Estimation Of The Fundamental Frequency Of The Speech signal Compressed By G.729 Algorithm Using PCC Interpolation

Transcription

1 (IJIRSE) Interntionl Journl o Innovtive Reserch in Science & Engineering ISSN (Online) 7-07 Estimtion O The Fundmentl Frequency O The Seech signl Comressed By G.79 Algorithm Using PCC Interoltion Zorn Milivojević *, Dro Brodić Dertment o Communiction, College o Alied Technicl Sciences, A. Medvedev No. 0, Niš, Serbi Dertment o Inormtics, University o Belgrde, Technicl Fculty Bor, Vojse Jugoslvije, Bor, Serbi * Correonding uthor E-mil: zorn.milivojevic@vtsnis.edu.rs Abstrct This er gives the estimtion o the undmentl requency (F 0 ) o the seech signl modeled by the G.79 method. The estimtion o the F 0 ws erormed by the Peing-Pes lgorithm with the imlemented Prmetric Cubic Convolution (PCC) interoltion. The eiciency o PCC ws tested or Keys, Greville nd Greville two-rmetric ernel. The window tht gives otiml results ws chosen ccording to the MSE results. At the end, the comrtive nlysis o the F 0 estimtion comressed by G.79 nd G.7. codec is mde. Keywords- Seech rocessing Seech comression Fundmentl requency VoIP I. INTRODUCTION ITU-T Telecommuniction stndrdiztion sector o ITU (Interntionl Telecommuniction Union) hs deined the stndrd H. or seciiction o comonents, rotocols nd rocedures which me multimedil communiction services in networs with cets communiction ossible []. Besides, H. deines udio nd video codecs (coders-decoders). G.79 nd G.7. re udio codecs intensively used in VoIP (Voice over IP) communictions [,]. G.7. rovides the seech comression to. b/s nd 6. b/s. In the irst cse ACELP (Algebric Code Excited Liner Prediction) lgorithm nd in the second MP-MLQ (Multiulse MAximum Lielihood Quntiztion) lgorithm re lied [,]. The rme size is 0 ms with 0 smles with smling o 8 Hz. Ech rme is being rocessed in the ollowing 7. ms. The rme is devided into smller rmes whose durtion is 7. ms. Delying (67-97 ms), delying vritions, cets losses nd echo inluence on the qulity o seech trsmitted by mens o cet networ. Evlution o MOS (Men Oinion Score) test is ccetble. (. bs) nd good.8 (6. bs) []. Recommendtion ITU -T Rec. G.79 contins the descrition o n lgorithm or the coding o seech signls t 8 bit/s using Conjugte-Structure Algebric-Code-Excited Liner- Prediction (CS -ACELP) []. The CS-ACELP coder is bsed on the Code-Excited Liner-Prediction (CELP) coding model. The coder oertes on seech rmes o 0 ms corresonding to 80 smles t smling rte o 8 Hz. For every 0 ms rme, the seech signl is nlyzed to extrct the rmeters o the CELP model (linerrediction ilter coeicients, dtive nd ixed-codeboo indices nd gins). These rmeters re encoded nd trnsmitted [6]. At the decoder, these rmeters re used to retrieve the excittion nd synthesis ilter rmeters. The seech is reconstructed by iltering this excittion through the short-term synthesis ilter. The short-term synthesis ilter is bsed on 0th order Liner Prediction (LP) ilter. The long-term, or itch synthesis ilter is imlemented using the so-clled dtive-codeboo roch. Ater comuting the reconstructed seech, it is urther enhnced by ostilter. This coder encodes seech nd other udio signls with 0 ms rmes. In ddition, there is loo-hed o ms, resulting in totl lgorithmic dely o ms. Evlution o MOS test is good.8 (6. bs) [7]. With G.7. nd G.79 coding the F0 estimtion is irstly being done, nd thn the estimted F0 is used in the ollowing stes o the seech signl coding. In the decoding rocess the seech signl is reconstructed which imlies renewl o the F 0 [8,9]. Ater the trnser o the seech signl by the VoIP service, on the recetion side there is need or the signl rocessing (seech nd seer recognition, echo cnceling, imrovement o the qulity nd the seech rticulteness etc.). The chrcteristic exmle is the correction o the seech signl qulity by reducing o dissonnt requencies [0,]. In the seech signl rocessing there is need or determintion o the F0. A number o lgorithms were develoed or determintion o the F 0 where the nlysis is erormed in the time nd requency domin [-]. The requently lied method or determintion o the F0 is bsed on the eing es o the mlitude chrcteristic in the seciic requency rnge. This method is used or nlyzing o the signl vlues in the sectrum on requencies on which the Discrete Fourier Trnsorm (DFT) ws clculted. Most oten the rel vlue o the F0 is not there on the requencies where DFT is clculted, but lies between the two sectrum smles. Tht cuses the requency estimtion error tht lies in the intervl [-(FS/(N) Hz, (FS/(N) Hz], where F S is the smling requency nd N is the DFT window size. One wy o reducing the error is determintion o the interoltion unction nd estimtion o the sectrum chrcteristics in n intervl between the two smles.

