An Efficient Time Domain Speech Compression Algorithm Based on LPC and Sub-Band Coding Techniques

Transcription

1 JOURNAL OF COMMUNICATIONS, VOL. 4, NO. 6, JULY An Efficient Time Domin Seech Comression Algorithm Bsed on LPC nd Sub-Bnd Coding Techniques P.Venkteswrn Det. of Electronics & Tele-Communiction Engg., Jdvur University, Kolkt (INDIA) Emil: Arindm Snyl 1, Snehsish Ds 2, R.Nndi 3, S.K.Snyl 4 Det. of Electronics & Tele-Communiction Engg., Jdvur University, Kolkt (INDIA) Emil: rindm_3110@yhoo.co.in 1, snehsish.ds@tcs.com 2, robnon@ieee.org 3, s_snyl@ieee.org 4 Abstrct Seech comression is mture technology with mny lictions. Over the st decde, huge dvnces hve been mde in the re of seech coding for reduced bit-rte trnsmission. With ercetul udio coding, the signl is coded efficiently using sychocoustic model, s in MPEG stndrds. In this er, two well known lgorithms, LPC (Liner Predictive Coding) nd subbnd coding re combined to reduce dt trnsmission rte from 128 Kbs to s low s 12 Kbs t minimum comlexity nd imlemented on digitl signl rocessor. The erformnce of the roosed lgorithm is lmost sme s the MPEG lgorithm without its comlexity. Index Terms Digitl signl rocessors, Levinson-Durbin lgorithm, Liner Predictive Coding, RELP, subbnd coding. I. INTRODUCTION Efficient seech coding [1] t low bit rtes find mny lictions for bndwidth conservtion nd enhnced rivcy in wireless cellulr nd stellite communictions. Digitl seech coders ttemt to reduce the bit rte while minimizing the quntiztion error. Time Domin coding methods such s delt modultion or sigm delt modultion oerte cross full bndwidth nd cn chieve comression rtios u to 2.5:1[2]. Frequency domin coding methods nlyze the signl in frequency domin nd reduce the quntiztion error using sychocoustic model. The time domin nd frequency domin methods cn be combined to yield the best of both the methods. In this work we hve studied the effects of combining time domin method, LPC coding, with frequency domin method, subbnd coding, to chieve comression rtios of u to 7:1 t minimum comlexity. Both the LPC nd subbnd coding re well known lgorithms. Subbnd coding is rticulrly suitble for seech comression s seech energy is mostly concentrted in the low frequency bnds. The min motivtion of the resent work is to develo n lgorithm which cn be esily embedded in low ower, ortble systems nd hnd held devices. As such, the lgorithm hs to be simle yet rovide erformnces mtching those of the other lgorithms in use. In this resent work, it hs been shown tht combintion of simle lgorithms like LPC nd subbnd coding cn mtch the comression rtios of the other comlex lgorithms using sychocoustic nlysis. A digitl signl rocessor is used to evlute the erformnce of the lgorithm in hrdwre, though the sme could hve been done on n FPGA ltform. The seech comressor used in the resent work hs synthesis by nlysis structure. The block digrm of the nlysis nd synthesis sections of the roosed seech comressor is shown in Fig.1. Bsed on the er An Efficient Time Domin Seech Comression Technique nd Hrdwre Imlementtion on TMSC Digitl Signl Processor, reliminry version ublished by Arindm Snyl, Snehsish Ds, P.Venkteswrn, S.K.Snyl nd R.Nndi which ered in the Proceedings of the 1 st IEEE Interntionl Conference on Signl Processing, Communictions nd Networking- ICSCN 2007, Ann University, Chenni, Indi, Feb Mnuscrit received July 26, 2008; revised December 30, 2008; cceted Jnury 21, Fig.1.. Trnsmitter (Anlysis Section)

