CVSD - A Tutorial

APPLICATION
Continuously Variable Slope Delta Modulation:

A Tutorial
1. Introduction
Virtually all means of wireless and wireline speech communication are, or are becoming, digital. Digital coding
of speech for transmission or storage clearly has advantages over traditional analog methods. In digital data
communication or storage systems, information is transmitted or recorded as a series of binary digits—the
receiver or player must only distinguish between a one or zero to exactly recover the original information. In
digital voice systems the information is the human voice. Digital speech coding algorithms are judged by their
ability to quantize (digitize) speech accurately for transmission and then perform the reverse at the decoder.
In other words, the original analog speech signal must be accurately recovered at the receiver. Sounds
simple, however, without compression quantized speech can require significantly more bandwidth than analog
speech. For wireline or wireless telecom applications, the speech extends from 300 Hz to 3300 Hz. If an
analog signal in this band is quantized using a linear Analog to Digital Converter (ADC) sampling just above
the Nyquist rate, say 8 KHz, with 256 quantization levels (8 bits) the resulting data rate is 64 Kbits per second.
Without exotic modulation/coding, the bandwidth (BW) for the digitally encoded signal is nearly twenty times
the original analog signal. Such BW consumption is not practical, particularly for wireless applications.
Engineers working in the field of Speech Coding have been actively searching for methods to reduce the
bandwidth consumed by quantized speech signals. Early algorithms attempted to take advantage of the
human ear’s adaptive dynamic range. The human ear has a built in ability to become more or less sensitive to
audible signals—the ear can hear sound pressure levels (SPL) as low as 0 dB SPL (threshold of human
hearing) to 120 dB SPL (on set of pain) yet at any one time the dynamic range of human hearing is generally
considered to be about 40 dB. In other words, we have a difficult time hearing someone whispering at a rock
concert. Some speech coding algorithms exploit this phenomena by using a greater number of progressively
smaller quantization levels for low amplitude signals and fewer, more coarse quantization levels for large
amplitude signals. This is known as non-uniform quantization. Non-uniform quantization is used in the Public
Switched Telephone Network (PSTN) where it is called -law Pulse Code Modulation (PCM). A slightly more
complex approach takes advantage of strong correlation between adjacent speech samples, quantizing the
amplitude difference (delta) between two samples as opposed to the entire sample amplitude. This difference
signal requires fewer quantization levels for the same signal quality and consequently, reduces the required
bandwidth. Algorithms employing this technique are classified under the broad category of differential
quantization or differential PCM (DPCM). Further bandwidth conservation is possible through more complex
algorithms. For example, combining adaptive quantization with DPCM results in the commonly used coding
algorithm, adaptive DPCM (ADPCM).
Delta modulation (DM) and Continuously Variable Slope Delta modulation (CVSD) are differential
waveform quantization techniques. Both employ two level quantizers (one bit). CVSD is basically DM with an
adaptive quantizer. Applying adaptive techniques to a DM quantizer allows for continuous step size
adjustment. By adjusting the quantization step size, the coder is abled to represent low amplitude signals with
greater accuracy (where it is needed) without sacraficing performance on large amplitude signals.
CVSD is used in tactical communications where “communication quality1” is required yet the option for
security must be available. MIL-STD-188-113 (16 Kb/s and 32 Kb/s), and Federal Standard 1023 (12 Kb/s
CVSD) are examples of a tactical communication systems using CVSD. With the tremendous worldwide
growth in wireless technology, secure communication is becoming important to everyone. In addition to point-
to-point communication, CVSD is commonly used in digital voice recording/messaging and audio delay lines.
This paper attempts to describe CVSD quantization, focusing on its application to coding of speech.
Before discussing the details of CVSD, the basics of uniform and non-uniform quantization (non-adaptive) will
be reviewed. Next, the subject of differential quantization will be explained, showing that DM and CVSD are
equivalent to one bit DPCM and ADPCM, respectively. Finally, some application suggestions for MX•COM
CVSD codecs will be presented.
1
Communication quality is a qualitative expression widely considered synonymous with “acceptable speech
communication.” It is not intended to imply “high-fidelity,” only that intelligible conversation can take place.
©1998 MX-COM Inc. www.mxcom.com Tel: 800 638-5577 336 744-5050 Fax: 336 744-5054 Doc. # 20830070.002
4800 Bethania Station Road, Winston-Salem, NC 27105-1201 USA All trademarks and service marks are held by their respective companies.
CVSD: A Tutorial 2 Application Note
2. Waveform Quantization [1, 3, 5, 9]

Waveform quantization is a process of assigning discrete levels to a sampled analog signal. In the context of
this paper the signal of interest is speech, although the concepts are applicable to signals in different
frequency bands. The term “quantization” implies a relationship between the amplitude of a discrete sample
and its numeric value. This relationship may be linear, non-linear, or differential.
2.1 Sampling [5, 9]

