S1.3
`
`LOW-DELAY CODE-EXCITED LINEAR-PREDICTIVE
`CODING OF WIDEBAND SPEECH AT 32 KBPS
`
`Erik Ordentlich
`Yair Shoham
`
`Signal Processing Research Dept.
`AT&T Bell Laboratories
`600 Mountain Ave.
`Murray Hill, NJ 07974
`
`ABSTRACT
`speech
`of multi-channeVmulti-user
`prospect
`The
`communication via the emerging ISDN has raised a lot of
`interest in advanced coding algorithms for 7KHz wideband
`speech. The capability of 32Kb/s wideband speech coding
`will support sustained high-quality stereo (or bilingual) voice
`transmission for audio-video teleconferencing over a basic
`rate 64Kb/s ISDN channel. In addition to the high-quality
`requirement, network applications are likely to dictate the use
`of algorithms with a very short coding delay. This paper
`reports on the use of the well known Codebook Excited
`Linear Predictive (CELP) [1-3] algorithm for 32Kb/s low(cid:173)
`delay (LD-CELP) coding of wideband speech. The main
`problem associated with wideband coding, namely, spectral
`is discussed. The paper proposes an
`noise weighting,
`enhanced noise weighting technique and demonstrates its
`efficiency via subjective listening
`tests.
`In
`these
`tests,
`involving 20 listeners and 8 test sentences, the average
`rating for the proposed 32Kb/s LD-CELP was essentially
`equal
`to
`that of the 64Kb/s standard (0.722) CCITI
`wideband coder.
`
`I. INTRODUCTION
`The prospect of high-quality multi-channel/multi-user
`speech communication via the emerging ISDN has raised a
`lot of interest in advanced coding algorithms for wideband
`speech. In contrast to the standard telephony band of 200 to
`3400 Hz, wideband speech is assigned the band 50 to
`7000Hz and is sampled at a rate of 16000 Hz for subsequent
`digital processing. The added low frequencies increase the
`voice naturalness and enhance the sense of closeness whereas
`the added high frequencies make the speech sound crisper
`and more intelligible. The overall quality of wideband
`speech as defined above
`is
`sufficient
`for
`sustained
`commentary-grade voice communication as required, for
`example,
`in multi-user
`audio-video
`teleconferencing.
`Wideband speech is, however, harder to code since the data .
`is highly unstructured at high frequencies and the spectral
`dynamic range is very high.
`In some network applications,
`there is also a requirement for a short coding delay which
`limits the size of the processing frame and reduces the
`efficiency of the coding algorithm. This adds another
`dimension to the difficulty of this coding problem.
`The study reported in this paper investigates the use of
`the well known Codebook Excited Linear Predictive (CELP)
`[1-3] algorithm for wideband speech coding. The study is
`reported in detail in [4]. Here, a very brief report is given,
`which highlights the problem, the solution and the main
`results.
`
`ll. CELP AND LOW-DELAY CELP
`The basic structure of conventional CELP [1-3] is shown
`in Figure 1. In the system, an excitation signal, drawn from
`an excitation codebook, excites an all-pole filter which is
`usually a cascade of an LPC-derived filter 1 I A (z) and a so(cid:173)
`called PWh filter 1 I B (z). The LPC polynomial is given by
`A (z) = Laiz-i and is obtained by a standard M 1h-order LPC
`i=O
`analysis of the speech signal. The pitch. filter is determined
`. p
`.
`q
`by the polynomial B (z) = Lbjz_,_ where P is the current
`"=()
`"pitch" Jag - a value
`that best represents
`the current
`periodicity of the input and bj's are the current pitch gains.
`Most often, the order of the pitch filter is q = 1 and it is
`rarely more than 3. Both polyno!J1ial A (z), B (z) are monic.
`algorithm
`implements
`a
`The CELP
`closed -loop
`(analysis-by-synthesis) search procedure for finding the best
`excitation. Each excitation vector is passed through the LPC
`and pitch filters in an effort to find the best match to the
`output, usually, in a weighted mean-squared error (WMSE)
`sense. As seen
`in Figure 1,
`the WMSE matching
`is
`accomplished via the use of a noise-weighting filter W (z ).
`The input speech s (n) is first pre-filtered by W (z) and the
`signal x (n) ( X (z )=S (z) W (z) )
`resulting
`serves
`as
`a
`reference signal in the closed-loop search. The quantized
`version of x (n ), denoted by y (n ), is a filtered excitation,
`closest to x (n) in an MSE sense. The filter used in the
`search
`loop
`is
`the
`synthesis
`filter
`weighted
`H(z) = W(z)I[B(z)A(z)].
