Performance comparison of 32 Kbps ADPCM coders in noise-free and noisy environments

MELECON '85 I Volume 11: Digi al Signal P ocessing A. Luque, A. R. Figuei as Vidal, V. Cappellini (eds.) Else ie Science Publishe s B. V. (No h·Holland) I© IEEE 1985 179 PERFOMANCE COMPARISON OF 32 KBPS ADPCM CODERS IN NOISE-FREE AND NOISY ENVIRONMENTS. * ** * En ique Masg au , F ancisco Ruiz- 1a nez and Jose B. Ma ino * E.T.S.I.Telecomunicaci6n,U.P.C., Ba celona, Spain. ** Cen o de In es igaci6n de I.N.T.E.L.S.A., Mad id, Spain. In his pape a compa a i e s udy o se e al schemes o Adap i e Di e en ial Pulse Code Modula ion (ADPCM) wi h backwa d adap a ion is p esen ed. The wo k he e epo ed shows a adeo be ween a good signal- o-noise a io and a low deg ada ion o he ecei ing quali y owing o ansmission e o s in noisy channels. On his second poin , he CC ITT s anda d exhibi s a supe io beha io o he es o he schemes, whe eas i s signal- o-noise a io will esul in 2 o 3 dB less han all he schemes i ansmission e o s a e no conside ed. In addi ion, a supe io pe omance (a ound 2 dB) o encode s using la ice p edic o s wi h espec o hose using ans e sal p edic o s is also shown. 1. In oducci6n Du ing he las yea s a g ea numbe o ADPCM encode s o speech signal and oice-band da a a a es o 24 and 32 Kbps. ha e appea ed. Thei good pe omances as well as hei popula i y ha e made he CCITT o es ablish a s anda d ADPCH a 32 Kbps /1/. Depending on he adap a ion ype o he p edic o , he ADPC -1' s a e classi ied in o wo la ge g oups. Namely, ADPCM's wi h o - wa d adap a ion o wi h backwa d adap a- ion. The i s one need o send an addi- ional side in o ma ion o he ecei e , since he p edic o coe icien s a e upda- ed in he ansmi e on he basis o he exac inpu signal, being his in o ma ion no a ailable a he ecei e . The ansmi- ssion speed inc ease ha his ep esen s make hen less a ac i e and, in ac , he CCITT s anda d belongs o he g oup o Backwa d Adap i e-DPCM. Some wo ks con aining compa a i e s udies be ween ADPC 1 encode s ha e appea ed in a sca e ed and no homogeneous way. Recen - ly, hese wo ks ha e been commen ed in e- e ence /2/, showing he di icul y c ea ed by hei unhomogenei y when pe omances ha- e o be compa ed. In his wo k we unde ake he compa a i e s udy o he mos usual schemes, be ween hem he CCITT s anda d, unde he same condi ions and including noise- ee and noisy ansmission channels. 2. ADPCM sys ems. Figu e l despic s he gene al scheme o an CH 2185-7/85/0000-0179(8)/$01.00 © 1985 IEEE ADPCM sys em which, as i is well-known, is composed a[ wo main blocks: he adap i e quan ize and he adap i e p edic o . The combina ion o di e en s uc u es o bo h blocks lead o a g ea a ie y o po- ssible ADPC/<1 sys ems. Nex , we will b ie ly desc ibe he di e en ypes o bo h blocks which will be conside ed in his pape . x ( n) +- e(n) eq(n) + Qan ize - ~(n) .. + + P edic o l x(n) Figu e l. Gene al scheme o an ADPCM encode 2.1 Adap i e Quan ize An adap i e quan ize consis in a in a- ian quan ize p eceded by a signal- scaling, which allows he adap a ion o£ he quan ize dynamic ange o he inpu si<;J- nal. Thus, a be e p o i o he quan ize bi s numbe is achie ed and, he e o e, a mo e accu a e ep esen a ion o he inpu signal. The e is a popula adap i e quan- ize in he li e a u e, de elopped by se e al au ho s (Jayan ,Goodman,Mi a)/3/ and ha will be conside d he e. The scale ac o is calcula ed on he basis o he p e ious ou pu le el o he quan ize and he p e ious scale ac o . To educe he p opaga ion e ec ha would be p oduced - 180 E. Masg au e al. in he ob en ion o scale ac o s in ecep- ion i a ansmission e oc occu eJ, a leakage ac oc S is used, so ha he ac- ual scale ac o c(n) 1s ob ained as ollows: c(n) = i•l. (n-l).c(n-l)B l ; B < 1. he e ~1. (n-1) i s a mul iplie assigned o l1e le.Jel i o he ou pu quan ize in he p e ious sample. The alues oE hese; mul iplie s ha e been ob ained as /4/ wi h a B ac o o 63/64. This quan ize is used in all ADPCM sys ems he e conside d, excep in he case o he CCITT s andacd, which ep esen s a slig h modi ica ion o ha one. In his case, he scale ac o p esen s wo wo king modes: a as mode, o a speech signal, and a slow mode o da a. The leakage ac o alue is o 31/32. De ails can be ound in ei:e en- ces /1,3/. 