We are IntechOpen, the world s leading publisher of Open Access books Built by scientists, for scientists. International authors and editors

Transcription

1 We are IntechOpen the world s leading publisher o Open Access boos Built by scientists or scientists M Open access boos available International authors and editors Downloads Our authors are aong the 54 Countries delivered to OP % ost cited scientists 2.2% Contributors ro top 500 universities Selection o our boos indexed in the Boo Citation Index in Web o Science Core Collection (BKCI) Interested in publishing with us? Contact boo.departent@intechopen.co Nubers displayed above are based on latest data collected. For ore inoration visit

2 Perceptual Echo Control and Delay Estiation Kirill Sahnov Eaterina Verteletsaya and Boris Sia Czech echnical University in Prague Czech Republic 4. Introduction Echo phenoenon has been always existed in telecounications networs. Generally it has been noticed on long international telephone calls. As technology advances and the data transission ethods tend ore to pacet-switching concepts the traditional echo proble reained up-to-date. An iportant issue in echo analysis is a round-trip delay o the networ. his is a tie interval required or a signal ro speaer s outh across the counication networ through the transit path to the potential source o the echo and then bac across the networ again on the receive path to the speaer s ear. he ain proble associated with IPbased networs is that the round-trip delay can be never reduced below its undaental liit. here is always a delay o at least two to three pacet sizes (50 to 80 s) (Choi et al. 2004) that can ae the existing networ echo ore audible (Gordy & Goubran 2006). hereore all Voice over IP (VoIP) networ terinals should eploy echo cancellers to reduce the aplitude o returning echoes. A ain paraeter o each echo canceller is a length o its coverage. he coverage eans the length o tie that the echo canceller stores its approxiation in eory. he adaptive ilter should be long enough to odel an unnown syste properly especially in case o VoIP applications (Nisar et al. 2009; Youhong et al. 2005). On the other hand it is nown that an active part o the networ echo path is usually uch saller copared to the whole echo path that has to be covered by the adaptive iltering algorith. hat is why the nowledge o the echo delay is iportant or using echo cancellers in pacet-switching networs. oday there is a wide aily o adaptive iltering algoriths that can exploit sparseness o the echo path to reduce high coputational coplexity associated with long echo paths (Dyba 2008; Hongyang & Dyba 2008; Khong & Naylor 2006; Hongyang & Dyba 2009). In this chapter we discuss nuerous ethods used or estiation o echo delay. Algoriths based on cross-correlation unction and adaptive ilters are used in the art. We will consider both types o the discuss their advantages and drawbacs. Aterwards we will pay our attention to the adaptive iltering techniques. We provide a study on dierent partial proportionate sparseness-controlled tie- and requency-doain adaptive ilters. he readers will get closer to an issue o echo cancellation which is relevant in nowadays telecounications networs. Ones will able to recognize iportant eatures and particular areas o ipleentation o various adaptive algoriths. Further we are giving a short introduction to the issue o echo control or telecounications networs. his description ephasises on two ost iportant aspects o perceptual echo control which are echo loudness and echo delay.

3 86 Adaptive Filtering Applications. Echo control issue In the very beginning o the telephone age all calls were ade through an analog pair o copper wires. he technology has progressively oved to digital circuit switched networs over the past several decades. oday ost o the phone traic is handled by the Public Switched elephone Networ (PSN) which provides end-to-end dedicated circuits. During the last years a ove to pacet-switched networs has been initiated to support voice traic over Internet Protocol (IP). he ain reason or the ove ro circuit-switched voice networs to pacet-switched networs is to enable convergence between data services and voice services. It is o econoical interest to be able to use the sae equipent or voice and data traic. Reduced cost o placing a phone call is another reason since the voice-pacet is treated and routed uch in the sae way as any other data pacet (note that Quality o Service plays a vital role in this process). hus conventional long distance taris have a tendency to be copletely eliinated in Voice over IP (VoIP) networs as well. Echo issue has long been recognized as a proble on telecounications networs though generally it has been noticed ostly on international telephone calls or when using speaer phones. As technology advances and the inoration transission ethods tend ore to pacet-switching concepts the traditional echo proble should be reviewed and updated. Previously unconsidered actors now play an iportant part in the echo characteristics. his section describes the echo delay proble which is oten encountered in pacet-switched networs. his proble is highlighted in relation to VoIP networs. More speciic details on the process o locating and eliinating echoes are included in conclusion to the chapter. Consider a siple voice telephone call where an echo occurs when you hear your own voice repeated. An echo is the audible lea-through o your own voice into your own receive path. Every voice conversation has always at least two participants. Fro the perspective o each participant there are two voice paths in every call: ransit path he transit path is usually depicted as x path. In a conversation the transit path is created when any person begins speaing. he sound is transitted ro the outh o the speaer to the ear o the listener. Receive path he receive path is also called the return and depicted as Rx path. In a conversation the receive path is created when a person hears the conversation coing ro the outh o another speaer. Fig. illustrates a siple diagra o a voice call between two persons A (Kirill) and B (Kate). Fro the user A s perspective the x path carries his voice to the user B s ear and the Rx path carries the user B s voice to the user A s ear. Fig.. A siple telephone call scenario

4 Perceptual Echo Control and Delay Estiation 87 here is one signiicant actor in the echo analysis and especially or the pacet-switching networs. It is a round-trip delay o the voice networ. he round-trip delay is the length o tie required or an utterance ro the user A s outh across the networ on the x path to the source o the lea and then bac across the networ again on the Rx path to the user A s ear. Let s deine two iportant stateents about echo nature which are the ollowing: he louder the echo (echo aplitude) the ore annoying it is he longer the round-trip delay (the later the echo) the ore annoying it is. able shows how tie delay can aect the quality o a voice conversation. One-Way Delay Range (s) Greater than 400 able. Eect o Delay on Voice Quality Eect on Voice Quality his is the expected range or national calls. here are no diiculties during conversation. his is the expected range or international calls using a terrestrial transport lin and IP telephony which includes only one instance o IP voice. his range is acceptable or ost users assuing the use o echo control devices. his is the expected range or a satellite lin. Delays in this range can interrupt the noral low o a conversation. A high-perorance echo canceller ust be used and careul networ planning is necessary. his is excessive delay and ust be avoided by networ planning. Fig. 2 shows how the echo disturbance inluenced by the two paraeters: delay and echo level. he etric called aler Echo Loudness Rating (ELR) denotes the level dierence between the voice and echo signals. he acceptable curve represents the liit or acceptable taler echo perorance or all digital networs. he act that the speaer A in Fig. hears an echo illustrates one o the basic characteristics o echo: perceived echo ost liely indicates a proble at the other end o the call. he proble that is producing the echo that A hears the leaage source is soewhere on B s side o the networ. I the person B was experiencing echo the proble would be on the user A s side. he perceived echo usually originates in the terinating side o the networ or the ollowing two reasons: Leaage happens only in analog circuits. Voice traic in the digital portions o the networ does not pass ro one path to another. Echo arriving ater very short tie about 25 illiseconds is generally iperceptible because it is ased by the physical and electrical side-tone signal. A hybrid transorer is oten ain source o the electrical signal leaage. he typical analog telephone terinal is 2-wire device: a single pair o conductors is used to carry both the x and Rx signals. For analog trun connections nown as 4-wire transission two pairs o conductors carry separate x and Rx signals. Digital truns (/E) can be virtual 4-wire lins because they also carry separate x and Rx signals. A hybrid is a transorer that is used to interace 4-wire lins to 2-wire lins. Fig. 4 shows a hybrid transorer in an analog tail circuit. Because a hybrid transorer is a non-ideal physical device a certain raction o the 4-wire incoing (Rx) signal will be relected into 4-wire outgoing (x) signal. A typical