`The filter W (z)
`is
`important for achieving a high
`perceptual quality in CELP systems and it plays a central
`role in the CELP-based wideband coder presented here, as
`will become evident later in this paper.
`In general, the CELP transmitter codes and transmits the
`following five entities: the excitation vector, the excitation
`gain, the pitch lag, the pitch tap(s), and the LPC parameters.
`The overall transmission bit rate is determined by the sum of
`all the bits required for coding these entities.
`The CELP coder uses a long block of future samples
`mainly for the LPC analysis and coding. This implies a long
`coding delay which i~ unacceptable in many applications.
`Low-Delay CELP [5] 1s based on the same CELP principle
`but it avoid~ t~e excessive coding delay by performing the
`LPC analysts m a backward-mode.
`In this mode LPC
`a~al~sis_ is performed on recent past quantized ;peech,
`ehmmaung the need for a long look-ahead data block and no
`LPC information is transmitted. If the pitch loop is not used
`and the gain is processed in a fully backward mode [5], all
`
`-9-
`
`CH2977-7/91/0000-0009 $1.00 © 1991 IEEE
`
`ZTE EXHIBIT 1031
`
`Page 1 of 4
`
`

`
`the transrmss10n bits are assigned to the excitation, which
`facilitates the use of an extremely short data frame. This, in
`tum, results in a very short coding delay.
`The LD-CELP proposed here for coding wideband speech
`at 32 Kb/s employs backward LPC. Two versions of the
`coder were studied. The first
`included pitch
`loop. The
`second did not, i.e., B (z)=l. The general structure of the
`coder is that of Figure I, excluding the transmission of the
`LPC information.
`
`(1)
`
`I -:;, Y2 < Y1
`
`-:;, 1
`
`III. PERCEPTUAL NOISE WEIGHTING ANALYSIS IN
`CELP SYSTEMS
`In closed-loop MSE waveform coding the quantization
`the error signal between the output and the target is white.
`As a result, the signal-to-noise ratio (SNR) is not uniform
`across the frequency range and the quantization noise tends
`to mask the low-energy regions of the input spectrum. This
`problem has been recognized and addressed in the context of
`CELP coding of telephony-bandwidth speech
`[6]. The
`solution was in a form of a noise weighting filter, added to
`the CELP search loop in a way shown in Figure I. The
`general form of this filter is:
`A(zly1)
`W(z) = ---
`A (z ly2)
`where A (z) is the LPC polynomial. The effect of y1 or Y2 is
`to move the roots of A (z) towards the origin, de-emphasizing
`the spectral peaks of l!A (z). With y1 and y2 as in (1), the
`response of W (z) has valleys (anti-formants) at the formant
`locations and the inter-formant areas are emphasized.
`In
`addition, the amount of an overall spectral roll-off is reduced,
`compared
`to
`the speech spectral envelope as given by
`1 I A (z).
`In the CELP system of Figure 1, the unweighted error
`signal E (z) = Y (z) - X (z) is white since this is the signal
`that is actually minimized. The final error signal is
`S(z)- s (z) = E (z) w-1 (z)
`(2)
`and has the spectral shape of w-1(z). This means that the
`noise is now concentrated in
`the formant peaks and
`is
`attenuated in between the formants. The idea behind this
`noise shaping is to exploit the auditory masking effect. Noise
`is less audible if it shares the same spectral band with a
`high-level tone-like signal. Capitalizing on this effect, the
`filter W (z) greatly enhances the perceptual quality of the
`CELP coder.
`As will be shown shortly, incorporating W(z) in the
`CELP coder decreases the MSE efficiency of the coder, that
`the minimum overall mean square error can not be
`is,
`achieved. This poses a particularly delicate problem in the
`case of wideband speech coding due to the high spectral
`dynamic range. Perceptually efficient spectral shaping is
`obtained at the price of an increased overall noise level.
`However, there is an apparent tradeoff of shaping vs. noise(cid:173)
`level and an excessive noise shaping usually deteriorates the
`performance.
`to
`The performance of the CELP coder is related
`predictability or the non-spectral-flatness of the speech input.
`This fact is not that obvious in the framework of a closed(cid:173)
`loop analysis-by-synthesis system since prediction is not
`explicitly performed. The CELP system can be heuristically
`analyzed with the aid of an equivalent "DPCM" coder
`shown in Figure 2. In this figure, the transmitter of Figure 1.
