Echo can be defined as the delayed and attenuated version of an original signal which gets reflected back to the beginning. It is produced by contemplation from points where the medium features, through which it passes, alterations. Echo is utile as it is employed in the echo sounder and radio detection and ranging for sensing and geographic expedition intent. But it is besides a threat in the telecommunication sector. It degrades the quality of service and hence doing it inconvenient for the user to pass on expeditiously.
Need of Echo Cancellation
Today voice quality is used as a criterion to measure the overall quality of a web. Ultimately, the hunt for improved voice quality led to intensive research in the country of echo cancellation. In a phone conversation, reverberation is the sound of your ain voice being heard once more to you after a hold. Strong and delayed echo signals can be really raging doing conversation impossible.
Public Switched Telephone Network ( PSTN ) , Mobile, and Voice over IP ( VoIP ) communications systems can acquire echo from a figure of beginnings, so network-based reverberation cancellers are critical for good quality of service. In today ‘s competitory market, the absence of an efficient echo cancellation method can turn out detrimental on the bearer ‘s ability to retain endorsers.
Types of Echo
Echo can be classified into two classs based on how it occurs: intercrossed reverberation and acoustic reverberation. Hybrid reverberation is caused by an electric resistance mismatch on the 4-wire to 2-wire transition in wire line webs. It is the primary network-induced reverberation in today ‘s webs. Acoustic reverberation is created as a consequence of deficient acoustic isolation between the earphone and the mike in little French telephones, or when acoustic moving ridges are reflected against a wall or enclosure, typically when utilizing a hands-free unit.
1.3.1 Hybrid Echo
In a wire line PSTN web, the endorser is linked to the local exchange centre by a 2-wire parallel connexion know as the local cringle. From the local exchange, a 4-wire digital nexus is used to link to the nomadic shift centre ( MSC ) . Due to the cost of telegraphing, two-wire circuits are used to link the telephone to the local telephone exchange instead than four-wire short pantss. Hence the send and receive waies use separate wire braces. Between the two waies is the loanblend, which converts the 4-wire interface to the 2-wire interface. The loanblend is a 4-port device where the 4th port is terminated with equilibrating electric resistance.
Figure 1 – Loanblends in a PSTN
To avoid signal contemplations in the loanblend, the equilibrating electric resistance of the intercrossed must fit the electric resistance of the 2-wire line terminated by the telephone. The electric resistance of the 2-wire line depends on many parametric quantities, such as the length and type of overseas telegram, every bit good as the electric resistance of the telephone sets at the client premises. In pattern, the balance of the loanblend is merely nominally achieved because the 2-wire cringle ‘s electric resistance can non be determined in progress. Therefore, a fraction of the signal is reflected back to the transmitter, which is heard as reverberation.
Figure 2 – Echo from the Hybrid
The grade of instability of the intercrossed determines the strength of the echo contemplation. This strength of the echo contemplation is expressed in footings of Echo Return Loss ( ERL ) .
1.3.2 Acoustic Echo
Acoustic reverberation occur when some of the sound from the talker portion of the telephone gets picked up and transmitted back by the mike. There are two typical beginnings of acoustic reverberation:
Acoustic isolation reverberation is generated when the earphone and mike are ill isolated from one another. In today ‘s radio webs, acoustic reverberation is common due to a proliferation of ill designed French telephones, headsets, and Bluetooth headsets.
Ambient acoustic reverberation is generated when a telephone conversation is held in an acoustically brooding environment. In this state of affairs, the French telephone mike foremost picks up the original sound watercourse, followed by the address that is reflected from the walls.
Figure 3 – Acoustic Echo Beginnings
1.4 Hybrid and Acoustic Echo Differences
The intercrossed echo way is stationary, which means that it is invariant over clip. Once the call way is established, the echo hold does non alter during the class of the call. Acoustic reverberation, on the other manus, varies based on a battalion of external factors like the place of the speaker in the room, or even head motions comparative to the French telephone, which makes the acoustic reverberation a extremely non-stationary signal.
Linearity is how good the wave form of echo signal lucifers the original signal. Hybrid reverberation is a additive signal, which means that a additive mathematical theoretical account constructed inside the echo canceller can accurately foretell the intercrossed echo signal. Acoustic reverberation is non a additive signal. First, nonlinearities might be created by the parallel circuitry. In the instance of the handset/headset it includes the mike, the mike amplifier, the speaker unit amplifier, and the speaker unit. More significantly, for radio and many VoIP calls, the voice codec processing introduces extra nonlinearities.
An echo signal is non a individual contemplation of the original signal, but is a back-to-back contemplation over a period of clip. Echos have a certain continuance, or scattering clip, which is the period of clip during which the reverberation contemplation occurs. A intercrossed reverberation has a typical scattering of less than 10 MS. However, since acoustic reverberation can be generated by contemplations from the environment ; acoustic reverberation is more diffusing, with scattering times of up to 100 MSs.
