@:/5.55 JOURNAL ON
`
`SELECTED AREAS IN
`COMMUNICATIONS
`
`(ISSN 0733-8716}
`
`31
`, Cir..-
`PE:1
`an
`NJ
`
`2 EEL
`
`Lo § 3
`
`-‘.
`3
`
`z:
`
`JUNE 1992
`
`VOLUME 10
`
`NUMBER 5
`
`ISACEM
`
`_A PUBLICATION OF THE IEEE CDMMUNICATIONS SOCIETY
`
` '
`
`SPEECH i_t_ND IMAGE comma
`Guest E_dito_r—N. I-lubing
`
`PAPERS
`
`.
`
`‘
`
`. .N. Joyrmr
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`_
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`' Signal Compression: Technology Targets and Research Directions [Invited Paper) .
`.5’. Wang, A. Sekey. and A. Gersho
`.
`.
`.
`.
`Objective Measure for Predicting Subjective Quality of Speech Coders [hioirerl Paper)
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`_
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`—-A Low-Delay CELP Coder for the CCITT I6 khfs Speech Coding Standard (hmfted Paper) .
`,
`.
`. . .
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`. .J’.-H. Chen, R. V. Cox, }’.~C. Lin, N. Jayanr, and M. J. Mefdiiter
`A High-Quality Muitirate Real-Time CELP Coder (Invited Paper) .
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`. P. Kroc.-n and K. Slt’tImiJ'lt'i'f.hflH
`Techniques for Improving the Performance of CELP—Typc Speech Coders [.-’noi'red Paper) .
`.
`.
`.
`.
`. I. A. Gerson arm‘ M. A. Ja.rimt'
`1 Two-Channel Conjugate Vector Quantizer for Noisy Channel Speech Coding .
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`. .
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`. T. Mariya
`‘Weighted Optimum Bit Allocations to Orthogonal Transforms for Picture Coding .
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`. .8. Macq
`_ Image Coding with the Discrete Cosine-III Transform .
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`. .
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`. 0. K. Ersoy and A. Notziro
`-.;Unified Variable-Length Transform Coding and Image-Adaptit-e Vector Quantization .
`.
`. L. Wang, M. Goldberg, rm.1’.S‘. S.-‘Him:
`I Adaptive Transform Tree Coding of images .
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`. W. /‘t. Pearlmon, P. Jakotdnr, and M. M. Lezmg
`I Spectral Entropy-Activity Classification in Adaptive Transform Coding .
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`. R. Master and U. F:-amlce
`[Shape-Gain Vector Quantization for Noisy Channels with Applications to Image Coding .
`.
`.
`.
`. J. Rosebrork and P. W. Be.r.r.-‘fair
`_ Stlbband Image Coding Using Entropy-Coded Quantization over Noisy Channels .
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`. N. Tmiabc and N. Fnruardfn
`"__;A Progressive Scheme for Digital Image Halfloning, Coding of Halftoncs. and Reconstruction .
`.
`. S. Koh’i‘a.c and D. Ana.rra.r.r:‘on
`‘Single Bit-Map Block Truncation Coding of Color Images .
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`. Y. Wu and D. C‘. Col!
`" ;I,I_1terframe Hierarchical Address-Vector Quantization .
`.
`.
`. ., .
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`. N. M. .Nnr.rrnbadr'. C‘. Y. Choc. and J. U. Roy
`-A Fast Feature-Based Block Matching Algorithm Using Integral Projections . .' .
`_
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`. J.-.S'. Kim and R.-H. Park
`PL Transformation for the Calculation of Filter Pairs for Perfect Reconstruction in Subbancl Coding with Linequincumt
`Subsampiing .
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`.
`. J. De Lrm-:ei'Hicm'e and G. Schamel
`
`,
`.
`if“ *:"°“ PAPER?
`I
`-. Illegrity of Public Telecommunication Networks .
`-
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`T
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`.
`
`RPX Exhibit 1034
`RPX Exhibit 1034
`RPX v. DAE
`RPX V. DAE
`
`

`
`iEi~:i-.' COMMUNICATIONS SOCIETY %'“
`
`®
`
`The field of interest of the IEEE Communications Society consists of all telecommunications including tetcphone. teiegraphy. facsimile. and point.
`point television. by electromagnetic propagation including radio; wire; aerial. underground. coaxial. and submarine cables; waveguides. communiq;
`satellites. and lasers; in marine. aeronautical. space. and fitted station services: repeaters. radio relaying. signal storage. and regeneration: tclecommuri
`(Lon error detection and correction: multiplexing and carrier techniques; communication switching systems: data communications; and communion
`t eory.