5 88 Adaptive Filtering Applications raction or a properly terinated hybrid in a PBX is about -25 decibels (db) eaning that the relected signal (the echo) will be a version o the Rx signal attenuated by about 25 db. For a PSN POS (Plain Old elephone Service) terination the expected value is between 2 and 5 db. Echo strength is expressed in db as a easureent called Echo Return Loss (ERL). hereore and ERL o 0 db indicates that the echo is the sae aplitude as the original source. A large ERL indicates a negligible echo. Reeber that an echo ust have both suicient aplitude and suicient delay to be perceived. For local calls with one-way delay ro 0 to 25 s an echo o strength o -25 db relative to the speech level o the taler is generally quiet enough to not be annoying. For a one-way delay in the range o 25 to 50 s the ERL should exceed 55 db to eliinate the perception o echo ro the end-user perspective as recoended in IU- recoendation G.68 on echo cancellation (IU- G ). In this case echo cancellation is required. Fig. 2. aler echo tolerance curves (IU- G ) 2. Echo delay estiation using cross-correlation he ollowing section presents a study o cross-correlation-based ie Delay Estiation (DE) algoriths. he ain purpose is to analyze a nuber o ethods in order to ind the ost suitable one or real-tie speech processing. As DE is an iportant topic during transission o voice signals over pacet-switching telecounication systes it is vital to estiate the true tie delay between x and Rx speech signals. We consider algoriths processing both in tie- and requency doains. An echo delay proble associated with IPbased transport networs is also included into the discussion. An experiental coparison o the perorance o nuerous ethods based on cross-correlation noralized crosscorrelation and a generalized cross-correlation unction is presented. 2. General scenario o delay estiation using cross-correlation unctions he nown proble associated with IP-based networs is that the round-trip delay can be never reduced below its undaental liit. here is always a delay o at least two to three pacet sizes (50 to 80 s) that can ae the existing networ echo ore audible. hereore all

6 Perceptual Echo Control and Delay Estiation 89 Voice over IP (VoIP) networ terinals should eploy echo cancellers to reduce the aplitude o returning echoes. A ain paraeter o each echo canceller is a length o coverage. Echo canceller coverage speciies the length o tie that the echo canceller stores its approxiation in eory. he adaptive ilter should be long enough to odel an unnown syste properly especially in case o VoIP applications. On the other hand it is nown that the active part o the networ echo path is usually uch saller copared to the whole echo path that has to be covered by the adaptive iltering algorith inside the echo canceller. hat is why the nowledge o the echo delay is iportant or using echo cancellers in pacetswitching networs successully. In general every counications syste includes a counications networ and counications terinals on the both sides o the networ. he counications terinals could be telephones sot phones and wireless voice counication devices. Fig. 3 illustrates how an echo assessent device can be arranged into the deined syste. he echo delay estiator has to onitor two parallel channels. An outgoing voice channel transits an original voice waveor ro the irst terinal through the counications networ to the second terinal. An incoing voice channel receives an echo waveor o the original signal returning ro the second terinal through the counications networ bac. his is a delayed and attenuated version o the original voice signal. Fig. 3. Arrangeent o echo assessent odule in the networ Fig. 4. General bloc diagra o delay estiator Fig. 4 illustrates a general bloc diagra o the echo delay estiator. he echo delay estiator coputes correlation between two voice channels or dierent set o delays in parallel anner (Carter 976). he delay-shit with the largest cross-correlation coeicient is selected as the delay estiate. Fig. 5 illustrates in a lowchart or steps perored when ipleenting a ethod o echo delay estiation utilizing cross-correlation algoriths. Once started ro bloc bloc 2 calculates the cross-correlation unction or a buer o

7 90 Adaptive Filtering Applications input saples o the Rx and x signals. Bloc 3 utilizes cross-correlation coeicients to copute the siilarities between the transitted signal and the received signal over a range o delays. For each particular delay the siilarity is obtained. Once the siilarities have been deterined or each delay within the range o delays bloc 4 chooses a delay that produces the greatest siilarity etric or the given input raes. Consequently bloc 5 indicates that the estiation process is copleted. Fig. 5. Flowchart or estiating echo delay value 2.2 Algoriths proceeding in tie-doain ie doain ipleentation o Cross-Correlation Function (CCF) and Noralized CCF (NCCF) is presented. he cross-correlation unction or a successive par o speech raes can be estiated by (Mueller 975) DL R x n y n n0... L 0... L xy nd DL R x n y n n0... L 0... L xy MIN MAX nd Here x(n) siply denotes a rae o the outgoing signal y(n) is related to a rae o the incoing signal. According to Fig. 4 the estiation o the CCF is done or a supposed range o delays. he tie-shit τ which is always in range o [τ in ; τ ax ] and causes the axial pea value o the CCF is declared as an estiate o the true echo delay D. Siilarly to the CCF an estiate o the NCCF is done (Buchner et al. 2006) (2) DL DL xn yn n nd R n0... L 0... L xymin E E x y xn yn n nd R n0... L 0... L xymax E E x y (3) Here E x and E y denotes a short-ter energy o the outgoing and the incoing raes. hese values are calculated using the ollowing equations

8 Perceptual Echo Control and Delay Estiation 9 DL 2 E x n x nd y (4) DL 2 (5) E y n nd Let us urther consider generalized cross-correlation algoriths which operate in the requency doain (Youn et al. 983; Zetterberg et al. 2005). 2.3 Algoriths proceeding in requency-doain More sophisticated way how to provide DE is to copute the cross-correlation unction in the requency doain. his process in literature is called Generalized Cross-Correlation (GCC) (Hertz 986). he idea behind this ethod is to peror pre-iltering o the input signals beore calculating CCF. It aes possible to iprove the accuracy o delay estiation. Note that the iltering procedure is perored in the requency doain. Let us describe this process in ore details. It is well nown that the siple cross-correlation unction R xy between signals x(n) and y(n) is related to the cross-power density unction (cross-power spectru) G xy by the general inverse Fourier transor relationship as (6) j 2 R G e d xy xy When x(n) and y(n) have been iltered with ilters having transer unctions H () and H 2 () the cross-power spectru between the ilter out-puts is given by g xy G H H G (7) 2 Consequently the Generalized Cross-Correlation Function (GCCF) between x(n) and y(n) is given by (Knapp & Carter 976) xy g j 2 R G e d xy (8) g xy (9) H H g 2 Here Ψ g is a generalized weighting unction. able 2 represents weighting unctions that were used or experients with speech signals (Wilson & Darrell 2006). he paraeter xy denotes a coplex coherence unction. It can be calculated as (ianshuang & Hongyu 996) xy xx G xy G G yy (0) () j 2 G R e d xx xx (2) j 2 G R e d yy yy

9 92 Adaptive Filtering Applications Here G xx () and G yy () are auto-power spectra o the outgoing and the incoing signal; R xx () and R yy () are auto-correlation unctions o the sae signals. Fig. 6 illustrates a bloc diagra o the ipleented generalized cross-correlation algorith where the Fast Fourier ransor (FF) is used or auto-spectra and cross-spectru calculation. Ater the crosspower spectru is estiated it is ultiplied by the corresponding GCC weighting unction. he inverse FF is used or obtaining the tie doain generalized-cross correlation unction. his operation is repeated or the speciied range o possible delays. Ater the whole process has copleted the tie shit with axiu corresponding pea value is declared as an estiation o the true delay. Processor Nae Weighting Function Cross-correlation ROH Gxy SCO G G PHA xx yy G xy Cps- G G H (ML) Ecart xy xx xy yy G 2 xy G xy G G xx xy Gyy Gxy HB G G G Wiener 2 able 2. Various GCC weighting unctions xy xx yy xy 2 Fig. 6. Diagra o the ipleented generalized cross-correlation algorith 2.4 Discussion over experiental results We used MALAB sotware as a siulation environent. he tie dierence between tie when the outgoing signal leaves the voice terinal and oent when the incoing signal containing the echo o the original signal arrives bac ro the networ is reerred to as a true echo delay. his value or the irst three igures that are presented below equals 6s (48 saples). For the purpose o DE it is also necessary to speciy tie interval through which

