`RPX v. DAE
`
`
`
`REGIONAL OFFICES
`
`Japan
`Japan Section
`c/o Japan Audio Society, Mori
`Building
`14-34, 1-chome, Jingumae,
`Shibuya-ku,
`Tokyo 158, Japan
`Tel: 81 -3-3403-6649
`Fax: 81-3—3404-6549
`
`’
`
`Europe
`AES Europe Region
`Zevenbunderslaan 142/9
`B-1190 Brussels, Belgium
`Tel: 322-345-7971
`Fax: 32-2-345-3419 ,
`
`United Kingdom
`British Section
`Audio Engineering Society Ltd.
`P. O. Box 645
`Slough, SL1 8BJ
`United Kingdom
`Tel: Bumharn 44 (0) 628 663725
`Fax: 44 (O) 628 667002
`
`AES REGIONS AND SECTIONS
`
`Eastern Region
`Sections: Atlanta, Boston,
`District of Columbia, New
`York, Philadelphia, Toronto.
`Student Sections: Duquesne
`University, Fredonia, Full Sail
`Center for Recording Arts,
`McGill University,
`Pennsylvania State University,
`University of Massachusetts-
`Lowell, University of Miami.
`
`Central Region
`Sections: Chicago, Detroit,
`lndianapoiis, Kansas City,
`Nashville, St. Louis, Upper
`Midwest, West Michigan.
`Student Sections: Belmont
`University, Hutchinson
`Technical College,
`Lawrence Technological
`University, Michigan
`Technological University,
`Middle Tennessee State
`University, Ohio University.
`
`Western Region
`Sections: Albuquerque/Santa
`Fe, Los Angeles, Pacific North-
`west, Portland, San Diego, San
`Francisco, Southern Arizona,
`Utah, Vancouver.
`Student Sections: American Riv-
`er, California State University-
`Chico, Cogswell Polytechnical
`College, Denver, Long Beach
`City College, Loyola Marymount
`University, San Francisco State
`University, University of South-
`ern Caliiornia.
`
`Europe Region
`Sections: Austrian, Belgian,
`Belgrade, British, Bulgarian,
`Central German, Croatian,
`Czech, Danish, Finnish, French,
`Greek, Hungarian, Iberian, Israel,
`Italian, Netherlands, North Ger-
`man, Norwegian, Polish, Roma-
`nian, Russian, South German,
`Slovakian Republic, Slovenian,
`Swedish, Swiss.
`
`International Region
`Sections: Adelaide,
`Japan, Melbourne, Mexican,
`Philippine, Sydney, Venezuelan.
`
`PURPOSE
`
`The Audio Engineering Society is organized for the pur-
`pose of: uniting persons performing professional services
`in the audio engineering field and its allied arts; collecting,
`collating, and disseminating scientific knowledge in the
`field of audio engineering and its allied arts; advancing
`such science in both theoretical and practical applications;
`and preparing, publishing, and distributing literature and
`periodicals relative to the foregoing purposes and policies.
`
`MEMBERSHIP
`Individuals who are interested in audio engineering may
`become members of the Society. Applications are consid-
`ered by the Admissions Committee.
`.
`Grades and annual dues are as follows: Full member,
`$65.00; Associate member, $65.00; Student member,
`$35.00. A membership application form may be obtained
`from headquarters.
`~
`Sustaining memberships are available to persons, corpo-
`rations, or organizations who wish to support the Society.
`A subscription to the Journal isincluded with all
`memberships.
`’
`R
`-
`
`AES-
`
`--
`
`-A 00
`
`0
`
`"'
`
`O
`
`O
`
`"
`
`(ISSN 0004-7554), Volume 42, Number 10, 1994 October
`Published monthly, except January/February and July/August when published bi-
`monthly, by the Audio Engineering Society, 60 East 42nd Street, New York, New
`York 10165-2520, USA, Telephone 212-661-2355. Fax 212-682-0477. Second-
`class postage paid at New York, New York, and at an additional mailing office.
`Postmaster: Send address corrections to Audio Engineering Society, 60 East 42nd
`Street, New York, New York 10165-2520.
`
`The Audio Engineering Society is not responsible for statements made by its
`contributors.