2 424 JOURNAL OF COMMUNICATIONS, VOL. 4, NO. 6, JULY 2009 inverse filter with trnsfer function (1/A (z)),which is n ll-ole filter, to recover the originl signl. The minimiztion of the men-squre vlue of the forwrd rediction error results in set of norml equtions, exressed in comct form s k 0 ( k) ( l k) 0 l 1, 2,..., ; (0) 1 Fig. 1.b. Receiver (Synthesis Section) The er is orgnized s follows: the RELP coder is described in Section II, the subbnd coding rincile is described in Section III, nd the hrdwre imlementtion nd results re discussed in Section IV. Section V brings u the conclusion nd ossible extensions to the resent work. II. RELP CODER Liner rediction [3] is n imortnt toic in digitl signl rocessing with mny rcticl lictions. In our roosed seech comressor, we hve used oneste forwrd liner redictor, which forms the rediction of the vlue n) by weighted liner combintion of the st vlues n-1), n-2),., n-) where is the order of the redictor. Hence the linerly redicted vlue of n) is ~ n) ( k) n k) where { (k)} reresent k 1 the rediction coefficients of the liner redictor. The difference between the vlue n) nd the redicted vlue is clled the forwrd rediction error, denoted s f ~ ( n ) x ( n ) x ( n ) n) k 1 ( k) n k) ( k) n k), k0 (0) 1 An equivlent reliztion of the rediction error filter is shown in Fig. 2. Fig. 2. Prediction-error filter Thus, we cn model the rediction error filter s n llzero FIR filter whose z-trnsform reltionshi is given by F (z) = A (z) X(z). At the receiver side, we hve n where N k l x ni ( l ) ( n) n l), N is the length of the cusl sequence n), i = l, k = 0 for l 0, nd i =0, k = l for l < 0 The Levinson-Durbin lgorithm is comuttionlly efficient lgorithm for solving the norml equtions. This lgorithm exloits the secil symmetry in the utocorreltion mtrix The desired recursion for the redictor coefficients in the Levinson-Durbin lgorithm is m m ( m) ( k) m ( m) (0) bt m1 bt m1 m1 b* m1 * m 1 1 ( k) ( m) ( m k) m bt m1 ( m 1) ( m 2),k=1,2,..,m-1; m=1,2,.,. t nd m denotes the trnsose of m. Thus, using the Levinson Durbin lgorithm, we cn construct the llzero rediction error filter t the trnsmitter nd subsequently, the ll-ole filter t the receiver to reconstruct the originl signl. The residul signl essentilly crries ll the informtion tht hs not been ctured by the LP nlysis, e.g., hse nd itch informtion, zeroes due to nsl sounds etc. In RELP, the residul encoding is bsed on sectrl rther thn wveform mtching. In ddition, RELP coders rely on the fct tht the low-frequency comonents of seech re ercetully imortnt. This residul signl is then subsequently divided into two subbnds for referentil encoding. Since only the residul is rocessed t the subsequent stges, RELP llows sufficient bit comression. III. SUBBAND CODING Subbnd coding [4] is n effective method to chieve dt comression in seech signls, where the signl energy is concentrted minly in the low frequency region. We hve divided the 8 KHz smled seech signl into two frequency regions below 2 KHz nd, (1)

3 JOURNAL OF COMMUNICATIONS, VOL. 4, NO. 6, JULY bove 2 KHz. Further, sub-division of the seech signl yields no comrble increse in seech comression rtio. The decimted signl is then referentilly encoded by llocting lrger number of bits to the low frequency region thn the high frequency region. We hve used QMF (Qudrture Mirror Filters) [5] to decimte nd subsequently interolte the signl. The filter chrcteristics of the low ss nd high ss QMFs s creted using the C5416 DSP chi re given in Fig. 3 Fig. 4.b. Interoltion by use of Polyhse filters IV. HARDWARE IMPLEMENTATION AND RESULTS Fig. 3.. Filter chrcteristics of low ss filter creted with C5416 Processor Fig.3.b. Filter chrcteristics of high ss filter creted with C5416 Processor With the choice of the bove filter chrcteristics, the comonent due to lising vnishes. Thus the lising resulting from decimtion in the nlysis section of the trnsmitter is erfectly cncelled by the imge signl sectrum tht rises due to interoltion t the receiver. To sve comuttions, we erformed the decimtion nd interoltion using olyhse filter structures s shown in Fig. 4. The incoming signl, in the form of 16 bit PCM smles, is ssed through redictor block which uses the history of revious smles to redict the next signl. This rediction is comred with the ctul signl nd the difference is mesured. The encoder in the trnsmitting block tkes this difference signl nd sses this through QMF (Qudrture Mirror Filter) bnk. This secilly designed filter slits the signl into two eqully divided subbnds [6] ccording to the frequency content. Through this rocess, the QMF lso down smles the clock rte to hlf the originl clock. In the subbnd rocessor, the 16-bit word is then ssed through quntizer routine which lloctes greter number of bits to the low frequency region thn the high frequency region. The decoding rocess is the inverse of the encoding rocess, where the bit strem is slit into two subbnd signls nd the inverse quntizer nd redictor is used to regenerte the originl signl. A 1-minute seech ws recorded using microhone nd the voice recorder rovided with Windows XP in 16- bit mono mode nd smling frequency 8 KHz, for our exeriments. This file recorded in.wv formt ws then rovided s n inut to the C5416 rocessor [7], in frmes of 320 smles. The comressed file s well s the recovered file ws stored in the hrd disk of comuter which ws interfced with the C5416 rocessor. The setu is shown in Fig. 5 below. Fig. 4.. Decimtion by use of Polyhse filters Fig. 5. The setu used in the resent work

4 426 JOURNAL OF COMMUNICATIONS, VOL. 4, NO. 6, JULY 2009 The sectr of the originl s well s the recovered signls were then viewed using sectrum nlyzer. While rocessing the originl file, the first 44 smles, which form the heder, were first extrcted nd lter were dded to the recovered file. For every 320 inut smles, corresonding frme contining the encoded dt ws written to the comressed file. The decoder t the receiving side generted the 320 smles from this frme. The first six smles, ech of 16-bits, contined the LPC rmeters. The next two 16 bit smles contined the gins in the two frequency subbnds. The remining smles contined the encoded seech dt. Tble I shows the number of bits with which 16- bit smle in low frequency bnd nd 16-bit smle in high frequency bnd re encoded, nd the corresonding comression rtios chieved. TABLE I Number of bits used in encoding the low frequency smles VARIOUS COMPRESSION RATIOS ACHIEVED Number of bits used in encoding the high frequency Originl file size Comressed file size Comression chieved smles kb 223 kb 76.2% (4.2 : 1) kb 194 kb 79.3% (4.8 : 1) kb 135 kb 85.6% (6.9 : 1) For comrison of the comressed signls with the originl signl, we hve given the signl sectrum nd the time domin reresenttions of the originl s well s the recovered signls (Fig h) for vrious vlues of the number of bits used to encode the high nd low frequency regions resectively. The time domin reresenttions show the first 2 ms of the seech signls. As cn be seen from the time-domin reresenttions, the recovered signls, nd the originl signl re identicl in the loction of the eks nd the energy distribution. Fig. 6 b. Time domin reresenttion of the originl signl (from 0 2s) Fig. 6. Sectrum of the originl signl Fig. 6 c. Sectrum of the 76.2% comressed signl

5 JOURNAL OF COMMUNICATIONS, VOL. 4, NO. 6, JULY Fig.6 d. Time domin reresenttion of the 76.2% comressed signl (from 0 2s) Fig. 6 g. Sectrum of the 85.6% comressed signl Fig. 6 h. Time domin reresenttion of the 85.6% comressed signl (from 0 2s) Fig. 6 e. Sectrum of the 79.3% comressed signl Fig. 6 f. Time domin reresenttion of the 79.3% comressed signl (from 0 2s) From the sectr in Fig. 6, we see tht the low frequency region of the seech signl contining most of the seech energy is lmost unffected due to bit comression. The high frequency sectrum gets distorted by the encoding of the seech signl. However, the informtion loss in the high frequency region does not ffect the overll qulity of the seech signl by drstic mount, nd the desired informtion cn very well be extrcted from the recovered signls. For rocessing the seech signl, we interreted the 16-bit integer smles s q15 numbers, nd then converted them to their equivlent floting oint numbers. In q15 formt, number between -1 nd ( ) cn be reresented by 16-bits, in which the first bit reresents the sign of the number nd the next 15-bits reresent the mgnitude of the number. By converting the numbers to their equivlent floting oint versions in this fshion, we restrict the numbers between -1 nd ( ), thus reducing the quntiztion error. After extrcting the residue through LP nlysis, the rnge of the numbers is so much reduced tht we cn encode ech 16-bit smle with less thn 4 bits even while mintining sufficient ccurcy. Also, we clculte correltion between two