10 Perceptual Echo Control and Delay Estiation 93 the value o the true delay is searched. o cover the 6s delay we choose the interval between 0 and 0s what corresponds to the axiu delay value o 60 saples. Aterwards we present the estiation results or a group o dierent delays. It helps to understand better perorance o the algoriths. Unortunately because o the nonstationary nature o huan speech the CCF is not reliable or all situations. Its perorance highly depends on nuerous actors i.e. signal strength signal-to-noise ratio (SNR) etc (Chen et al. 2006). he NCCF is not so sensitive to the sudden changes in the signal s aplitude. It outperors the CCF when we wor with low level signals. he advantages o the algoriths proceeding in the requency doain copared to the algoriths operating in the tie doain are accuracy and reduced coputational coplexity. Fig. 7 illustrates the outputs o the GCC algoriths which were presented in able 2. (a) ROH - weighting unction (b) SCO - weighting unction (c) PHA - weighting unction (d) CPS-M (M=2) - weighting unction

11 94 Adaptive Filtering Applications (e) H - weighting unction () ECKAR - weighting unction (g) HB - weighting unction (h) WIENER - weighting unction Fig. 7. ie delay estiation using GCCF able 3 and 4 provides us along with the ollowing results. he joint coparison was done in ters o the estiation accuracy o the algoriths. he group o delays was chosen or this experient. Delay values are consistent with the ones reerenced in the corresponding IU- recoendation G.3 (IU- G ). Once the respective cross-correlation unction was calculated its axiu pea value is detected using the searching procedure described in Fig. 4. SCC is related to the Standard CC unction.

12 Perceptual Echo Control and Delay Estiation 95 [s] SCC ROH SCO PHA CPS-2 H ECKAR HB WIENER able 3. Mean values o estiated delays [s] SCC ROH SCO PHA CPS-2 H ECKAR HB WIENER able 4. Root ean square deviation o estiated delays he abscissa o the largest pea value is the estiated delay. Note that 50 trial speech records or each processor were evaluated to obtain the ean value and the Root Mean Square Deviation (RMSD) paraeter (Anderson & Woessner 992). Not only dierent speech signals but various hybrid ipulse response odels have been used. he results or delays ro 5 to 300 s are presented in the corresponding tables. able 3 contents the ean values whether able 4 illustrates the estiated RMSD values. 3. Echo delay estiation using adaptive ilters In this section we introduce ethods or extracting an echo delay between speech signals using adaptive iltering algoriths. We now that tie delay estiation is an initial step or any speech processing applications. Conventional techniques that estiate a tie dierence o arrival between two signals are based on the pea deterination o the generalized cross-correlation between these signals. o achieve a good precision and stability in estiation the input sequences have to be ultiplied by an appropriate weighting unction. Regularly the weighting unctions are dependent on the signals power spectra. he spectra are generally unnown and have to be estiated in advance. An ipleentation o the tie delay estiation via the adaptive least ean squares is analogous to estiating the Roth generalized cross-correlation weighting unction. he estiated paraeters using the adaptive ilter have a saller variance because it avoids the need or the spectru estiation. In the ollowing we discuss proportionate and partial-update adaptive techniques and consider their perorance in ter o delay estiation.

13 96 Adaptive Filtering Applications ie delay estiation (DE) has always been and reains a popular research topic. It is nown that the ain proble associated with IP-based networs is the round-trip delay which can never be reduced below its undaental liit. A nuber o eorts were ade in order to iprove the DE precision. Various ethods based on the Generalized Cross- Correlation (GCC) were proposed. he GCC algoriths arrange a pre-ilter to obtain the odiied signal spectru or optial tie delay estiation. o speciy the ilter s characteristic it requires a priori nowledge o the statistics o the received signals. However the eiciency o the algoriths decreases considerably when little or no prior nowledge about the signal statistics is nown. Fro the tie when Widrow proposed an adaptive iltering technique based on Least Mean Squares (LMS) (Widrow 2005; Hayin 200) an adaptive theory also ound an application to delay estiation. An adaptive ipleentation o the tie delay estiation via Widrow s LMS algorith is usually reerred to as DLMS. Coparing to the GCC algoriths the adaptive iltering techniques do not require a priori inoration o the signal statistics because the estiation o the signal spectru is no longer needed. he adaptive iltering algoriths deterine the tie delay in an iterative anner. here are coparative studies which provide coparison o the LMS versus the generalized cross-correlation (Zetterberg et al. 2005). Generally the tie doain ipleentation o any adaptive ilter is associated with high coputational coplexity. It directly depends on the length o the adaptive ilter. In order to reduce the coputational load o the DLMS (Eadzadeh et al. 2008) we oer using adaptive iltering algoriths with reduced coputational coplexity. hey provide savings coparing to the conventional adaptive algoriths. In the ollowing we discuss each o the algoriths in greater details. First a general scenario or the adaptive tie delay estiation using a siple Noralized Least Mean Squares (NLMS) adaptive iltering algorith is presented. Aterwards we introduce the proportionate and partial-updated algoriths proceeding in the tie doain. A new partial-updated proportionate NLMS algorith is outlined. A coparison between the DE algoriths is ade in context o the networ echo delay estiation. 3. General scenario o delay estiation using adaptive ilters raditionally the NLMS algorith is used or the echo canceller ipleentation. It applies a Finite Ipulse Response (FIR) adaptive ilter with adjustable weights or odelling the unnown echo path s ipulse response. he NLMS algorith iniizes the least ean squares dierence between two signals: the reerence incoing signal and the iltered original (outgoing) signal (see Fig. 7). Fig. 7. ie delay estiation using adaptive ilter

14 Perceptual Echo Control and Delay Estiation 97 Basically the NLMS algorith is a siple extension o the Widrow s LMS algorith. According to Fig. 7 the adaptive ilter weights are tuned iteratively using a special rule. he iportant paraeter controlling the whole adaptation process is reerred to as a step-size paraeter. It should be varied in tie in order the algorith to be able to trac nonstationary changes in the echo path s ipulse response. Unlie the LMS algorith the NLMS algorith s step-size is adjusted according to the instantaneous short-tie energy o the input signal. he adaptive iltering process using the NLMS algorith can be described by the ollowing set o equations w n w n e 0 n n x x n x n (3) w x nw n w n... w n 2 n xn xn... xn L n n n... n 2 L L L e n y n n n y n w n x n i w x μ i0 where w is a vector containing L coeicients o the adaptive ilter; x is a vector consisting o L saples o the input signal x(n); e(n) is a dierence between the reerence signal y(n) and the adaptive ilter output during the n th iteration; 0 is the ixed NLMS step-size paraeter ro the interval (0;). Fig. 8 illustrates the basic principle o the NLMS adaptive algorith. While the signal x(n) corresponding to the outgoing voice signal (that notation is used in the previous section) is reerred to as the ar-end signal the signal y(n) corresponding to the incoing voice signal is reerred to as the near-end signal. Basically the near-end signal y(n) is the delayed and attenuated version o the ar-end signal x(n). (4) Fig. 8. Principle o the noralized least ean squares algorith