`-
`I
`_
`_
`In addition to the above. this JOURNAL or the IEEE TRANSACTIONS ON COMMUNICATIONS contains papers pertaining to anaiog and digital si
`processing and modulation. audio and video encoding techniques. the theory and design of transmitters. receivers. and repeaters for communications j
`optical and sonic media. the design and analysis of computer communication systems, and the development of COI't'lI1'I|lI'_lIl'.?ttI|.l:2Il1‘SOI‘l_.\lI'3I'e. Contributions-‘
`theory enhancing the understanding of communication systems and techniques are included. as are discussions of the social implications of the tlettelcipm
`of communication technology. All members of the IEEE are eligible For membership in the Society upon payment ofthe annual Society membership
`$15.00. Members may receive this JOURNAL or the IEEE TRANSACTIONS ON COMMUNICATIONS upon payment ofao additional Si 5.00 ($30.00 total)
`both publications upon payment ofan additional $30.00 {$45.00 total). For information onjoining. write to the IEEE at the address below. Membercopfg
`Trnn.rttcIion.rfJoitrntrt.r ore for personal use onty.
`
`'
`
`IEEE COMMUNICATIONS SOCIETY
`JOURNAL Editorial Board 1992
`
`S. B. WEINSTEIN. Director of Publicotioris
`Bellcore. Rm. 2L-28'?
`445 South Street
`Morristown. NJ 07960-1910
`
`A. M. BUSH
`National Science Foundation
`1800 G. Street. NW
`Washington. DC 20550
`
`N_ K_ CHEUNG
`Betlcore. Rm. NVC 32-2:9
`331 Newman S rings Rd.
`new-=u=.~: m»
`J_ I.-_ HAYES
`Dep. Elec. Eng.
`Concordia Univ.
`1455 De Maisonneuve.
`Montreal. Quebec.
`Canada H3G IM8
`
`M. KAWASHIMA
`Fujitsu Inst. Comp. Sci.
`1- I7-25 Shinkamatn
`Ota-Ku
`Tokyo 144-00. Japan
`
`5‘ """“'“°”°
`3:91‘ tSy5tl'J&. C°mP""" Eng’
`c:i.:..:i"...ii*.-...
`g;‘:a"g‘a-I?{'§- 536
`N. MAXEMCHUK
`AT&T Bell Labs.
`600 Mountain Ave.
`Murray Hill. NJ 0'i974
`
`W. H. TRANTER. Editor-in-Chief
`Dep. Elec. Eng.
`Univ. Missouri-Rolla
`Rolla. MO 6540i-0249
`tranter@ ee.ttmr.edu
`Senior Editors
`T. S. RAi=vAi=oit'r
`The Bradley Dep. Elec. Eng.
`6 I5 Whittemore Hall
`Virginia Polytech. Inst. St
`State Univ.
`BIB.CkSlJI.l!‘g, VA
`A. A. ROBROCK
`391:3...
`la
`.
`t
`ocquevi c
`20154 MIlaI1.Itaiy
`W. D. SINCOSKIE.
`I-lellcore, Rm. 2Q-286
`445 South Street
`Morristown. NJ 0'i960-l9l0
`
`SUE L. MCDONALD. Associate Edit-,3
`Bellcore. Rm. 2P-291
`445 South Street
`Motristown, NJ 07960-5910
`sue @ thumper.be|lcore.cont
`
`D. L. TENNENHOUSE
`M.I.'I'.. Rm. NE43-533
`545 Technology Square
`Cambridge. MA 02139
`
`R. Ynrsuaosiii
`Fujitsu Labs.
`l0l5 Kamiiiodzinaka Nzikahara-lru_..'
`Kawasaki 211. Japan
`
`II
`
`,3
`
`THE INSTITUTE OF ELECTRICAL AND ELECTRONICS ENGINEERS. INC.
`Ofllcers
`
`.
