REAL-TIME SPEECH COMPRESSION BY USING CODE EXCITED LINEAR PREDICTION ALGORITHM

Transcription

1 62 Intenational Jounal on Intelligent Electonic Systems, Vol.3, No.1, Januay 2009 Abstact REAL-TIME SPEECH COMPRESSION BY USING CODE EXCITED LINEAR PREDICTION ALGORITHM S.Pabu 1, S.Nandakuma 2 1 School of Computing Sciences, VIT Univesity, Velloe, Tamilnadu 2 School of Electical Sciences, VIT Univesity, Velloe, Tamilnadu E- Mail: 1 spabu@vit.ac.in, 2 snandakuma@vit.ac.in A lot of effot has been spent ove the last few yeas in the development of digital speech coding methods and thei subsequent standadization. Algoithms have evolved which povide good quality speech at sub 8 kbps bit ates although at a much computational expense. Speech compession is poposed based on code excited linea pediction algoithm and implementation in DSP algoithm. Algoithm based on thee-stage technique which involves simulate, evaluate, debug and implementation in G.723 low delay code excited linea pediction (LD-CELP)[4] algoithm. Fist stage, the algoithm is evaluated via simulation to detemine whethe it meets the design citeion. Then, it is implemented in eal-time based an object oiented appoach. Afte the algoithm is thooughly tested, it is futhe efined to obtain tighte and faste coding. This technique can be applied to othe eal-time DSP algoithms. A simulation esult shows that bette speech quality is obtained. The techniques descibed in this pape ae applicable to any othe speech codec. Key wods: : Low delay Code excited Linea Pediction (LD-CELP) algoithm, DSP, Speech Compession, G Standad I. INTRODUCTION A Selection of a Low Bit Rate Vocode. The thee most impotant appoaches ae wavefom coding, tansfom coding, and paametic coding. Wavefom Coding basically does compession of sample at the tansmitting end and de-compession of sample at the eceiving end. Paametic Coding elies on speech chaacteistics. Tansfom Coding as the name indicates compesses speech by employing a tansfomation technique. Nowadays, most successful Vocode and hybid codes make use of linea pediction in ode to estimate the needed paametes. Indeed, LPC in[5] (Linea Pediction Coding) is the most successful method fo encoding at low bit ate and is used in many applications. The compaison chat Table 1 summaizes the pefomance of vaious algoithms. As can be seen G.723.1[1] tops out as one of the bette algoithms offeing communication quality speech at a elatively low bit ate. Mean Opinion Scoe (MOS) is a subjective measue of the pefomance of the algoithm. An intoduction to LPC and CELP [4] is given as these ae assumed fo G (ACELP/MP-MLQ) Table 1. Compaison of Speech Coding AlgoithmB. CELP Standad Coding Type Bit Rate(kbps) MOS Algo. Delay(ms) ITU-G.711 PCM ITU-G.721 ADPCM ITU-G ITU-G.726 ACELP/ MP- MLQ 6.3, VBR- 2.0,3.2,4, 16,24,32,40 ADPCM ITU-G.728 LD-CELP Standad Coding Type Bit Rate(kbps) MOS Algo. Delay(ms) ITU-G.711 PCM ITU-G.721 ADPCM ITU-G ITU-G.726 ACELP/ MP- MLQ 6.3, VBR- 2.0,3.2,4, 16,24,32,40 ADPCM ITU-G.728 LD-CELP CELP stands fo Code Excited Linea Pediction and stats fom basic LPC coding. It is the most commonly used code in telephony. Thee ae many ways in which the esidual and the LP coefficients ae used: CELP [9] is the exteme in tems of complexity. The LP paametes ae estimated as befoe, and used to fom the synthesis filte. Howeve, the synthetic speech is obtained fom a synthetic esidual that diffes fom LPC: it is obtained fom a codebook. The esidual is theefoe vecto quantized and chosen fom a pe computed set of excitations. Only the index of the excitation used fom the codebook has to be sent to the eceive. Vecto quantization exploits the high coelation between the paametes and is applied to both the coefficients and the excitation. The system looks like Fig 1. G i C(n) + + E(n) Code book + + Gain âz -M Pitch Fig. 1. Block Diagam of CELP? az -k Synthesis filte