2 (IJIRSE) Interntionl Journl o Innovtive Reserch in Science & Engineering ISSN (Online) 7-07 This rocedure gives the reconstruction o the sectrum on the bse o DFT. The sectrum rmeters re then determined by nlytic rocedures (dierentition, integrtion, extreme vlues,...). Determintion o sectrl chrcteristic vlues between DFT smles is ten s the locl convolution o the neighboring DFT smles nd interoltion ernel (iece-wise olynomil interoltion) [6-8]. Two methods re requently used or interoltion: ) the cubic B-sline interoltion [6] nd b) the Prmetric Cubic Convolution (PCC) interoltion [7]. From the oint o view o the stness o the mximum osition estimtion the liction o PCC interoltion ernel is more suitble, becuse it is ossible to ind the mximum osition directly (by mens o ormul) using the smled dt without convolution being lied. The detiled nlysis o the F0 estimtion by mens o PCC interoltion is described in the er [9]. The results o the liction o PCC interoltion or determining o the F0 in the conditions o liction o some window in the rocessing o the discrete seech signl re resented in [0]. Through some simultion rocedures lgorithm eiciency nlyses hve been done where, s qulity mesure o n lgorithm, the Men Squre Error (MSE) hs been used. The best results were shown by the lgorithm with the imlemented Blcmn window. The results o the F0 estimtion o the seech signl modeled by SYMPES (Systemtic Procedure to Model Seech Signls vi Predeined Enveloe nd Signture Sequences ) method re shown in []. The ssessment o the F0 in MP codec is given in []. Further, in [] the estimtion o the F0 in seech signl comressed by G.79 lgorithm is described. In ddition, the uthor in [] estblished the ormule or nlyticl clcultion o the F0 without convolution. It is mde by multile seed-u rocedure or the estimtion o F 0. Further in this er there will be resented some results o the F0 estimtion in udio nd seech signls comressed by G.79 lgorithm in order to determine the F0 using PCC interoltion ernels (Keys, Greville nd two rmetric Greville ernel) nd some window unctions (Hmming, Hnning, Blcmn, Rectngulr, Kiser nd Tringulr window). The estimtion o F0 ws erormed on the bse o the nlyticl exression rom [9] or Keys ernel nd rom [] or Greville nd two-rmetric Greville ernels. As mesure o the qulity o interoltion lgorithm MSE (Men Squre Error) will be lied. On the bse o minimum vlues o MSE otimum ernel rmeters nd the corresonding window unction will be determined. Aterwrd the comrison o the estimtion o F 0 between G.79 nd G.7. [] codecs will be imlemented. This er is orgnized s ollows: In Section II there is descrition o the PCC lgorithm. In Section II.A there re deinitions o interoltion ernels. In Section II.B the lgorithm or determintion o otiml ernel rmeters is resented. In Section III numericl MSE results in the estimtion o F 0 o the seech signl modeled by the G.79 method re resented. The comrtive nlysis o the results nd the choice o the otiml ernel nd window unction re shown in Section IV. Section V gives the Conclusion. II. ALGORITHM OF FUNDAMENTAL FREQUENCY ESTIMATION Algorithm or the estimtion o F 0, bsed on the lgorithm rom [9], is shown in Fig.. This lgorithm is relized in the ollowing stes: Ste : Audio or seech signl s(n) is coded by G.79 coder. Ste : Coded signl is decoded by G.79 decoder nd generises signl x(n). Ste : Window w(n) whose length is N lies to decoded signl x(n). Ste : Sectrum X() is clculted by using DFT: DFT xn X The sectrum is clculted in discrete oints =0,...,N-, where N is the length o DFT. The rel sectrum o signls x(n) is continuous, wheres DFT deines the vlues o the sectrum t some discrete oints. Ste : By using e icing lgorithm, the osition o the mximum o the rel sectrum tht is between - th nd (+)-th smles is determined, where the vlues X() nd X(+) re the highest in the seciied domin. Ste 6: The osition o the mximum o the sectrum is clculted by PCC interoltion. The reconstructed unction is: L X r i, r i i L where i =X(i), r() is the ernel o interoltion nd L the number o smles tht rticite in interoltion. Ste 7: By dierentition X r () nd zero djustment the osition o mximum is determined it resents the estimted F0, which is mred s e.