`ARVID G. LARSON. Vice Presidetrt. Professiorroi Acti‘vtti’es
`MERRILL W. BUCKLEY. J R.. President
`J. T. CAIN. Vice Prestdeirt. Pitbticntion Activities
`MARTHA SLOAN. President-Eiect
`LUIS T. GANDiA. Vice President. Regional Activities
`KARSTEN E. DRANGEID. Secretary
`MARCO W. MIGLIARO. Vice President. Standards Activities
`THEODORE W. HISSEY. J R.. Treasurer
`FERNANDO ALDANA. Vice President. Tecitnicoi Activities
`EDWARD A. PARRISH. Vice President. Educational Activities
`FREDERICK T. ANDREWS. JR.. Director. Dtotliiott III-—Commimr‘cotions Technology Division
`
`Headquarters Staff
`ERIC I-IERZ. Executive Director and General Manager
`THOMAS W. BARTLETT. Associate Genera! Monager—Ft'norice and Adniinr'.itrati‘on
`WILLIAM D. CRAWLEY. Associate Generai Mctnager—Progrnni.r
`JOHN H. POWERS. Associate General Ma-troger— Volunteer Services
`
`DONALD CHRISTIANSEN, Editor, IEEE Spectrum
`IRVING ENGELSDN, Staff Director. Technical Activities
`LEO FANNING. Staff Director, Professional Activities
`W. R. HABINGREITHER. Staff Director. Customer Service Center
`PHYLLIS HALL. Staff Director. Publishing Services
`
`MEi.vii~i I. OLKEN. Staff Director. I-‘ictd Services
`EDwARD ROSENBERG. Controiier
`ANDREW G. SALEM, Staff Director. Staridards
`RUDOLF A. STA MPFL. Staff Director. Educattotioi AcrivItie.t
`
`Publications Department
`Publications Managers: ANN H. BURGMEYER. GAIL S. FERENC
`Managing Editor: VALERIE CAMMARATA
`Associate Editor: SHARVN L. PERRY
`
`r
`
`IEEEJOURNALON SELECTED AREAS IN COMMUNICATIONS is published nine timesa year in January. February. April. May. June. August.SepternbBT_ti-
`October. and December by The Institute of Electrical and Electronics Engineers. Inc. Responsibility for the contents rests upon the authors and not upon
`IEEE. the SocietyfCouncil.‘or its members. IEEE Headquarters: 345 East 47 Street, New York. NY 1001?. NY Telephone: 212-705 + extension: Informs"-=
`tion -7900; General Manager -7910; Public Information -7867; Publishing Services -‘.’560; Sgectrum -‘I556. Telecopiers: NY (Headquarters) 212-'l'S2-41929.:
`gdY (Pubtiqations) 212305-76%: NY ;!I'§le§: 23g-4-ll {international messages only). [E E Service Center (for orders, subscriptions, address chqn es.
`ucationa Activities.
`egion
`ection
`tu ent crvices. Standards}: 445 Hoes Lane. PO. Bo
`I331. Facatawa . NJ 08855-1331. NJ Tele hone:
`in or-'5
`matiori: 908-931-0060; 903-56 + extension: Controller -5365; Technical Activities -3900; IEEEx\V8SIJtl1gll0l'| Officyc [for US. professional 3Ct.!l’ltllIcS)I 1323.-
`L Street. NW. Suite 1202. Washington. DC 20036-5104. Washington Telephone: 202~'iB5-0017. Price,’Puhli'cation Information: Individual copies: IEEB:
`members S I 0.00 (first copy only). nonmembers $20.00 per copy. (Note: Add $4.00 postage and handling charge to any order from $1.00 to $50.00. including
`prepaid orders.) Member and nonmember subscription prices available on request. Available in microfiche and microfilm. Copyright and Reprint Pei-nits}
`sions:Abstracting is permitted with credit to the source. L'braries ar
`'ttedt ptiotoco
`be
`it th i'mits of the U.S. C
`' ht I.
`f
`' ate set)
`patrons: 1) those post-19?? articles that carry a code at the bottortfdjfihg first ‘page. pl'0\["’IydEd’IIIl]; perfchpy fee indicated I?1Pl.yI‘Id.gC(!tieaI: piiui
`Iirrouglh the;
`Copyright Clearance Center. 29 Congress Street. Salem. MA 0l9'i0; 2) pre-l978_ articles with_out_fee. Instructors are permitted _to photocopy isolated;
`articles for noncommercial classroom use without fee. For all other copying. reprint. or republication permission. write to Copyri his and Permisstom
`Department, IEEE PublishinqServices. 445 I-Ioes-Lane. P. O. Box 1331. Piscataway. NJ 088554 33 I . Copyright (=3 1992 by The Institute of Electrical art'fl__-
`A
`Electronics Engineers. Inc.
`I rights reserved. Seoond«class postage paid at New York, NY. and at additional mailing offices. Postmaster: Send addrc5§l
`changes to IEEE JOURNAL ON ELECTED AREAS [N COMMUNICATIONS. IEEE.-145 Hoes Lane. P.0. Box I331. Piscataway. NJ 08855-1331. GST.
