A study on the efficient compression algorithm of the voice/data integrated multiplexer

Transcription

1 A study o the efficiet compressio algorithm of the voice/data itegrated multiplexer Gyou-Yo CHO' ad Dog-Ho CHO' * Dept. of Computer Egieerig. KyiigHee Uiv. Kiheugup Yogiku Kyuggido, KOREA PHONE : FAX : gycho@ms.kyughee.ac.kr Abstract The compressio techique is required to icrease the efficiecy of commuicatio chael. The LZW(Lempe1 Ziv Welch) algorithm is used as a compressio algorithm i commuicatio chael. The LZW algorithm, however, has some redudacies over Hageul text ad voice. Thus, i this paper, the VDI-LZW(Voice Data Itegrated LZW) algorithm that decreases the redudacies of the LZW algorithm is suggested for efficiet compressio of Hageul text ad voice. The VDI-LZW algorithm uses both the codeword of the complete Hageul ad the variable legth codeword methodology i data compressio. O the other had, i voice compressio, it uses the differetial method that reduces the codeword size as well as the legth of repeated strig. The, the repetitio ratio is icreased. Accordig to the simulatio results, it ca be see that the performace of the proposed compressio algorithm is better by 4% to 18% i data compressio ratio ad is better by 35% to 44% i voice compressio ratio tha that of the covetioal modified LZW algorithms. I. Itroductio Recetly, due to the progress ad icreasig variety of telecommuicatio service, the amout of data trasfemd through the commuicatio chael is beig icreased rapidly. Therefore, the issues o the efficiet trasmissio mechaism of various media stream is actively ivestigated. While there are may studies o the compressio of data traffic, compressio of voice stream is ot much studied ulike the speech codig such as waveform coder ad vocoder. So, i this paper, the simple commo compressio algorithm which could be used to maximize the efficiecy of the voice/data itegrated multiplexer is studied. May techiques about data compressio are beig studied owaday. For example, statistical ad uiversal compressio techiques are beig studied vigorously. The former are based o the probability of data geeratio, ad the latter are idepedet the probability of data geeratio. Also, LZW compressio techique based o uiversal compressio techique is recommeded with V.42 bis compressio algorithm by CCITT, because of its good efficiecy. However, this LZW compressio techique has some redudacies. That is. the size of compressed data is larger tha primary data i early stage ad the bit stream of output codeword icludes.5 garbage bit i average case. Also, because the codeword is suitable oly to Eglish, LZW algorithm has some redudacies i 2 bytes Hageul which is Korea character. To reduce the redudacy of output bit stream, the VV-LZW (Variable-to-Variable LZW) algorithm is suggested. Moreover, to reduce the redudacy of Hageul, LZWH(L2W Hageul) is suggested. However, these algorithms could ot completely reduce those redudacies. O the other had, the covetioal data compressio. algorithms have ot take ito accout voice traffic. Whe the voice traffic is sampled usig 64k bps PCM ecoder, its repetitio cycle becomes more tha 8 samplesc21. Thus, the covetioal data compressio algorithms which have 2 characters i strig is iefficiet for voice traffic which has log repetitio cycle. Therefore, i this paper, the differetial method is applied to the PCM coded voice traffic, ad the compressio is doe based o LZW compressio algorithm. I this paper, i order to achieve efficiet compressio ratio i the voice/data itegrated multiplexer, the ew compressio algorithm which could compesate for these iefficiecies of the covetioal compressio algorithm is proposed ad aalyzed through computer simulatio. Followig this itroductio, i sectio 2, the LZW compressio algorithm ad the covetioal modified compressio algorithms are surveyed. Also, i sectio 3, the ew compressio algorithm which reduces the redudacies of the covetioal algorithm is proposed. Moreover, i sectio 4, the ewly proposed compressio algorithm is compared with the covetioal algorithms through the computer simulatio. Fially, coclusios are made i sectio 5. II. The Covetioal Compressio Algorithms ad Problems 1. LZW algorithm LZW(Lempe1 Ziv Welch) is a algorithm that uses coversio table mappig icomig strig ito the codeword. I LZW codeword table, the codeword has the prefix of strig. That is, whe there is a strig wk composed of a strig w ad a character K, K ad w are i the table. Here, K is called as exted character ad w is called as prefix strig~ll,~31,~41,~51,~81,~11. LZW codeword table icludes the strigs compressed previously. It meas that the strigs of codeword table have bee traied already, ad that the codewords table reflects the statistical property of the data stream previously applied. 1.1 The ecodig procedure of LZW algorithm The ecodig procedure of LZW algorithm is as follows. 1) The iitializig step of codeword table : Register 256 characters as the iitial codewords of $ IEEE 1438 Authorized licesed use limited to: Korea Advaced Istitute of Sciece ad Techology. Dowloaded o September 11, 29 at 11:52 from IEEE Xplore. Restrictios apply.

