�

�

�

��M

ii nxnn wy (1)

� generating an estimation error ( ) by comparing thisoutput ( ) with the desired response ( ) (2) (Fig.1),

e ny n d n

),()()( nye �= ndn ( )2

),()(2)()1( nnμenn xww �� (3)

w

w

( ) is tap – weight vector,( +1) is tap – weight vector update [1, 2].

This algorithm is obtained from the conventional LMSrecursion (3) by replacing ( ) with its sign. This leads to thefollowing recursion

n M

n M

e n

2.2The Sign LMS algorithm

).())((sign2)(1)( nxnenn �� ww ( )4

Because of the replacement of ( ) by its sign,implementation of this recursion may be cheaper thanconventional LMS recursion [1, 13].

The Signed-Regressor algorithm is obtained from theconventional recursion (3) by replacing the tap-input vector( ) with the vector sign( ( )), where the sign function is

applied to the vector ( ) on an element-by-element basis.The Signed-Regressor recursion is then [1, 13]

e n

n nn

x xx

2.3The Signed - Regressor LMS algorithm

)).()sign((2)()1( nnenn xww �� (5)

2.4The Sign - Sign LMS Algorithm

The Sign-Sign algorithm combines the Sign andSigned-Regressor recursions, resulting in the followingrecursion [1, 13]

554

Application of variations of the LMS adaptive filter for voice communications with control system

Technical Gazette 18, 4(2011) 553-560,

J Va uš, V Stýskala. ň .

3Settings of a step size factor μ

The step size factor must be selected properly tocontrol the stability and the convergence speed of theadaptive filter with the LMS algorithms [2] and the LMSalgorithms variations. Determination of settings of a stepsize factor is described in the next text.

For determination, when the LMS algorithm remainsstable it is necessary to find the upper bound of , thatguarantees the stability of the LMS algorithm (7) [1]

μ

μ

μmax

tr[ ] – trace of , which means the sum of the diagonalelements of ,

– Toeplitz autocorrelation matrix calculated from thevector of the input signal ( ), size is × .

The Toeplitz autocorrelation matrix is calculated byequation (8)[2]

R RR

Rx R

R

n M M

" " indicates the optimum Wiener solution for the Wienerfilter [2].

Determining of the optimal value of the step size factoris important in conducting an algorithm LMS. When

selecting factor it becomes a compromise between thetwo aspects. On the one hand, large values of can leadquickly to the optimal settings of the LMS algorithm forspeech signal processing. On the other hand, high valuemay increase an estimate error of the speech signalprocessing in further steps. A small value , on the contrary,ensures the stability and the convergence of the LMSalgorithm [1]. As a result a small value slows down theconvergence of the LMS algorithm and, consequently,increases the inaccuracies in the filtration of non-stationarysignals [3]. The following equation is used for thecalculation of the optimal factor (10), (Tab ) [1]

0

opt

opt

opt . 1

μ

μ

μ

μ

μ

μ

μ

Convergence behavior of the LMS algorithm is directlylinked to the eigenvalue spread of the autocorrelation matrix

and the power spectrum of ( ). The convergence of theLMS algorithm is directly related to the flatness in thespectral content of the underlying input process. E[ ( )]converges to zero when remains within the range offormula (9). E[ ( )] is expectation of the weight – errorvector ( ) = ( ) – .

R x

v

vv w w

n

n

nn n

μ

0

w(n+1)=w(n) + 2� sign(e(n))sign(x(n)). (6)

Figure 1 Block diagram of the LMS adaptive filter [3]

,][3

1max

Rtr�� (7)

.)()(E nn TxxR �� (8)

,1

max� conv (9)

λmax

0

0

– maximum eigenvalue of the autocorrelation matrixof the input vector ( ).

The above range does not necessarily guarantee thestability of the LMS algorithm. The convergence of theLMS algorithm requires the convergence of the mean of

( ) towards and also the convergence of the variance ofthe elements of ( ) to some limited values [1]. Vector iscalculated by the Wiener Hopf equation and the superscript

R

x

w w

w w

n

n

n

-

,][)1( Rtropt

�+�

M

M� (10)

tr[ ] – trace of , which is the mean sum of the diagonalelements of ,

– the misadjustment parameter.