3 (IJIRSE) Interntionl Journl o Innovtive Reserch in Science & Engineering ISSN (Online) 7-07 The qulity o the lgorithm or the estimtion o F0 cn be lso exressed by MSE: e MSE where is true F0 nd e is the estimtion o F 0. Figure. Algorithm or the estimtion o F 0. A. Interoltion Kernel Next, we give deinitions o the interoltion ernels which re tested in this er: ) Keys interoltion ernel [7,8]: otherwise 0,, 8,, r For L= rom Eq. () osition o mximum is determined: 0, b - b - 0, mx c b c where: c b b) Greville interoltion ernel [6]:

4 (IJIRSE) Interntionl Journl o Innovtive Reserch in Science & Engineering ISSN (Online) i 0, i 9, i, 0 i r For L=6 rom Eqs. () nd (7) osition o mximum is determined ccording to Eq. (), where: c b c) Greville two-rmetric cubic convolution ernel (GP) [6]:.. 0, 0 9, 6 0 9, 0 r For L=8 rom Eqs. () nd (9) osition o mximum is determined ccording to Eq. (), where: c b In the Eqs. ()-(0) there re nd rmeters. The otiml vlues o these rmeters will be determined by the miniml vlue o MSE, or Keys, Greville nd GP ernel. For the irst two o them: MSE ot min rg nd or the GP ernel:

5 (IJIRSE) Interntionl Journl o Innovtive Reserch in Science & Engineering ISSN (Online) 7-07 rg minmse ot, ot, The detiled nlysis in [9-] showed tht the miniml vlue o MSE deends on the liction o window by which signl rocessing x(n) is crried out in time domin. MSE will be deined or: ) Hmming, b) Hnning, c) Blcmn, d) Rectngulr, e) Kiser nd ) Tringulr window. B. Interoltion Kernel Prmeters Algorithm or determintion o interoltion ernel rmeters nd is relized in the ollowing stes: Ste : signl x(n), which ws reviously coded nd decoded by G.79 lgorithm, is modiied by the window unction w(n) whose length is N, Ste : sectrum X() is determined by the liction o DFT, Ste : reconstruction o the continul unction tht reresents sectrum X() is erormed by the liction o PCC interoltion, Ste : MSE is clculted or vrious vlues o rmeters nd deending on the imlemented window, Ste : ot nd ot re determined or which miniml vlue o MSE is obtined. C. Test Signls PCC lgorithm o the estimtion o F 0 will be lied to: i) simultion Sine test signl, nd ii) Seech test signl. Simultion Sine signl or testing o PCC lgorithm is deined in [9]: st K M i g0 s i sini o g t i NM where 0 is undmentl requency, i nd i re hse nd mlitude o the i-th hrmonic, resectively, K is the number o hrmonics, M is the number o oints between the two smles in sectrum where PCC interoltion is being mde. The Seech test signl is obtined by recording o seer in the rel coustic mbient. For urther comrtive nlysis by interoltion it is suitble or the F 0 o the Seech test signl nd the one o the Sine test signl to be equl. PCC lgorithm will be lied to: ) uncoded simultion Sine nd Seech test signls, nd b) by G.79 lgorithm coded nd decoded Sine nd Seech test signls. III. EXPERIMENTAL RESULTS AND COMPARISON A. Testing Prmeters In the simultion rocess 0 nd i re rndom vribles with uniorm distribution in the rnge [G (97.99 Hz), G (78.99 Hz)] nd [0,]. Signl requency o smling is S =8 Hz nd the length o window is N=6, which ssures the nlysis o subsequences tht lst ms. The results resented urther in this er relte to O =- 0.6Hz (requencies between the eighth nd ninth DFT comonents). Number o requencies in the seciied rnge or which the estimtion is done is M=00. The Sine test signl is with K=0 hrmonics. All urther nlyses will relte to ) Hmming, b) Hnning, c) Blcmn, d) Rectngulr, e) Kiser nd ) Tringulr window. B. Exerimentl Results ) Keys Kernel Alying n lgorithm or determintion o rmeters o Keys interoltion ernel some digrms re drwn MSE(), the minimum vlue MSE Kmin determined, nd on the bse o it the otimum vlue o Keys ernel ot determined or: ) Hmming (Fig. ), b) Hnn, c) Blcmn, d) Rectngulr, e) Kiser, nd ) Tringulr window unctions. Vlues MSE Kmin nd ot re resented in Tble I (uncoded Sine test signl MSE Kmin, G.79 coded Sine test signl MSE KGmin ) nd Tble II (Seech test signl MSE KSPmin, G.79 coded Sine test signl MSE KGSPmin ). For the comrison, the results o the G.7. in Tbles I - II re given s well. These results re obtined under identicl circumstnces s in [].