`
`EDITORIAL STAFF
`
`Daniel R. von Recklinghausen Editor
`Patricia M. Macdonaid Executive Editor
`Ingeborg M. Stochmal
`Copy Editor
`Abbie J. Cohen Senior Editor
`Patricia L. Sarch Art Director
`G. Franklin Montgomery
`Consulting Technical Editor
`Flavia Elzinga Advertising
`Gerri M. Calamusa Associate Editor
`Stephanie Paynes Writer
`Robert O. Fehr Editor Emeritus
`
`;;
`
`George L. Augspurger
`Jeffrey Barish
`Jeff)’ Ballck
`SarenBech
`Durand Begault
`piet J_ Berkhout
`Barry A' messer
`‘ John Bradiey
`“°“" "' B“°"°’5
`Mahlon o. Burkhard
`,;
`.
`R‘°"“"d C‘ cm‘
`3?; Marvin Camras
`Duane H. Cooper
`" Robert R. Cordell
`i
`Andrew Duncan
`John M. Eargle
`Edward J. Foster
`Mark B. Gardner
`Earl R. Geddes
`
`'
`
`REVIEW BOARD
`John T. Mullin
`Richard A. Greiner
`Marlin P0l0n
`David Griesinger
`Malcolm 0. J. I-Iawkstord D-Pfels
`John M. Hollywood
`?a"'°!°g°°"
`1-
`|-
`H t
`l3“°'5 Umsell
`A32 "fi‘°.nKa?zg:an
`Kees Schouhamer
`James M. Kates
`Immmk
`D B Keele Jr
`‘er
`David L Kle
`'
`. 9*’
`Paul W. Klipsch
`W Marsh ll Leach Jr
`‘
`3 _
`_
`’
`Stanley P. Lipshitz
`_
`James F‘ M°G'" _
`J‘ G‘
`McKmgh‘
`G w M N ll
`“V -
`C 3 Y
`D-J- M93793
`Robert A. M009
`James A. Moorer
`
`R? iteardi-i §m”r?'"
`°
`‘
`3
`Herman J. M. Steeneken
`John Strawn
`G H B b Th
`d
`“““°"
`-
`-( ° )
`Floyd E30019
`Em“ |__ Torick
`John vanderkoo
`-
`Dame|R_von
`Recklinghausen
`Rhonda Wilson
`Wieslaw V. Woszczyk
`
`V
`
`COPYRIGHT
`Copyright © 1994 by the Audio Engi-
`neering Society, Inc. It is permitted to
`quote from this Journal with custom-
`ary credit to the source.
`
`:} COPIES
`Individual readers are permitted to
`photocopy isolated articles for re-
`search or other noncommercial use.
`Permission to photocopy for internal
`or personal use of specific clients is
`granted by the Audio Engineering So-
`ciety to libraries and other users regis-
`tered with the Copyright Clearance
`Center (CCC), provided that the base
`fee of $1.00 per copy plus $0.50 per
`page is paid directly to CCC, 21
`Congress Street, Salem, MA 01970,
`USA. 0004-7554/93. Photocopies of
`individual articles may be ordered
`from the AES Headquarters office at
`$5.00 per article.
`
`REPRINTS AND REPUBLICATION
`Multiple reproduction or republica-
`tion of any material in this Journal re-
`quires the permission of the Audio
`Engineering Society. Permission may
`also be required from the author(s).
`Inquiries should be sent to the AES
`Editorial office.
`‘
`
`SUBSCRIPTIONS
`The Journal is available by subscrip-
`tion. Annual rates are $125.00
`surface mail, $170.00 air mail. For
`further information, contact AES
`Headquarters.
`
`BACK ISSUES
`Selected back issues are available:
`From Vol. 1 (1953) through Vol. 12
`(1964), $10.00 per issue (members),
`$15.00 (nonmembers); Vol. 13 (1965)
`to present, $6.00 per issue (mem-
`bers), $11.00 (nonmembers). For in-
`formation, contact AES Headquarters
`office.
`.
`
`MICROFILM
`(1971 Jan-
`1
`Copies of Vol. 19, No.
`uary) to the present edition are avail-
`able on microfilm from University Mi-
`crofilms International, 300 North
`Zeeb Rd., Ann Arbor, MI 48106,
`USA.
`
`ADVERTISING
`Contact the AES Editorial office.
`
`MANUSCRIPTS
`For information on the presentation
`and processing of manuscripts, see
`Information for Authors.