`Registration No. 125634188.
`
`

`
`This material may be protected by Copyright law (Title 17 U.S. Code)
`
`

`
`(‘HEN H ni.:
`
`.i'\ LOW-DEL.-\Y CELP CODER
`
`R3l
`
`speech coding standards now already exist for specific ap-
`plications. To avoid proliferation of different 16 kb/s
`standards and the potential difficulty of interworking be-
`tween diiferent standards, "in June 1988 the CCITT de-
`
`cided to investigate the possibility of establishing a single
`16 kb/s speech coding standard for universal applica-
`tions. The CCITT‘s intended applications include video-
`phone, cordless telephone, digital satellite systems, Dig-
`ital Circuit Multiplication Equipment (DCME), Public
`Switched Telephone Network (PSTN), Integrated Service
`Digital Network (ISDN), digital leased lines, voice store
`and forward systems, voice messages for recorded an-
`nouncements,
`iancl-digital mobile
`radio,
`packetized
`speech, etc. [[2].
`Because of the variety of applications this 16 kb/s
`standard has to serve, the CCITT determined a stringent
`set of performance requirements and objectives for the
`candidate coding algorithms. Not every application would
`need every requirement to be met. Yet, to be accepted by
`the CCITT as a universal 16 kb/s standard, a candidate
`algorithm must meet all of the requirements. The major
`requirement was that the speech quality should be roughly
`comparable to that of G.'i'21 while the one—way encoder!
`decoder delay should not exceed 5 ms (the objective was
`52 ms) [12]. In other words, the CCITT was looking for
`a toll—quality low-delay speech coder at 16 kb/s.
`The CCITT specifies the speech quality requirements
`in terms of qdu, or the quantization distortion unit. By
`definition, one encoding with a 64 lcb/s G.7ll PCM co-
`dec introduces 1 qdu of distortion (CCITT Recommen-
`dation (3.1 13). For asynchronous tandem connections, the
`qdu of individual speech coders is Supposed to be addi-
`tive. For example, N asynchronous tandeming stages of
`G71 1 codees should result in N qdu. Another example is
`that a single encoding ofa 32 kb/s G.72l ADPCM codec
`is rated at 3.5 qdu, and therefore two ADPCM encodings
`should be rated at ? qdu and four encodings at 14 qdu.
`The CCITT performance requirements for the 16 kb/s
`standard specify that, for a clear channel (i.e., no hit er-
`rors), a candidate coder should produce 4 qdu or less for
`a single encoding and 14 qdu or less for three asynchro-
`nous tandeming stages. In effect, this says that a candidate
`coder can be slightly worse than G321 ADPCM for a
`single encoding, and three asynchronous encodings of the
`candidate coder should match the speech quality of four
`asynchronous encodings of G.?2l ADPCM. For noisy
`channels, the CCITT required that, for random bit errors
`at a bit error rate (BER) of 104 or 10-2, a candidate coder
`should produce decoded speech quality not worse than that
`of G.'r'2l ADPCM under the same conditions.
`In addi-
`
`tion, the coder should pass network signaling tones such
`as DTMF and CCITTSignaiing Systems No. 5, 6, and 7.
`It was not so difficult to meet each one of these quality
`requirements individually. However,
`in 1988.
`it was a
`major challenge to create a 16 kb/s coder that would meet
`all of these requirements simultaneously. Furthermore, the
`addition of the 5 ms low-delay requirement made such an
`attempt even more difficult.
`
`Because of the low—delay requirement. none of the 16
`kb/s coders mentioned above (CE-l_.P, MPLPC, AFC,
`ATC, and SBC-ADPCM) could be used in their current
`form. With all these well-established coders ruled out, the
`only hope seemed to be backiirrii'rI-ridripriirc predictive
`coders which derive their predictor coefiicients from pre-
`viously quantized speech and thus do not need to buffer a
`large frame of speech samples.
`(The G.'?2l ADPCM
`coder belongs to this category.)
`Prior to the CClTT‘s recent standardization effort at 16
`kb/s, several researchers had previously reported their
`work on low-deiay speech coding at 16 kb/s. Jayant and
`Rarnamoorthy I13}, [14] used an adaptive posttilter to en-
`hance l6 kb/s ADPCM speech and achieved a mean
`opinion score (MOS) [15] of 3.5 at nearly zero coding
`delay. Cox er in‘. [16] combined SBC and vector quanti-
`zation (VQ) and achieved an MOS of roughly 3.5-3.3’ at
`a coding delay of 15 ms. Berouti er at. {l?] reduced the
`coding delay of an MPLPC coder to 2 ms by reducing the
`frame size to 1 ms. However,
`the speech quality was
`equivalent to 5.5—bit log PCM—a significant degradation.