R RR

M

Table 1 Calculation of the step size factors , , of the LMSadaptive filter for input signal with different SSNR values

μ μopt max convμx( )n

The misadjustment parameter is defined as the ratioof the steady – state value of the excess mean-square error(MSE) to the minimum mean square (MSE) error .ξ ξexcess min

M

].[min

Rtr��

�� excess

M (11)

The misadjustment is a dimensionless parameter,that provides a measure of how close the LMS algorithm isto optimality in the mean - square – sense. The closer is to0, the more accurate is the adaptive filtering action, which isbeing performed by the LMS algorithm. The values of themisadjustment parameter are usually the 10 %, 20 % and30 % (Tab. 4), (Tab. 5), (Tab. 6), (Tab. 7), (Tab. 8). The valueof = 10 % means that the adaptive system has an MSEonly 10 percent greater than [11].

The dynamic time warping (DTW) criterion is analgorithm for measuring similarity between two sequenceswhich may vary in time or speed. A well known application

M

M

M

M

ξmin

4Using of the DTW criterion for determination of the adaptivefilter length M

555Tehni ki vjesnikč 18, 4(2011) 553-560,


has been automatic speech recognition, to cope withdifferent speaking speeds [8].

The correct determination of the adaptive filter lengthis very important. When the length of the adaptive filter

is low, the speech signal processing is inaccurate as a resultof the adaptive filter's small number of parameters. A highvalue of the adaptive filter length leads to inaccuratespeech signal processing by the influence of the estimatorvariance increase. The proposed method in this work usedthe DTW criterion for determining the value of the filterlength of the LMS adaptive filter.

The DTW criterion is used to compare the twosequences of vectors: reference vector = [ (1), . . . ( )] oflength and test vector = [ (1), . . . ( )] of length [5].The value of the LMS adaptive filter length M is determinedby setting values of the length in intervals {0 to 150} andcalculating of the minimum distance (similarity) betweenthe reference vector = [ (1), . . . , ( )] of the length (thedesired signal ( ) (Fig. 2)) and the test sequence = [ (1), .. . , ( )] of the length (the error signal ( ) (Fig. 6), (Fig.10), (Fig. 11)). Words are almost never represented by thesequence of the same length [7]. The distance betweenthe sequences and is given as minimum distance overthe settings of all possible paths (all possible lengths) [5, 7,8]. When the value of distance d was , the isolatedword . This value was

M M

M

M

p p PP o o T T

M

r r R Rd n o

o T T e n

R T

PO

RO

O R

�

dd

< 0,2was recognised < 0,2


Figure 2 The desired speech signal ( ) of the Czech isolated word"jeden" (one) to the LMS adaptive filter variations. Waveform,

spectrogram (frequency time analysis) andperiodogram of the power spectral density estimate.

d n

Figure 3 The additive noise ( ) (in input signal ( )= ( )+ ( )) tothe LMS adaptive filter variations. Waveform, spectrogram (frequencytime analysis) and periodogram of the power spectral density estimate

n n n d n n na ax

.

determined from experiments with isolated wordrecognition (Tab. 3) [10] for comparison of the distances dbetween isolated words.

Minimum distance computation

� ),,(min),( RR OO c

CDD � (12)

is simple, when normalization factor is not a function ofthe path and it is possible to write = for

N

N Nc

c

� � � ).()((),((min

1),(

1

kWkrrktdN

D c

Kc

kcc

Cc �

�

�� oRO (1 )3

5The segmental signal to noise ratio calculation

In the speech signal processing in a real environmentmust be expected in the speech signal interference. Thestandard benchmark for measuring of the level of noise inthe signal is the signal to noise ratio criterion - (Signal-to-Noise Ratio). In analyzing of the speech signal it isexpected that the speech signal is disturbed by two basictypes of interference. These types of interference aredescribed as additive noise or convolution distortion. Theadditive noise (Tab. 2) is added to the speech signal either asbackground noise environment where the speech is sensedas the noise or the speech signal transmission path. Theadditive noise is added to speech signal processing and thespeech coding operations in the digital signal processorswith fixed-point.