`
`
`
`(cid:3) (cid:55)(cid:75)(cid:76)(cid:86)(cid:3)(cid:80)(cid:68)(cid:87)(cid:72)(cid:85)(cid:76)(cid:68)(cid:79)(cid:3)(cid:80)(cid:68)(cid:92)(cid:3)(cid:69)(cid:72)(cid:3)(cid:83)(cid:85)(cid:82)(cid:87)(cid:72)(cid:70)(cid:87)(cid:72)(cid:71)(cid:3)(cid:69)(cid:92)(cid:3)(cid:38)(cid:82)(cid:83)(cid:92)(cid:85)(cid:76)(cid:74)(cid:75)(cid:87)(cid:3)(cid:79)(cid:68)(cid:90)(cid:3)(cid:11)(cid:55)(cid:76)(cid:87)(cid:79)(cid:72)(cid:3)(cid:20)(cid:26)(cid:3)(cid:56)(cid:17)(cid:54)(cid:17)(cid:3)(cid:38)(cid:82)(cid:71)(cid:72)(cid:12)(cid:3)
`This material may be protected by Copyright law (Title 17 U.S. Code)
`
`PAPERS
`
`ISO-MPEG-1 Audio: A Generic Standard for Coding
`of HighQuality Digital Audio*
`
`KARLHEINZ BRANDENBURG, AES Fellow
`
`FhG—lIS, Erlangen, Germany
`
`AND
`
`GERHARD STOLL
`
`Institutfiir Rundfunktechnik, Munich, Germany
`
`Coauthors:
`
`Yves—Frang:ois Dehéry, C.C.E.T. T, France
`James D. Johnston, AES Member, AT&T, Bell Laboratories, USA
`Leon v.d. Kerkhof, Philips, Netherlands
`Ernst F. Schréder, AES Member, Thomson Consumer Electronics, Germany
`
`The standardization body ISO/IEC/JTC1/SC29/WG1l (Moving Pictures Expert Group,
`MPEG) was drafting a standard for compressing the high bit rate of moving pictures and
`associated audio down to 1.5 Mbit/s. The audio part of the proposed standard is described.
`Three layers of the audio coding scheme with increasing complexity and performance
`were defined. These layers were developed in collaboration mainly with AT&T, CCETT,
`FhG/University of Erlangen, Philips, IRT, and Thomson Consumer Electronics. The
`generic coding system is suitable for different applications, such as storage on inexpensive
`storage media or transmission over channels with limited capacity (such as digital audio
`broadcasting or ISDN audio transmission).
`
`0 INTRODUCTION
`
`The necessity to specify a generic Video and audio
`coding scheme for many applications dealing with digi-
`tally coded video and audio and requiring low data rates
`has led the ISO/IEC standardization body to establish
`the ISO/IEC JTC 1/SC29/WG1 1 , called MPEG (Moving
`Pictures Experts Group). This group had the task to
`compare and assess several digital audio low-bit-rate
`coding techniques in order to develop an international
`standard for the coded representation of moving pic-
`tures, associated audio, and their combination when
`used for storage and retrieval on digital storage media
`
`* Presented at the 92nd Convention of the Audio Engi-
`neering Society, Vienna, Austria, l992‘Ma_rch.24-27; revised
`1994 July 15.
`.
`
`
`
`780
`
`(DSM). The DSM targeted by MPEG include CD-ROM,
`DAT, magneto—optical disks, and computer disks, and
`it is expected that MPEG-based bit-rate reduction tech-
`niques will be used in a variety of communication chan-
`nels such as ISDN and local area networks and in broad-
`casting applications. The international standard ISO/IEC
`11172 “Coding of Moving Pictures and Associated
`Audio for Digital Storage Media at up to about 1.5 Mbit/s”
`was finalized in November 1992 and consists of three
`
`parts: system, video, and audio [1]. The system part
`(11172-1) deals with synchronization and multiplexing
`of audio—visual information, whereas the video (11172-
`2) and audio (11172-3) parts address the video and the
`audio bit-rate reduction techniques, respectively. This
`standard is also known as the MPEG-1 standard.
`
`MPEG-2 Audio is the consequent extension from two
`to five audio channels providing backward compatibility
`
`J. Audio Eng. Soc., Vol. 42, No. 10, 1994 October
`
`
`
`PAPERS
`
`CODING OF HIGH-QUALITY DIGITAL AUDIO
`
`to MPEG-1. The main aspects are high quality of five
`( + 1) audio channels, low bit rate and backward compat-
`ibility—-the key to insuring that existing 2-channel de-
`coders will still be able to decode compatible stereo
`information from five (+ 1) multichannel signals.