15 98 Adaptive Filtering Applications 3.2 ie-doain adaptive algoriths Knowing the adaptive theory it is trivial that the delay estiation can be achieved by selecting the largest value ro the adaptive ilter weights vector w. here is only one issue that has to be taen into account. he adaptive ilter needs soe tie in order to converge to the optial perorance. he existing adaptive algoriths dier ro each other with dierent convergence properties and coputational eory requireents. he robust ast converging algoriths are priarily used in the acoustical echo cancellation applications. hey tae a lot o coputational resources. In our case it is not necessary to apply the coplex algoriths because the adaptive ilter is not directly used or the purpose o echo cancellation but or the delay estiation. hereore the reduced coplexity adaptive iltering algoriths becae the subject o our interest Proportionate adaptive iltering algoriths In the ollowing we provide a reader along with a principal o the proportionate adaptive technique. he proportionate noralized least ean squares (PNLMS) algorith proposed in (Duttweiler 2000) has been developed or use especially in the telephone networ environent. For hybrid echo cancellers it is reasonable to assue that the echo path has a sparse character (i.e. any IR s (Ipulse Response) coeicients are close to zero). Although there are studies and research on the ultiple relection echo paths a typical echo path ipulse response in the practical counication networs has only one relection which eans all the active coeicients are occupied in a continuous area o the whole echo span. hereore we do not need to update all the ilter weights to achieve a noinal perorance. One can adjust and operate only with non-zero active coeicients. Basically this is a ain idea behind the PNLMS algorith and other subsequently derived proportionate adaptive ilters (Hongyang & Doroslovaci 2006; Paleologu et al. 2006). Proportionate approaches achieve their higher convergence rate by using the act that the active part o networ echo path is usually uch saller (4-8s) copared to s o the whole echo path that has to be covered by the adaptive ilter. In case o voice transission over the pacet-switching networ these nubers ay be ore considerable. In the PNLMS algorith the adaptive step-size paraeters are assigned to all the ilter coeicients. hey are calculated ro the last estiate o the ilter weights in such a way that a larger coeicient receives a larger increent. As a result the convergence rate can be increased the act that the active taps are adjusted aster than non-active coeicients. hereore or the sparse IR the PNLMS algorith converges uch aster coparing to the NLMS. his eature is an advantage especially when it is necessary to estiate the long echo delays. he PNLMS algorith can be described using the ollowing equations: wn wn n e n n 0 n n n G x x G x (5) n diagg0n... g n L G (6) where G(n-) is a diagonal atrix adjusting the step-size paraeters 0 is an overall stepsize paraeter. he diagonal eleents o G(n) are estiated as ollows: n ax ax w n... w n w n l p 0 L l (7)

16 Perceptual Echo Control and Delay Estiation 99 l g n l L i0 n n 0 ll (8) i Paraeters p and ρ are positive nubers with typical values or p = 0.0 and or ρ = 5/L. he irst ter in (7) ρ prevents w l (n) ro stalling when it is uch saller than the largest coeicient and p regularizes the updating when all coeicients have zero values at initialization. In spite o the sparse syste identiication which is a vital requireent or the ast converging adaptive ilters there is another requireent. It is directly addressed to the adaptive ilter ipleentation. he algorith should have reasonable power concerns. Unortunately the PNLMS algorith has drawbacs. One o the is an increase in the coputational coplexity by 50% copared to the NLMS algorith. Furtherore the PNLMS algorith shows the slow convergence rate ater the ast initial start. It is because o the slow convergence rate dedicated to the sall coeicients (Gay 998). he increased coputational coplexity can be reduced by the way o selective partial-updating. In turn the slow convergence o the PNLMS in the stable state can be iproved by switching ro the PNLMS to NLMS equations ater the ast initial convergence has been achieved (Benesty & Gay 2002). Recently researchers proposed the partial-update techniques or using along with the adaptive algoriths. he partial-update algoriths reduce coputational coplexity by updating only a subset o the ilter weights per iteration. hey can be relevant or applications requiring ast real-tie processing. Unortunately there is another side to the coin. he ewer coeicients you are going to update per iteration there is ore isalignent presented in the algorith. hereore a certain trade-o should be ade when selecting the nuber o coeicients to update. he ollowing subsection presents several partial-update algoriths along with the proposed odiication o the partialupdate PNLMS algorith Partial-update adaptive iltering algoriths he partial-update algoriths can be seen to exploit the sparseness o the echo path in two dierent ways. It is nown that when the unnown syste s ipulse response is sparse any o the adaptive ilter s weights can be approxiated to zero. Alternatively the sparseness ay be present in the weight update vector as a consequence o the distribution o the input saples in the (Lx) input vector x n =[x(n) x(n-) x(n-l+)]. In both these cases exploiting the sparseness properties can reduce coplexity and iprove perorance o the adaptive algorith (Fevrier & Geland 999). Soe o the irst wor on the partial-update algoriths was done by Douglas (Douglas 997). It presents the periodic and the sequential updating schees or the Max-NLMS algorith. However these partial-update algoriths show slow convergence 2properties copared to the ull-update algoriths. he reason is inconsistent updating schees. More recently the partial-updating concept was developed by Aboulnasr (Aboulnasr 999). It leads to the M-Max NLMS algorith and supporting convergence analysis (Aboulnasr & Mayyas 998). Another bloc-updating schee or the NLMS algorith was studied by Schertler (Schertler 998). he latter wor was published by Dogancay and anriulu. hey consider approaches or ore robust Aine Projection Algorith (APA) (Dogancay & arinulu ). Further we give a suary o the general partial-update algoriths along with the proposed one.

17 00 Adaptive Filtering Applications M-Max NLMS he algorith selects a speciied nuber o the coeicients providing the largest reduction in the ean squared error per iteration (Naylor & Sherlier 2003). Only M out o the total L ilter coeicients are updated. hose M coeicients are the ones associated with the M largest values within the ollowing vector x(n - i + ) ; i = ;...; L. he update equations or this algorith are w n w i in e 0 n n i x x n x n i M axia o x ni w n 0 i L i 0 otherwise (9) (20) One o the eatures o the M-MAX-NLMS algorith is that it reduces the coplexity o the adaptive ilter by selectively updating the coeicients while aintaining the closest perorance to the ull-update NLMS algorith. We present isalignent curves or the algorith in the up-ollowing section. he isalignent value was calculated using the equation orula Misalignent n 0 log 0 L j j L j j j n h n w n Here w j is the j -th coeicient o the adaptive ilter while h j is the j -th coeicient o the siulated echo path s IR. Selective-partial-update NLMS his algorith opposed to the M-Max NLMS has a bloc structure. An objective behind the latter is the sae: it reduces coputational costs by updating a subset o the ilter coeicients. But irst the vector x(n) and the coeicient vector w(n) are arranged into K blocs o length M = L/K where L is an integer as in n n n n h 2 2 (2) x x x... x 2 K (22) n n n n w w w... w 2 K (23) he coeicient vector s blocs w (n) w 2 (n) w K (n) represent candidate subsets that can be updated during the current iteration. For a single-bloc updating schee the constrained iniization proble which is solved by the NLMS algorith can be written as w nw i in e 2 0 nx n i x i n (24) he selection o the bloc that has to be updated is ade by deterining the bloc with the sallest squared-euclidian-nor update. According to Eq. (24) that justiication can be described by the ollowing ters

18 Perceptual Echo Control and Delay Estiation i arg x n arg x n (25) in j ax j j M j M Generalization ro the single-bloc to the ultiple-bloc updating schee is done through the ollowing. Suppose that only the B (B<K) blocs with the largest agnitudes are selected to be updated. Let a vector I B = {i i 2 i B } denote the subset o B blocs out o { 2 K} to be updated. hus the equation or the B-bloc updating schee is w nw I I n B B e 2 0 nx n IB x I B n (26) where x IB and w IB are deined as ollows n n n... n x IB x x x i i2 ib (27) n n n... n w IB w w w i i2 ib (28) he coputational and eory requireents o the selective-partial-update NLMS algorith are alost identical to those o the selective-bloc-update algorith proposed in (Aboulnasr & Mayyas 998). Nevertheless siulation results illustrated in the next section shows that this approach does not lead to the reasonable trade-o between perorance and siplicity. he algorith s eiciency is weaer than the one o the M-Max NLMS algorith. As an alternative approach a sparse-partial-update NLMS algorith applies ore relevant selection criterion. Sparse-partial-update NLMS his algorith utilize a so-called sparse-partial (SP) weight selection criterion (Jinhong & Doroslovaci 2008). he adaptive ilter weights are updated based on the largest product o the ultiplication o x(n) and w(n). he SP-NLMS single-bloc update equations are given by w n w i in e 0 n n i x x n x n i M axia o x ni w n i w n 0 i L i 0 otherwise (29) (30) Hongyang and Dyba recently suggested a generalization or updating B blocs out o K i.e. w n w B Bn en n B x x n x n I I 0 I I I B i : is one o the B largest aong... n n n n I w x w x (32) B B i i i i Siple-partial-update PNLMS his proposed algorith exploits the sparseness o the counication channel to speed up the initial convergence and eploys the partially updating schee to reduce the B B (3)