SNR

Table 2 The Segmental Signal to Noise ratio values calculated forthe speech signal of Czech word "jeden" (Fig. 2) to additive noise ( )

(Fig. 3) and to additive white noise ( ).n n

n na

w

Type of noise Values of Segmental SNR

additive noise na(n) SSNRa=6,7 dB

additive white noise 1 nw1(n) SSNRw1=18,2 dB

additive white noise 2 nw2(n) SSNRw2=3 dB

additive white noise 3 nw3(n) SSNRw3=–1,8 dB

For implementation of experiments are used theadditive noises with a calculated egmental ( ignal to

oise atio) – (Tab. 2) for speech signal processing[9]

S SNR SN R SSNR

,1 1

0�

�

�

�L

iiiVAD

KSNRSSNR (14)

LKVAD SNR

d n 2)

– is the number of segments of speech signal,– the number of segments in speech activity,

–is information about speech activity .

The proposed method with the DTW criterion was usedin two channel structures of variations of the LMS adaptivenoise canceller (Fig. 4).

Aprimary input contains the desired signal ( ) (Fig.

i i

6Using the proposed method with the DTW criterion for theLMS adaptive noise cancelling from the speech signal6.1MATLAB simulation

556 Technical Gazette 18, 4(2011) 553-560,

adaptive noise canceller (the Signed LMS algorithm, theSigned – Regressor LMS algorithm and the Signed – SignedLMS algorithm) (Tab. 6, Tab. 7 Tab. 8).and

and the additive noise ( ). A noise reference input isassumed to be available containing ( ), which iscorrelated with the original corrupting noise ( ).As shownin 4, the LMS adaptive filter receives the referencenoise, filters it, and subtracts the result from the primaryinput.

n nn" n

n nFig.


From the point of view of the adaptive filter, theprimary input ( ( )+ ( )) acts as its desired response andthe system output acts as its error. The noise canceller output( ) (Fig. 6) is obtained by subtracting the filtered reference

noise ( ) from the primary input.Adaptive noise cancelinggenerally performs better than the classical approach sincethe noise is subtracted out rather than filtered out [11].

The proposed DTW method was used for optimalsettings values of the filter length and a step size factorof the adaptive filter with the LMS algorithm in theapplication of the suppression of additive noise from thespeech signal. Optimal values of the filter length of theLMS adaptive noise canceller and distance betweendesired speech signal ( ) (Fig. 2) and error signal ( ) (Fig.6) from the LMS adaptive noise canceller are calculated inTab. 5.

The proposed method for optimal adjustment of a step

d n n n

e nn n

M

Md

d n e n

μ

μ

μ

μ

size factor and the filter length of the LMS adaptivefilter in an application for the suppression of additive noisefrom the speech signal was applied in the next steps [10, 12]:

Calculation of a step size factor optimal value (10)from the input signal ( ) to variations of the LMS adaptivenoise canceller ( = 10 %, = 20 %, = 30 %), ( = 6,7dB, = 18,2 dB, = 3 dB, = –1,8 dB),(Tab. 1).

As the reference vector is used the desired signal ( )(Fig. 2).

As a test vector was chosen the error signal ( ).Further were calculated the distance (12), (13)

between the signals ( ) and ( ) for settings of the filterlengths of variations of the LMS adaptive noise cancellerin interval {1 to 150}(Fig. 5).

As the optimal value of the LMS adaptive noisecanceller length was chosen for example value = 17 forminimum distance (Fig. 5), (Tab. 5) betweentwo compared signals ( ) (Fig. 2) and ( ) (Fig. 6) (= 6,7 dB, = 5,9×10 ).

The same procedure was used for the calculation ofoptimal filter length of other variations of the LMS

opt

opt

a

w1 w2 w3

a

1

M

x nSSNR

SSNR SSNR SSNR

d n

e nd

d n e nM

M M

d n e n SSNR

M

1.

2.

3.4.

5.

6.

M M M

P

O

d = 9,9×10-2

-3


Table 3 Using the DTW criterion for recognition of the isolated Czech words (numbers one - ten, one - ten) from a single speaker

jeden–jeden(one-one)

jeden–dva(one-two)

jeden–tři(one-three)

jeden–čtyři(one-four)

jeden–pět(one-five)

jeden–šest(one-six)

jeden–sedm(one-seven)

jeden–osm(one-eight)

jeden–devět(one-nine)

jeden–deset(one-ten)

d=0 d=0,713 d=1,218 d=1,415 d=0,552 d=1,917 d=1,46 d=1,071 d=0,553 d=1,268

dva–jeden(two-one)

dva–dva(two- two)

dva–tři(two-three)

dva–čtyři(two-four)

dva–pět(two-five)

dva–šest(two-six)

dva–sedm(two-seven)

dva–osm(two-eight)

dva–devět(two-nine)

dva–deset(two-ten)

d=0,713 d = 0 d=0,406 d=0,568 d=0,373 d=1,165 d=0,791 d=0,39 d=0,37 d=0,592

Figure 4 Block diagram of the LMS adaptive noise canceller [6, 11]

Figure 5 Calculation of the filter length = 17 of the LMS adaptivenoise canceller, (first iteration cycle), ( =6,7 dB and =5,9 10 )

and distance between e(n) (Fig. 6) and ( ) (Fig. 2)( = 6,7 dB, =5,9 10 , (Tab. 5)).