2 codeword table. That is, iitialize each code by ASCII character values whose rages are from to 255. Also, iitialize the codeword variable cl to ) Read the first iput character from the iput data,ad set it to the prefix (w). 3) Read the the ext iput character, ad set it to exted character (K). i) Whe o iput exists i the ext sequece, fid the codeword matched to w ad ecode it to the biary code of [log(cl+l)l bits. (Here, ~log(cl+l)l meas a miimum iteger that is bigger tha log(cl+l).) 4) Whe wk exists i the codeword table, replace wk with a ew prefix (w) ad go to step 3). 5) Whe wk does ot exist i the codeword table, fid the codeword matched to w ad ecode it to the biary code of [Iog(cl+l)l bits, the go to followig step. i ) Assig cl that is the codeword of wk to the codeword table, ad register as a ew strig. ii ) Icrease cl by 1 for the ext codeword assigmet. iii) Replace K with a ew prefix w, ad go to the step 3). 1.2 The decodig procedure of LZW algorithm The decodig procedure usually follows the reverse order of ecodig procedure. While the decodig procedure is activated, it has the same codeword table as the ecodig procedure. Also, whe ecoded codewords are released, each codeword is decoded ito a prefix strig ad a exted character by the cotets of a codeword table. The, the exted character is outputed to stack ad the prefix is decoded agai to a prefix ad a exted character. Here, whe a exted character is separated from a prefix ad outputed, it uses stack to chage the order of output characters sice the decodig order of exted characters is the reverse order of the ecodig. 2. VV(Variab1e to VariablekLZW algorithm As metioed above, i LZW algorithm, all codewords of strigs geerated durig the ecodig procedure are ecoded to a biary code of rlog(cl+l)l bits. That is, whe the value of cl is from 256 to 511, the value of codeword is ecoded ito 9 bits. Also, whe the value of cl is from 512 to 123, the value of codeword is ecoded ito 1 bits. Thus, this algorithm has much redudacies. To remove this redudacy of a output bit stream, the variable legth LZW algorithm is used. I VV-LZW, a limit value is chose amog the values from codeword to cl-1. For the codeword value betwee ad limit value, it is ecoded to a biary code of ([log cf-1) bits. Also, for the codeword value betwee limit value ad cl-1, it is ecoded to a biary code of [log(cl+l)l bits. 3. LZWH(LZW Haged) algorithm Beig applied to the Hageul text data, of course, LZW algorithm has a redudacy. Although there are various Hageul types, KSC 561, the Korea stadard IIageul type for admiistrative computer etwork, is oly cosidered. I KSC 561, the umber of the code assiged to Hageul is , ad it is possible to express Haged by usig 12 bits. However, LZW algorithm uses 16 bits. Thus, there are some redudacies i applyig LZW algorithm to the KSC 561 Hageul code. LZWH algorithm is proposed to reduce those redudacies. The first step of this algorithm is to select 47 characters used most frequetly amog the 235 KSC Hageul characters. The, register these characters as iitial codewords. If Haged codes are etered, they are ecoded by the Haged compressio procedure. Otherwise, they are ecoded by the geeral compressio procedure. 4. The problem of voice compressio LZW ecodig algorithm takes ito accout oly Eglish text. So, while it is very efficiet i data compressio, it is iefficiet i voice compressio. Thus, to compesate for this iefficiecy, the ew voice compressio algorithm is required. a. The New Ecodig Algorithm From the previous discussio, it could be see that the redudacy for output bit stream could be removed, but the redudacy for Haged code has still remaied i the case of applyig VV-LZW algorithm. Also, it could be see that the redudacy for Hageul code could be removed, but the redudacy for output bit stream has still remaied i the case of applyig LZWH algorithm. Moreover, It is't suitable to use the previous data compressio algorithm for voice compressio. The