19 02 Adaptive Filtering Applications coputational coplexity. A selection procedure is perored in accordance with the estiated agnitude o the channel s ipulse response. he S-PNLMS algorith or single -bloc update is deined as ollows. Arrange x(n) and w(n) into K blocs o length M = L/K in the sae way as it is done in Eq. (22 23). hen let G i (n) denote the corresponding MxM bloc o the diagonal weighing atrix G(n). he recursion or updating adaptive ilter weights is given by w n w n n e n n n n n G x x G x i i 0 i i i i i (33) where the bloc selection is done according to the ollowing It is dierent to in j j M n i arg G (34) n i x G x (35) arg in j j j j M which is used with the SPU-PNLMS algorith. It is apparently ro the siulations that the S-PNLMS has siilar perorance to the SP-NLMS and outperors the SPU-PNLMS algorith. he S-PNLMS or updating B blocs out o M has these update equations w n w n n e n n B B B B n n n G x x G x I I 0 I I I I I B B B (36) i: n is one o the Blargest aong n... n I G G G (37) B i M Further we provide the coparison results or the presented algoriths and deonstrate their perorance while estiating the predeined echo delay. able 5 illustrates the coputational coplexity o the ull-update algoriths and shows saving achieved by the partially updating schees. ALG. MUL. ADD. DIV. NLMS 3L+ 3L- M-Max-NLMS 3M+ 3M- SPU-NLMS 3M*B+ 3M*B - SP-NLMS 3M*B + 3M*B - PNLMS 6L+ 4L-2 L+ M-Max-PNLMS 6M + 4M -2 M + SPU-PNLMS 6M*B + 4M*B -2 M*B + SP-PNLMS 6M*B + 4M*B -2 M*B + S-PNLMS 6M*B + 4M*B -2 M*B + able 5. Coparison o coputational coplexity

20 Perceptual Echo Control and Delay Estiation Discussion over experiental results o evaluate the perorance o the adaptive iltering algoriths we ipleented the in MALAB environent. he woring scenario is as ollows. he ilters have to estiate the predeined echo path s ipulse responses speciied in the IU- Recoendation G.68 (IU- G ). hey consist o 024 taps. he overall step-size paraeter 0 is chosen to be 0.. he control paraeters ρ and p are chosen to be 0.00 and 0.0 respectively. In the irst part o the experient we loo at the isalignent curves o the M-Max- SPU- SPand S-PNLMS algoriths. hey are illustrated in Fig. 9 below. he SPU-updating schee produces the worst results. he proposed S-criterion considerably outperors it especially in ters o the initial convergence speed. he rest o the algoriths have nearly the sae convergence and tracing perorance. (a) Max-PNLMS algorith (b) SPU-PNLMS algorith

21 04 Adaptive Filtering Applications (c) SP-PNLMS algorith (d) S-PNLMS algorith Fig. 9. Misalignent curves o the partial-update algoriths All the algoriths except the M-Max-PNLMS show poor results when the M value equals 64. It can be explained by the act that the active part o the IR is approxiately 6s long. his value corresponds to 28 saples or sapling requency o 8Hz thereore 64 saples are not enough to cover the active region copletely. Regarding to the dissiilar selection criterion the M-Max-PNLMS algorith can deal relatively well with that proble. he Max-updating orula does not count with the sparse character o the IR. It perors selection according to the distribution o the values o the input vector. Otherwise its drawbac is lower initial convergence speed coparing to the SP-PNLMS algorith. he second part o our experient concerns the perorance o the adaptive algoriths versus the ones based on the generalized cross-correlation unction. hey are copared in the context o the tie delay estiation. he tie dierence between tie when the outgoing

22 Perceptual Echo Control and Delay Estiation 05 signal leaves the voice terinal and the oent when the incoing signal containing the echo o the original signal arrives bac ro the networ is reerred to as a true echo delay. able 6 concludes the results or the AF and GCC algoriths or dierent delay values. Fig. 0 shows the estiated echo delay diagra or the echo path that has a 60 s pure delay. It can be observed that the outputs ro the adaptive ilter in the steady state have uch saller variance than the results obtained ro the algorith based on the generalized cross-correlation unction. (a) DE using S-PNLMS algorith Fig. 0. Estiated echo delay curves (b) DE using S-PNLMS algorith

23 06 Adaptive Filtering Applications [s] ROH SCO CPS-2 HB M-Max SPU SP S saples able 6. Mean values o estiated echo delays 4. Partial proportionate and sparse control or ulti-delay ilters aing into the account the act that the generalized cross-correlation algoriths operate in the requency doain and use advantages o the ast Fourier transor reasonable coputational savings or adaptive iltering algoriths can be achieved as well. he basic operation underlying requency doain adaptive ilters is the transoration o the input signal into a ore desirable wave or beore its adaptive processing. his is accoplished by the Discrete Fourier ransor (DF) whereby the input signal is transored to the requency doain as shown in Fig.. Fig.. Frequency doain adaptive ilter coniguration Frequency Doain (FD) adaptive ilters have priarily two advantages copared to tie doain ipleentations (Dohnal 995). he irst advantage is the potentially large savings in the coputational coplexity. he Fast Fourier ransor (FF) is an eicient ipleentation o the DF which provides these savings. A second advantage is that the DF structures generate signals that are approxiately uncorrelated (orthogonal) (Shyn 992). As a result a tie-varying step-size paraeter is used or each adaptive weight thereby allowing a ore unior convergence rate across the adaptive ilter. hus the overall convergence rate o the algorith ay be iproved soeties approaching that achievable with Recursive Least Squares (RLS) algoriths without a siilar increase in the coputational coplexity (Widrow & Stearns 985). Many other attractive eatures and variation o the FD adaptive ilters can be ound in the literature (Khong et al. 2007).

24 Perceptual Echo Control and Delay Estiation Sparse partial-update proportionate ulti-delay ilter he ulti-delay requency doain adaptation algorith generates its individual step-size control actors that are utually orthogonal. his advantage put the Multi-Delay bloc Frequency doain (MDF) algorith closer to the analytical RLS algorith. It diers ro the algoriths processing in the tie doain with better convergence properties and lower coputational coplexity. he MDF algorith decreases the processing delay associated with requency doain adaptive ilters. he delay value is directly proportional to the nuber o the ilter coeicients. he ulti-delay approach solves this proble by partitioning the tie-doain ilter o length L into K sub-ilters each o length N with L=KN. Consequently the delay value is reduced by a actor o L/N copared to the ull length approach. he diagra o the MDF algorith is shown in Fig. 2. his algorith uses the FF o size N FF. It equals to the sallest power o two integers larger than or equal to 2L/K=2KN/K=2N. he MDF algorith wors as ollows. Let us irst deine as the rae index. he speech signals are operated on the rae-by-rae basis. he irst step o the MDF algorith is to convert the ost recent overlapped input saples to the requency doain as ollows X K diagff x x... x x x... x 0 N 0 N he input vectors incoing to each o the sub-ilter blocs can be obtained via the rae index shiting without invoing any coputation as ollows: (38) Fig. 2. Bloc diagra o the MDF algorith