`was redrawn in a different way. Virtual predictor and
`quantizer were added to make the circuit look like a standard
`differential coder. Note that the weighting filter is now
`
`represented by ~ W (z) where W (z) is assumed to be a
`rational monic filter, that is, the leading terms of both the
`numerator and denominator are strictly 1.0. The virtual
`predictor
`
`(3)
`
`p (z) = 1 _ B(z)A(z)
`'
`W(z)
`It is easy to show
`plays an important role in this analysis.
`that is guarantees the fundamental DPCM relation, namely,
`R(z)- R (z) = Y(z)- X(z) = E(z)
`(4)
`Note that for P,(z) to be a valid predictor the leading term of
`its impulse response must be strictly 0.0. This is guaranteed
`by B (z ),A (z ), W (z) being all monic. Note, also,
`that if
`W (z) = 1, the predictor reduces to that of a standard DPCM
`system.
`to view the quantization
`The relation (4) allows us
`operation as done by a virtual vector quantizer Q (r) p, (see
`Fig. 2) although, in reality, it is done in output domain by
`matching Y (z)
`to X (z ). The performance of such a
`quantizer, in terms of MSE SNR, is given by the well known
`expression
`
`(6)
`
`ae = a, ETb
`(5)
`where ae is the variance of the quantization error, a, is the
`variance of the virtual residual R (z), and b is the coding rate
`in bits per sample. Experimental results indicate that, for a
`low bit rate (b < 1 as in the usual CELP case) and an
`uncorrelated input (such as R (z )), the quantizer constant E is
`close to 1.0. Henceforth, we neglect this constant.
`the usual
`To proceed with
`this analysis we make
`assumption that the output Y (z) is very close to X (z ), i.e.,
`the quantization error is small. So, using Y (z) = X (z ), we
`get
`R (z) =X (z)- [ 1 - B (z)A (z) w-1(z) ]X (z)
`= B(z)A (z) w-1(z)X(z) = ~S(z)B(z)A (z)
`The result R (z) = ~ S (z) B (z) A (z) is rather striking.
`It tells
`us
`that
`the
`virtual
`residual
`is
`approximately
`independent of the weighting filter up to some fixed gain.
`Moreover, it tells us that R (z) is approximately a minimum(cid:173)
`inverse-filtering
`that
`is
`residual obtained by
`energy
`characterized by maximum LPC prediction gain:
`as
`Pc=(cid:173)a,
`where as is the variance of the input speech. Note that P c
`characterizes the input signal only and it has nothing to do
`with the coding system.
`Using the above, the variance of the error E(z) is given
`by ae =~(as I Pc )Tb. Since E(z) is statistically white,
`the average-magnitude-spectrum of the final error signal
`D (z) = S(z)- S(z) = E(z) ~-Iw-1 (2)
`is given by
`
`(7)
`
`(8)
`
`MD(ro) = ~ rb I w-1 (ro) I
`Pc
`The MSE performance is, therefore, linearly related to the
`prediction gain of the signal. The output noise is affected
`only by the monic part of the weighting filter. Henceforth,
`we assume without loss of generality, that W(z) is monic.
`The power of the output error signal D (z) is given by
`a2
`1t
`
`(9)
`
`aa = --f- T2b f I w-2 (ro) I drol2rc
`
`Pc
`
`-1t
`
`(10)
`
`-10-
`
`Page 2 of 4
`
`

`
`Using the standard logarithmic inequality we get
`02
`1l
`
`o2 = -+ T2b f I w-2(ro) I dro/2rc ~
`a; ,_:G[ f Jog-~ w-2(ro) I dro/2rc + 1]
`
`Pc
`
`-11
`
`(11)
`
`Since W (z) is rational, monic and stable, the first term
`inside the brackets is zero, which leaves us with
`02
`2 >
`s 2-2b
`od--2
`Pc
`with equality if and only if W(z) = 1. Therefore, applying a
`weighting filter always increases the overall noise level.
`
`(12)
`
`A smoothed version of the speech input spectrum may be
`represented by
`
`o,
`Ss(ro) = B (ro)A (ro)
`Combining (9) and
`(13) we get an expression for a
`frequency-dependent signal to noise ratio in the log domain:
`
`(13)
`
`(14)
`
`SNR (ro) =log I W(ro)
`
`B(ro)A (ro)
`
`I + b log2
`
`(16)
`
`in between the formants. The conclusion was that the
`formant and tilt problems ought to be decoupled. The
`approach taken was to use W(z) only for formant modeling
`and to add another section for controlling the tilt only. The
`general form of the new filter is
`W'(z) = W(z)P(z)
`(15)
`where P(z) is responsible for the tilt only. Various forms of
`P (z) were studied. A detailed discussion of these forms can
`be found in [ 4]. In this paper we focus on a solution that has
`been found, after careful listening to coded speech, to work
`the best.