`This standard, which is expected in November 1994,
`is based on standards and recommendations from inter-
`
`national organizations such as_ ITU-R, SMPTE, and
`EBU. International standardization bodies will insure
`
`the highest audio signal quality by extensive testing.“
`For audio reproduction the loudspeaker positions left,
`center, right, left and right surround are used, according
`to the 3/2-standard.
`
`1 STANDARDIZATION AND QUALITY
`ASSESSMENTS WITHIN MPEG-1 AUDIO
`
`Since 1988 ISO/MPEG has been undertaking the stan-
`dardization of compression techniques for video and as-
`sociated audio. The main topic for standardization in
`MPEG was video coding together with audio coding
`for DSM. On the other hand the audio coding standard
`developed by this group was the first international stan-
`dard in the field of digital audio compression and is
`expected to be followed in different applications. Beside
`several subgroups such as video, system, test, imple-
`mentation, requirement, and DSM, the audio subgroup
`of MPEG had the responsibility for developing a stan-
`dard for coding of PCM audio signals with sampling
`rates of 32, 44.1, and 48 kHz at bit rates in a range of
`32-192 kbit/s per mono and 64-384 kbit/s per stereo
`audio channel. The operating modes are
`
`- Single channel
`* Dual channel, like bilingual
`- Stereo
`
`0 Joint stereo (combined coding of left and right chan-
`nels of a stereophonic audio program)
`
`Table 1 gives a short general View of the milestones
`of the MPEG-AUDIO group. This group asked for pro-
`posals for the audio coding standard in mid-1989, and 14
`proposals were submitted for this purpose. The original
`proposals were grouped into four clusters according to
`algorithmic similarities. The clustered candidate algo-
`rithms were called ASPEC, ATAC, MUSICAM, and
`SB/ADPCM.
`
`A number of subjective tests were performed [2]—[4]
`since mid-1990 to assess the audio quality of the ISO/
`MPEG/Audio coding standard. During this time period
`several improvements have been made to meet the pres-
`ent audio quality. The important milestones in the devel-
`opment of the standard have been the official tests orga-
`nized by the Swedish Broadcasting Corporation in
`Stockholm under the auspices of ISO and EBU. In July
`1990 large listening tests and objective evaluations, such
`as basic audio quality at different bit rates, sensitivity to
`transmission bit errors , encoder and decoder complexity ,
`and coding delay, were performed-jon prototype real-
`time implementations of the four clustered algorithms.
`
`J. Audio Eng. Soc., Vol. 42, No. 10, 1994 October
`
`Both the ASPEC and MUSICAM proposals have shown
`a very high subjective quality at bit rates of about 100
`kbit/s per channel.
`.
`Due to the result that the proposals of the ASPEC
`and MUSICAM groups have been subjectively nearly
`equally rated, and were judged relatively close in their
`overall performance, the official decision was as fol-
`lows [5]:
`
`the MPEG standardization committee decided to
`.
`.
`.1
`approve a collaborative development of the draft audio I
`coding standard between the ASPEC and MUSICAM
`groups, because the ASPEC codec was slightly superior
`with respect to the audio quality, especially for lower
`bit rates (64 kbit/s/channel), and the MUSICAM codec
`was slightly superior with respect to implementation
`complexity and decoding delay. The decision was that
`MUSICAM should be the basis for the low-complexity
`first layer, and algorithmic refinements including contri-
`butions of ASPEC should be used in the subsequent
`layers.
`
`Table 1. Milestones of ISO/MPEG—Audio group during the
`development of audio part of the International Standard 11 172.
`
`Date
`
`1988 December
`
`1989 January to 1990 March
`1989 May
`
`1989 June
`
`1989 October
`
`1989 December
`
`1990 May
`
`1990 June
`
`1990 August
`
`1990 December
`
`1991 May
`
`1991 June
`
`1991 November
`
`1991 December
`
`1992 November
`
`Activities
`
`First audio meeting in
`Hanover
`Preparation of tests
`Determining requirements and
`weighting procedure
`Proposal of 14 algorithms to
`be tested
`Clustering of proponents into
`four groups
`Detailed description of four
`clustered proposals:
`ASPEC, ATAC,
`MUSICAM, and
`SB—ADPCM
`Exchange of tapes with coded
`audio sequences between
`four clusters
`Subjective and objective tests
`at SR, Stockholm
`Presentation of results and
`decision to follow a layer
`concept
`First draft of part 3, “Audio
`Coding” of International
`Standard ISO 11172 was
`prepared.