6 (IJIRSE) Interntionl Journl o Innovtive Reserch in Science & Engineering ISSN (Online) 7-07 () (c) (d) Figure. MSE() or Keys ernel nd Hmming window: ) uncomressed Sine test signl, b) G.79 comressed Sine test signl, c) uncomressed Seech test signl, d) G.79 comressed Seech test signl. TABLE I. MINIMUM MSE AND OPT FOR SINE TEST SIGNAL (KEYS KERNEL). Window Uncoded signl Signl coded by G.7. lgorithm Signl coded by G.79 lgorithm ot MSE Kmin ot MSE KGmin ot MSE KGmin Hmming Hnning Blcmn Rectngulr Kiser Tringulr TABLE II. MINIMUM MSE AND OPT FOR SPEECH TEST SIGNAL (KEYS KERNEL). Window Uncoded signl Signl coded by G.7. lgorithm Signl coded by G.79 lgorithm ot MSE KSPmin ot MSE KGSPmin ot MSE KGSPmin Hmming Hnning Blcmn Rectngulr Kiser Tringulr On the bse o the results resented in Tbles I - II it is obvious tht: ) t Sine test signl the gretest recision o F 0 estimtion is when Blcmn window (MSE Kmin =0.000) is lied. At G.7. coded Sine test signl the gretest recision o estimtion is in Rectngulr (MSE KGmin =0.8) window. When G.7. coding is lied, the recision in F 0 estimtion is lowered or MSE KGmin /MSE Kmin =0.8/0.00=9. times. At G.79 coded Sine test signl the gretest recision o estimtion is in Rectngulr ( MSE KGmin =0.8) window. When G.79 coding is lied, the recision in F 0 estimtion is lowered or MSE KGmin /MSE Kmin = 0.8/0.00=8.6 times.

7 (IJIRSE) Interntionl Journl o Innovtive Reserch in Science & Engineering ISSN (Online) 7-07 b) t Seech test signl the gretest recision is in Tringulr window (MSE KSPmin =0.077). At G.7. coded Seech signl the gretest recision is in Rectngulr window (MSE KGSPmin =0.67). When coding is lied, the recision in F 0 estimtion is lowered or MSE KGSPmin /MSE KSPmin =0.67/0.077=.7 times. At G.79 coded Seech test signl the gretest recision o estimtion is in Rectngulr (MSE KGmin =0.) window. When G.79 coding is lied, the recision in F 0 estimtion is lowered or MSE KGmin /MSE Kmin = 0./0.077=.707 times. c) t G.7. coded Seech signl in reltion to coded Sine signl inccurcy in F 0 estimtion is greter or MSE KGSPmin /MSE KGmin =0.67/0.8=.7 times. At G.79 coded Seech signl in reltion to coded Sine signl inccurcy in F 0 estimtion is greter or MSE KGSPmin /MSE KGmin =0./0.8=.07 times. ) Greville Kernel Alying lgorithm or determintion o Greville interoltion ernel some digrms MSE() re drwn, minimum vlue MSE Gmin determined, nd on the bse o it otimum vlue o Greville ernel rmeters ot determined or: ) Hmming (Fig. ), b) Hnn, c) Blcmn, d) Rectngulr, e) Kiser, nd ) Tringulr window. Vlues MSE min nd ot re resented in Tble III (uncoded Sine test signl MSE Gmin, coded Sine test signl MSE GGmin ) nd Tble IV (Seech test signl MSE GSPmin, coded Seech test signl MSE GGSPmin ). For the comrison, the results o the G.7. in Tbles III - IV re given s well. These results re obtined under identicl circumstnces s in []. () (c) (d) Figure. MSE() or Greville ernel nd Hmming window: ) uncomressed Sine test signl, b) G.79 comressed Sine test signl, c) uncomressed Seech test signl, d) G.79 comressed Seech test signl. TABLE III. MINIMUM MSE AND OPT FOR SINE TEST SIGNAL (GREVILLE KERNEL). Window Uncoded signl Signl coded by G.7. lgorithm Signl coded by G.79 lgorithm ot MSE Gmin ot MSE GGmin ot MSE GGmin Hmming Hnning Blcmn Rectngulr Kiser Tringulr TABLE IV. MINIMUM MSE AND OPT FOR SPEECH TEST SIGNAL (GREVILLE KERNEL).