2 Sidha Pattanaik et al : Possible use of Smat Antennas fo Development of In CELP, the pitch is not encoded in the codebook: a longtem pedicto is used instead to educe the complexity of the codebook. Finding the appopiate LP coefficients and excitation equies a seach though the codebooks. Each enty is evaluated by calculating a peceptual eo: the best one is chosen. The incease in efficiency and quality is big, but the complexity is high: seaching the codebook can be long, and equies a lot of computations. II. OBJECTIVE The objective of this pape is to design achitectue fo a low bit ate codec G ecommendation (ACELP/MP-MLQ) [2]. The desciption of the speech coding algoithm is made in tems of bit-exact, fixed point mathematical opeations. A. Simulation of G Encode This code has two bit ates associated with it. These ae 5.3 and 6.3 kbit/s. The highe bit ate has geate quality. The lowe bit ate gives good quality and povides system designes with additional flexibility. Both ates ae a mandatoy pat of the encode and decode.this code was optimized to epesent speech with a high quality at the above ates using a limited amount of complexity. Music and othe audio signals ae not epesented as faithfully as speech, but can be compessed and decompessed using this code. This code encodes speech o othe audio signals in 30 msec fames. In addition, thee is a look ahead of 7.5 msec, esulting in a total algoithmic delay of 37.5 msec. All additional delays in the implementation and opeation of this code ae due to: I) Actual time spent pocessing the data in the encode and decode ii) Tansmission time on the communication link iii) Additional buffeing delay fo the multiplexing potocol. The desciption of the speech-coding algoithm of this Codec is made in tems of bit-exact, fixed-point mathematical opeations B Encode pinciple The code is based on the pinciples of linea pediction analysis-by-synthesis coding and attempts to minimize a peceptually weighted eo signal. The encode opeates on blocks (fames) of 240 samples each. That is equal to 30 msec at an 8 khz sampling ate. Each block is fist high pass filteed to emove the DC component and then divided into fou sub fames of 60 samples each. Fo evey subfame, a 10th ode Linea Pediction Code (LPC) [5,7] filte is computed using the unpocessed input signal. The LPC filte fo the last subfame is quantized using a Pedictive Split Vecto Quantize (PSVQ). The unquantized LPC coefficients ae used to constuct the shot-tem peceptual weighting filte, which is used to filte the entie fame and to obtain the peceptually weighted speech signal. Fame Hig h Pass Filte LPC Fo mat Pec eptu al Wei ghti ng Pitch Est. LSP Qua ntize Imp ulse Res pons e Calc Simulated Decode LSP LSP Dec Intep ode olato Mem oy Upda te Zeo Input Respons e Pitch Decode Ha mon ic Nois Pitch e - Pedi Sha p g cto _ + Excitatio n Decode MP- MLQ/ ACEL P Fig. 2. Block diagam of the Speech Encode Fo evey two subfames (120 samples), the open loop pitch peiod, LOL, is computed using the weighted speech signal. This pitch estimation is pefomed on blocks of 120 samples. The pitch peiod is seached in the ange fom 18 to 142 samples. Fom this point the speech is pocessed on a 60 samples pe subfame basis. Using the estimated pitch peiod computed peviously, a hamonic noise shaping filte is constucted. The combination of the LPC synthesis filte, the fomant peceptual weighting filte, and the hamonic noise shaping filte is used to ceate an impulse esponse. The impulse esponse is then used fo futhe computations. Using the pitch peiod estimation, LOL, and the impulse esponse, a closed loop pitch pedicto is computed. A fifth ode pitch pedicto is used. The pitch peiod is computed

3 64 Intenational Jounal on Intelligent Electonic Systems, Vol.3, No.1, Januay 2009 as a small diffeential value aound the open loop pitch estimate. The contibution of the pitch pedicto is then subtacted fom the initial taget vecto. Both the pitch peiod and the diffeential value ae tansmitted to the decode. Finally the non-peiodic component of the excitation is appoximated. Fo the high bit ate, Multipulse Maximum Likelihood Quantization (MP-MLQ) excitation is used, and fo the low bit ate, an algebaiccode-excitation (ACELP) is used. III. DECODER The G speech decode is divided into the following modules: Decode, which includes the initialization outines as well as the decode Line Spectum Analysis, which includes LSP decode and intepolation Linea Pedictive Analysis, which includes LPC synthesis and Fomant post filte Adaptive and Fixed Excitation, which includes decoding of pitch infomation, Excitation decode, pitch post filte, and fame intepolation Miscellaneous utility functions, which include gain scaling. A. Decode pinciples The decode opeation is also pefomed on a fame-byfame basis. Fist the quantized LPC indices ae decoded, then the decode constucts the LPC synthesis filte. Fo evey subfame, both the adaptive codebook excitation and fixed codebook excitation ae decoded and input to the synthesis filte. The adaptive post filte consists of a fomant and a fowad-backwad pitch post filte. The excitation signal is input to the pitch post filte, which in tun is input to the synthesis filte whose output is input to the fomant post filte. A gain scaling unit maintains the enegy at the input level of the fomant post filte. B. Vecto Quantization This section is to intoduce vecto quantization employed in gain and LSP quantization. Befoe intoducing vecto quantization, lets look at what scala quantization is. In Scala Quantization one epesents the values by fixed subset of epesentative values. Fo example, if we have 16 bit values and we send only 8 most signifcant bits, we get an appoximation of the oiginal data at the expense of pecision. In this case the fixed subset is all the 16-bit numbes divisible by 256, i.e 0, 256, 512,...Vecto Quantization (VQ) is a genealization of scala quantization to highe dimensions. This genealization opens up a wide ange of possibilities and techniques not pesent in the scala case. Unlike scala quantization, VQ is usually applied to signals that have aleady been digitized. It is pimaily used fo data compession and patten ecognition. In VQ, an input patten o wod is matched to a set of stoed pattens o wods, and the best match is chosen. The index of the 'best match' can then be tansmitted, theeby educing the amount of data that needs to be tansmitted. LSP Dec ode LSP Inte polat o Pitc Pitc Synt Fo h + h hesi man Dec Post s t ode filte Filte Post filte Exci tatio n Dec ode Gai n Scal ling Fig. 3. Block diagam g speech decode Mathematically, a vecto quantize, Q of dimension k and size N is defined as a mapping fom a vecto in k- dimensional Euclidean space, Rk, into a finite set of vectos, o codewod. That is: Q : Rk C [1] Whee C = (y1,y2,...,yn) is the codebook with codewod yi??rk fo i??{1, 2,...,N}. Vecto Quantization equies both an encode and decode. The encode E maps the input Rk into the index set, I: E : Rk I [2] The encode thus outputs an index to the codewod that offes lowest distotion. In this case the lowest distotion is found by evaluating the Euclidean distance between the input vecto and each codewod in the codebook. Once the

4 Sidha Pattanaik et al : Possible use of Smat Antennas fo Development of closest codewod is found, the index of that codewod is sent though a channel. The decode D simply maps the index set I into the epoduction set C: D : I C [3] The decode is a simple table lookup.it does not need to know anything about the patition cells of Rk. A Voonoi o neaest neighbo vecto quantize is a special class of vecto quantize in which the patition is completely detemined by the codebook and a distotion measue. The neaest neighbo vecto quantize is, in fact, the most common type of vecto quantize in pactice. d(x,y)= k ( xi - yi )2 [4] i=1 The most common distotion measue used in neaest neighbo vecto quantizes is mean squae eo, which is defined by the Euclidean distance between vectos. When lage data sets ae involved, especially when consecutive points ae coelated in some way then this method finds its use. Speech compession schemes like CELP use this scheme to quantize the excitation vectos. IV. SIMULATION OF G DECODER The G does no seaching, so is computationally much less complexity than the code. The decoding pocess basically deals with extaction of the paametes using the codebook indices and econstuction of the speech signal. This chapte explains the complete pocess of G decode and its simulation using MATLAB. A. Decode oveview Fist the fame paametes LP coefficients, adaptivecodebook vecto, fixed-codebook vecto and gain ae decoded. Then these decoded paametes ae used to econstuct the speech signal. This econstucted speech signal is enhanced by a post pocessing pocess. The LP decoding pocess is done pe each fame wheeas the following steps ae epeated fo each sub fame decoding of the adaptive-codebook vecto decoding of the fixed-codebook vecto decoding of the adaptive and fixed-codebook gains Computation of the econstucted speech. B. Decoding of LP coefficients The eceived indices L0, L1, L2, L3 and the code books lspcb1, lspcb2 ae used to econstuct the quantized LSP coefficients. The sum of the value of lspcb1 codebook at index L1 and the value of lspcb2 codebook at the index L2 fo the fist five coefficients and at L3 fo the next five coefficients ae obtained. These coefficients ae eaanged fo a minimum distance of J. The eaangement pocess is done twice. Fist with a value of J = , then with a value of J = Afte this eaangement pocess, the quantized LSF coefficients fo the cuent fame ae obtained using the 4th ode MA pediction filte as explained in the encode. The selection of MA pediction filte of the available two filtes is done based on the value of L0. Afte computing the LSF coefficients (ùi), the coesponding filte is checked fo stability based on the following conditions 1. Aange the coefficient ùi in inceasing value. 2. if ùi < then ùi = if ùi+1 - ùi then ùi+1 = ùi , fo I = 1,,9 4. if ù10 > then ù10 = The LSP coefficients ae the cosine of the LSF coefficients. C. Decoding of adaptive and fixed codebook gains The eceived gain-codebook indices GA and GB along with the codebooks ae used to decode the gains. Then the estimated fixed codebook gain is obtained using a 4th ode MA pediction filte. The excitation u(n) is computed using v(n), c(n), gp and gc using the following equation U(n) = gp * v(n) + gc * c(n) [5] D. Computing the econstucted speech The excitation u(n) is passed though a 10th ode synthesis filte whose coefficients ae given by the LP coefficients a. The output fom this LP filte is the econstucted speech. E. Post pocessing Post-pocessing consists of thee functions namely adaptive post-filteing, high-pass filteing and signal upscaling. The adaptive post-filte is the cascade of thee filtes: a long-tem postfilte Hp(z), a shot-tem postfilte Hf (z) and a tilt compensation filte Ht(z), followed by an adaptive gain contol pocedue. The postfilte coefficients ae updated evey 5 ms subfame. The postfilteing pocess is oganized as follows. Fist, the econstucted speech is invese filteed to poduce the esidual signal ^(n). This signal is used to compute the delay T and gain gt of the long-tem postfilte Hp (z). The signal ^(n) is then filteed though the long-tem postfilte Hp (z) and the synthesis filte1/[gf Â(z /gd)]. Finally, the output signal of the synthesis filte 1/[gf Â(z /gd)] is passed though the tilt compensation filte Ht (z) to geneate the post-filteed econstucted speech signal sf(n). Adaptive gain contol is then applied to sf(n) to match the enegy of s^(n).