`Verification of three layers by
`"
`subjective testing, again at
`SR in Stockholm
`Layers I and II are frozen;
`Layer III and ‘joint stereo
`coding’ are still under
`discussion
`Second verification of Layer
`III and first checking of
`Joint Stereo Coding by
`subjective testing at NDR
`in Hanover
`Draft of International
`Standard (DIS) ready for
`balloting at national
`standardization bodies
`International Standard ISOI
`IEC 11172-3 accepted by
`national standardization
`bodies
`
`781
`
`
`
` BFIANDENBUFIG AND STOLL
`
`PAPERS
`
`A three-layer coding algorithm has been defined. These
`three layers were tested again in April 1991 by the Swed-
`ish Broadcasting Corporation [3], and a last verification
`test for the very low bit rate of 64 kbit/s/channel and
`“joint stereo coding” was carried out by the University
`of Hanover under the auspices of NDR in November
`1991 [4]. In November 1991 the final proposal, consist-
`ing of three modes of operation called “Layers,” was
`adapted by ISO/MPEG [6].
`
`2 BASIC STRUCTURE OF A GENERIC AUDIO
`CODING SCHEME USING PERCEPTUAL
`CRITERIA
`
`perceptual audio coding
`
`The basic structure of a
`scheme is shown in Fig. 1.
`1) A time—frequency mapping (filter bank) is used
`to decompose the input signal into subsampled spectral
`components. Depending on the filter bank used, these
`are called subband values or frequency lines.
`2) The output of this filter bank, or the output of a
`parallel transform, is used to calculate an estimate of the
`actual (time—dependent) masking threshold using rules
`known from psychoacoustics.
`3) The subband samples or frequency lines are quan-
`tized and coded with the aim of keeping the noise, which
`is introduced by quantizing, below the masking thresh-
`old. Depending on the algorithm, this step is done in
`very different ways. The complexity varies from block
`companding to analysis-=by-synthesis systems using ad-
`ditional noiseless compression.
`4) A frame packing is used to assemble the bit stream,
`which typically consists of the quantized and coded
`mapped samples and some side information, such as bit
`allocation information.
`Depending on the focus on either low frequency reso-
`lution together with high time resolution or high fre-
`quency resolution which leads to only limited time reso-
`lution, the systems are usually called subband coders or
`transform coders.
`
`2.1 Filter Banks
`
`The following list provides a short overview] over the
`most common filter banks used for coding of high-qual-
`ity audio signals:
`1) QMF—Tree Filter Banks: Different frequency reso-
`lution at different frequencies is possible. Typical QMF-
`tree filter banks use from 4 to 24 bands. The computa-
`tional complexity is high.
`2) Polyphase Filter Banks: These are equally spaced
`filter banks which combine the filter design flexibility
`of generalized QMF banks with low computational com-
`plexity [7]. It is possible to design the prototype filter
`in a way that achieves both good frequency resolution
`(stop-band attenuation better than 96 dB) and good con-
`trol of possible time-domain artifacts. A polyphase filter
`bank using 32 bands is used for Layers I and II of the
`ISO/MPEG audio coder.
`3) DFT, DCT with Sine-Taper Window: These were
`the first transforms used in transform coding of audio
`signals. They implement equally spaced filter banks with
`128 -512 bands at a low computational complexity. They
`do not provide critical sampling, that is, the number of
`time—frequency components is greater than the number
`of time samples represented by one block length. An-
`other disadvantage of these transforms are possible
`blocking artifacts.
`4) Modified Discrete Cosine Transform (MDCT, us-
`ing time-domain aliasing cancellation as proposed in
`[8])2 This transform combines critical sampling with a
`good frequency resolution provided by a sine window
`(compared to a sine-taper window) and the computa-
`tional efficiency of a fast FFT-like algorithm. Typically
`128-512 equally spaced bands are used.
`5) Hybrid Structures (such as polyphase and MDCT):
`Using hybrid structures as first proposed in [9] it is
`possible to combine different frequency resolutions
`different frequencies with moderate implementation
`complexity. A hybrid system consisting of a polyphase
`filter bank and an MDCT is used in Layer III.
`ISO/MPEG/AUDIO
`Bitstream
`
`PFarE:‘:g
`
`Digital Audio
`3'9"“ (POM)
`
`Time
`Frequency -
`Mapping
`
`
`
`ISO/MPEG/AUDIO
`Bitstream
`
`( a)
`
`782
`
`
`
`
`
`
`Digital Audio
`Signal (PCM)
`
`
`
`Frequ er1}c/Time
`
`
`
`PAPERS
`
`Theoretically MDCT and polyphase filter banks be-
`long to the same class of time—frequency domain map-
`pings, called lapped orthogonal transform.
`
`3 GENERIC CODING CONCEPT
`
`In view of a number of totally different applications,
`a concept of a generic coding system was envisioned.
`Depending on the application, three layers of the coding
`system with increasing comp1ex_ity and performance can
`be used. A standard ISO decoder is able to decode bit-
`
`stream data which have been encoded in any of the
`layers. There will also be standard ISO Layer X de-
`coders, which are able to decode Layers X and X — n.
`The ISO/MPEG/Audio coding technique offers to deal
`with a much higher dynamic range, due to the scaling
`technique used, than Compact Disc or DAT,
`that is,
`conventional 16-bit PCM.
`
`In all three layers the input PCM audio signal is con-
`verted from the time domain into a frequency domain.
`This is done by a polyphase filter bank consisting of 32
`subbands [7].
`In Layers I and II a filter bank creates 32 subband
`representations of the input audio stream, which are then
`quantized and coded under the control of a psycho-
`acoustic model from which a blockwise adaptive bit
`allocation is derived.
`
`LayerI is a simplified version of the MUSICAM cod-
`ing scheme, most appropriate for consumer applications
`such as digital home recording on tapes, Winchester
`discs, or on magneto—optical disks, that is, for those
`applications for which very low data rates are not
`mandatory.
`Layer II introduces further compression with respect
`to Layer I by redundance and irrelevance removal on
`the scale factors, and uses more precise quantization.
`Layer II is nearly identical with the MUSICAM scheme
`[10], [1 1], with the exception of the frame header. This
`header has been added to the MUSICAM frame during
`the ISO/MPEG/Audio development work. Layer II has
`numerous applications in both consumer and profes-
`sional audio, such as audio broadcasting, television, re-
`cording, telecommunication, and multimedia [12].
`Layer III consists of a combination of the most effec-
`tive modules of the ASPEC [13] and MUSICAM coding
`schemes. An additional frequency resolution is provided
`by the use of a hybrid filter bank. Every subband is
`thereby further split into higher-resolution frequency
`lines by a linear transform that operates on 18 subband
`samples in each subband. In Layer III, nonuniform quan-
`tization, adaptive segmentation, and entropy coding of
`the quantized values are employed for a better coding
`efficiency. The application of this layer is appropriate
`most of all in telecommunication, in particular with nar-
`row-band ISDN and in the field of professional audio
`with high weights on very low bit rates.
`Joint stereo coding can be added as an additional fea-
`ture to any of the layers. This. technique exploits the
`redundancy and irrelevance of typical stereophonic pro-
`gram material and can be used to. increase the audio
`
`J. Audio Eng. Soc., Vol. 42, No. 10, 1994 October
`
`CODING OF HIGH-QUALITY DIGITAL AUDIO
`
`quality at low bit rates or reduce the bit rate for stereo-
`phonic signals [14], [15]. The increase of encoder com-
`plexity is small and requires negligible additional de-
`coder complexity. Joint stereo coding does not enlarge
`the overall coding delay.
`
`3.1 Psychoacoustic Models
`
`The psychoacoustic model calculates the minimum
`masking threshold necessary to determine the just no-
`ticeable noise level for each band in the filter bank. The
`
`difference between the maximum signal l_evel and the
`minimum masking threshold is used in the bit or noise
`allocation to determine the actual quantizer level in each
`subband for each block. Two psychoacoustic models are
`given in the informative part of the standard. While they
`can both be applied to any layer of the MPEG/Audio
`algorithm, in practice model 1 will be used for Layers I
`and II, and model 2 for Layer III. In both psychoacoustic
`models the final output of the model is a signal-to—mask
`ratio for each subband (Layers I and II) or group of
`bands (Layer III). The psychoacoustic models are only
`necessary in the encoder. This allows decoders of sig-
`nificantly less complexity. It is therefore possible to im-
`prove even later the performance of the encoder, relating
`the ratio of bit rate to subjective quality. For some appli-
`cations which are not demanding a very low bit rate, it
`is even possible to use a very simple encoder without
`any psychoacoustic model.
`
`3.1.1 Psychoacoustic Model 1
`
`A high frequency resolution, that is, small subbands
`in the lower frequency region, and a lower resolution in
`the higher frequency region with wide subbands should
`be the basis for an adequate calculation of the masking
`thresholds in the frequency domain. This would lead to
`a tree structure of the filter bank. The polyphase filter
`network used for the subband filtering has a parallel
`structure which does not provide subbands of different
`widths. Nevertheless, one major advantage of the filter
`bank is given by adapting the audio blocks optimally to
`the requirements of the temporal masking effects and
`inaudible preechoes. The second major advantage is
`given by the small delay and complexity. To compensate
`for the lack of accuracy of the spectrum analysis of the
`filter bank, a 512-point fast Fourier transform (FFT) for
`Layer 1, and a 1024-point FFT for Layer II are used in
`parallel to the process of filtering the audio signal into
`32 subbands [16]. The output of the FFT is used to
`determine the relevant tonal,
`that is, sinusoidal, and
`nontonal,
`that is, noise maskers, of the actual audio
`signal. It is well known from psychoacoustic research
`that the tonality ofa masking component has an influence
`on the masking threshold. For this reason it is worth-
`while to discriminate between tonal and nontonal compo-
`nents. The individual masking threshold for each masker
`above the absolute masking threshold are calculated de-
`pending on frequency position, loudness level, and to-
`nality. All the individual masking thresholds, including
`the absolute threshold, are added to the so-called global
`masking threshold. For each subband the minimum
`
`783
`
`
`
`BRANDENBURG AND STOLL
`
`PAPERS
`
`value of this masking curve is determined. Finally the
`difference between the maximum signal level, calculated
`by both the scale factors and the power density spectrum
`' of the FFT, and the minimum masking threshold is cal-
`culated for each subband and each block. The block-
`
`length for Layer I is determined by 12 subband samples,
`corresponding to 384 input audio PCM samples, and for
`Layer II by 36 subband samples, corresponding to 1152
`input audio PCM samples. This difference of maximum
`signal level and minimum masking threshold is called
`signal-to-mask ratio (SMR) and is the relevant input
`function for the bit allocation.
`
`3.1.2 Psychoacoustic Model 2
`The frequency—domain representation of the data is
`calculated via FFT with a window length of 1024 sam-
`ples. The calculation is done every 576 samples, that
`is, synchronous to the hybrid filter bank. The separate
`calculation of the frequency—domain representation is
`necessary because the hybrid filter bank values cannot
`easily be used to get a magnitude—phase representation
`of the input sequence. The magnitude—phase representa-
`tion is necessary to calculate the tonality of the current
`input block for every frequency component.
`The tonality estimation works using a simple polyno-
`mial predictor, as described in [9]. The basic idea is to
`use the predictability of the signal as an indicator for
`tonality. The prediction is done in the magnitude~phase
`domain. The values stores from the last two blocks are
`.used to predict the magnitude and phase of each fre-
`quency line for the current block. The Euclidian distance
`between estimated and actual values in the magni-
`tude—phase domain is normalized to the maximum pos-
`sible distance. The normalized value is called “chaos
`
`measure” and can assume values between 0 (the rotating
`phasor prediction had 0 distance from the actual value)
`and l (the predicted value has the maximum distance
`from the actual value). A logarithmic mapping is used
`to map the chaos measure range between 0.5 and 0.05
`to tonality values of between 0 and 1.
`The magnitude values of the frequency-domain repre-
`sentation are converted to a one-third critical band en-
`
`ergy representation. A convolution of these values with
`the cochlea spreading function follows. The next step
`in the threshold estimation is the calculation of the just
`masked noise level in the cochlea domain using the to-
`nality index and the convolved spectrum. A correction
`for the dc gain of the convolution has to be applied. The
`last step to get the preliminary estimated threshold is
`the adjustment for the absolute threshold. As the sound
`pressure level of the final audio output is not known in
`advance, the absolute threshold is assumed to be some
`
`amount below the LSB for the frequencies around 4 kHz.