`Taniguchi er of. [18] developed an ADPCM coder with
`rnultiquantizers where the best quantizer was selected once
`every 2.5 ms. The coding delay of their real-time proto-
`type codec was 8.3 ms. With the help of postfiltering, the
`coder produced speech quality “nearly equivalent to a
`7-bit p.~1aw PCM" [18]. but this was achieved with a non-
`
`standard 6.4 kHz sampling rate and a resulting nonstan-
`dard bandwidth of the speech signal. Gibson et al. [19],
`[20] studied backward-adaptive predictive tree coders and
`predictive treliis coders which should have low coding de-
`lays. Unfortunately, they did not report the exact coding
`delay or the subjective speech quality. Iyengar and Kabal
`[21] also developed a backward-adaptive predictive tree
`coder with a 1 ms coding delay and a level of speech qual-
`ity equivalent to T-bit log PCM. Watts and Cuperman [22]
`proposed a vector generalization of ADPCM with a delay
`between 1 and 1.5 ms. They did not report the subjective
`speech quality of the coder.
`Since 1988, when the CCITT announced its intention
`
`to standardize a 16 itbfs low—delay speech coder, there
`has been a great deal of research activity in the area of
`low-delay speech coding at 16 kb/s [23]—[38]. In re-
`sponse to the CClTT’s standardization effort, we have
`created a 16 kb/s coder called low-delay CELP, or LD-
`CELP, which achieves high speech quality with a one-
`way coding delay less than 2 ms [24], [29], [32], [33],
`[35], [38].
`
`The LD-CELP coder is a predictive coder that com-
`bines: 1) high—order backward-adaptive linear prediction;
`2) backward gain—adaptive vector quantization [39], [40]
`for excitation; 3)
`the analysis-by-synthesis excitation
`codebook search of CELP; and 4) adaptive postfiltering
`[14], [41]. The low coding delay is achieved by using
`backward-adaptive prediction to avoid the long speech
`buffer required by forward—adaptive prediction, and by
`using a small excitation vector size of oniy five sampies,
`or 0.625 ms (assuming the standard 8 kHz sampling rate).
`
`

`
`832
`
`IEEE JOURN.-\L ON SELECTED ARE.-\S IN C().\'iML‘N|(?.-\'|‘IOi\".'§. VOL.
`
`ill. NO. 5. JUNE lull}
`
`With the processing delay and transmission delay also in-
`cluded,
`the total one-way coding delay can be less than
`2 ms. This not only surpasses the CCITT delay require-
`ment of 5 ms but actually meets" the objective of 2 ms.
`This LD-CELP coder was submitted by AT&T to the
`CCITT and has been the only candidate coder since 1989.
`This coder has been implemented in real—time hardware
`using the AT&T WEE” DSP32C floating—point digital sig-
`nai processor, and the resulting hardware prototype
`LD-CELP coder has been used in the ofhcial CCITT lab-
`
`oratory tests.
`In the standardization process, there were two phases
`of laboratory testing. The first phase of testing was con-
`ducted in late 1989 and early I990. while the second phase
`was in early I991. The l_.D-CELP coder submitted for the
`first phase of testing (called the Phase i coder from here
`on) met all of the CCITT’s performance requirements ex-
`cept fo1' the requirement of three asynchronous tandems.
`Based on the Phase 1
`test results,
`the Speech Quality
`Experts Group (SQEG} of the CCITT indicated that the
`LD-CELP coder could be standardized for point-to-point
`applications but not for networking applications where
`tandeming may occur. unless the code1' could be improved
`to meet
`the tandeming performance requirement
`in the
`Phase 2 test.
`
`In late 1990 to early 1991, we improved the LD-CELP
`coder‘s tanderning performance significantly and pro-
`duced what we called the Phase 2 coder. The hardware
`
`prototype Phase 2 coder was then tested in the second
`phase of laboratory testing in 1991. From the Phase 2 test
`results, the SQEG concluded that the 16 kb/s LD—CELP
`coder “has at petformrmce eqnivrrieiit to or better than
`G. 72!." and “meets (iii the speech qnriiity requirements
`set by Sttidy Group XV and tested by Study Group XH "
`[42]. Therefore.