M

SSNR

d n

SSNR

a

a

μ

μ

1

1

-3

-3

×

×d = 9,9×10

-2

The output signal ( ) from the(first iteration cycle) ( )

only for the next conditions: = 6,7 dB ( = 5,9 10 ,

= 17, = 10 %), ( = 1,08 10 , =10, = 20 %), ( =

14,996 10 , = 7,

e n

SSNR

M M

M

LMS adaptive noisecanceller was recognized d < 0,2

a 1

3

‒3

2

3

μ

SSNR

e n

SSNR

M M

SSNR

e n

SSNR

M M

SSNR M

M M

M

Me n

SSNR M

×

×

× = 30 %), (Tab. 5). When the additivewhite noise values in the speech signal were higher( ), the speech signal was not recognized.

The output signal ( ) from the(first iteration cycle) (

) only for the next conditions: = 6,7 dB ( =

5,9×10 , = 8, = 10 %), ( = 1,08×10 , = 10, = 30%), (Tab. 6). When the additive white noise values inthe speech signal were higher, the speech signal was notrecognized.

The output signal ( ) from the(first iteration cycle)

( ) for next conditions: = 6,7 dB (

= 5,9×10 , = 2, = 10 %), ( = 1,08×10 , = 2, = 20

%) and for = 18,2 dB ( = 6,4×10 , = 2, = 10

%), ( = 1,18×10 , = 2, = 20 %), ( = 1,63×10 , = 2,= 30 %) (Tab. 7), but the filter lengths of the LMS Sign-

Sign adaptive noise canceller was very low ( = 2).The output signal ( ) from the

(first iteration cycle)( ) only in one case for next conditions:

= 6,7 dB ( = 5,9×10 , = 12, = 10 %), (Tab. 8).

M M

M

M M

M M

M

M

M

M

μ μ

μ

μ

μ

μ

μ

μ μ

μ

‒

‒

‒3 ‒2

‒3

‒3

‒2 ‒2

‒3

2

w

a 1

2

w

a 1

w1 1

2 3

a 1

LMS Sign adaptivenoise canceller

LMS Sign-Signadaptive noise canceller

LMS Sign-Reggressoradaptive noise canceller

was recognized

wasrecognized

wasrecognized

d

d

d

<0,2

< 0,2

< 0,2

2

-2



6.2Implementation of the LMS adaptive noise canceller on theTMS320C6713 DSK

The proposed method with DTW criterion fordetermining the filter length of the adaptive filter with theLMS algorithm was used in an application to suppressadditive noise ( ) from the speech signal ( ). It wasimplemented with a two channel structure of the LMSadaptive noise canceller on the TMS320C6713 DSP(Digital Signal Processor), Starter Kit (DSK) (Fig. 7) [10]and it was programmed in the C programming languagethrough the CCS studio version 3.1 [13].

The input signal ( ) is composed of the desired signal( ) and the additive noise ( ). The segmental signal to

noise ratio of the input signal ( ) was = 6,7 dB.Application of the proposed method with the DTW

criterion with implementation of the LMS adaptive noise

M

n n x n

x nd n n n

x n SSNRa

canceller on the TMS320C6713 DSK was carried out inseveral steps:

1. step – calculation of optimal value of a step sizefactor from the input signal ( ) to the TMS320C6713DSK (calculated in MATLAB).

2. step – calculation of the filter length values of theLMS adaptive noise canceller (Tab. 4), (calculated inMATLAB).

3. step – empirically was found, that the factor for theLMS adaptive noise canceller, implemented on theTMS320C6713 DSK allows settings only in the range10 to 10 The filter length of the LMS adaptivenoise canceller can be set only in the range = 16 to = 52.This indicates the possibility of only partial use of theproposed method in practical implementation of the LMSadaptive noise canceller on the TMS320C6713 DSK.