`A more detailed description of the estimation of the
`masking threshold using spreading convolution can be
`found in [17].
`.
`The final step in the calculation of the threshold is
`preecho control. Preechoes are audible if the backward
`masking of the signal is not sufficient to mask the error
`signal, which was spread in time due to the limited
`784
`
`time resolution of the synthesis filter bank. This is only
`possible if there is a sudden increase in signal energy,
`at least for part of the signal bandwidth: From this a
`sufficient (but not necessary) condition for the absence
`of audible preechoes can be derived. The estimated
`masking threshold is restricted not to exceed the prelimi-
`nary estimated threshold of the last block. This condition
`
`on the final estimated threshold may reduce the estimated
`threshold by a large amount. To keep the actual quanti-
`zation noise below this modified threshold, additional
`bits need to be available to the quantization and coding
`loop. Layer III contains an intelligent buffer manage-
`ment scheme (called bit reservoir) in order to make the
`additional bits available when needed. This technique
`was taken from OCF (see [18]).
`
`4 LAYER I AND LAYER ll CODiNG SCHEME
`
`Block diagrams of the Layer 1 and Layer II encoders
`are given in Fig. 2. The coding technique for these layers
`is based on a subband splitting of the input PCM audio
`signal by a polyphase analysis filter bank into 32 equally
`spaced subbands, a dynamic bit allocation derived from
`a psychoacoustic model, block companding of the sub-
`band samples, and the bit-stream formatting [10], [l 1],
`[19]. The individual steps of the encoding and decoding
`process are explained in detailed form in the following
`sections.
`
`4.1 Filter Bank
`
`The prototype QMF filter is of order 511. It is opti-
`mized in terms of spectral resolution and rejection of
`side lobes, which is better than 96 dB. This rejection is
`necessary for a sufficient cancellation of aliasing distor-
`tions. This filter bank provides a reasonable tradeoff
`between temporal behavior on one side and spectral ac-
`curacy on the other. A time—frequency mapping provid-
`ing a high number of subbands facilitates the bit-rate
`reduction due to the fact that the human ear perceives
`the audio information in the spectral domain with a reso-
`lution corresponding to the critical bands of the ear, or
`even lower. These critical bands have a width of about
`
`100 Hz in the low—frequency region, that is, below 500
`Hz, and widths of about 20% of the center frequency at
`higher frequencies. The requirement of having a good
`spectral resolution is unfortunately contradictory to the
`necessity of keeping the transient impulse response, the
`so-called pre- and postecho, within certain limits in
`terms of temporal position and amplitude compared to
`the attack of a percussive sound. The knowledge of the
`temporal masking behavior [20] gives an indication of
`the necessary temporal position and amplitude of the
`preecho generated by a time—frequency mapping in such
`a way that this preecho, which normally is much more
`critical compared to the postecho, is masked by the origi-
`nal attack. In conjunction with the dual-synthesis filter
`bank located in the decoder, this filter technique pro-
`.vides a global transfer function optimized in terms of
`perfect impulse response perception.
`-
`In the decoder the dual-synthesis filter bank recon-
`
`J. Audio Eng. Soc., Vol. 42, No. 10, 1994 October
`
`
`
`PAPERS
`
`CODING OF HIGH-QUALITY D|GlTAL AUDlO
`
`structs a block of 32 output samples. The filter structure
`is extremely efficient for implementation in a low—com-
`plexity and non-DSP-based decoder and requires gener-
`ally fewer than 80 integer multiplications or additions
`per PCM output sample. Moreover, the complete analy-
`sis and synthesis filter gives an overall time delay of
`only 10.5 ms at a 48-kHz sampling rate.
`
`4.2 Determination and Cooling of Scale Factors -
`' The calculation of the scale factor for each subband
`is performed for a block of 12 ‘subband samples. The
`maximum of the absolute value of these 12 samples is
`determined and quantized with a word length of 6 bits,
`covering an overall dynamic range of 120 dB per sub-
`band with a resolution of 2 dB per scale factor class. In
`
`Layer I a scale factor is transmitted for each block and
`each subband that has no O-bit allocation.
`
`Layer II uses an additional coding to reduce the trans-
`mission rate for the scale factors. Due to the fact that
`
`in Layer II a frame corresponds to 36 subband samples,
`that is, three times the length of a Layer I frame (see
`Fig. 3), three scale f