Figure 4 – Hybrid Echo Waveform
Figure 5 – Acoustic Echo Waveform
COMPONENTS OF AN ECHO CANCELLER
Figure 6 – Hybrid Echo Canceller Functional Blocks
2.1 Adaptive Linear Filter
In this undertaking FIR ( Finite Impulse Response ) filter is used and its coefficients are optimized utilizing the NLMS ( Normalized Least Mean Square ) Algorithm. The signal Rin is fed into the adaptative filter. The filter with the aid of the algorithm develops an estimation of the reverberation. This estimation is so subtracted from the echo signal Sin. The consequence is so fed to the filter through which it adapts its filter coefficients. This procedure is repeated until the mistake signal is minimized.
2.2 Double Talk Detector
The reverberation canceller besides consists of a dual talk sensor. Its map is to halt the filter version when both companies are speaking at the same clip. In the absence of dual talk sensor the dual talk may confound the reverberation canceller taking to divergence in the echo canceller algorithm.
2.3 Non-Linear Processor
The Non-Linear Processor ( NLP ) is used to take the residuary reverberation. As the name suggest it filters the non-linear portion of the reverberation which is n’t removed by the additive adaptative filter. Even the NLP, along with the adaptative filter, is non active during dual talk status.
ADAPTIVE LINEAR FILTER
The adaptative filter is one that self-adjusts its transportation map harmonizing to an optimising algorithm. Its frequence response varies with clip, as a map of the input signal. Here we use Finite Impulse Response ( FIR ) Filters whose coefficients are updated utilizing NLMS algorithm.
3.1 FIR Filters
FIR filters are all zero filters, additive, clip invariant and stable. The Z-transform of the impulse response yields the transportation map of the FIR filter:
Therefore the filter is of the signifier,
FIR filters are clearly bounded-input bounded-output ( BIBO ) stalls, since the end product is a amount of a finite figure of finite multiples of the input values, so can be no greater than times the largest value looking in the input.
3.2 NLMS Algorithm
The Normalized Least Mean Square ( NLMS ) Algorithm is used to update the FIR filter coefficients. The coefficient updating equation is derived below:
The block diagram, shown below, is the base of all adaptative filter realisation. The thought behind it is that a variable filter extracts an estimation of the coveted signal.
Figure 7 – Adaptive FIR NLMS Filter
As seen from the diagram, error signal is the difference between the end product and the coveted signal
( 1.1 )
One of the most widely used nonsubjective map in adaptative filtering is the mean-square mistake ( MSE ) defined as,
( 1.2 )
Each component of the input signal vector consists of a delayed version of the same signal, i.e.
( 1.3 )
In this instance the signal is the consequence of using an FIR filter to the input signal,
( 1.4 )
is the input vector stand foring tapped-delay line
and is the tap-weight vector.
( 1.5 )
P is the cross-correlation vector between the desired and input signals and R is the input signal correlativity matrix.
By comparing the gradient vector to zero, the optimum values for the tap-weight coefficients that minimize the nonsubjective map can be evaluated which gives,
( 1.6 )
Substituting the above status we get,
( 1.7 )
If good estimations of matrix R and of vector P are available so a steepest-descent algorithm can be used to seek the Wiener solution as follows,
( 1.8 )
where represents an estimation of the gradient vector of the nonsubjective map with regard to the filter coefficients.
One possible solution is to gauge the gradient vector by using instantaneous estimations for R and P as follows:
The ensuing gradient estimation comes out to be,
( 1.9 )
The ensuing gradient-based algorithm is known as the least-mean-square ( LMS ) algorithm, whose updating equation is,
( 1.10 )
where the convergence factor I? should be chosen in a scope to vouch convergence.
Typically, one loop of the LMS requires N + 2 generations for the filter coefficient updating and N + 1multiplications for the mistake coevals.
To increase the convergence velocity of the LMS algorithm without utilizing estimations of the input signal correlativity matrix, a variable convergence factor is a natural solution. The normalized LMS algorithm normally converges faster than the LMS algorithm, since it utilizes a variable convergence factor taking at the minimisation of the instantaneous end product mistake.
The updating equation of the LMS algorithm can use a variable convergence factor in order to better the convergence rate. In this instance, the updating expression is expressed as
( 1.11 )
where must be chosen with the aim of accomplishing a faster convergence.
A possible scheme is to cut down the instantaneous squared mistake every bit much as possible. The motive behind this scheme is that the instantaneous squared mistake is a good and simple estimation of the MSE. On work outing we get,
( 1.12 )
Using this variable convergence factor, the updating equation for the LMS algorithm is so given by,
( 1.13 )
Normally a fixed convergence factor is introduced in the updating expression in order to command the misadjustment, since all the derivations are based on instantaneous values of the squared mistakes and non on the MSE.