4. step – settings of the filter length = 21, settingsvalues of the step size factor 10 10 10 (Tab.9) and implementation of the LMS adaptive noise cancelleron the TMS320C6713 DSK. The input signal is( )= ( )+ ( ), where ( ) (Fig. 2) is isolated Czech wordjeden (one) and ( ) is additive noise with = 6,7 dB.

5. step – calculation of distance between ( ) (forexample (Fig. 10 Fig. 11) and ( ) (Fig. 2) for settings ofthe filter length = 21 and step size factor = 6,7dB) (Tab. 9). The optimal settings values of a step size factor

and the filter length were and for theLMS adaptive noise canceller, implemented on theTMS320C6713 DSK.

== .

= , = , =

" "

and(

μ

μ

μ

μ

μ μ μ

μ

μ

x n

M

MM M

M

x n d n n n d nn n SSNR

d e nd n

M SSNR

M

‒ ‒

‒ ‒ ‒

8

8

12

10 12

a

= 21 10M

a

=μ‒8


Figure 6 The error signal ( ) from the LMS adaptive noise canceller(first iteration cycle), ( = 17, = 5,9×10 , = 6,7 dB,

word , (Tab. 5)),(simulated in MATLAB) [10].

e n

M SSNRμ3

a

‒

d = 9,9 10 was recognized‒2

×

Table 4 Calculation of distance , the filter length and step size factorof the LMS adaptive noise canceller (first iteration cycle), ( =10 %,

= 20 %, = 30 %, = 6,7 dB), (calculated in MATLAB).

d M

SSNR

μ M

M M a

M M = 10 % M = 20 % M = 30 %

� �1 = 0,103 �2 = 0,188 �3 = 0,26

M M = 21 M = 40 M = 99

d d = 0,184 d = 0,265 d = 0,307

Table 5 Calculation of the length of the adaptive filter and distance between the desired speech signal ( ) to theand the error signal ( ) from the LMS adaptive noise canceller (first iteration), calculated by way of the draft method with the DTW criterion

(simulated in MATLAB).

M d d ne n

LMS Adaptive Noise Canceller

SSNRa=6,7 dB SSNRw1=18,2 dB SSNRw2=3 dB SSNRw3=–1,8 dB

M =10 % �1=5,9 10× -3; M=17

d=9,9 10× -2

�1=6,4 10× -3; M=43

d=5,421 10× -1�1=5,4 10× -3; M=149

d=9,142 10× -1�1=4,09 10× -3; M=74

d=1,473

M =20 % �2=1,08 10× -2; M=10

d=9,8 10× -2

�2=1,18 10× -2; M=99

d=5,409 10× -1�2=1 10× -2; M=103

d=1,127�2=7,5 10× -3; M=74

d=1,42

M =30 % �3=14,996 10× -3; M=7

d=9,87 10× -2

�3=1,63 10× -2; M=99

d=5,405 10× -1�3=1,38 10× -2; M=103

d=1,111�3=1,04 10× -2; M=74

d=1,374

Table 6 ( )

( )

.

Calculation of the length of the adaptive filter and distance between the desired speech signal to theand the error signal from the LMS Sign adaptive noise canceller (first iteration), calculated by way of the draft method with the DTW criterion

(simulated in MATLAB)

M d d ne n

LMS Sign Adaptive Noise Canceller

SSNRa=6,7 dB SSNRw1=18,2 dB SSNRw2=3,1 dB SSNRw3=–1,8 dB

M = 10 % �1=5,9×10-3; M=8

d=0,103

�1=6,4×10-3; M=2

d=0,53�1=5,4×10-3; M=2

d=0,7411�1=4,09×10-3; M=2

d=0,5568

M = 20 % �2=1,08×10-2; M=10

d=0,109

�2=1,18×10-2; M=2

d=0,5189�2=1×10-2; M=2

d=0,4053�2=7,5×10-3; M=2

d=0,5580

M = 30 % �3=14,996×10-3; M=10

d=0,113

�3=1,63×10-2; M=2

d=0,51�3=1,38×10-2; M=2

d=0,3238�3=1,04×10-2; M=117

d=0,4526

The speech error signal (Fig. 10) from the LMSadaptive noise canceller implemented on theTMS320C6713 DSK as ( ) (first

e n( )

(

) < 0,2was recognized d

iteration cycle of the LMS adaptive noise canceller) forsettings of parameters ( = 6,7 dB), = 1×10 , = 21).SSNR Ma μ

‒8



7Using the LMS adaptive noise canceller in voicecommunication with a control Bus system

The proposed method with the DTW criterion was usedfor optimal settings parameters of the LMS adaptive noisecanceller, implemented on the TMS320C6713 DSK,applied in voice communications with the control Bussystem (Fig. 7). The control Bus system was used in theapplication of visualization operational control of thetechnical functions of the building with the visualizationsoftware Promotic.