2.2 Adap i e P edic o . The p edic o unc ion is ob aining he mos exac p edic ion o he ac ual inpu sample, so ha he di e ence be ween he wo alue , o p edic ion e o , be mini- mum. The dynamic ange educ ion o he signal o be quan ized allows a mo e accu- a e ep esen a ion o his, and he e o e, a smalle quan iza ion e o . Since he adap a a ion is o backwa d ype, inpu signal p edic ion mus be done on he basis o he p e ious econs uc ed sam- ples, which di e om he eal ones in he quan iza ion e o . The e o o be minimized is he quan ized p edic ion e co , which is he only a ailable in he ecep ion. All he p edic o s usec;l a e o FIR ype I excep in he case o CCITT s anda d, which is he IIR ype. The algod hm he e consi- de ed a e he ollowing: a). T ans e sal LMS no malized (LMS). b). G adien Adap i e La ice (GAL). c). G adien Adap i e La ice wi h Two Pa co s (GAL2). d). Exac Leas Squa es Adap i e La ice (LSAL). e). CCITT s anda d: Recu si e Signed G adien . Equa iond o upda ing he coe icien s in e e y case a e b ie ly exp essed as o- llows: a) LMS no malized. am(n+l)=am(n) + P(n).p e (n) x(n-m) q wi h A p x(n)= E a (n) ~(n-m) m=l m x(n) P(n) ~(n) + e (n) q ( l-w)x 2( n) + wP(n-1) p: p edic oc o de . b) GAL The la ice p edic o s uc uce is shown in Figu e; 2. 2 2 P (n)=(l-w)~e (n)+ (n-l))+wP (n-1) m m m m x(n ) Figu e 2. La ice p edic o es uc u e c) GAL2 In his case, wo di e en s Pa co coe - icien s, Ke and K , a e conside ed. m m Ke +l(n+l)=Ke +l(n)+ a em+l(n) m(n-1) m m P (n-l) m . e em+l(n)=em(n)-K m+l(n) m(n-1) m+l(n)= m(n-1)-K m+l(n)em(n) 2 P (n-1)=(1-w) (n-l)+wP (n-2) m m m e 2 e P (n)=(l-w)e (n)+wP (n-1) m m m d) LSAL K (n+l)=wK (n)+(l-y 1 (n) )e (n) (n-1) m m m- m m Ke (n+l) m Km(n+l) P (n-1) m • 32 KBPS ADPCM Code s in Noise-F ee and Noisy En i onmen s 181 e e P 1 (n+l)=P (n+l) - m+ m P l(n+l)=P (n) - m+ m K2 (n+l) m P ( n-l) m K2 ( n+l) m ( (1- Y 1 ( n+ 1) ) ( n) ) 2 m- m ym(n+l}=ym_ 1 (n+l)+----~~~----~----- P (n) m e) CCITT s anda d. In his case he p edic o is o IIR ype wi h wo ze os and six poles.The esp es- sions, which will no w i e he e, can be ound in e e ence /1/. 3. Empi ical Resul s. As a es signal he ollowing sen ence was u e ed in spanish: 'Es a es una p ueba de senal de oz'. To do ha , encode s and de- code s o e e y p oposed sys em ha e been simula ed~ aking in o accoun he lack ca- se as well as he p esence oE ansmission e o s. In he e alua ion o he ob ained esul s se e al measu es we e used. The global signal- o-noise a io, i.e., a e aged along he whole speech eco d ( 200 blocks); he SNRSE:G, ob ained ~y a e aging he mean alue o he SNR o all blocks; he p edic o gain GP, wich ela es he inpu signal powe o he p edic ion e o powe a he quan ize inpu ; and he sa u a ion a e P o pe cen age o samples in which he quan ize is sa u ed. Fi s ly, he ou sys ems con aining FIR p edic o s we e un o di e en s alues o hei con e gence pa ame e s, o o de s l o 7, and se e al u e ances including males and emales speake s. I was obse ed ha a ou h o de was a good elec ion, being ile mo e adcqua es pa ame e alues shown in Table I. w p 0.15 0.98 4 GAL 0.01 0.99 4 GAL2 0.01 0.99 4 LSAL 0.01 0.99 4 Table I. Pa ame e choices o FIR algo i hms Figu es 3 and 4 despic he ajec o ies o he SNR a e aged in e e y block ( 256 by- es / block) o each o he cases cspcci ied in Table I and male speake . Figu e 3. SNR ajec o ies