`In the proposed system, the weighting filter (15)
`incorporates two-pole section
`p (z) = --2--.-.
`1 + LPiS'z-'
`i=l
`The coefficients Pi are found by applying the standard LPC
`algorithm to the first three correlation coefficients of the
`current-frame LPC inverse filter ( A (z) ) sequence ai. The
`parameter S is used to adjust the spectral tilt of P (z ). The
`values S = 0.7, Yt=0.95 and Y2=0.8 were found to yield the
`best perceptual perfonnance.
`Figure 3 demonstrates the effect of the enhanced noise(cid:173)
`shaping filter. The dashed curve in the figure shows a typical
`spectrum of a conventional inverse filter w-1 (z ). The solid
`inverse
`filter
`the spectrum of an enhanced
`curve
`is
`w-1(z)P- 1(z) for the same underlying LPC filter. As seen,
`for
`the same general
`tilt,
`the enhanced filter has
`less
`pronounced
`formants especially at
`lowest and highest
`formants. Compared to the old filter, the enhanced filter
`attenuates the noise by about 5 dB at the lowest and highest
`formant, while achieving the right overall spectral tilt.
`The CELP coder with the noise weighting as in (15, 16)
`was implemented with various pitch loops, i.e., various
`orders forB (z) and various number of bits for the pitch taps.
`Interestingly, the best system- in terms of both perceptual
`and MSE performance - was the one without a pitch loop,
`i.e., B (z) = 1.
`
`V. SUBJECTIVE PERFORMANCE TEST
`The performance of the wideband LD-CELP was assessed
`by comparing it to the 0.722 CCITT standard wideband
`coder [7]. This coder operates at 64 Kb/s, twice the rate of
`LD-CELP. The LD-CELP algorithm employed an adaptive
`two-pole tilt filter with S=0.7, in combination with a formant
`filter W(z) defined by y1=0.98 and y2=0.8. Backward LPC
`analysis was used, of order 32. The coder did not use a
`pitch loop. The coding block size was 5 and the codebook
`size was 1024.
`included 4 male and four female
`The test material
`utterances. Each utterance was coded by the 0.722 and by
`the LD-CELP to form a pair of utterances. The order in a
`pair was set at random. However, each pair was played in
`both orders to eliminate biased decisions. Twenty listeners
`took part in the test. The listener was asked to vote for the
`better sounding utterance in his judgement, or, to split his
`vote equally, if no preference could be made. The final
`scores were defined as the percent number of votes for each
`system per given condition.
`Table 1 shows the experimental results. As shown, the
`overall scores are 51.72 and 48.28 for the LD-CELP and the
`0.722, respectively. This means that, on the average, the
`··. two
`systems
`performed
`alike, which
`is
`extremely
`encouraging, recalling that the LD-CELP rate is half that of
`the 0.722. Also, the coding delay of the 0.722 coder is 1.5
`
`Averaging this expression over the frequency range shows
`that the only non-zero term left is b log 2 ( recall that all the
`filters are monic). Since b Jog2 is a small value, the SNR is
`bound to be negative in some frequency regions, mostly at
`high frequencies.
`It is also clear from
`(14)
`that
`the
`weighting filter W (z) cannot improve this situation globally.
`It can only improve the SNR in some regions at the price of
`reducing the SNR in others.
`
`IV. PERCEPTUAL NOISE WEIGHTING IN CELP CODING
`OF WIDEBAND SPEECH
`-
`is
`The spectral dynamic range of wideband speech
`considerably higher then that of telephony speech where the
`added high-frequency region of 3400 to 6000 Hz is usually
`in
`the
`near the bottom of this range. Expression (14)
`previous analysis tells us that, in this case, the unweighted
`SNR tends to be highly negative at high frequencies. While
`coding of the low-frequency region seems to be easier,
`coding of the high-frequency region poses a severe problem.
`The auditory system is quite sensitive in this region and the
`quantization distortions are clearly audible in a form of
`crackling and hiss. Noise weighting is,
`therefore, more
`crucial, in wideband CELP. The balance of low to high
`frequency coding is more delicate. The major effort in this
`study was towards finding a good weighting filter that would
`allow a better control of this balance.