Software My Voice, linked with software Promotic(Fig. 7), was used for the speech recognition in voicecommunication with the control Bus system. By use of thesoftware My Voice operational technical functions in thebuildings can be done through voice control. In voicecommunication with the control Bus system were usedvoice commands, for example switch on/off lights, increaseof temperature, decrease of temperature, reducing oftemperature by degree Celsius, turn on/off .one boiler

" /zhasnout světla", "zvýšit teplotu","snížit teplotu", "snížit teplotu o 1 °C" ("teplota mínus

" " " " "" "

For this experiment were chosen Czech words as voicecommands " " and " ". The aim of thisexperiment was to determine the success of the recognitionof selected voice commands ("bojler", "jedna") withadditive noise, by the use of software My Voice in voicecommunication with the control Bus system, withsubsequent conditions:

" " " "

" " " "

" "

In the Czech language were used the following voicecommands: rozsvítit

jedna ), zvýšit teplotu o 1 °C ( teplota plus )zapnout/vypnout ) etc.

a) in the case partial use of the proposed method withsettings of optimal parameters ( , ) of the LMS adaptivenoise canceller, (as voice command was used of the isolatedCzech word bojler , ( boiler )) andb) in the case without partial use of the proposed method,without settings of optimal parameters ( , ) of the LMSadaptive noise canceller (as voice command was used of theisolated Czech word jedna , ( one )).

ad a) The voice command "bojler" was used forsimulation of turning on/off of the boiler (Fig. 8).Conditions of this experiment were following:1. 100 was spoken voice command bojler without theLMS adaptive noise canceller implemented on theTMS320C6713 DSK.

without additive noise 99 % successfulspeech recognition,

with additive noise 81 % successfulspeech recognition.

jednabojler

jedna bojler

M

M

μ

μ

�

�

measurement 1

measurement 2

×


Table 7 Calculation of the length of the adaptive filter and distance between the desired speech signal ( ) to theand the error signal ( ) from the LMS Sign-Sign adaptive noise canceller (first iteration), calculated by way of the draft method with

the DTW criterion (simulated in MATLAB).

M d d ne n

LMS Sign-Sign Adaptive NoiseCanceller


M = 10 % �1=5,9×10-3; M=2

d=0,136

�1=6,4×10-3; M=2

d=0,118

�1=5,4×10-3; M=2

d=0,229�1=4,09×10-3; M=2

d=0,276

M = 20 % �2=1,08×10-2; M=2

d=0,179

�2=1,18×10-2; M=2

d=0,148

�2=1×10-2; M=2

d=0,305�2=7,5×10-3; M=2

d=0,384

M = 30 % �3=14,996×10-3; M=2

d=0,215�3=1,63×10-2; M=2

d=0,194

�3=1,38×10-2; M=2

d=0,355�3=1,04×10-2; M=2

d=0,414

Table 8 Calculation of the length of the adaptive filter and distance between the desired speech signal ( ) to theand the error signal ( ) from the LMS Sign-Reggressor adaptive noise canceller (first iteration), calculated by way of the draft method

with the DTW criterion (simulated in MATLAB).

M d d ne n

LMS Sign-Reggressor AdaptiveNoise Canceller


M = 10 % �1=5,9×10-3; M=12

d=0,172

�1=6,4×10-3; M=146

d=0,407�1=5,4×10-3; M=144

d=0,591�1=4,09×10-3; M=3

d=0,538

M = 20 % �2=1,08×10-2; M=11

d=0,21�2=1,18×10-2; M=146

d=0,304�2=1×10-2; M=37

d=0,349�2=7,5×10-3; M=2

d=0,33

M = 30 % �3=14,996×10-3; M=77

d=0,228�3=1,63×10-2; M=129

d=0,249�3=1,38×10-2; M=4

d=0,23�3=1,04×10-2; M=2

d=0,261

Table 9 Calculation of distance d for the set of the filter length = 21and step size factor with = 6,7 dB (the LMS adaptive noise

canceller was implemented on the TMS320C6713 DSK, (first iteration)).