25 08 Adaptive Filtering Applications X X... K (39) he described approach suggests that only one FF is needed per rae iteration in order to transor the input vector into the requency doain. It iplies a signiicant coputation saving. he output and the error requency-doain vectors can be expressed and calculated as K d last N ters o FF X W (40) E FF yd N zeros N ters where W () is the th coeicient vector and d() is the desired vector. he coeicient update equations to iniize the Mean Square Error (MSE) are the ollowing irst hal o FF S t MDF X E (42) MDF MDF (4) S S X X (43) Φ t N zeros FF (44) K W W Φ (45) where =2 K and μ is the bloc step-size control paraeter 0<<< is the orgetting actor. Further we exaine how the partial update technique ay be incorporated into the MDF algorith. For requency-doain selection the nown approach is to select requency bins corresponding to the largest agnitude Fourier transor o the tap-input over all the sub-ilter blocs =2 K (Khong et al2008). However the selection is better to peror on the coeicient-bloc basis (Deng & Dyba 2008). It is aster and taes less coputation power. hereore irst we need to calculate etrics or each o the sub-ilter in order to peror selection procedure. he nuber o etrics equals the nuber o sub-ilter blocs K. One particular requency-doain etric Q ay or exaple represent the su o the requency bins in the corresponding sub-ilter bloc 2 N Q X j (46) j here is a little bit dierent selection approach which shows better convergence perorance copared to the last one. Deine the 2L x vector () which consists o the concatenated Fourier transor o the input signal across all sub-ilters as ( ) X 0... X K... 2L (47) Each eleent in the 2L x 2L diagonal selection atrix P()

26 Perceptual Echo Control and Delay Estiation 09 l p i l corresponds to M ax ia o l L l 0 otherwise (48) he tap-selection routine is done by ultiplying this atrix with the corresponding input in every th sub-ilter as it is shown below P X X (49) Otherwise i it is supposed that the algorith odels a sparse networ ipulse response the sparse partial updating schee should be involved. he control etrics are calculated as ollows (=2 K) Q X W (50) Here we suggest using a new type o etric in order to choose sub-ilter blocs or updating. Note that it should be calculated in the tie-doain as ollows N L i i0 Q t (5) 2 j0 w L w We have to reeber that the iniu nuber M o the sub-ilter blocs to be updated ust equal or be greater than two (2 M<K). his is because the pure delay o the echo path is never nown. hus i only one sub-ilter is chosen it can not cover all the possibilities o the variation o echo path. Another iportant question or requency doain adaptive ilters is the FF length? Fro our point o view it is reasonable using the FF o 256 saples long. his is an optial length which can provide satisactory results or the high range o MDF algoriths. We have tried to apply dierent values but they did not bring suicient iproveents. For exaple i the echo path consists o 024 taps the adaptive ilter should be divided in 024/(256/2) sub-ilters. he nuber K o subilters will equal 8. Beore we present a proposed requency doain algorith let us resue the ollowing. he partial update requency-doain algoriths ay use the values o the calculated bloc etrics as selection criteria. Each type o etrics has its own eatures. In case o (3 8 9) only K (K<<2L) etrics are sorted in order to select M (2 M<K) active subilters. Since K is a sall nuber copared to the entire nuber 2L o subilters coeicients in ( ) the sorting overhead ay be reduced by actor 2L/N. We have spoen about the eatures o sparse echo paths and the possibility o iproving their convergence rate by assigning the individual step-size paraeters proportionally to the coeicient agnitudes. In the art there is a big aily o various MDF algoriths. Researchers are trying to iprove algoriths perorance ro one hand and to reach coputational savings ro the other hand. Further we present our approach how to deal with this contradictory proble. First we describe the basic idea o sparseness easure and show how to incorporate it into the MDF raewor. he concept o sparseness reers to a representational schee where only a ew units (out o a large assive) are eectively used to represent typical data vectors. In eect this iplies ost units taing values close j

27 0 Adaptive Filtering Applications to zero while only ew tae signiicantly non-zero values. Nuerous sparseness easures have been proposed and used in the literature to date (Benesty et al 2006). Such easures quantiy how uch energy o a vector is paced into only a ew coponents. On a noralized scale the sparsest possible vector (only a single coponent is non-zero) should have a sparseness o one whereas a vector with all eleents equal should have a sparseness o zero. In our approach we use s sparseness easure based on the relationship between the L nor and the L 2 nor: j w j L w 2 L i0 L j0 w j j L w 2 (52) Here L is the length o the adaptive ilter in the tie doain. his unction evaluates to unity i and only i w contains only a single non-zero coponent and taes a value o zero i and only o all coponents are equal (up to signs) interpolating soothly between the two extrees. We use this sparseness easure ξ o the tie doain estiated ilter coeicients to adaptively control the nuber o taps that need to be updated M L (53) t Note that the total nuber o corresponding requency doain coeicients is 2L or an L- long tie doain adaptive ilter thereore M 2M 2L (54) t I selection is perored on the bloc-based basis then the corresponding nuber o subilters in requency doain equals 2L M round roundk 2 N Fig. 3 shows sparseness easure or sparse and dispersive ipulse responses. It is also shown the nuber o adaptive ilter coeicients that should be updated i partial updating schee is considered. he convergence process by itsel can be divided into two stages i.e. beore and ater the convergence o the ipulse response (or soe tie beore and ater the convergence o the considerable/large coeicients). It has been showed in the tie doain adaptive iltering that it is better using PNLMS/IPNLMS at the initial stage or ast convergence and aterwards switching to NLMS or assuring better isalignent. Here or recognizing between two stages in adaptation process we propose to use the ollowing paraeter (55) 0 log 0 L j0 w j L (56)

28 Perceptual Echo Control and Delay Estiation Fig. 3. Sparseness easure and its ipaction on length o adaptive ilter he quantity in Eq. (56) represents an energy easure in [db] within an estiated ipulse response. Fig. 4 deonstrates the curve or this paraeter or two requency doain algoriths. It can be observed that ater 0.5 seconds estiation o IR energy easure or IPMDF stops luctuating. Consequently this act can be used or switching between dierent adaptation schees. Fig. 4. Estiated isalignent and energy easure or MDF and IPMDF algoriths

29 2 Adaptive Filtering Applications Beore representing the proposed MDF schee declare the ollowing stateents: I. Utilization o sparseness easure 0<ξ()< (or real IR) a. ξ()<0.7 i. IR is considered dispersive ii. M t = L i ully updated schee is chosen (ostly during initial period) iii. L/K M t < L i partial updated schee is chosen b. ξ() 0.7 i. IR is considered sparse ii. Mt = L.(- ξ()) II. Utilized updating scheas a. non-partial b. based selection (coeicient-based) 2L-vector c. μ - based selection (bloc-based) K-vector III. Utilization o IR energy easure η a. η() - η(-) Δ η switching to the MDF algorith b. η() - η(-) > Δ η switching to the IPMDF (SC-IPMDF) algorith Finally our proposed algorith can be described as: M w G st stage: η() - η(-) Δ η M i 0.7 2L or M K w L Gt φt diagg N gn... gn N w t g N l else N L else 0.7 2L L 2 L 2 j0 L 2 L w const i N L j N l const L w j L w 2 or M roundk i od 0 to M axia o l 0... i l corresponds l pl 0 otherwise X ~ P X φ irst hal o FF X ~ S t j 2L MDF E w w L G φ const t t

30 Perceptual Echo Control and Delay Estiation 3 0 od else otherwise K o M axia to corresponds i w L w L j j N i i L t t φ G w w 2 nd stage: η() - η(-) < Δ η & Counter η > η K M or L M i K round M or L M else od i otherwise L l axia o M to corresponds l i p l l P ~ X X S FF MDF t E X φ ~ o hal irst N t FF zeros φ Φ K Φ W W const 0 od else otherwise K M axiao to corresponds i w L w L j j N i i K Φ W W 4.2 Discussion over experiental results o copare the perorance o the proposed algorith with reerence MDF IPMDF SC- IPMDF adaptive iltering algoriths we ipleent all our adaptive ilters in MALAB environent. hese MDF-based ilters estiate the ipulse response o the predeined echo paths which are speciied by the IU- Recoendation (IU- G ). he tested ipulse responses are 024 taps long. All experients are perored using pre-recorded real speech signals. he nuber K o sub-ilters equals 8. he ollowing values or control paraeters are used: initial = 0. α initial = = 0-3 Δ η = 0.05 db η = 8. he