`The starting point of this investigation was the weighting
`filter of the conventional CELP as in (1). The goal was to
`find a good set (Yt ,y2) for best perceptual performance. It
`was found that, similar to the narrow-band case, the values
`y1 =0.9 , Y2=0.4 produced reasonable results. However the
`performance was not yet satisfactory. It was found that the
`filter W (z) as in (1) has an inherent limitation in modeling
`the
`fonnat
`structure
`and
`the
`required
`spectral
`tilt
`concurrently. The spectral tilt is more or less controlled by
`the difference '{I -y2. The tilt is global in nature and it is not
`possible to emphasize it separately at high frequencies. Also,
`changing the tilt affects the shape of the formants of W (z ).
`A pronounced tilt is obtained along with higher and wider
`formants, which puts too much noise at low frequencies and
`
`-11-
`
`Page 3 of 4
`
`

`
`msec. whereas that of the LD-CELP is only 0.94 msec. In
`the ISDN environment, this means stereo, multi-user or
`bilingual communication over a basic-rate channel without
`loss of quality. When the scores are broken up in terms of
`utterance gender, one can notice an asymmetry in quality: the
`LD-CELP does better on males (62.81% preference) whereas
`the 0.722 does better on females (59.38% preference). The
`reason for this is not yet understood.
`The results show that 32Kb/s LD-CELP an an excellent
`candidate
`for wideband
`speech coding
`in high-quality
`communication networks. Although LD-CELP
`is more
`complex than the 0.722, future advances in DSP technology
`will render this disadvantage insignificant.
`
`I
`I
`
`1/
`w-1 (z)
`'Y1 =0.95 'Y2=0.4
`
`'Y1 =0.95 'Y2=0.8
`'Yp=0.7
`w-1 (z)p-1 (z/yp)
`
`\//
`
`I
`
`\
`
`\.
`
`X
`
`y
`
`Analysis
`
`-20~--~~~~~~~~--~--~--~--~--~
`8000
`0
`
`Frequency (Hz)
`
`Figure 3. Effect of the enchanced noise-shaping filter.
`
`Dashed line: Spectrum of a conventional inverse
`filter w-I (z).
`Solid line: Spectrum of an enhanced inverse
`filter w-1(z)P-1(z).
`
`CHANNEL
`
`Utterances
`All
`Male
`Female
`
`LD-CELP (%) G.722 (%)
`48.28
`51.72
`37.19
`62.81
`40.62
`59.38
`
`/\ s
`
`Table 1.
`
`Listening test results for the proposed LD-CELP and the
`CCfiT-G.722 wideband coders.
`
`RECEIVER
`
`Figure 1. CELP Coder
`
`Figure 2. "DPCM" Equivalent of a CELP Coder
`
`- 12-
`
`REFERENCES
`[1] B.S. Atal, M.R. Schroeder , "Stochastic Coding of Speech
`Signals at Very Low Bit rates", Proc. IEEE Int. Conf.
`Comm., May 1984, P. 48.1
`[2] M.R. Schroeder, B.S. Atal, "Code-Excited Linear Predictive
`(CELP): High Quality Speech at Very Low Bit Rates", Proc.
`IEEE Int. Conf. ASSP., 1985, pp. 937-940.
`P. Kroon, E.F. Deprettere "A Class of Analysis-by-Synthesis
`Predictive Coders for High-Quality Speech Codihg at Rate
`Between 4.8 and 16 Kb/s.", IEEE J. on Sel. Area in Comm.
`SAC-6(2), Feb. 1988, pp. 353-363.
`E. Ordentlich, "Low Delay Code Excited Linear Predictive
`(LD-CELP) Coding of Wide Band Speech at 32Kbit/sec.",
`MS Thesis, EE Dept., MIT, March 1990.
`
`[3]
`
`[4]
`
`[5]
`
`J.H. Chen, "A Robust Low-Delay CELP Speech Coder at 16
`Kb/s", Proc. GLOBECOM-89, Vol. 2, Nov. 89, pp. 1237-40.
`[6] B.S. Atal, M.R. Schroeder, "Predictive Coding of Speech
`Signals and Subjective Error Criteria", IEEE Tr. ASSP, Vol.
`ASSP-27, No.3, June 1979, pp. 247-254.
`P. Mermelstein, "G.722, a New CCITT Coding Standard for
`Digital Transmission of Wide band Audio Signals". IEEE
`Comm. Mag., pp. 8-15, Jan. 1988.
`
`[7]
`
`Page 4 of 4