8 Window (IJIRSE) Interntionl Journl o Innovtive Reserch in Science & Engineering ISSN (Online) 7-07 Uncoded signl Signl coded by G.7. lgorithm Signl coded by G.79 lgorithm ot MSE GSPmin ot MSE GGSPmin ot MSE GGSPmin Hmming Hnning Blcmn Rectngulr Kiser Tringulr On the bse o the results resented in Tbles III - IV it is obvious tht: ) t Sine test signl the gretest recision o F 0 estimtion is when Blcmn (MSE Gmin =0.000) window is lied. At G.7. coded Sine test signl the gretest recision o estimtion is in Rectngulr window (MSE GGmin =0.7). When coding is lied, the recision o F 0 estimtion is lowered or MSE GGmin /MSE Gmin =0.7/0.000=. times. At G.79 coded Sine test signl the gretest recision o estimtion is in Tringulr window (MSE GGmin =0.96). When coding is lied, the recision o F 0 estimtion is lowered or MSE GGmin /MSE Gmin =0.96/0.000= times. b) t Seech test signl the gretest recision is in Kiser window (MSE GSPmin =0.0). At G.7. coded Seech signl the gretest recision is in Rectngulr window (MSE GGSPmin =0.78). When G.7. coding is lied, the recision o F 0 estimtion is lowered or MSE GGSPmin / MSE GSPmin =0.78/0.0=0. times. At G.79 coded Seech signl the gretest recision is in Blcmn window (MSE GGSPmin =0.0). When G.79 coding is lied, the recision o F 0 estimtion is lowered or MSE GGSPmin / MSE GSPmin =0.0/0.0=8.00 times. c) t G.7. coded Seech signl in reltion to coded Sine signl inccurcy o the F 0 is greter or MSE GGSPmin /MSE GGmin =0.78/0.7=.8 times. At G.79 coded Seech signl in reltion to coded Sine signl inccurcy o the F 0 is greter or MSE GGSPmin /MSE GGmin =0.0/0.96=.0 times. ) GP Kernel For window (Blcmn, Tringulr, Kiser nd Blcmn) which showed the best results t G.79 coding with Greville ernel, n nlysis ws erormed by mens o GP ernel. Three-dimensionl MSE(,) grhics re drwn (Figs..,., 6., nd 7.), the shit o minimum MSE min in (,) level determined, nd ot nd ot vlues determined nd resented in Tble V. Furthermore, it gives the results or the G.7. codec rom []. In Figs.b,.b, 6.b, nd 7.b the ositions o MSE min =MSE( ot, ot ) minimum in (,) lne or Greville (oint A) nd GP (oint B) interoltion ernel, re shown. Vector AB shows the osition chnge o minimum (MSE( ot, ot )). () Figure. Sine test signl without comression with the liction o Blcmn window: ) MSE(α,β) or GP PCC interoltion, b) ositions o min (MSE(α ot,β ot )) in lne (αβ) or Greville (oint A) nd GP PCC (oint B) interoltion.

9 (IJIRSE) Interntionl Journl o Innovtive Reserch in Science & Engineering ISSN (Online) 7-07 () Figure. Sine test signl with G.79 comression with the liction o Tringulr window: ) MSE(α,β) or the liction o GP PCC interoltion, b) Positions o min (MSE(α ot,β ot )) in lne (αβ) or Greville (oint A) nd GP PCC (oint B) interoltion. () Figure 6. Seech test signl without comression with the liction o Kiser window: ) MSE(α,β) or GP PCC interoltion, b) ositions o min (MSE(α ot,β ot )) in lne (αβ) or Greville (oint A) nd GP PCC (oint B) interoltion. () Figure 7. Seech test signl with G.79 comression with the liction o Blcmn window: ) MSE(α,β) or the liction o GP PCC interoltion, b) Positions o min (MSE(α ot,β ot )) in lne (αβ) or Greville (oint A) nd GP PCC (oint B) interoltion. TABLE V. MINIMUM MSE, OPT AND OPT (GP KERNEL). ot ot MSE GPmin Sine test signl Uncoded signl Blcmn win MSE GPmin=0.000 G.7. coded Rectng. win MSE GPGmin=0.70 G.79 coded Tring. win MSE GPGmin=0.676 Seech test signl Uncoded signl Kiser win MSE GPSPmin=0.098 G.7. coded Rectng. win MSE GPGSPmin=0.898 G.79 coded Blcmn. win MSE GPGSPmin=0. On the bse o the results resented in Tble. V it is obvious tht:

10 (IJIRSE) Interntionl Journl o Innovtive Reserch in Science & Engineering ISSN (Online) 7-07 ) t Sine test signl the gretest recision o F 0 estimtion (Greville interoltion ernel) is when Blcmn (MSE Gmin =0.000) window is lied. When GP interoltion (MSE GPmin =0.000) is lied, the recision o F 0 estimtion is incresed or MSE Gmin /MSE GPmin =0. 000/0.000=.7 times. b) t G.7. coded Sine test signl the gretest recision o estimtion (Greville interoltion ernel) is in Rectngulr window (MSE GG =0.7). When GP interoltion (MSE GPmin =0.7) is lied, the recision o F 0 estimtion is incresed or MSE GGmin /MSE GPmin =0. 7/0.7=.87 times. c) t G.79 coded Sine test signl the gretest recision o estimtion (Greville interoltion ernel) is in Tringulr window (MSE GG =0.96). When GP interoltion (MSE GPmin =0.676) is lied, the recision o F 0 estimtion is incresed or MSE GGmin /MSE GPmin =0. 96/0.676=.7 times. d) t Seech test signl the gretest recision o the F 0 estimtion (Greville interoltion ernel) is when Kiser (MSE GSPmin =0.0) window is lied. When GP interoltion (MSE GPSPmin =0.098) is lied, the recision o F 0 estimtion is incresed or MSE GSPmin / MSE GPSPmin =0.0/0. 098=.8 times. e) t G.7. coded Seech test signl the gretest recision o estimtion (Greville interoltion ernel) is in Rectngulr window (MSE GGSPmin =0.78). When GP interoltion (MSE GPSPmin =0.898) is lied, the recision o F 0 estimtion is incresed or MSE GGSPmin /MSE GPSPmin =0.78/0.898=.78 times. ) t G.79 coded Seech test signl the gretest recision o estimtion (Greville interoltion ernel) is in Blcmn window (MSE GGSPmin =0.0). When GP interoltion (MSE GPGSPmin =0.) is lied, the recision o F 0 estimtion is incresed or MSE GGSPmin /MSE GPGSPmin =0.0/0.=.78 times. g) t G.7. coded Seech signl in reltion to G.7. coded Sine signl, the inccurcy o F 0 estimtion is greter or MSE GPGSPmin /MSE GPGmin =0.898/0.7=.66 times. h) t G.79 coded Seech signl in reltion to G.79 coded Sine signl, the inccurcy o F 0 estimtion is greter or MSE GPGSPmin /MSE GPGmin =0.676/0.=.6 times. IV. COMPARATIVE ANALYSIS Comrtive recision nlysis o the estimted F 0 o the Sine test signl nd the Seech test signl, without nd with G.79 comression will be erormed on the bse o the miniml vlues o MSE. The miniml vlue o MSE is determined on the bse o digrm in the Fig. (Keys), Fig. (Greville) nd Figs. - 7 (GP) nd resented in the Tble I (MSE Kmin, MSE KGmin ), Tble II (MSE KSPmin, MSE KGSPmin ), Tble III (MSE Gmin, MSE GGmin ), Tble IV (MSE GSPmin, MSE GGSPmin ) nd Tble V (MSE GPmin, MSE GPGmin, MSE GPSPmin, MSE GPGSPmin ) resectively. Furthermore the comrtive nlysis o the F 0 estimtion with G.79 nd G.7. lgorithms will be mde. Comring the vlues MSE min rom Tbles. I - V it cn be concluded tht: ) the otimum choice or Sine test signl (without comression) is Blcmn window or ll interoltion ernels. GP interoltion ernel, which genertes 66.9% less thn Keys nd.6% less thn Greville ernel, showed the best results. b) the otimum choice or Seech test signl is GP ernel with Kiser window, which genertes 9.7% less thn Keys ernel (Tringulr window) nd 6.08% less thn Greville ernel (Kiser window). c) the otimum choice or Sine test signl coded by G.7. lgorithm is Rectngulr window or ll interoltion ernels. GP interoltion ernel, which genertes.6% less thn Keys nd 6.% less thn Greville ernel, showed the best results. d) the otimum choice or Seech test signl coded by G.7. lgorithm is rectngulr window or ll interoltion ernels. GP interoltion ernel, which genertes 7.% less thn Keys nd.0% less thn Greville ernel, showed the best results. e) the otimum choice or Sine test signl coded by G.79 lgorithm is GP interoltion ernel with Tringulr window, which genertes.% less thn Keys (Rectngulr window) nd.7% less thn Greville ernel (Tringulr window), showed the best results ) the otimum choice or Seech test signl coded by G.79 lgorithm is GP interoltion ernel with Blcmn window, which genertes 6.66% less thn Keys (Rectngulr window) nd.9% less thn Greville ernel (Blcmn window), showed the best results. g) comring MSE or GP ernel or uncoded Seech test signl (Kiser window, MSE GPSPmin =0.098), G.7. coded Seech signl (Rectngulr window, MSE GPGSPmin =0.898) nd G.79 coded Seech signl (Blcmn window, MSE GPGSPmin =0.), in reltion to MSE GPGSPmin / MSE GPSPmin =0.898/0.098=.66 nd MSE GPGSPmin / MSE GPSPmin =0./0.098=.78 hs been obtined. It is cler tht the ccurcy o the G.79 lgorithm is considerbly higher comred to the G.7. lgorithm. Precisely, it is higher or.66/.78 =. times.

11 (IJIRSE) Interntionl Journl o Innovtive Reserch in Science & Engineering ISSN (Online) 7-07 In ccordnce to the derived conclusion, the liction o the lgorithm or urther rocessing o G.7. coded signl with lgorithms bsed on the estimted F 0 (utomtic veriiction o seer, recognition o the Seech etc.) would not bring stisctory results [], while rocessing o G.79 coded signl cuses signiicntly smller error. The obtined results recommend using PCC lgorithm with GP ernel in rerocessing signls comressed by G.79 method or urther rocessing by lgorithms which require recise determintion o the F 0. V. CONCLUSION This er resented the nlysis o the F 0 estimtion results o the Seech signl comressed by G.79 lgorithm which hs been intensively used in VoIP services. The estimtion o the F 0 hs been mde by Peing Pes lgorithm with imlemented PCC interoltion. Exeriments hve been erormed with Keys, Greville nd Greville two-rmetric GP ernels. In order to minimize MSE some windows hve been imlemented. The detiled nlysis hs shown tht the otiml choice is Greville two-rmetric ernel nd the Blcmn window imlemented in PCC lgorithm. In reltion to Keys nd Greville ernels, Greville two-rmetric ernel genertes 6.66% nd.9% less MSE, resectively. Comring the obtined results to the results o the estimtion o the F 0 in the Seech signl tht is not modeled by G.79 method, reltion o miniml MSEs.78 hs been obtined. The obtined results recommend using o PCC lgorithm with GP ernel in rerocessing o signls comressed by G.79 method or urther rocessing by lgorithms which require recise determintion o the F 0 (utomtic veriiction o the seer, recognition o the seech, etc.). REFERENCES [] ITU-T Rec. H., Audiovisul nd Multimedi Systems - Terminl or low bit-rte multimedi communiction, 009. [] ITU-T Rec. G.79, Coding o Seech t 8 Kbit/s Using Conjugte-Structure Algebric-Code-Excited Liner-Prediction (CS -ACELP) 996. [] ITU-T Rec. G.7., Dul-rte Seech Coder or Multimedi Communictions Trnsmitting t. nd 6. bit/s, 996. [] S. Krntzis nd F. Pvlidiou, VoIP: A comrehensive survey on romising technology, Comuter Networs, vol., , 009. [] V. Reguer, F. Pliz, W. Godoy nd E. Rernndez, On the imct o ctive queue mngement on VoIP qulity o service, Comuter Communictions, vol.,. 7 87, 008. [6] Texs Instruments, Imlementtion o G.79 on the TMS0Cx, Aliction Reort, SPRA66, Mrch 000 [7] P. Thigo nd N. Sergio, Eicient Serch in the Adtive Codeboo or ITU-T G.79 Codec, IEEE Signl Processing Letters, VOL. 6, NO. 0, , October 009. [8] G. Kosr nd T. Sheltmi, Voice trnsmission over d hoc networ dting otimum roches to mximize the erormnce, Comuter Communictions, vol.,. 6 69, 009. [9] A. Zolghdr, E. Gooddrzi nd M. Moosvinezhd, Rel-time imlementtion o G.7. seech coder, Irnin Journl o Inormtion Science nd Technology, vol., No.,. -, 00. [0] B. Joen, S. Kng, S. Be nd K. Sung, Filtering o Dissonnt Frequency Bsed on Imroved Fundmentl Frequency Estimtion or Seech Enhncement, IEICE Trns. Fundmentls, vol. E86-A, No. 8, , August 00. [] Z. Milivojevic nd D. Blnesovic, Enhncement o the ercetive qulity o the noisy seech signl by using o DFF-FBC lgorithm, Fct Universittis, Ser.: Elec. Energ. vol., no.,. 79-9, December 009. [] L. Qiu, H. Yng nd S. Koh, Fundmentl requency determintion on instntneous requency estimtion, IEEE Signl Process., vol.,. -, 99. [] A. Cheveigne nd H. Kwhr, YIN, undmentl requency estimtor or seech nd music, J. Acoust. Soc. Am., Vol., No., , Aril 00. [] T. Murmi nd Y. Ishid, Fundmentl requency estimtion o seech signl using MUSIC lgorithm, Acoust. Sci. & Tech. vol.,. 9-97, 00. [] A. Kch, F. Grenez nd K. Benmhmmed, Time requency nlysis nd instntneous requency estimtion using two-sided liner rediction, IEEE Signl Processing, vol. 8,. 9-0, 00. [6] R. Keys, Cubic convolution interoltion or digitl imge rocessing, IEEE Trns. Acoust., Seech & Signl Process, vol. 9,. - 60, 98. [7] K. Pr nd R. Schowengerdt, Imge reconstruction by rmetric cubic convolution, Comuter Vision, Grhics & Imge Processing, vol.,. 8-7, 98. [8] S. Reichenbch, Two-Dimensionl Cubic Convolution, IEEE Trnsctions on imge rocessing, Vol., No. 8, , August 00. [9] H. Png, S. Be nd K. Sung, Imroved Fundmentl Frequency Estimtion Using Prmetric Cubic Convolution, IEICE Trns. Fundmentls. vol. E8-A, , 000. [0] Z. Milivojevic, M. Mirovic nd S. Milivojevic, An Estimte o Fundmentl Frequency Using PCC Interoltion Comrtive Anlysis, Inormtion technology nd Control, vol.,. -6, 006. [] Z. Milivojevic nd M. Mirovic, Estimtion o the undmentl requency o the seech signl modeled by the SYMPES method, Int. J. Electron. Commun. (AEÜ), vol. 6, , 009. [] Z. Milivojevic, M. Mirovic nd S. Milivojevic, Fundmentl Frequency Estimtion o The Seech Signl Comressed by MP Algorithm Using PCC Interoltion, Advnces in Electricl nd Comuter Engineering Journl, Vol. 0, no.,. 8-, 00. [] Z. Milivojevic nd D. Brodic, Estimtion o the Fundmentl Frequency o the Seech Signl Comressed by G.7. Algorithm Alying PCC Interoltion, Journl o ELECTRICAL ENGINEERING, Vol. 6, No., 0, 8 89.