o LMS and GAL2 algo i hms Figu e 4. SNR ajec o ies o GAL and LSAL algo i hms. I , I~.:/ E ~--------~----~--' Figu e 5. SNR e sus p edic o o de o FIR cases. In Figu e 5 i can be obse ed ha he imp o emen o he SNR when inc easing he p edic o o de eaches i s maximum om o de 4. Abo e his alue, he imp o emen is almos null and e en a deg ada ion o he SNR occu s in he cases o he simples algo i hms such as GAL and LMS. This e ec is due o he coe icien s noise (p opo - ional o he o de ) being g a e han he imp o emen in oduced by he p edic ion wi h a highe o de . In Figu e 5 , he supe io i y o he la icce p edic o ype - 182 E. Masg au e al. o e he ans e sal one is shown. On he con a y, one can be susp issed when con- side ing he small di e ences ha he i s ones exhibi . In pa icula , he op imal cha ac e is ic o LSAL algo i hm could make us expec an imp o emen o e he es o he algo i hms. In his case, he pe omance o GAL2 and LSAL a e almos he same, being his ac al eady e i ied in o he wo ks /5/. Nex , he compa a i e esul s i>b ained in he ou FIR ype cases o 4 o de a e shown in Table II oge he wi h hose co esponding o ha p oposed by he CCITT. SNR SNRSEG GP P SNRSEG P£ =10- 3 LMS 29.96 18.3 13.35 .2 GAL 31.15 18.63 14.43 .13 GAL2 31.30 18.7 14.92 .14 LSAL 31.37 18.99 15.15 .17 CCITT 27.9 16.5 13.95 .49 14.5 Table II. Pe omance Compa ison o he i e ADPCM sys em. All magni udes in dB excep P. As i can be obse ed, he esul s ob ained in case o noise- ee channels a e clea ly be e hose encoded by a la ice p edic- o , he wo s pe omance co esponding o he CCITT s anda d. Nex , a noisy en i onme~~s was simula ed wi h an e o a e o 10 , si ua ion wich equi:ces a leas s abili y in he decoding p ocess. Unde hese condi ions, all he FIR algo i hms exhibi ed a g ea endency o di e ge, wich makes i s use impossible. On he con a y, he CCITT s anda d shows a s able beha io , and he deg:cada ion o i s pe omances consis s in a loss o a:cound 2 dB (in SNRSEG) wi h espec o he noise- ee channel case. I is ob ious ha in he case o he CCITT s anda d he quali y le el has been :ceduced in he op imal ansmission condi ions in bene i o a be e beha io in ad e se condi ions. This is ca ied ou by he use o a smalle leakage pa ame e in he quan- ize and he in oduc ion o a simila pa- ame e in he p edic o coe icien s up- da e. The p edic o ze:cos a e also cons- ain s o inne pa s o he uni ci cle so !1a he impulse esponse o he o e all sys em o one sample o e o:c. This cons- ain s is compensa ed by in oducing poles in he p edic o (ze:cos in he o e all sys- em esponse). 4. Conclussions. A compa a i e s udy o he pe omances o se e al ADPCM sys ems has been p esen ed. The main consequencences de i ed om his s udy a e he ollowing: la ice p edic o s p o ide pe omance (a ound 2 dB in SNR) ans e sal one. a be e han he -among he la ice algo i hms, he di e- ences a e small, indis iguishable in an un eliable audi ion, being he so called GAL2 a easonable elec ion o quali y- complexi y. - he ou h o de seem he mos ap opia e o all FIR algo i hms. - he CCITT s anda d encode shows a good comp omise be ween SNR and low deg ada ion unde ansmission e o:c condi ions. Unde:c hese las condi ions, in he decoding p:cocess, he FIR algo:ci hms a e no s able and hey canno wo k wi hou p e ious modi ica ions. 5. Re ences. /1/. D a Recomenda ion G. 7zz. Ad-hoc G oup on 32 Kbps ADPCM Algo i hms. Tempo a y Documen . No embe 83. /2/.J.D.Gibson. 'Adap i e P edic ion o Speech Encoding'.IEEE ASSP Magazine, pp 12-16. July 1984. /3/.N.S.Jayan ,P.Noll. Digi al Coding o Wa e o ms.Cap.4.P en ice-Hall,l984. /4/.D.Mi a,B.Go z. 'An Adap i e PCM Sys em Designed o Noisy Channels and Digi al Implemen a ions'. BSTJ,Vol.57,No.7, pp. 2727-2763, Sep embe 1978. /5/ .R. S.Medaugh. 'A Compa ison o Two Fas Linea P edic o s'. Ph.D .• Boulde 1981. •