Besides a parametric quantity I? should be included, in order to avoid big measure sizes when becomes little.
The coefficient updating equation is so given by,
( 1.14 )
DOUBLE TALK DETECTOR
Another of import constituent of an echo canceller is the dual talk sensor. The dual talk sensor senses whether the near-end signal is corrupted by the far-end signal. When such a instance is detected its map is to stop dead the version of the filter to avoid the divergency of the adaptative algorithm.
The basic dual talk sensing procedure starts with calculating a sensing statistic
and comparing it with a preset threshold. Different methods have been proposed to organize
the sensing statistic. The Geigel algorithm has proven successful in line reverberation
cancellers. However, the Geigel algorithm is n’t that efficient and dependable in the acoustic reverberation cancellation applications. For AEC ‘s the cross-correlation based methods are found to be more utile.
The general process for managing dual talk is described by the undermentioned four stairss:
A sensing statistic is formed utilizing available signals such as ten, vitamin D and vitamin E and the estimated filter coefficients.
The sensing statistic, L is compared to a preset threshold, T, ( a invariable ) , and dual talk is declared if L & lt ; T.
Once doubletalk is declared the sensing is held for a minimal period of clip Thold. While the sensing is held the filter version is disabled.
If L a‰? T consecutively over a clip Thold the filter resumes version while the comparing of I? to T continues until L & lt ; T once more.
The clasp clip, Thold, in stairss 3 and 4 is indispensable to stamp down sensing dropouts due to the noisy behaviour of the sensing statistic.
A nonlinear processor, ( NLP ) , is a signal processing circuit or algorithm that is placed in the address way after echo cancellation in order to supply farther fading or remotion of residuary reverberation signals that can non be removed wholly by an echo canceller.
A non-linearity, a deformation, or added noise signals are illustrations of signals that can non be to the full cancelled by an echo canceller. Therefore, these signals are typically removed or attenuated by a nonlinear processor.
Few non-linear reverberations are:
clipped address signals
hapless quality speakerphones
pulse codification transition ( PCM )
Non-linear reverberations dispute whirl processor to develop an accurate reverberation estimation.
ECHO CANCELLER STANDARD
The International Telecommunication Union ‘s telecommunication standardisation sector ( ITU – Thymine ) has developed criterions specifying how telecommunication webs operate and interwork.
G.168 is recommendation for Digital Network Echo Cancellers that describes the features of the echo canceller, every bit good as the research lab trials that should be performed on an echo canceller to measure its public presentation under conditions likely to be experienced in the web.
As per ITU-T,
Echo cancellers have the undermentioned cardinal demands:
Rapid convergence ;
Low returned echo degree during individual talk ;
Low divergency during dual talk and off terminal address ;
Assured dual talk sensing and off terminal address sensing ;
Proper operation during facsimile and low velocity ( & lt ; 9.6 Kbit/s ) voice set informations
Few trials defined are:
Test No. 2 – Convergence and steady province residuary and returned echo degree trials
Test No. 3 – Performance under conditions of dual talk
Test No. 4 – Leak rate trial
Test No. 5 – Infinite return loss convergence trial
Test No. 6 – Non-divergence on narrow-band signals
Test No. 7 – Stability trial
SIMULATION USING MATLAB AND SIMULINK
Figure 8 – Simulink Model
An sound file was taken as input. The input signal was added with its attenuated and delay version. The ensuing signal was fed to the NLMS block input. The coveted signal was taken to be the input itself. The weights and mistake of the NLMS were shown utilizing the show block in simulink. The end product was stored as an sound file. A drifting range was used to expose the input, desired, mistake and end product signal. The above simulink theoretical account was simulated for 10 seconds.
Figure 9 – Floating Scope Output during Simulation
Figure 10 – Input signal Voice Signal Plot
Figure 11 – Voice Signal and Echo Plot
Figure 12 – Filter Output Plot
Figure 13 – ERLE Plot
Figure 14 – Mistake Plot
ITU EVALUATION TESTS
In order to measure the effectual working of the algorithm, few of the ITU trials were conducted.
Test No. 2 – Convergence and steady province residuary and returned echo degree trials
“ This trial is meant to guarantee that the echo canceller converges quickly for all combinations of input signal degrees, echo waies, and certain echo way alterations and that the returned echo degree is sufficiently low. ”
As seen from Figure 12 the filter converged quickly and therefore call offing the reverberation. Besides from Figure 13 high ERLE value was observed which indicates better public presentation. As per ITU-T lower limit ERLE value of 30dB is required. As seen from the secret plan maximal ERLE value of 55dB was observed.
Therefore the theoretical account has successfully satisfied ITU-T G.168 Test No. 2.
Test No. 6 – Non-divergence on narrow-band signals
“ This trial has the object of verifying that the echo canceller will stay converged for subscriber-originated narrow-band signals after holding converged on a broadband signal. ”
In this trial a single-talk CSS ( composite beginning signal ) was given as input. In time-domain, it consists of 3 parts, the active voice portion, the random noise portion and the silence ( intermission ) portion. Plot of the signal is shown below in Figure 14.
Figure 14: ITU-T Single-Talk CSS Signal Plot
Figure 15: CSS Signal and Echo Plot
Figure 16: Filter Output Plot
Figure 17: Mistake Plot
Figure 17: ERLE Plot
As seen from Figure 16 we observed that the filter did n’t diverge on using narrow-band signal. Besides from Figure 17 maximal value of ERLE was found to be 5dB. As per ITU-T the value should be less than equal to 6dB.
Therefore the theoretical account has successfully satisfied ITU-T G.168 Test No. 6.
Test No. 7 – Stability Trial
On detecting the secret plans from the above two trials we can corroborate that the filter is found to be stable for both wide-band and narrow-band signals.
Therefore the theoretical account has successfully satisfied ITU-T G.168 Test No. 7.
MATLAB CODE IMPLEMENTING NLMS
clear all ;
mule = .01 ; % Larger values for fast convergence
max_run = 200 ;
for run=1: max_run ;
lights-outs = 20 ; % Adaptive Filter Taps
freq = 2000 ; % Signal Frequency
tungsten = nothing ( 1, lights-outs ) ; % province of adaptative filter
clip = .2 ; % length of simulation ( sec )
samplerate = 8000 ; % samples/sec
samples = time*samplerate ;
max_iterations = samples-taps+1 ;
loops = 1: max_iterations ; % Vector of loops
t=1/samplerate:1/samplerate: clip ;
rand ( ‘state ‘ , amount ( 100*clock ) ) ; % Reset Randome Generator
noise=.02*rand ( 1, samples ) ; % noise added to signal
s=.4*sin ( 2*pi*freq*t ) ; % Pure Signal
x=noise+s ; % input to adaptive filter
echo_amp_per = .4 ; % Echo per centum of signal
% rand ( ‘state ‘ , amount ( 100*clock ) ) ; % Reset Randome Generator
echo_time_delay = .064 ;
echo = echo_amp_per* [ zeros ( 1, echo_delay ) ten ( echo_delay+1: samples ) ] ;
for i=1: max_iterations ;
Y ( I ) =w*x ( one: i+taps-1 ) ‘ ;
vitamin E ( run, I ) =echo ( I ) -y ( I ) ;
mule ( I ) = .5/ ( x ( one: i+taps-1 ) *x ( one: i+taps-1 ) ‘+ .01 ) ;
tungsten = w + 2*mule*e ( run, I ) *x ( one: i+taps-1 ) ;
% % Mean Square Error
mse=sum ( e.^2,1 ) /max_run ;
% Ouput of System
out=b ( 1: length ( y ) ) -y ;
subplot ( 3,1,1 ) , secret plan ( B ) ;
rubric ( ‘Signal and Echo ‘ ) ;
ylabel ( ‘Amp ‘ ) ;
xlabel ( ‘Time sec ‘ ) ;
subplot ( 3,1,2 ) , secret plan ( out ) ;
rubric ( ‘Output of System ‘ ) ;
ylabel ( ‘Amp ‘ ) ;
xlabel ( ‘Time sec ‘ ) ;
subplot ( 3,1,3 ) , semilogy ( mse ) ;
rubric ( ‘LEARNING CURVE mu=.01 reverberation delay=64ms runs=200 ‘ ) ;
ylabel ( ‘Estimated MSE, dubnium ‘ ) ;
xlabel ( ‘Number of Iterations ‘ ) ;
subplot ( 3,1,2 ) , semilogy ( loops, vitamin E ( 1, : ) .^2 ) ;
subplot ( 3,1,3 ) , semilogy ( loops, vitamin E ( 2, : ) .^2 ) ;
Figure 18 – Matlab Simulation Result demoing the filter meeting
With the universe shriveling into a planetary small town because of superior communications, telephones, both conventional and hands-free sets, occupy a outstanding place in work outing people ‘s communicating demands. One of the major jobs in a telecommunication application over a telephone system is echo. The Echo cancellation algorithm presented in this study successfully attempted to happen a package solution for the job of reverberations in the telecommunications environment. The proposed algorithm was wholly a package attack without using any DSP hardware constituents. The algorithm was capable of running in any Personal computer with MATLAB package installed. In add-on, the consequences obtained were converting. The sound of the end product address signals were extremely satisfactory and validated the ends of this undertaking.