proposed VDI-LZW algorithm could reduce these two redudacies which VV-LZW ad LZWH have. Sice the separate chaels are used for each of the voice ad data streams i the voice/data itegrated multiplexer, VDI-LZW algorithm applies the voice ecodig algorithm to the voice chael ad the data ecodig algorithm to the data chael. After ecodig is doe, ecoded bits are moved to multiplexer ad trasmitted to the receiver. The, the receiver demultiplexes multiplexed voice ad data, ad decodes the data ad voice by usig voice ad data decodig algorithm, ad trasmits voice ad data to the voice ad the data chael respectively. 1. Iitializig step I the iitializatio step, register 256 ASCII codes ad 47 complete Haged characters beig used most frequetly i the codeword table. The, iitialize cl with 726 to idicate the ext character that the codeword to be registered should be iserted ito the 726th row off the codeword table. I the case of voice, register the codeword of 64 codes ad the iitialize cl with 65 to idicate the ext character that the codeword to be registered should be iserted ito the 65th row of the codeword table. Fig. 1 shows the cofiguratio of the codeword table for the data ad voice respectively. First of alll, the frequecies of 235 complete Haged characters appearig i the file cotaiig both Hageul ad Eglish characters are computed. The, 47 characters which has higher frequecy are selected. From the aalysis result for the PCM sampled 8 bits voice samples, it could be see that Authorized licesed use limited to: Korea Advaced Istitute of Sciece ad Techology. Dowloaded o September 11, 29 at 11:52 from IEEE Xplore. Restrictios apply.

3 99% of the differeces betwee adjacet samples values is less tha 32. Therefore, the umber of codeword required is oly 64. So, it is possible to express codeword based o 6 bits ASCII Hageul codes 1 bits codewords 11 bits codewords 12 bits codewords 128 i)u bits codewords 8 bits codewords (a) Data codewords table (b) Voice codewords table Fig. 1 The cofiguratio of the codeword table for the data ad the voice 2. Ecodig Step 2.1 Data ecodig I the case of complete Hageul, the first byte is assiged from BO to C8 ad the secod byte is assiged from A1 to FE i hexadecimal. So, the discrimiatio algorithm betwee the complete Hageul code ad ormal ASCII code operates as follows. 1) If the first byte is bigger tha BO ad smaller tha C8, read the secod byte. i If appropriate code value is foud, perform biay search ad retur the appropriate code value. If appropriate code value is ot foud, that is, if it is ot icluded i 47 characters, retum a geeral ASCII codes for each byte. ii) If the secod byte is ot icluded i rage from A1 to FE, retur a geeral ASCII codes for each byte. 2) If the first byte is smaller tha BO or bigger tha C8, retur a geeral ASCII code. 2.2 Voice ecodig I the case of the voice, to apply the differetial method, the followig procedure is performed. 1) After subtractig cur that is the value of the curret sample from old that is the value of the previous sample, save it as temp. Here, temp, cur ad old are variables. 2) Ivestigate the size of temp. i 5 temp < +31, Proceed to the step 3). ii) -32 < temp <, After takig 2 s complemet of temp expressed as 6 bits, proceed to the step 3). iii) temp > -16, Output 32 to idicate that temp is smaller tha -32. The, after dividig temp by 2 ad calculatig its absolute value, proceed to thc step 3). iv) temp < 159, Output 31 to idicate that temp is bigger tha +31. The, after dividig temp by 2, proceed to the step 3). 3) Retur the value of temp. 2.3 The output step of ecoded codeword To reduce the size of ecoded codewords, the followig procedures are performed i the output step of ecoded codeword. Whe ecoded codewords are released, the followig formula must be cofirmed i order to request the right strig. Here, if the umber of codewords curretly assiged is cl ad the limit value that ca reduce the size of bits is limit, followig equatio is derived. (cl+ limit) < 2r1dc1+1) -1 (11, Here, if = 2rlo~(cl+1)1, equatio (1) becomes as follows limit = d-cl-1. (2) Thus, whe codewords are chaged ito bit stream based o the compressio algorithm, if the codeword is smaller tha limit value, output the codeword expressed with rlog(cl+l)l-l bits. Otherwise, output the codeword expressed [log(cl+l)l bits. For example, whe cl is 75, the value of limits is calculated by usig equatio (2). That is, limit = = 273. Therefore, if a codeword is betwee ad 273, ecode ad output it based o 9 bits. However, if a codeword is betwee 274 ad 749 durig ecodig, add 274 to the codeword. The, ecode ad ouput based o 1 bits. Durig decodig, a codeword bit is read i 9 bits. The, if this value is smaller tha 274, the value is take as a codeword. Also, if it is bigger tha 274, codeword bit is read 1 bits ad the subtractio of 274 from read value is doe. Whe cl is 75, the example of output bit stream is show i Table I. Table I. The example of output codeword whe cl is 75 ecode codeword 1 2 output codeword bits 1 1 decimal value 1 L lv. Simulatio ad Discussios ( ) 549( ) 123( 274~749 ) I this sectio, for the performace evaluatio of the ew compressio algorithm proposed i this paper, VDI-LZW algorithm is implemeted usig C laguage i 486 PC. Also, i order to compare it with covetioal algorithms, LZW, VV-LZW, ad LZWH are implemeted by C laguage i Authorized licesed use limited to: Korea Advaced Istitute of Sciece ad Techology. Dowloaded o September 11, 29 at 11:52 from IEEE Xplore. Restrictios apply.

4 PC. Here, the text data such as papers ad reports which represets typical data stream trasferred through the commuicatio chael are cosidered. Also, for the voice data, real coversatioal speech data o the telephoe lie based o 641tbps PCM ecoder is cosidered.. Table U. The compressio ratio of data % of Hageul LZW VV-LZW 7% 56% 59% compressio ratio LZWH I I I 2.5 VDI-LZW I I 2.59 The performaces of the various compressio algorithms for text data are show i Table U. Here, it could be see that the ew compressio algorithm(vd1-lzw) has better by 4% to 22% tha that of covetioal modified LZW algorithms. Also, Fig 2 shows the graphical expressio of Table U. aoo d 1.6 $ x c I.2 1 H d a 5 OD Fig. 4 The probability of differetial code after the differetial operatio I the case of the voice, the differetial method is used to icrease the coipressio efficiecy. The differetial method coverts the PCM ecoded voice sample ito the differetial code. Fig. 3 ad 4 show the probability distributios before ad after the coversio. It could be see from Fig. 3 ad 4 that the maximum probability of the differece after the coversio is about.36 while the maximum probibality of the PCM code before the coversio is about.2. From the above results, it could be see that the covergece rate of data is icreased after coversio. Therefore, the code repetitio rate could be icreased durig the ecodig step of LZW algorithm by usig the differetial coversio. a 5 aoo 57b1b 111b8 1b43b Data Size (byte) VDl-iir Fig. 2 The compressio ratio diagram of text data Each sample of Fig. 2 shows that the differece of compressio ratio amog LZW, VV-LZW, LZWH ad VDI-LZW is big i the sample 1 ad is small i the sample 4. The reaso for this pheomeo is that the IIageul ratio of sample 1 is larger tha that of sample SO 8 &Io 6 4 coversio zoo f am a asm Size(byle) Fig 5. The compressio ratio diagram of voice Table RI. The compressio ratio of voice Table III ad Fig. 5 show the compressio ratio of each algorithm. It could be see that the performace of VDI-LZW algorithm is superior by 35% to 44% to that of the covetioal algorithm. Also, it could be see from Fig. 5 that the compressio ratio is high i the sample 2 of Fig 5 ad low i the sample 4. This pheomeo could be explaied by the 1441 Authorized licesed use limited to: Korea Advaced Istitute of Sciece ad Techology. Dowloaded o September 11, 29 at 11:52 from IEEE Xplore. Restrictios apply.