`“the SQEG recommends’
`that
`the
`I6 kb / 5 LD- CELF codec can be .5‘i(t.Jt(iat‘(iiZ€d as (I CC1TT
`G Series Recommenriatian as regards to its speech quai-
`ity” [42]. According to the current standardization sched-
`ule, this 16 kb/s LD-CELP coder is expected to be stan-
`dardized by the first half of 1992.
`In this paper, we will describe the 16 kb/s LD-CELP
`coding algorithm,
`its
`implementation, and its perfor-
`mance. Section II introduces system concepts and pro-
`vides an overview ofthe LD-CELP coder. Section III de-
`
`the LD-CELP coding algorithm. Section IV
`scribes
`discusses the implementation issues. Section V describes
`the subjective and objective performance, and Section VI
`gives some concluding remarks.
`
`II. SYSTEM CONCEPTS AND Ovr-:RvtEw
`
`In this section, we review the conventional CELP al-
`gorithm [l] and then give an overview of the LD-CELP
`algorithm and point out the differences between conven-
`tional CELP and LD-CELP. Along the way. we also dis-
`cuss the issue of coding delay.
`A. Review 0fCon-uentioiiai CELP
`
`A typical example of the conventional CELP speech
`coder is shown in Fig. l. The CELP coder is based on the
`
`“source—filter” speech production model [43], with the
`short—term synthesis filter modeling the vocal tract and the
`excitation VQ, together with the long-term synthesis fi1_
`ter, modeling the glottal excitation. The CELP coder syn-
`thesizes specch by passing a gain-scaled excitation se— '
`quencc through long—term and short-term synthesis filters,
`Both synthesis filters are all-pole filters containing either
`a long-term or a short-term predictor in a feedback loop_
`Basically,
`the CELP coder encodes speech frame—by-
`frame, and within each frame it attempts to find the best
`predictors, gain, and excitation such that a perceptually
`weighted mean-squared error (MSE} between the input
`speech and the synthesized speech is minimized.
`The long-term predictor is often referred to as the pitch
`predictor, because its main function is to exploit the pitch
`periodicity in voiced speech. Typically, a one—tap pitch
`predictor is used.
`in which case the predictor transfer
`function is:
`
`Pi(z) = fiz _"’
`
`(1)
`
`where p is the bulk delay or pitch period, and .6 is the
`predictor tap. The short—term predictor is sometimes re-
`ferred to as the LPC predictor, because it is also used in
`the well-known LPC {linear predictive coding) vocoders
`which operate at 2.4 kb/s or below. The LPC predictor
`is typically a 10th-order predictor with a transfer function
`of:
`
`It}
`
`‘Pita =
`
`a.-z
`
`(2)
`
`through din are the predictor coefficients. The
`where ti.
`excitation VQ codebook contains a table of codebook vec-
`tors (or cotievectors) of equal length. The codevectors are
`typicaily populated by Gaussian random nttmbers with
`possible center clipping.
`In the actual encoding process, the encoder first buffers
`an input speech frame of about 20 ms or so, and then
`performs linear predictive analysis [43] (or LPC niiniysis)
`on the buffered speech. The resulting LPC parameters are
`then quantized. The pitch predictor parameters, including
`the pitch period and the predictor tap. are then determined
`either in an open-loop fashion [1] or in a closed-loop fash-
`ion [44]. The quantized LPC parameters and pitch pre-
`dictor parameters are both sent as side infonnation to the
`decoder. This scheme is called for'1vard—adciptive predic-
`tion.
`
`The input speech frame is further subdivided into sev-
`eral equa1—length .sttbft'ames, or vectors, typically of size
`4 to 8 ms. Then, for each vector. the encoder passes each
`candidate codevector
`in the excitation VQ codebook
`through the gain scaling unit and the two synthesis filters.
`and then compares the corresponding filtered output vec-
`tor with the input speech vector and computes the asso-
`ciated perceptually weighted MSE distortion. The en-
`coder repeats this process for all candidate excitation
`codevectors and then identifies the codevector that mini-
`
`mizes the perceptually weighted MSE distortion. This
`
`

`
`t_‘llE|\' H :r!.: A LO\\r’—DEL.-KY CELP CODI.-ZR
`
`Encode
`and
`multiplex
`
`LPC
`analysis &
`quantization
`
`Excitation
`VQ
`codebook
`
`Long-term
`synthesis
`filter
`
`{bl
`
`Short—term
`
`synthesis
`fitter
`
`Fig.
`
`I.
`
`(a) Typical conventional CELP encoder. ([1,) Corresponding CBLP
`decoder.