6 428 JOURNAL OF COMMUNICATIONS, VOL. 4, NO. 6, JULY 2009 smles bit-wise which results in greter ccurcy, thus ermitting us to comress the seech gretly. Also better results cn be obtined if we ly re-emhsis filter to enhnce the high frequency region of the seech before erforming LP nlysis. Such signl would then hve to be de-emhsized t the receiving end by deemhsis filter fter erforming inverse LP filtering. V. CONCLUSION Seech coding is n essentil liction of digitl signl rocessing in modern dy telehony nd mobile communictions, which emloy high dt comression rtios. In this er, we hve shown simle method of effective seech comression technique combining the well known LPC nlysis nd subbnd coding. Using only RELP would chieve comression rtio of 60% [8], while combintion of RELP nd subbnd coding results in comression rtio s high s 85.6%. We hve reduced the trnsmission bndwidth to 12 Kbs from 128 Kbs, or bndwidth reduction fctor of 11:1, which is comrble to the MPEG (12:1 comression rtio) [9] without its comlexity. The comression rtios chieved re lso slightly better thn similr roject undertken t Communictions University of Mrylnd, College Prk, Eson Plo Alto, R&D lb, in which the comression rtios chieved re between 57.78% to 82.45% [10]. We hve lso resented the results of our exeriments on seech comression which shows good imrovement on Bndwidth with ccetble distortion level. Imlementtion of this system with C5416 Digitl Signl Processor is lso given. The number system used in this work is q15 formt which is lso suitble for FPGA (Field Progrmmble Gte Arry) or ASIC (Aliction Secific Integrted Circuit) imlementtion. Also, the lgorithm used cn redily be imlemented in VHDL or Verilog. Thus, this seech comression method cn be extended to ny low ower hnd held device, which cn store lrge number of udio files in comressed formt, nd decomress nd ly file when required. The seech comression method is lso fr less comlex thn stndrd comression lgorithms, thus requiring less gte count, nd hence less re nd ower, during FPGA or ASIC imlementtion. REFERENCES [1] J. D. Gibson : Seech Coding Methods, Stndrds, nd Alictions. IEEE Circuits nd Systems Mgzine. Vol. 5, No. 4, Fourth Qurter 2005, [2] Toms Krtchovil nd Mrtin Sevcik : Frequency Anlysis of Low Bit Rte Modern Digitl Audio Comression Algorithms. Institute of Rdio Electronics, Brno University of Technology. [3] P. Strobch : Liner Prediction Theory-A Mthemticl Bsis for dtive Systems. Sringer-Verlg, [4] John G. Prokis nd D. J. Mnolkis : Digitl Signl Processing, PHI, 3/e, (July 2003). [5] S. Sridhrn.: Hndbook of Neurl Networks for Seech Processing. 1/e, Ed. S. Ktgiri, Artech House, Chter 4(199). [6] Neet Awsthy, J.P.Sini nd D.S.Chuhn : Sectrl Anlysis of Seech: A New Technique. Interntionl Journl of Signl Processing, Vol. 2, no.1, 2005, ISSN [7] Texs Instruments.:TMS320C54X Otimizing C/C++ Comiler User s Guide., Literture Number SPRU103G. [8] Biren Mohnty, Some Studies On Excittion Methods Of RELP Vocoder For Seech Coding At Low Bit Rte, M.E Tel. E. Thesis Work,.62, Jdvur University, Kolkt , Indi. [9] Clrk Novk, Sectrl Bnd Reliction nd AACPlus Coding-An Overview, Telos Systems, Clevelnd, Ohio. [10] Ali Zndifr nd Shhrooz Shhrni, ENEE 6244 Project 2:Liner Prediction for Seech Coding nd Recognition, Communictions University of Mrylnd, College Prk, Eson Plo Alto, R&D Lb URL. P. Venkteswrn is working s Reder in the Det. of Electronics & Tele-Communiction Engg. (ETCE), Jdvur University (JU), Kolkt, Indi. He hs ublished over 30 ers in vrious Ntionl / Interntionl Journl / Conference Proceedings. His fields of interest re Comuter Communiction, Microcomuter Systems nd Digitl Signl Processing (DSP). He is Member of IEEE (USA), nd Present Secretry of Clcutt Chters of IEEE Communictions Society (ComSoc) & IEEE Circuits And Systems Society (CASS). Arindm Snyl hs comleted his Bchelor s degree from Det. of ETCE, JU in He is currently ursuing his Msters degree t Indin Institute of Technology (IIT), Khrgur in Microelectronics nd VLSI disciline. His current reserch interest is RFIC. Snehsish Ds hs comleted his Bchelor s degree from Det. of ETCE, JU in He is currently n Assistnt Systems Engineer with Tt Consultncy Services (TCS) Limited, Kolkt. Dr. Rbindrnth Nndi is Professor in the Det. of ETCE, JU. His res of interest re ASP, DSP, Comuter Communiction. He hs uthored more thn 110 reserch ers in Ntionl / Interntionl Journls nd some in Conferences / Seminrs. He served s the Hed of ETCE Det., JU during nd served s the Chir of IEEE Clcutt Section during He hs tught in vrious Institutes brod. Dr. Slil Kumr Snyl is Professor nd Hed of the Det. of ETCE, JU. He hs ublished more thn 100 reserch ers in Interntionl / Ntionl Journls / Conference Proceedings. His current reserch interests include Anlog Signl Processing (ASP) & DSP, Tunble Microstri Antenn, Communiction Engineering, VLSI Circuits & Systems Design. He served s the Chir of IEEE Clcutt Section during