31 4 Adaptive Filtering Applications perorance o each algorith is studied using the noralized isalignent paraeter which can be estiated as ollows 2 hw MIS 0 log in [ db] 0 h 2 where h is a true ipulse response o length L. Another criterion is Echo Return Loss Enhanceent (ERLE) which is used in real-lie environent to evaluate perorance ERLE y d y 2 0 log [ db] 0 where y() is a desired signal (echo) and d() is adaptive ilter s output. Note that any reliable adaptive ilter with disabled residual echo suppressor has to achieve ERLE o -5dB within second ater starting convergence process (IU- G ). Fig illustrate an application o μ-based selection etric. his ind o etric is used or sub-ilter selection when estiated sparse easure paraeter ξ equals or larger than 0.7. his value was deined experientally during ultiple trials or nuerous types o echo path. We suggest using μ-based selection etric as an individual bloc step-size paraeter. It helps accelerating a speed o convergence by allocating larger step-size values or currently updated sub-ilters. I you loo at the diagra illustrated in Fig. 7 careully you will notice that the energy which is available or adaptation is concentrated around the sparse region o the echo path. hus this act can be used or selecting sub-ilters to be updated along with setting the step-size paraeter or these sub-ilters. When estiation o sparse easure ξ is saller than 0.7 we suggest switching to the based selection etric. 2 (57) (58) Fig. 5. Sparse ipulse response and estiated μ-based selection etrics

32 Perceptual Echo Control and Delay Estiation 5 Fig. 6. Dispersive ipulse response and estiated μ-based selection etrics Fig. 8 deonstrates the noralized isalignent and ERLE paraeters obtained or real speech signals. he proposed partially updated schee or MDF shows the siilar perorance coparing to the other three ully updated requency doain algoriths. During our uture wor we are going to enhance the above described algorith and propose a new class o partial sparse-controlled robust algoriths which will wor reliably even in double-tal situation. We will apply all the nowledge which were presented within this particular chapter. Further to conclude the chapter let us provide suary o aterial and ae several contributions according to the proposed algoriths. Fig. 7. Step-size paraeter estiated using μ-based selection etric

33 6 Adaptive Filtering Applications Fig. 8. Misalignent and ERLE curves 5. Conclusion he irst section outlines a basic principle o echo control in pacet-based networs. It explains why it is so iportant to provide onitoring during telephone conversations. When delivering the VoIP service in the pacet-switching networ it is iportant to have the value o the echo delay under control. he increasing transission delay associated with pacet data transission can ae a negligible echo ore annoying. hereore it is suggested using the echo assessent algorith. It is purpose is to add an additional attenuation to a particular voice channel (which in ters eans to activate an echo canceller) so as to reove the unwanted echo in tie. In the second section we consider an opportunity o using cross-correlation or estiating echo delays. hat section provides readers along with up-to-date correlation-based DE algoriths which we use to estiate the echo path delays. he proble o long delays taen place in the pacet-switching networ is considered as a topic o interest. he experients show that the algoriths precision decreases with increasing transission delays. he generalized cross-correlation algoriths operating in the requency doain provide ore reliable result coparing to the standard cross-correlation and noralized cross-correlation algoriths. As an alternative to correlation-based ethods techniques which use adaptive iltering algoriths can be also applied. hereore the third section presents nuerous partial-update algoriths and their application to delay estiation. he echo assessent is based on the reduced coplexity partial-update adaptive ilters. he experients show a reliable perorance o these algoriths. However their precision suers during the initial stage o convergence. According to the IU- Recoendation G.68 this period should not last ore than one second. he Multi-Delay bloc Frequency doain (MDF) adaptive algorith can easily

34 Perceptual Echo Control and Delay Estiation 7 outperor all existing tie doain algoriths. Moreover taing into the account the act that the generalized cross-correlation algoriths operate in the requency doain and use advantages o the ast Fourier transor urther coputational savings or the adaptive ilters are achieved in the requency doain. hereore the ourth section deals with partial proportionate and sparse-controlled adaptive iltering algoriths woring in the requency doain. What we claied within this section is: a new etric or peroring partial updating; a new approach or designating transitions between MDF and IPMDF-based updating scheas; a ethod or estiating step-size control paraeter; a new partially updated sparseness-controlled iproved proportionate ulti-delay ilter; all the approaches are suitable or ipleentation whether in tie or requency doains. he proposed algorith has both a perorance copared to the IPMDF and SC-MDF algoriths and reduced coputational coplexity along with the adjustable step-size paraeter. Although the preerred ebodients o the proposed algorith have been described it will be understood by those silled in the art that various changes ay be ade thereto without departing ro the ain scope o the invention or the appended clais. 6. Acnowledgent his wor was supported by the Grant Agency o the Czech echnical University in Prague grant No. SGS 0/275/OHK3/3/3 and by Grant he Ministry o Education Youth and Sports No. MSM Reerences Choi B.-K.; Moon S.; Zhi-Li Z. (2004) Analysis o Point-o-Point Pacet Delay In an Operational Networ Proceedings o INFOCOM 2004 wenty-third Annual Joint Conerence o the IEEE Coputer and Counications Societies vol. 3 pp ISBN Hong Kong March Gordy J.D.; Goubran R.A. (2006) On the Perceptual Perorance Liitations o Echo Cancellers in Wideband elephony IEEE ransactions on Audio Speech and Language Processing vol. 4 issue pp ISSN: Nisar K.; Hasbullah H.; Said A.M. (2009) Internet Call Delay on Peer to Peer and Phone to Phone VoIP Networ Proceedings o ICCE 09 International Conerence on Coputer Engineering and echnology vol. 2 pp ISBN Singapore Januar Dyba R.A. (2008) Parallel Structures or Fast Estiation o Echo Path Pure Delay and heir Applications to Sparse Echo Cancellers Proceedings o CISS nd Annual Conerence on Inoration Sciences and Systes pp ISBN Princeton NJ March Hongyang D.; Dyba R.A. (2008) Eicient Partial Update Algorith Based on Coeicient Bloc or Sparse Ipulse Response Identiication Proceedings o CISS nd Annual Conerence on Inoration Sciences and Systes pp ISBN Princeton NJ March Khong A.W.H.; Naylor P.A. (2006) Eicient Use O Sparse Adaptive Filters Proceedings o ACSSC 06 Fortieth Asiloar Conerence on Signals Systes and Coputers pp ISBN Paciic Grove CA October 29 -Noveber