5 66 Intenational Jounal on Intelligent Electonic Systems, Vol.3, No.1, Januay 2009 F. High-pass filteing and up scaling LPpa = Rnn: [11x1 double] The output fom the adaptive gain contol unit sf'(n) is Win: [180x1 double] applied to a high-pass filte with a cut-off fequency of 100 Hz. This is followed by upscaling by a facto of two to LagWin: [11x1 double] compensate fo the down scaling pefomed in the encode ECWin: [11x1 double] to pevent fixed point oveflow. Thus the completely pocessed speech signal is obtained. Fom the simulation FStat: 60 esults we found that the output fom the decode is as WStat: [ ] simila as the oiginal speech and nomal heaing cannot pickup any distubances. SFRef: 4 V. RESULTS AND PERFORMANCE ANALYSIS LMem: 120 ENCODER PERFORMANCE LSFpa = ECWin: [11x1 double] The esults show excitation, esidual and the codebook vectos that fom the excitation. As we aleady know, an appoximation of esidual that is used to econstuct the speech is excitation. This excitation is composed of two pats adaptive codebook vecto and fixed codebook vecto. The above diagam shows that, adaptive codebook vecto ties to follow the esidual in a slow manne wheeas the fixed codebook vecto has the ability to spot sudden changes. In essence, the wavefom shown above combined with LP paametes aleady coded signifies the opeation of G encode. name =MALE.WAV Waning: Function call modenc invokes inexact match C:\MATLAB7\wok\mail send with encode and dec\modenc.m. Matlab simulation of speech code Stat the encoding pocess Mean: [10x1 double] Pcof: VQ: {[3x256 double] [3x256 double] [4x256 double]} Fix: [1x1 stuct] IntC: [ ] lsfq: [10x1 double] TVpa = LSubfame: [ ] PWpa: [1x1 stuct] POLSubfame: [ ] PitchOLpa: [1x1 stuct] HNWpa: [1x1 stuct] ente the input file--->male.wav SineDetpa =c: [ ] ente the output file--->>> x.dat NTh: 14 > In path at 115 cth: In addpath at 95 In G7231Code at 9 PMin: [18 18] Pitchpa = PMode: [4x1 double] > In SetCodePa at 132 PMax: [ ] In G7231Code at 25 LOffs: {[-1 0 1] [ ]} In MODENC at 15 CBookTh: 58 b: {[5x85 double] [5x170 double]} HPFilt = b: [1-1] CL: [ ] a: [ ] Tamepa: [1x1 stuct] Mem: [] POffs: [-2 2] emem: [146x1 double]

6 Sidha Pattanaik et al : Possible use of Smat Antennas fo Development of Fame: 17 MPpa = g: [24x1 double] Gid: {{1x2 cell} {1x2 cell} {1x2 cell} {1x2 cell}} Np: [ ] gioffs: [-2 1] LTh: 58 ModV: [ ] CL: [ ] nck: [31x7 double] Clippa = MinTh: MinVal: -1 MaxTh: MaxVal: Fame: 18 Fame: 19 Fame: 20 Fame: 21 Fame: 22 Fame: 23 Fame: 24 Fame: 25 Fame: 26 G723.1 data file: C:\MATLAB7\wok\mail send with encode and dec\x.dat Elapsed time is seconds. WAVE file: C:\MATLAB7\wok\mail send with encode and dec\male.wav Numbe of samples : (2.746 s) Sampling fequency: 8012 Numbe of channels: 1 (16-bit intege) Fame: 1 Fame: 2 Fame: 3 Fame: 4 Fame: 5 Fame: 6 Fame: 7 Fame: 8 Fame: 9 Fame: 10 Fame: 11 Fame: 12 Fame: 13 Fame: 14 Fame: 15 Fame: 16 A. Decode Pefomance Fig. 4. 1Encode Pefomance The decode pat econstucts the same excitation fom the adaptive codebook vecto and fixed codebook vecto. This is fist synthesized to fom a econstucted speech. This is used to get the esidual. Post filte opeates upon this esidual. As can be seen thee is not much diffeence in the