5 fact that sample 2 has smaller differece amog samples tha sample 4. Table N represets the umber of chaged samples whe the voice compressed with the VDI-LZW algorithm is decoded ito the PCM code. I Table N, it could be see that the average ratio of the chaged samples is.26%. Such a mior loss of voice sample due to the compressio is robust to the huma ears. Table lv. The umber of compressio sample 1 sample 2 sample 3 sample 4 total file size(byte1 samples chaged due to the chaged sample error rate(%) Both the compressio ratios whe the silece period exists ad the compressio ratio whe the silece period does ot exist are show i Fig 6. From this figure, the sample 1 has little silece period, ad the sample 2 icludes 4% silece period. It could be see that the compressio ratio of the VDI-LZW algorithm is icreased as silece period is larger sice code repetitio ratio is icreased as the differece betwee adjacet samples is reduced. S O O A _ 25 U U E 2 E I applied to complete Hageul code ad the variable legth codeword methodology i data compressio. Also, i voice compressio, it uses the differetial coversio method to miimize the codeword size ad the legth of repeated strig. Accorgig to the simulatio results, it could be see that the performace of VDI-LZW algorithm is superior to that of covetioal compressio algorithms by 4% to 18% i data compressio ratio. I additio, it could be see that the performace of VDI-LZW algorithm is superior to that of covetioal compressio algorithms by 35% to 44% i voice compressio ratio. Referece 111 CCITT Recommedatio V.42 bis, "Data Compressio Procedures for Data Circuit Termiatio Equipmet usig Error Correctio Procedures", CCITT, CCITT Recommedatio G.711, "Pulse Code Modulatio (PCM) of Voice Frequecies", CCITT, [31 Timothy Bell, "Modelig for Text Compressio", ACM Computig Serveys. V1.21, No.4. December Terry A.Welch,"A Techique for High-Performace Data Compressio", EEE Computer 17.6, PP [51 Edward R. Fiala,"Data Compressio with Fiite Widows", Computer of ACM, VOL 32, N.4, April [SI Y.J.JON, "A Study o the Desig of Compressio Techique i Complete Hageul Text", Iformatio Sciece CO V1.2, No.1, Sprig [71 S.I.Jug, "A Study o the improvemet of the LZW compressio algorithm", The Proceedigs Of The Korea Istitue of Commuicatio Scieces 9-8 Vol. 15 No.8 [SI HELD, "Data Compressio", WILEY, [91 Alle G, Robert M. Gray, "Vector Quatizatio ad Sigal Compressio", KLUWER ACADEMIC PUBLISHERS, [lo] JACOB ZIV, ABRAHAM LEMPEL, "Compressio of Idividual Sequeces via Variable-Rate Codig", IEEE TRAN. ON INFORMATION THEORY, VOL. IT-24, N.5, September 1978 a 6 a oo Fig 6. The compressio ratio of voice traffic V. Coclusios The LZW algorithm has several redudacies i applyig it to I-Iageul text ad voice. I order to reduce these redudacies, LZWH algorithm is suggested. However, LZWH algorithm could ot reduce the redudacies related to the voice traffic, because it does ot take ito accout the voice. Thus, i this paper, the ew VDI-LZW algorithm which could reduce the redudacies of Hageul test ad voice data for efficiet voice/data itegrated multiplexer is suggested. The proposed VDI-LZW algorithm uses both the codeword 1442 Authorized licesed use limited to: Korea Advaced Istitute of Sciece ad Techology. Dowloaded o September 11, 29 at 11:52 from IEEE Xplore. Restrictios apply.