`
`is
`process is called a cl'osed—foop .rearci'r {sometimes it
`called an andIysfs-by—s_yr1rhesis procedure). Vector quan-
`lization of the excitation using a closed-loop search is the
`main feature of CELP;
`it
`is also the main reason why
`CELP coders outperform other linear predictive coders
`such as APC and MPLPC.
`
`Theoretically, it is possible to jointly optimize the ex-
`citation codevector and the gain.
`In practice, however,
`almost all conventional CELP coders separately quantize
`the gain subsequent
`to the closed-loop excitation VQ
`codebook search. There are at least two reasons for. this.
`
`First, performing such sequential optimization produces
`a much lower computational complexity than a joint op-
`timization approach. Second, with the gain typically
`quantized to 5 b or so, the resolution in gain quantization
`is high enough that the performance difference between
`the two approaches is essentially negligible.
`
`In conventional CELP. five different kinds of infor-
`mation are encoded and sent to the decoder. These are:
`
`l) the LPC parameters; 2) the pitch period; 3) the pitch
`predictor tap; 4) the excitation gain; and 5) the excitation
`“shape“ codevectors. The decoder decodes such infor-
`mation and reproduces speech frame—by—frame by exciting
`the two cascaded synthesis filters with the scaled excita-
`tion vector sequence. Usually, a postfiiter of the type pro-
`posed in [4]] is used in a CELP decoder to enhance the
`perceptual quality of decoded speech.
`
`B. Corfiirg Defrtry
`
`In the Introduction. we mentioned that a conventional
`
`CELP coder typically has a one—way coding delay of 50
`to 60 ms. The one~way coding delay is defined as the
`elapsed time from the instant a speech sample arrives at
`
`

`
`IEEE JOURNAL ON SI-Il.l-LC'|‘ED AREAS IN COMl\-lUNI("ATl(}NS. VOL.
`
`ltl. NO. 5. JUNIE W92
`
`the encoder input to the instant when that same speech
`sample appears at the decoder output, less any delay added
`by other communication equipment (such as modems) in
`between the encoder—decode‘r pair. and the signal propa-
`gation delay which depends on distance. In other words,
`it
`is as if the encoder and the decoder were connected
`
`back-to-back by wires at the same physical location with-
`out any equipment in between. This definition makes the
`coding delay dependent only on the coding algorithm. not
`on other equipment or communication distance.
`By such definition,
`the coding delay of CELP coders
`can be roughly determined in terms of the speech frame
`size used. The coding delay consists of three kinds of de-
`lay: 1) algorithmic buffering delay; 2) processing delay;
`and 3) bit transmission delay. These delays are explained
`below. First, due to the forward-adaptive LPC analysis,
`the CELP encoder first has to buffer one frame of speech
`samples before the encoding of the first sample in that
`frame can be started. Such buffering introduces at least
`one frame worth of buffering delay. Second, assume the
`real—time hardware is just fast enough to run the coder in
`real—time (which is usually the case). Then,
`it will take
`almost one frame worth of processing delay to perform
`the encoding and decoding of the buffered speech frame.
`Third. suppose the encoder does not start sending bits cor-
`responding to a given speech frame until the encoding of
`the entire frame is completed and the decoder does not
`start decoding a speech frame until all of the bits of that
`frame have been received, then one additional frame of
`
`bit transmission delay will be introduced. This is because
`it is one frame worth of time from the instant the first bit
`of the frame is sent to the instant the last bit of the frame
`
`is received. provided that the bit rate of the communica-
`tion channel is the same as the bit rate of the speech coder.
`Hence, the total one-way coding delay of a CELP coder
`is roughly three frames.
`Of course, the above delay analysis is oversimplified.
`In practice,
`it is possible to reduce the processing delay
`without using a faster processor. This can be achieved by
`sending bits out “on-the-fly“ at the encoder as soon as
`certain bits become available, and by decoding bits “on-
`the-fly" at the decoder as soon as the received bits are
`sufficient to start decoding the first speech sample in the
`frame. (Some constraints on bit timing have to be care-
`fully observed, of course.) Having a fast processor can
`further out down the processing delay. The use ofa faster
`communication channel (e.g.,
`in multiplexed communi-
`cation systems) can also reduce the bit transmission de-
`lay. On the other hand. for ease of implementation, some-
`times a designer may choose to have a coding delay longer
`than three frames. Therefore, the actual coding delay is
`likely to vary between two and four frames, depending on
`the actual implementation and application. If we take 2.5
`to 3 frames as the average, then a CELP coder with a 20
`ms frame size is likely to have a coding delay of 50 to 60
`ms. This long coding delay is mainly due to the large 20
`ms frame buffer that is required by the forward adaptation
`of LPC predictor coefficients.
`
`C. Over-New of LD-CELP
`
`In LD-CELP. we reduce the coding delay by making
`the LPC predictor “backward-adaptive." (See [15, ch. 4,
`6] for a comprehensive discussion of forward and back-
`ward adaptation.) This means that
`the predictor coeffi-
`cients will not be derived from the input speech samples
`yet
`to be coded, but rather from previously quantized
`speech samples. Since previously quantized speech sain-
`ples are also available at the decoder, the decoder can also
`derive the predictor coefiicients using the same procedure
`used in the encoder. Thus, there is no need to transmit
`
`any side information bits to specify the predictor coeffi-
`cients. More importantly, with backward adaptation, there
`is no need to buffer about 20 ms of input speech for for»
`ward-adaptive LPC analysis. Therefore,
`the excitation
`vector (subframe) becomes the basic buffer unit, and the
`one-way coding delay becomes roughly three times the
`vector size (or vector dimension}. Since the vector di-
`
`mension is much smaller than 20 ms, the coding delay is
`greatly reduced.
`From the very beginning, our goal was to bypass the 5
`ms delay requirement and to achieve the delay objective
`of2 ms or less. The following simple analysis shows that
`we also needed to make the excitation gain backward-
`adaptive in order to achieve this goal. With the 8 kHz
`standard sampling rate, a delay of2 ms corresponds to 16
`samples. Since the coding delay is roughly three times the
`vector dimension, to achieve a one-way delay of 16 sam
`ples or less, the largest vector dimension we can use is
`five samples (0.625 ms). With a vector dimension of five
`samples and a bit rate of two bits/sample, we only have
`10 bits to encode the excitation gain as well as the exci-
`tation shape codevector. The excitation gain typically re-
`quires four to five bits of resolution in conventional CELP.
`In our case, however, it would be highly inefficient to use
`40-50% of the total bit rate to encode the slowly varying
`and somewhat redundant gain term. To achieve better
`coding efiiciency and speech quality under the 2 ms delay
`constraint, we use backward g(ifll-ftditpffve vector quan-
`tization [40] for the excitation. By making the excitation
`gain backward-adaptive. we derive the excitation gain
`from the gain information embedded in previously quan-
`tized excitation, and there is no need to send any bits to
`specify the excitation gain, since the decoder can derive
`the same gain in the same manner.
`A simplified block diagram of the 16 kb/s LD-CELP
`coder is shown in Fig. 2. In this coder. the excitation vec-
`tor has a dimension of five samples, so the one-way cod-
`ing deiay is less than 2 ms. The pitch predictor in con-
`ventional CELP is eliminated, and a 50th—order LPC
`predictor is used. The LPC predictor coefficients are up-
`dated once every four speech vectors (2.5 ms) by per-
`forming LPC analysis on previously quantized speech.
`The excitation gain is updated once every vector by using
`a 10th-order adaptive linear predictor in the logarithmic
`gain domain. The coefficients of this log-gain predictor
`are updated once every four vectors by performing linear
`
`

`
`CHEN 2:
`
`.m’.: A LOW-DELAY CELP CODER
`
`Excitation
`VQ
`codebook
`
`gain
`adaptation
`
`Fig. 2.
`
`(El) 16 kb/s |ow—de|ay CELP encoder. (b) 16 kb/s low-delay CELP
`decoder.
`
`predictive analysis on the logarithmic gains of previously
`quantized and scaled excitation vectors. The 10th-order
`perceptual weighting filter is also updated once every four
`vectors, but it is updated by performing an LPC analysis
`on the input speech. To reduce the codebook search com-
`plexity,
`the 10-bit excitation VQ codebook is a product
`code of a three-bit gain codebook and a seven—bit shape
`codebook.
`
`Ideally, we should be able to achieve better coder per-
`formance if we update all predictors and filters once a
`vector. However, more frequent updates result in higher
`computational complexity. To make it possible for real-
`time implementation with the DSP32C, we found it nec-
`essary to reduce the update rate to once every four vec-
`tors. Fortunately,
`the spectral envelope of speech does
`not change very rapidly with time, so the less frequent
`updates of predictors does not cause noticeable degrada-
`tion in speech quality.
`It should be emphasized that, although all predictors
`
`and filters are updated once every four vectors (20 sam-
`ples),
`in order to achieve a low coding delay the basic
`buifer size of the LD-CELP coder is still limited to only
`one vector