7 68 Intenational Jounal on Intelligent Electonic Systems, Vol.3, No.1, Januay 2009 esiduals of both encode and decode stage. CL: [ ] Mat lab simulation of speech code nck: [31x7 double] Stat the decoding pocess Clippa = ente the input file--->waning: Function call moddec MinTh: invokes inexact match C:\MATLAB7\wok\encode and dec\moddec.m. MinVal: -1 inputfile=.dat,outputfile=.wav MaxTh: ente the input file--->>> x.dat MaxVal: ente the output file--->m1.wav Fame: 1 > In path at 115 Fame: 2 In addpath at 95 Fame: 3 In G7231Decode at 8 Fame: 4 In MODDEC at 6 Fame: 5 In MODENC at 12 Fame: 6 LSFpa = ECWin: [11x1 double] Fame: 7 Mean: [10x1 double] Fame: 8 Pcof: Fame: 9 VQ: {[3x256 double] [3x256 double] [4x256 double]} Fame: 10 Fix: [1x1 stuct] Fame: 11 IntC: [ ] Fame: 12 lsfq: [10x1 double] Fame: 13 Pitch pa = PMode: [4x1 double] Fame: 14 PMin: [18 18] Fame: 15 PMax: [ ] Fame: 16 LOffs: {[-1 0 1] [ ]} Fame: 17 CBookTh: 58 Fame: 18 b: {[5x85 double] [5x170 double]} Fame: 19 CL: [ ] Fame: 20 POffs: [-2 2] Fame: 21 emem: [149x1 double] Fame: 22 MPpa = g: [24x1 double] Fame: 23 Gid: {{1x2 cell} {1x2 cell} {1x2 cell} {1x2 cell}} Fame: 24 Np: [ ] Fame: 25 gioffs: [-2 1] Elapsed time is seconds. LTh: 58 ModV: [ ]

8 Sidha Pattanaik et al : Possible use of Smat Antennas fo Development of Fig. 5. Decode Pefomance VI. CONCLUSION Achitectue fo G encode and decode has been designed and it's pefomance fo vaious input speech signals ae veified using MATLAB simulation. The veification was done using 16-bit linea PCM samples fom a wav file and the output witten to a wav file. The input speech signal is compessed efficiently by the encode and it is econstucted to the fom of oiginal speech by the decode. The econstucted speech esembles the oiginal input speech signal. The esults obtained indicate the best pefomance of G achitectue designed fo speech compession. REFERENCES [2] ITU-T Recommendation G.723.1, "Dual Rate Speech Code fo multimedia Communications Tansmitting at 5.3and 6.3 kbitls," Ma [3] Implementation of G on the TMS320C54x Application epot fom TI Spa656 Mach [4] Implementation of G on the TMS320C54x Application epot fom TI Spa656 Mach 2000 [5]. J. Wang and J. D. Gibson, Pefomance compaison of intafame and intefame LSF quantization in packet netwoks, Poc IEEE Wokshop on Speech Coding,Delavan, WI, USA, Septembe 2000 [6]. Motivation fom a Full-Rate Specific Design to a DSP Coe Appoach fo GSM Vocodes - Shevin Sheidaei,Hamid Nooi,Ahmad Akbai,Hosein Pedam Voiceage.com Fee libay povide fo G [7]. J.Cambell,T.Temain and V.Welch,1990, The DoD 4.8 Kbps Standad (poposed Fedeal Standad 1016), in Advances in Speech Coding,ed.B.Atal, V.Cupeman and A.Gesho, Kluwe Academics Publishes. [8] Real time imlpementation and evaluation of vaiable ate celp codes ETSI Telecommunication U.Vigo Apatado,62,36280 Vigo SPAIN. Pedam Voiceage.com Fee libay povide fo G S.Pabu eceived the B.E degee i n C o m p u t e S c i e n c e a n d engineeing fom Sona College of Technology, Salem, India in 2002 and the M.Tech degee in Remote S e n s i n g a n d G e o g a p h i c a l Infomation Systems (GIS) fom College of Engineeing Guindy, Anna Univesity, Chennai, India, in Pesently, he is woking with School of Computing Sciences, VIT Univesity, Velloe, India as a Assistant Pofesso (S.Gade). [1] J.Cambell,T.Temain and V.Welch, The DoD 4.8 Kbps Standad (poposed Fedeal Standad 1016), in Advances in Speech Coding,ed.B.Atal, V.Cupeman and A.Gesho, Kluwe Academics Publishes, 1990