MSSNRμ a

settings of factor �(implemented on the

TMS320C6713 DSK)�=1×10–12 �=1×10–10 �=1×10–8

calculation of distancevalues d (in MATLAB)

d=4,27×10–1 d=3,48×10–1d=8,97×10–2

settings of the filterlength M

M=21 M=21 M=21

Figure 7 Implementation of the LMS adaptive noise canceller for the voice communications with the control Bus system.



2. 100 was spoken voice command "bojler" with theLMS adaptive noise canceller implemented on theTMS320C6713 DSK with partial use of the proposedmethod


ad b) The isolated word " " was used as part ofvoice command for increasing of temperature +1 °C (Fig.9). Conditions of this experiment were the following:1. 100 spoken voice command "jedna" without the LMSadaptive noise canceller implemented on theTMS320C6713 DSK

without additive noise 98 % successfulspeech recognition,


2. 100 spoken voice command jedna with the LMSadaptive noise canceller implemented on theTMS320C6713 DSK without partial use of theproposed method

with additive noiseof the LMS adaptive noise canceller

( 00 spoken voice command jedna ) 64 % successfulspeech recognition,

of the LMS adaptive noise canceller(100 spoken voice command jedna ) 96 % successfulspeech recognition,

of the LMS adaptive noise canceller(100 spoken voice command jedna ) 99 % successfulspeech recognition.

�

�

�

�

measurement 3

jedna (one)

measurement 1

measurement 2

measurement 31 adaptation

2 adaptation

3 adaptation

" "

" "

" "

" "

1

×

×

×

×

×

×

8Conclusion

Acknowledgments

In experimental section of this paper was described theway of using of the proposed method in variations of theadaptive filter with the LMS algorithm in application ofsuppressing noise from the speech signal by simulations inMATLAB software. The proposed method was partiallyused in the practical implementation of the LMS adaptivenoise canceller on the TMS320C6713 DSK. Thisimplementation was used in voice communication with Bussystem for controlling of operating – technical functions inbuildings. Application of variations of the LMS adaptivefilter for voice communications with control system are alsosuitable for manufacturing technologies for on-line qualitycontrol e.g. [14, 15, 16, 17] and for analysis fo materialproperties for rapid prototyping.

This paper has been supported by the VŠB TU grant No.SP2011/12. The authors are thankful for the support.


Figure 8 Evaluation of recognition of isolated Czech word - voicecommand "bojler" (boiler) with software MyVoice.

Figure 9 Evaluation of recognition of isolated Czech word - voicecommand "jedna" (one) with software MyVoice.

Figure 10 The error signal ( ) from the LMS adaptive noise canceller,implemented on the TMS320C6713 DSK ( = 21, = 10 ,

= 6,7 dB, (first iteration cycle), word (Tab. 9).

e n

MSSNR

μ−8

d = 8,97×10was recognized

2−

Figure 11 The error signal ( ) from the LMS adaptive noise cancellerimplemented on the TMS320C6713 DSK ( = 21, = 10 , = 0,427,

= 6,7 dB, (first iteration cycle), word (Tab.9).

e n

M dSSNR

12μ

−

was not recognized



9References

[1] Farhang – Borounjeny, B. Adaptive Filters, Theory andapplications, John Wiley & Sons, Chichester 2005, ISBN0–471–98337–3, pp. 139-168.

[2] Poularikas, D. A.; Ramadan, M. Z. Adaptive filteringprimer with MATLAB, Taylor & Francis Group, 2006, ISBN0-8493-7043-4, pp. 101-122.

[3] Jan,nakladatelství VUTIUM, Brno, 2002, ISBN 80-214-1558-4(2. uprav. vydání), pp. 287-308.