35 8 Adaptive Filtering Applications Hongyang D.; Dyba R.A. (2009) Partial Update PNLMS Algorith or Networ Echo Cancellation Proceedings o ICASSP 2009 IEEE International Conerence on Acoustics Speech and Signal Processing pp ISBN aipei April IU- Recoendation G.3 (2003) aler Echo and its Control IU- Recoendation G.68 (2002) Digital networ echo cancellers Carter G.C. (976) ie Delay Estiation Ph.D. dissertation University o Connecticut Storrs C pp Mueller M. (975) Signal Delay IEEE ransactions on Counications vol. 23 issue pp Buchner H.; Benesty J.; Gansler.; Kellerann W. (2006) Robust Extended Multidelay Filter and Double-tal Detector or Acoustic Echo Cancellation IEEE ransactions on Audio Speech and Language Processing vol. 4 issue 5 pp ISSN Youn D.H.; Ahed N.; Carter G.C. (983) On the Roth and SCOH Algoriths: ie- Doain Ipleentations In Proceedings o the IEEE vol. 7 issue 4 pp ISSN Zetterberg V.; Pettersson M.I.; Claesson I. (2005) Coparison Between Whitened Generalized Cross-correlation and Adaptive Filter or ie Delay Estiation Proceedings o MS/IEEE OCEANS vol. 3 ISBN Washington D.C. Septeber Hertz D. (986) ie Delay Estiation by Cobining Eicient Algoriths and Generalized Cross-correlation Methods IEEE ransactions on Acoustics Speech and Signal Processing vol. 34 issue pp.-7 ISSN Knapp C.; Carter G.C. (976) he Generalized Correlation Method or Estiation o ie Delay IEEE ransactions on Acoustics Speech and Signal Processing vol. 24 issue 4 pp ISSN Wilson K.W.; Darrell. (2006) Learning a Precedence Eect-Lie Weighting Function or the Generalized Cross-Correlation Fraewor IEEE ransactions on Audio Speech and Language Processing vol. 4 issue 6 pp ISSN ianshuang Q.; Hongyu W. (996) An Ecart-weighted adaptive tie delay estiation ethod IEEE ransactions on Signal Processing vol. 44 issue 9 pp ISSN X Chen J.; Benesty J.; Huang Y.A. (2006) he SCO Weighted Adaptive ie Delay Estiation Algorith Based on Miniu Dispersion Criterion Journal o EURASIP on Applied Signal Processing vol ISBN Anderson M.P.; Woessner W.W. (992) Applied Groundwater Modeling: Siulation o Flow and Advective ransport Acadeic Press (2nd Edition ed.) ISBN USA Widrow B. (2005) hining about thining: the discovery o the LMS algorith IEEE Magazine on Signal Processing vol. 22 issue pp ISSN Hayin S. (200) Adaptive Filter heory Fourth Edition Prentice-Hall ISBN USA Zetterberg V.; Pettersson M.I.; Claesson I. (2005) Coparison Between Whitened Generalized Cross-correlation and Adaptive Filter or ie Delay Estiation Proceedings o MS/IEEE OCEANS vol. 3 ISBN Washington D.C. Septeber

36 Perceptual Echo Control and Delay Estiation 9 Eadzadeh A.A.; Lopes C.G.; Speyer J.L. (2008) Online tie delay estiation o pulsar signals or relative navigation using adaptive ilters Proceedings o 2008 IEEE/ION Position Location and Navigation Syposiu pp ISBN Monterey CA May Duttweiler D.L. (2000) Proportionate noralized least-ean-squares adaptation in echo cancellers IEEE ransactions on Speech and Audio Processing vol. 8 issue 5 pp ISSN Hongyang D.; Doroslovaci M. (2006) Proportionate adaptive algoriths or networ echo cancellation IEEE ransactions on Signal Processing vol. 54 issue 5 pp ISSN: X Paleologu C.; Benesty J.; Ciochina S. (200) An iproved proportionate NLMS algorith based on the l0 nor Proceedings o ICASSP '0 200 IEEE International Conerence on Acoustics Speech and Signal Processing pp ISSN Dallas X March Gay S.L. (998) An eicient ast converging adaptive ilter or networ echo cancellation Proceedings o the hirty-second Asiloar Conerence on Signals Systes & Coputers vol. pp ISBN Paciic Grove CA Noveber Benesty J.; Gay S.L. (2002) An iproved PNLMS algorith Proceedings o ICASSP ' IEEE International Conerence on Acoustics Speech and Signal Processing vol. 2 pp ISSN Minneapolis MN USA April Fevrier I.J.; Geland S.B.; Fitz M.P. (999) Reduced coplexity decision eedbac equalization or ultipath channels with large delay spreads IEEE ransactions on Counications vol. 47 issue 6 pp ISSN Douglas S.C. (997) Adaptive ilters eploying partial updates IEEE ransactions on Circuits and Systes II vol. 44 issue 3 pp ISSN Aboulnasr.; Mayyas K. (999) Coplexity reduction o the NLMS algorith via selective coeicient update IEEE ransactions on Signal Processing vol. 47 issue 5 pp ISSN X Aboulnasr.; Mayyas K. (998) MSE analysis o the M-Max NLMS adaptive algorith Proceedings o IEEE International Conerence on Acoustics Speech and Signal Processing vol. 3 article ID 0.09/ICASSP pp ISBN Seattle WA May Schertler (998) Selective bloc update o NLMS type algoriths In Proceedings o IEEE International Conerence on Acoustics Speech and Signal Processing vol. 3 pp ISBN Seattle WA May Dogancay K.; anriulu O. (200) Adaptive iltering algoriths with selective partial updates IEEE ransactions on Circuits and Systes II vol. 48 issue 8 pp ISSN Naylor P.A.; Sherlier W. (2003) A short-sort M-Max NLMS partial-update adaptive ilter with applications to echo cancellation Proceedings o ICASSP ' IEEE International Conerence on Acoustics Speech and Signal Processing vol. 5 pp ISBN Hong Kong April Jinhong W.; Doroslovaci M. (2008) Partial update NLMS algorith or sparse syste identiication with switching between coeicient-based and input-based selection Proceedings o CISS nd Annual Conerence on Inoration Sciences and Systes vol.3 pp ISBN Princeton NJ March

37 20 Adaptive Filtering Applications Shyn J. J. (992) Frequency-Doain and Multirate Adaptive Filtering IEEE ransactions on Signal Processing Magazine vol. 9 pp ISSN: Widrow B.; Stearns S.D. (985) Adaptive Signal Processing Prentice-Hall ISBN USA Khong A. W. H.; Naylor P. A.; Benesty J. (2007) A low delay and ast converging iproved proportionate algorith or sparse syste identiication EURASIP Journal on Audio Speech and Music Processing vol Article ID pages Khong A. W. H.; Xiang L.; Doroslovaci M.; Naylor P. A. (2008) Frequency doain selective tap adaptive algoriths or sparse syste identiication Proceedings o ICASSP 2008 International Conerence on Acoustics Speech and Signal processing pp ISBN Las Vegas NV March 3-April Deng H.; Dyba R. (2008) Eicient Partial Update Algorith Based on Coeicient Bloc or Sparse Ipulse Response Identiication Proceedings o CISS 2008 Conerence on Inoration Sciences and Systes pp ISBN Princeton NJ March Benesty J.; Huang Y. A.; Chen J.; Naylor P. A. (2006) Adaptive algoriths or the identiication o sparse ipulse responses In: Selected Methods or the Cancellation o Acoustical Echoes the Reduction o Bacground Noise and Speech Processing Hänsler E.; Schidt G. Springer pp ISBN Berlin

38 Adaptive Filtering Applications Edited by Dr Lino Garcia ISBN Hard cover 400 pages Publisher Inech Published online 24 June 20 Published in print edition June 20 Adaptive iltering is useul in any application where the signals or the odeled syste vary over tie. he coniguration o the syste and in particular the position where the adaptive processor is placed generate dierent areas or application ields such as: prediction syste identiication and odeling equalization cancellation o intererence etc. which are very iportant in any disciplines such as control systes counications signal processing acoustics voice sound and iage etc. he boo consists o noise and echo cancellation edical applications counications systes and others hardly joined by their heterogeneity. Each application is a case study with rigor that shows weaness/strength o the ethod used assesses its suitability and suggests new ors and areas o use. he probles are becoing increasingly coplex and applications ust be adapted to solve the. he adaptive ilters have proven to be useul in these environents o ultiple input/output variant-tie behaviors and long and coplex transer unctions eectively but undaentally they still have to evolve. his boo is a deonstration o this and a sall illustration o everything that is to coe. How to reerence In order to correctly reerence this scholarly wor eel ree to copy and paste the ollowing: Kirill Sahnov Eaterina Verteletsaya and Boris Sia (20). Perceptual Echo Control and Delay Estiation Adaptive Filtering Applications Dr Lino Garcia (Ed.) ISBN: Inech Available ro: Inech Europe University Capus SeP Ri Slava Krautzea 83/A 5000 Rijea Croatia Phone: +385 (5) Fax: +385 (5) Inech China Unit 405 Oice Bloc Hotel Equatorial Shanghai No.65 Yan An Road (West) Shanghai China Phone: Fax: