Prolog to
`Speech Coding: A Tutorial Review
`
`A tutorial introduction to the paper by Spanias
`
`If the subtle sounds of human speech are to travel
`the information highways of the future, digitized speech
`will have to be more efficiently transmitted and stored.
`Designers of cellular communications systems, wireless
`personal computer networks, and multimedia systems are
`all searching for improved techniques for handling speech.
`Since its awkward beginnings in the 1930's, speech
`coding has developed to become an essential feature of ev(cid:173)
`eryday telephone system operations. Speech coding is now
`finding applications in cellular communications, computer
`systems, automation, military communications, biomedical
`systems, and almost everywhere that digital communication
`takes hold.
`Speech coding involves sampling and amplitude quanti(cid:173)
`zation of the speech signal. The aim is to use a minimum
`number of bits, while preserving the quality of the recon(cid:173)
`structed speech at the receiving end. Coding research is now
`taking aim at low-rate (8 to 2.4 kbits/s) and very-low-rate
`(below 2.4 kbits/s) techniques.
`The entire gamut of speech coding research is covered in
`this paper. An extensive list of references gives the reader
`access to the speech coding literature. The paper has tutorial
`information to orient applications engineers, and it nicely
`summarizes coder developments for research experts.
`The meaning of the words we speak often changes with
`the smallest inflection of our voices, so better speech quality
`is an essential goal for coding research. The paper lays out
`the quality levels of reconstructed speech, ranging from
`the highest quality broadcast, wide-band speech produced
`by coders at 64 kbits/s, to the lowest quality, synthetic
`speech, currently produced by coders that operate well
`below 4 kbits/s. A section on speech quality points out
`that subjective testing can be lengthy and costly. Speech
`quality has been gauged by objective measures, beginning
`with the signal-to-noise ratio, but these measures do not
`account for human perception.
`The bulk of this paper is devoted to explaining and
`reviewing a wide variety of speech coders. First of all,
`the paper discusses waveform coders. Waveform coders, as
`opposed to vocoders, compress speech waveforms without
`making use of the underlying speech models.
`
`Scalar quantization techniques include familiar classical
`methods such as pulse-code modulation (PCM), differential
`PCM, and delta modulation.
`Vector quantization techniques make use of codebooks
`that reside in both the transmitter and receiver. The paper
`attributes much of the progress recently achieved in low(cid:173)
`rate speech coding to the introduction of vector quantization
`techniques in linear predictive coding. Highly structured
`codebooks allow significant reduction in the complexity of
`high-dimensional vector quantization.
`Sub-band and transform coders rely on transform-domain
`representations of the voice signal. In sub-band coders,
`these representations are obtained through filter banks. Sub(cid:173)
`band encoding is used in medium-rate coding. Fourier
`transform coders obtain frequency-domain representations
`by using unitary transforms. Perhaps the most successful of
`the early transform coders is the adaptive transform coder
`was developed at Bell Laboratories.
`The paper describes analysis-synthes~ methods that use
`the short-time Fourier transform, and also various methods
`that use sinusoidal representations of speech. Multiple
`sine waves have been successfully used in many different
`speech coding systems. For example, one sinusoidal analy(cid:173)
`sis-synthesis system performed very well with a variety of
`signals, including those from multiple speakers, music, and
`biological sounds, and this system also performed well in
`the presence of background noise. Sinusoidal coders have
`been used for low-rate speech coding, and have produced
`high-quality speech in the presence of background noise.
`Another coder that belongs to this class is the multiband ex(cid:173)
`citation coder which recently became part of the Australian
`mobile satellite and International Mobile Satellite standards.
`Since 1939, vocoder systems have tried to produce in(cid:173)
`telligible human speech without necessarily matching the
`speech waveform. Initially, simple models were used to
`produce low-rate coding. The result was synthetic, buzzy(cid:173)
`sounding reconstructed speech. More recently, sophisticated
`vocoders have provided improved quality at the cost of
`increased complexity. The paper briefly describes channel
`and formant vocoders, and the homomorphic vocoder, but
`focuses mostly on linear predictive vocoders.
`
`PROCEEDINGS OF THE IEEE. VOL. X2. NO. 10. OCTOBER 1994
`
`1539
`
`0018-9219/94$04.00 © 1994 IEEE
`
`Ex. 1047 / Page 1 of 44
`Apple v. Saint Lawrence
`
`

`

`Linear predictive coders use algorithms to predict the
`present speech sample from past samples of speech. Usually
`8 to 14 linear predictive parameters are required to model
`the human vocal tract. The analysis window is typically
`20-30 ms long and parameters are generally updated every
`10-30 ms. Real-time predictive coders were first demon(cid:173)
`strated in the early 1970's. The paper describes a linear
`predictive coding algorithm that has become a U.S. federal
`standard for secure communications at 2.4 kbits/s. The U.S.
`Government is currently seeking an improved algorithm to
`replace that standard.
`In analysis-by-synthesis methods, the reconstructed and
`original speech are compared, and the excitation parameters
`are adjusted to minimize the difference before the code is
`transmitted.
`Hybrid coders determine speech spectral parameters by
`linear prediction and optimize excitation using analysis(cid:173)
`by-synthesis techniques. These hybrid coders combine the
`
`features of modem vococ.lers with an ability to exploit
`the properties of the human auditory system. The pa(cid:173)
`per describes several analysis-by-synthesis linear predic(cid:173)
`tive coding algorithms. The coder used in the British
`Telecom International skyphone satellite-based system is
`based on one of these algorithms (MPLP). Another of
`these algorithms (RPE-LTP) has been adopted for the
`full-rate GSM Pan-European digital mobile standard. The
`U.S. Department of Defense has adopted another algorithm
`(CELP) described in the paper, for possible use in a new
`secure telephone unit. The 8-kbits/s algorithm (VSELP)
`adopted for the North American Cellular Digital System
`is also described, as is the LD-CELP coder selected by the
`CCITT as its recommendation for low-delay speech coding.
`
`-Howard Falk
`
`1540
`
`PROCEEDINGS OF THE IEEE. VOL. 82, NO. 10. OCTOBER 199-1
`
`Ex. 1047 / Page 2 of 44
`
`

`

`Speech Coding: A Thtorial Review
`
`ANDREAS S. SPANIAS, MEMBER, IEEE
`
`The past decade has witnessed substantial progress towards
`the application of low-rate speech coders to civilian and military
`communications as well as computer-related voice applications.
`Central to this progress has been the development of new speech
`coders capable of producing high-quality speech at low data
`rates. Most of these coders incorporate mechanisms to: represent
`the spectral properties of speech. provide for speech waveform
`matching, and "optimize" the coder's performance for the human
`ear. A number of these coders have already been adopted in
`national and international cellular telephony standards.
`The objective of this paper is to provide a tutorial overview of
`speech coding methodologies with emphasis on those algorithms
`that are part of the recent low-rate standards for cellular commu(cid:173)
`nications. Although the emphasis is on the new low-rate coders, we
`attempt to provide a comprehensive survey by covering some of the
`traditional methodologies as well. We feel that this approach will
`not only point out key references but will also provide valuable
`background to the beginner. The paper starts with a historical
`perspective and continues with a brief discussion on the speech
`properties and performance measures. We then proceed with de(cid:173)
`scriptions of waveform coders, sinusoidal transform coders, linear
`predictive vocoders, and analysis-by-synthesis linear predictive
`coders. Finally, we present concluding remarks followed by a
`discussion of opportunities for future research.
`
`I.
`
`INTRODUCTION
`Although with the emergence of optical fibers bandwidth
`in wired communications has become inexpensive, there is
`a growing need for bandwidth conservation and enhanced
`privacy in wireless cellular and satellite communications.
`In particular, cellular communications have been enjoying
`a tremendous worldwide growth and there is a great deal of
`R&D activity geared towards establishing global portable
`communications through wireless personal communication
`networks (PCN's). On the other hand, there is a trend
`toward integrating voice-related applications (e.g., voice(cid:173)
`mail) on desktop and portable personal computers-often
`in the context of multimedia communications. Most of these
`applications require that the speech signal is in digital
`format so that it can be processed, stored, or transmitted
`under software control. Although digital speech brings
`flexibility and opportunities for encryption, it is also as(cid:173)
`sociated (when uncompressed) with a high data rate and
`
`Manuscript received July 6, 1993; revised March 4, 1994. Portions of
`this work have been supported by Intel Corporation.
`The author
`is with
`the Department of Electrical Engineering,
`Telecommunications Research Center, Arizona State University, Tempe,
`AZ 85287-5706 USA.
`IEEE Log Number 9401511.
`
`hence high requirements of transmission bandwidth and
`storage. Speech Coding or Speech Compression is the field
`concerned with obtaining compact digital representations
`of voice signals for the purpose of efficient transmission or
`storage. Speech coding involves sampling and amplitude
`quantization. While the sampling is almost invariably done
`at a rate equal to or greater than twice the bandwidth of
`analog speech, there has been a great deal of variability
`among the proposed methods in the representation of the
`sampled waveform. The objective in speech coding is to
`represent speech with a minimum number of bits while
`maintaining its perceptual quality. The quantization or
`binary representation can be direct or parametric. Direct
`quantization implies binary representation of the speech
`samples themselves while parametric quantization involves
`binary representation of speech model and/or spectral pa(cid:173)
`rameters.
`With very few exceptions, the coding methods discussed
`in this paper are those intended for digital speech communi(cid:173)
`cations. In this application, speech is generally bandlimited
`to 4 kHz (or 3.2 kHz) and sampled at 8 kHz. The simplest
`nonparametric coding technique is Pulse-Code Modulation
`(PCM) which is simply a quantizer of sampled amplitudes.
`Speech coded at 64 kbits/s using logarithmic PCM is
`considered as "noncompressed" and is often used as a
`reference for comparisons. In this paper, we shall use the
`term medium rate for coding in the range of 8-16 kbits/s,
`low rate for systems working below 8 kbits/s and down to
`2.4 kbits/s, and very luw rate for coders operating below
`2.4 kbits/s.
`Speech coding at medium-rates and below is achieved
`using an analysis-synthesis process. In the analysis stage,
`speech is represented by a compact set of parameters which
`are encoded efficiently. In the synthesis stage, these param(cid:173)
`eters are decoded and used in conjunction with a reconstruc(cid:173)
`tion mechanism to form speech. Analysis can be open-loop
`or closed-loop. In closed-loop analysis, the parameters are
`extracted and encoded by minimizing explicitly a measure
`(usually the mean square) of the difference between the
`original and the reconstructed speech. Therefore, closed(cid:173)
`loop analysis incorporates synthesis and hence this process
`is also called analysis by synthesis. Parametric representa(cid:173)
`tions can be speech- or non-speech-specific. Non-speech(cid:173)
`specific coders or waveform coders are concerned with the
`
`PROCEEDINGS OF THE IEEE. VOL. 82. NO. 10. OCTOBER 1994
`
`15.11
`
`0018-9219/94$04.00 © 1994 IEEE
`
`Ex. 1047 / Page 3 of 44
`
`

`

`faithful reconstruction of the time-domain waveform and
`generally operate at medium rates. Speech-specific coders
`or voice coders (vocoders) rely on speech models and are
`focussed upon producing perceptually intelligible speech
`without necessarily matching the waveform. Vocoders are
`capable of operating at very-low rates but also tend to
`produce speech of synthetic quality. Although this is the
`generally accepted classification in speech coding, there
`are coders that combine features from both categories.
`For example, there are speech-specific waveform coders
`such as the Adaptive Transform Coder [303] and also
`hybrid coders which rely on analysis-by-synthesis linear
`prediction. Hybrid coders combine the coding efficiency
`of vocoders with the high-quality potential of waveform
`coders by modeling the spectral properties of speech (much
`like vocoders) and exploiting the perceptual properties of
`the ear, while at the same time providing for waveform
`matching (much like waveform coders). Modem hybrid
`coders can achieve communications quality speech at 8
`kbitsls and below at the expense of increased complexity.
`At this time there are at least four such coders that have
`been adopted in telephony standards.
`
`A. Scope and Organization
`In this paper, we provide a survey of the different
`methodologies for speech coding with emphasis on those
`methods and algorithms that are part of recent communica(cid:173)
`tions standards. The paper is intended both as a survey and
`a tutorial and has been motivated by advances in speech
`coding which have enabled the standardization of low(cid:173)
`rate coding algorithms for civilian cellular communications.
`The standardizations are results of more than fifty years of
`speech coding research. Until recently, low-rate algorithms
`were of interest only to researchers in the field. Speech
`coding is now of interest to many engineers who are
`confronted with the difficult task of learning the essentials
`of voice compression in order to solve implementation
`problems, such as fitting an algorithm to an existing fixed(cid:173)
`point signal processor or developing low-power single-chip
`solutions for portable cellular telephones. Modem speech(cid:173)
`coding algorithms are associated with numerical methods
`that are computationally intensive and often sensitive to ma(cid:173)
`chine precision. In addition, these algorithms employ math(cid:173)
`ematical, statistical, and heuristic methodologies. While
`the mathematical and statistical techniques are associated
`with the theory of signal processing, communications, and
`information theory, many of the heuristic methods were
`established through years of experimental work. Therefore,
`the beginner not only has to get a grasp of the theory
`but also needs to review the algorithms that preceded the
`standards. In this paper we attempt to sort through the
`literature and highlight the key theoretical and heuristic
`techniques employed in classical and modem speech-coding
`algorithms. For each method we give the key references
`and, when possible, we refer first to the article that the
`novice will find more accessible.
`The general notation adopted in this paper is as follows.
`The discrete-time speech signal is denoted as s( n ), where
`
`n is an integer indexing the sample number. Discrete(cid:173)
`time speech is related to analog speech, sa(t), by s(n) =
`Sa ( nT) = sa ( t) lt=nT, where T is the sampling period.
`Unless otherwise stated, lower case symbols denote time(cid:173)
`domain signals and upper case symbols denote transform(cid:173)
`domain signals. Bold characters are used for matrices and
`vectors. The rest of the notation is introduced in subsequent
`sections as necessary.
`The organization of the paper is as follows. The first
`section gives a brief description of the properties of speech
`signals and continues with a historical perspective and
`a review of performance measures. In Section II, we
`discuss waveform coding methods. In particular, we start
`with a general description of scalar [55], [82], [152] and
`vector quantization [81], [98], [115], [192] methods and we
`continue with a discussion of waveform coders [48], [52].
`Section III presents sinusoidal analysis-synthesis meth(cid:173)
`ods [205] for voice compression and Section IV presents
`vocoder methods [11], [162], [163]. Finally, in Section V
`we discuss analysis-by-synthesis linear predictive coders
`[96], [100], [123], [272] and in Section VI we present
`concluding remarks. Low-rate coders, and particularly those
`adopted in the recent standards, are discussed in more detail.
`The scope of the paper is wide and although our literature
`review is thorough is by no means exhaustive. Papers with
`similar scope [12], [23], [82], [83], [96], [104], [109], [150],
`[154], [155], [157], [191], [270], [279]; special journal and
`magazine editions on voice coding [18], [19], [131], [132],
`[134]-[136], [138], [139]; and books on speech processing
`[62], [86], [90], [91], [99], [113], [152], [199], [232], [234],
`[236], [251], [275] can provide additional information.
`There are also six excellent collections of papers edited
`by Jayant [156], Davidson and Gray [61], Schafer and
`Markel [269], Abut [1], and Atal, Cuperman, and Gersho
`[9], [10]. For the reader, who wants to keep up with the
`developments in this field, articles appear frequently in
`IEEE TRANSACTIONS and symposia associated with the areas
`of signal processing and communications (see references
`section) and also in specialized conferences, workshops,
`and journals, e.g., [133] [137], [140], [291].
`
`B. Speech Properties
`Before we begin our presentation of the speech coding
`methods, it would be useful if we briefly discussed some
`of the important speech properties. First, speech signals are
`nonstationary and at best they can be considered as quasi(cid:173)
`stationary over short segments, typically 5-20 ms. The
`statistical and spectral properties of speech are thus defined
`over short segments. Speech can generally be classified as
`voiced (e.g., Ia!, Iii, etc), unvoiced (e.g., Ish!), or mixed.
`Time- and frequency-domain plots for sample voiced and
`unvoiced segments are shown in Fig. l. Voiced speech
`is quasi-periodic in the time domain and harmonically
`structured in the frequency domain while unvoiced speech
`is random-like and broadband. In addition, the energy of
`voiced segments is generally higher than the energy of
`unvoiced segments.
`
`1542
`
`PROCEEDJ!'JGS OF THE IEEE. VOL. 82, NO. 10, OCTOBER 1994
`
`Ex. 1047 / Page 4 of 44
`
`

`

`~---­
`
`-
`
`nme domain speech segment
`TAPIT11d:I01C
`
`M
`
`~
`
`~ ~
`
`~
`
`1\
`
`~
`
`~I
`
`1.0
`
`! 'a 0.0
`
`~
`
`-1.0
`
`!50
`
`iii"
`
`.......... --1 ..... i 20
`
`il'
`::;;
`
`16
`Tlma(mS)
`
`24
`
`32
`
`2
`
`Frequency (KHz)
`
`1.0
`
`nme domain speech segment
`
`-1.0
`
`16
`Time (mS)
`
`24
`
`32
`
`Frequency (KHz)
`
`Fig. 1. Voiced and unvoiced segments and their short-time spectra.
`
`The short-time spectrum1 of voiced speech is character(cid:173)
`ized by its fine and formant structure. The fine harmonic
`structure is a consequence of the quasi-periodicity of speech
`and may be attributed to the vibrating vocal chords. The
`formant structure (spectral envelope) is due to the inter(cid:173)
`action of the source and the vocal tract. The vocal tract
`consists of the pharynx and the mouth cavity. The shape
`of the spectral envelope that "fits" the short-time spectrum
`of voiced speech, Fig. 1, is associated with the transfer
`characteristics of the vocal tract and the spectral tilt (6
`dB/octave) due to the glottal pulse [261]. The spectral
`envelope is characterized by a set of peaks which are called
`formants. The formants are the resonant modes of the vocal
`tract. For the average vocal tract there are three to five
`formants below 5 kHz. The amplitudes and locations of
`the first three formants, usually occurring below 3 kHz, are
`quite important both in speech synthesis and perception.
`Higher formants are also important for wideband and
`unvoiced speech representations. The properties of speech
`are related to the physical speech production system as
`follows. Voiced speech is produced by exciting the vocal
`tract with quasi-periodic glottal air pulses generated by
`the vibrating vocal chords. The frequency of the periodic
`pulses is referred to as the fundamental frequency or pitch.
`Unvoiced speech is produced by forcing air through a
`constriction in the vocal tract. Nasal sounds (e.g., In!) are
`due to the acoustical coupling of the nasal tract to the vocal
`tract, and plosive sounds (e.g., /p/) are produced by abruptly
`releasing air pressure which was built up behind a closure
`in the tract.
`
`1 Unless otherwise stated the term spectrum implies power spectrum
`
`More information on the acoustic theory of speech pro(cid:173)
`duction is given by Pant [75] while information on the
`physical modeling of the speech production process is given
`in the classic book by Flanagan [86].
`
`C. Historical Perspective
`Speech coding research started over fifty years ago with
`the pioneering work of Homer Dudley [66], [67] of the
`Bell Telephone Laboratories. The motivation for speech
`coding research at that time was to develop systems for
`transmission of speech over low-bandwidth telegraph ca(cid:173)
`bles. Dudley practically demonstrated the redundancy in
`the speech signal and provided the first analysis-synthesis
`method for speech coding. The basic idea behind Dudley's
`voice coder or vocoder (Fig. 2) was to analyze speech in
`terms of its pitch and spectrum and synthesize it by exciting
`a bank of ten analog band-pass filters (representing the
`vocal tract) with periodic (buzz) or random (hiss) excitation
`(for voiced and unvoiced sounds, respectively). The channel
`vocoder received a great deal of attention during World
`War II because of its potential for efficient transmission of
`encrypted speech. Formant [223] and pattern matching [68]
`vocoders along with improved analog implementations of
`channel vocoders [221], [292] were reported through the
`1950's and 1960's. In the formant vocoder, the resonant
`characteristics of the filter bank track the movements of the
`formants. In the pattern-matching vocoder the best match
`between the short-time spectrum of speech and a set of
`stored frequency response patterns is determined and speech
`is produced by exciting the channel filter associated with
`the selected pattern. The pattern-matching vocoder was
`
`SPANIAS: SPEECH CODING
`
`1543
`
`Ex. 1047 / Page 5 of 44
`
`

`

`Analysis
`
`Pnch Channel
`
`Synthesis
`
`A total of ten channels
`
`Fig. 2. Dudley's channel vocoder [67].
`
`essentially the first analysis-synthesis system to implicitly
`employ vector quantization.
`Although early vocoder implementations were based on
`analog speech representations, digital representations were
`rapidly gaining interest due to their promise for encryption
`and high-fidelity transmission and storage. In particular,
`there had been a great deal of activity in Pulse-Code
`Modulation (PCM) in the 1940's (see [156] and the ref(cid:173)
`erences therein). PCM [228] is a straightforward method
`for discrete-time, discrete-amplitude approximation of ana(cid:173)
`log waveforms and does not have any mechanism for
`redundancy removal. Quantization methods that exploit
`the signal correlation, such as Differential PCM (DPCM),
`Delta Modulation (DM) [153], and Adaptive DPCM were
`proposed later and speech coding with PCM at 64 kbits/s
`and with ADPCM at 32 kbits/s eventually became CCITT2
`standards [32].
`With the flexibility offered by digital computers, there
`was a natural tendency to experiment with more sophisti(cid:173)
`cated digital representations of speech [266]. Initial efforts
`concentrated on the digital implementation of the vocoder
`[112]. A great deal of activity, however, concentrated on the
`linear speech source-system production model developed by
`Fant [75] in the late 1950's. This model later evolved into
`the familiar speech production system shown in Fig. 3. This
`model consists of a linear slowly time-varying system (for
`the vocal tract and the glottal model) excited by periodic
`impulse train excitation (for voiced speech) and random
`excitation (for unvoiced speech).
`The source-system model became associated with Au(cid:173)
`toregressive (AR) time-series methods where the vocal tract
`filter is all-pole and its parameters are obtained by Linear
`Prediction analysis [189]; a process where the present
`speech sample is predicted by the linear combination of pre(cid:173)
`vious samples. Itakura and Saito [143], [264] and Atal and
`Schroeder [ 14] were the first to apply Linear Prediction (LP)
`techniques to speech. A tal and Hanauer [II] later reported
`
`2 International Consultative Committee for Telephone and Tele(cid:173)
`graph currently called International Telecommunications Union-Tele(cid:173)
`communication Standardization Sector (ITU-TSS)
`
`illil ___.
`
`• ___.
`
`VOCAL
`TRACT
`FILTER
`
`SYNTHETIC
`
`SPEECH
`
`Fig. 3. The engineering model for speech synthesis.
`
`an analysis-synthesis system based on LP. Theoretical and
`practical aspects of linear predictive coding (LPC) were
`examined by Markel and Gray [199] and the problem
`of spectral analysis of speech using linear prediction was
`addressed by Makhoul and Wolf [190].
`LP is not the only method for source-system analysis.
`Homomorphic analysis, a method that can be used. for
`separating signals that have been combined by convolution,
`has also been used for speech analysis. Oppenheim and
`Schafer were strong proponents of this method [229], [230].
`One of the inherent advantages of homomorphic speech
`analysis is the availability of pitch information from the
`cepstrum [41], [227].
`The emergence of VLSI technologies along with ad(cid:173)
`vances in the theory of digital signal processing during
`the 1960's and 1970's provided even more incentives for
`getting new and improved solutions to the . speech coding
`problem. Analysis-synthesis of speech usmg the Short(cid:173)
`Time Fourier Transform (STFT) was proposed by Flanagan
`and Golden in a paper entitled "Phase Vocoder" [87]. In
`addition, Schafer and Rabiner designed and simulated an
`analysis-synthesis system based on the STFT [26?], [268),
`and Portnoff [240], [242], [243] provided a theoretical basis
`for the time-frequency analysis of speech using the STFT.
`In the mid- to late 1970's there was also a continued activity
`in linear prediction [304], [310], transform coding [303],
`and sub-band coding [52]. An excellent review of this work
`is given by Flanagan et al. [82], and a unified analysis of
`transform and sub-band coders is given by Tribolet and
`Crochiere [303]. During the 1970's, there were also parallel
`efforts for the application of linear prediction in military
`secure communications (see the NRL reports by Kang et al.
`
`1544
`
`PROCEEDINGS OF THE IEEE, VOL. 82, NO. 10, OCTOBER 1994
`
`Ex. 1047 / Page 6 of 44
`
`

`

`[162]-[166]. A federal standard (FS-1015) which is based
`on the LPC-10 algorithm, was developed in the early 1980's
`(see the paper by Tremain [301]).
`Research efforts in the 1980's and 1990's have been
`focused upon developing robust low-rate speech coders ca(cid:173)
`pable of producing high-quality speech for communications
`applications. Much of this work was driven by the need
`for narrow-band and secure transmission in cellular and
`military communications. Competing methodologies pro(cid:173)
`moted in the 1980's included: sinusoidal analysis-synthesis
`of speech proposed by McAulay and Quatieri [205], [206],
`multiband excitation vocoders proposed by Griffin and Lim
`[ 117], multipulse and vector excitation schemes for LPC
`proposed by Atal et al. [13], [272], and vector quantization
`(VQ) promoted by Gersho and Gray [98], [99], [115],
`and others [47] [192]. Vector quantization [1] proved to
`be very useful in encoding LPC parameters. In partic(cid:173)
`ular, Atal and Schroeder [17], [272] proposed a linear(cid:173)
`prediction algorithm with stochastic vector excitation which
`they called "Code Excited Linear Prediction" (CELP).
`The stochastic excitation in CELP is determined using a
`perceptually weighted closed-loop (analysis-by -synthesis)
`optimization. CELP coders are also called hybrid coders
`because they combine the features of traditional vocoders
`with the waveform-matching features of waveform coders.
`Although the first paper [ 17] on CELP addressed the
`feasibility of vector excitation coding, follow-up work [37],
`[100], [170], [171], [176], [177], [276], [315] essentially
`demonstrated that CELP coders were capable of produc(cid:173)
`ing medium-rate and even low-rate speech adequate for
`communications applications. Real-time implementation of
`hybrid coders became feasible with the development of
`highly structured codebooks.
`Progress in speech coding, particularly in the late 1980's,
`enabled recent adoptions of low-rate algorithms for mo(cid:173)
`bile telephony. An 8-kbit/s hybrid coder has already been
`selected for the North American digital cellular standard
`[ 1 00], and a similar algorithm has been selected for the
`6.7-kbit/s Japanese digital cellular standard [102], [103],
`[217], [314]. In Europe, a standard that uses a 13-kbit/s
`regular pulse excitation algorithm [307] has been completed
`and partially deployed by the "Group Speciale Mobile"
`(GSM). Parallel standardization efforts for secure military
`communications [ 169] have resulted in the adoption of a
`4.8-kbit/s hybrid algorithm for the Federal Standard 1016
`[9]. In addition, a 6.4-kbit/s improved multiband excitation
`coder [121] has been adopted for the International Maritime
`Satellite (INMARSAT-M) system [322] and the Australian
`Satellite (AUSSAT) system. Finally, we note that there
`are plans to increase the capacity of cellular networks by
`introducing half-rate algorithms in the GSM, the Japanese,
`and the North American standards.
`
`and acoustic interference. In general, high-quality speech
`coding at low rates is achieved using high-complexity
`algorithms. For example, real-time implementation of a
`low-rate hybrid algorithm must be typically done on a
`digital signal processor capable of executing 12 or more
`million instructions per second (MIPS). The one-way delay
`(coding plus decoding delay only) introduced by such
`algorithms is usually between 50 to 60 ms. Robust speech
`coding systems incorporate error correction algorithms to
`protect the perceptually important information against chan(cid:173)
`nel errors. Moreover, in some applications coders must
`perform reasonably well with speech corrupted by back(cid:173)
`ground noise, nonspeech signals (such as DTMF tones,
`voiceband data, modem signals, etc), and a variety of
`languages and accents.
`In digital communications, speech quality is classified
`into four general categories, namely: broadcast, network or
`toll, communications, and synthetic. Broadcast wideband
`speech refers to high-quality "commentary" speech that can
`generally be achieved at rates above 64 kbits/s. Toll or
`network quality refers to quality comparable to the classical
`analog speech (200-3200 Hz) and can be achieved at rates
`above 16 kbits/s. Communications quality implies some(cid:173)
`what degraded speech quality which is nevertheless natural,
`highly intelligible, and adequate for telecommunications.
`Synthetic speech is usually intelligible but can be unnat(cid:173)
`ural and associated with a loss of speaker recognizability.
`Communications speech can be achieved at rates above 4.8
`kbits/s and the current goal in speech coding is to achieve
`communications quality at 4.0 kbits/s. Currently, speech
`coders operating well below 4.0 kbits/s tend to produce
`speech of synthetic quality.
`Gauging the speech quality is an important but also very
`difficult task. The signal-to-noise ratio (SNR) is one of
`the most common objective measures for evaluating the
`performance of a compression algorithm. This is given by
`
`SNR = 10log10
`
`"'I: 1
`82(n)
`{
`-::-:M~--1 n_==_O __ _
`n~O (s(n) - s(n) )2
`
`}
`
`(1)
`
`where s( n) is the original speech data while s( n) is the
`coded speech data. The SNR is a long-term measure for
`the accuracy of speech reconstruction and as such it tends
`to "hide" temporal reconstructi