[4] Hayikin, S.; Widrow, B. Least – Mean – Square AdaptiveFilters, WILEY INTERSCIENCE A JOHN WILEY &SONS, New Jersey 2003, ISBN 0–471–21570–8

, J. udijní opora,http://www.fit.vutbr.cz/study/courses/ZRE/public/, VUTBrno, (5.3.2011)

[6] Hayikin, S. Adaptive filter theory, PRENTICE HALL,New Jersey 2002, ISBN 0-13-090126-1, pp. 231-258

, J.; Sovka, P.; Pollák, P.; Hanžl, V.; , R.Technologie

, pp.75-165http://en.wikipedia.org/wiki/Dynamic_time_warping(8. 3. 2011)

[9] Sovka, P.; Pollák, P. slicového

, pp. 89 - 94[10] V , J. Voice communication with control system,

Dissertation thesis, VŠB TU Ostrava, 2010[11] Widrow, B.; Walach, E. Adaptive Inverse Control: A Signal

Processing Approach, Published by John Wiley & Sons, Inc.,Hoboken, New Jersey 2008. ISBN 978-0-470-22609-4, pp.59-87.

, J. Application of optimal settings of the LMS adaptivefilter for speech signal processing, IMCSI

N 978-83-60810-27-9, pp. 767 – 774.[13] Chassaing, R.; Reay, D. Digital Signal Processing and

Applications with the TMS320C6713 and TMS320C6416DSK, John Wiley & Sons, Inc. New Jersey 2008, ISBN978–0–470–13866–3, pp. 319-353.

, J. Hloch, S. Using the acoustic sound pressure levelfor quality prediction of surfaces created by abrasive waterjet.// International Journal of Advanced ManufacturingTechnology 48 1-4(2010), pp 193-203

[15] , J. et al. An investigation of surfaces generated byabrasive waterjets using optical detection. // Strojniskivestnik-Journal of mechanical engineering. vol. 53, (2007),pp. 224-232.

topographical parameters investigation created by abrasivewaterjet. International Journal of Surface Science andEngineering 3, 4(2009), 360-373.DOI: 10.1504/IJSURFSE.2009.027421.

J.; Hloch, S.; Kozak, D. Surface geometricparameters proposal for the advanced control of abrasivewaterjet technology. // The International Journal of AdvancedManufacturingTechnology 41, 3-4(2009), 323-328.

[18] Hloch, S. et al. Using waterjet in reverse logistic operations indiscarded munitions processing // Tehnicki vjesnik-Technicalgazette. 18, 2(2011), pp. 267-271

[19] Pilipovic, A.; Raos, P.; Sercer M. Experimental analysis ofproperties of materials for rapid prototyping. // InternationalJournal of Advanced Manufacturing Technology 40 1-2(2009), pp 105-115

ures evaluation of based on conceptual models using thepotential method. // Tehnicki Vjesnik - Technical Gazette 16,4(2009), p . 35-46.

[21] Hreha, P. et al. Water jet technology used in medicine. //Tehnicki Vjesnik - Technical Gazette 17, 2(2010), p . 237-240.

J. Číslicová filtrace, analýza a restaurace signálů,

[5] Černocký Zpracování řečových signalů – st

[7] Uhlíř Čmejlahlasových komunikací, nakladatelství ČVUT

Praha 2007, ISBN 978–80–01–03888–8

Vybrané metody čízpracování signálů, vydavatelství ČVUT, Praha, 2003, ISBN80–01–02821–6

aňuš

[12] VaňušTechnology, Wisła

2010, Poland, ISB

[14] Valíček ;

, , .Valíček

[16] Valíček, J.; Hloch, S. Optical measurement of surface and

//,

[17] Valíček,

,

.

, ,. .

[20] Ivandić, Z.; Ergić, T.; Kljajin, M. Welding robots kinematicstruct

,p

, p

[8]

[22] Gaćeša, B.; Milošević-Mitić, V.; Maneski, T.; Kozak, D.;Sertić, J. Numerical and experimental strength analysis o

, p

ffire-tube boiler construction. // Tehnicki Vjesnik - TechnicalGazette 18, 2(2010), p . 237-242.

Authors’addresses:

Jan Vaňuš

Vítězslav Stýskala

Faculty of Electrical Engineering and Computer ScienceVŠB - Technical University of Ostrava17 Listopadu, 708 33 Ostrava-PorubaCzech Republictel.: +420 59 699 1509e mail: [email protected]

Faculty of Electrical Engineering and Computer ScienceVŠB - Technical University of Ostrava17 Listopadu, 708 33 Ostrava-PorubaCzech Republictel.: +420 59 699 1509e mail: [email protected]

-

-


application of variations of the lms adaptive filter …

Documents