Before a signal can be quantized it must first be sampled. Sampling is the act of instantaneously capturing the
level of a continuous signal at some predetermined rate. This predetermined rate is called the sampling
frequency. As the sampling rate increases, the sampled signal begins to approximate the original continuous
signal. As the sampling frequency decreases, samples move further apart in time, eventually the original
signal cannot be reconstructed from the sampled version. The limit on how far apart these samples can be,
without losing information, is the basis for Shannon's sampling theorem.
Shannon's sampling theorem [5] was originally stated as follows: "If a function f(t) contains no frequencies
higher than f cycles per second it is completely determined by giving its ordinates at a series of points spaced
(1/2f) seconds apart." A mathematical version of this theorem can be obtained by convolving the Fourier
transform of the signal to be sampled with the Fourier transform of an infinite sequence of impulse functions
1
FS ( ) [ F( ) * S( )] (1)
2
or,

1
FS ( )
2 S( ) F( )d (2)

Where, FS ( ) sampled signal spectrum

F( ) original signal spectrum
S( ) spectrum for a sequence of impulses
2f (radian frequency).
The result of the convolution in equation (2) is the original signal spectrum repeated at multiples of the
sampling frequency, see Figure 1. Notice that if the sampling frequency is less than twice the bandwidth of the
original signal the replicas centered at multiples of the sampling frequency will overlap and distort the original.
This undesirable phenomenon is aliasing. To avoid aliasing,
fS 2B . (3)
Where, fS = sampling frequency in Hz

B = bandwidth of original signal in Hz.
f
0
(a)
f
-2F S -F S 0 FS 2F S
(b)
Figure 1: Fourier transform of (a) continuous time signal and (b) sampled signal.
Shannon's sampling theorem is mathematically accurate. Although, in most cases it is not practical to
sample a signal at exactly twice it's highest frequency. Band-limiting is necessary to avoid aliasing. For fs=2B
the band-limiting filter must have a so called “brick wall” roll-off at frequency B. A filter that matches this
requirement is physically unrealizable. Several factors contribute to the actual sampling frequency used.
Generally there is a compromise between the complexity of the band-limiting filter versus the cost of the
analog to digital converter (ADC). As fS becomes closer to 2B, the band-limiting filter requires more stages to
give the desired roll-off. As fS increases, the required conversion time necessitates a faster ADC. Cost for
ADCs is inversely proportional to conversion time, as conversion time decreases, cost rises.
2.2 Uniform Quantization [3, 4, 9]

Figure 2 illustrates the transfer characteristic of a seven level uniform quantizer. To represent the quantized
output q(n) as a binary number would require 3 bits. Once a signal has been sampled it is discrete in time.
However, the amplitude remains continuous. The quantized version of a signal is obtained by applying the
sampled signal to a binary encoder where the discrete voltage levels are assigned to the nearest binary
numbers. The complete process, combining sampling and amplitude quantization, is known as PCM.
q(n)
s(t)
-4 -3 -2 -1 1 2 3 4
-1
-2
-3
Figure 2: Uniform quantizer transfer characteristic.
The noise introduced by PCM is primarily due to the rounding to the “nearest” binary number. If a signal is
quantized to eight levels (3 bits) using a quantizer transfer characteristic similar to the one shown in Figure 2,
the finest resolution is the full scale magnitude divided by eight. In general,
A
. (4)
2n
Where, A = full scale amplitude

n = number of bits per sample
= difference between quantization levels (i.e. finest resolution).
The uncertainty error due to rounding is based on the assumption that there is a continuous range of
values within , all of which are equally likely to have been the actual value of the original signal (uniform
distribution). The error signal is the difference between the original signal amplitude and the quantized
sample value. As a result, the mean square uncertainty error can be solved for by finding the expected value
of error squared over the range of -/2 to /2

2
1
E ( ) d .
2 2
(5)
2
Where, E = expected value operator

= difference between actual and quantized signal levels.
Performing the integration in equation (5) yields an expression which is equivalent to quantization noise power
2
2
N OUT . (6).
12
Taking the square root of both sides gives the root mean square (RMS) noise voltage

N OUT . (7)
2 3
The maximum signal-to-noise ratio (SNR) for a uniformly quantized signal can be calculated by finding the
ratio of the full scale quantization level to the noise voltage in equation (7). The full scale quantization level is
simply the total number of quantization levels multiplied by the minimum quantization increment
(2 n )
SOUTMAX . (8)
2
Where, n = number of bits in quantizer.

The ratio of equations (7) and (8) yields the SNR
2n
SNR . (9)
3
In dB,
SNRdB 4.77 6.02 n . (10)
Equation (10) is an objective measure of quality in systems employing uniform quantization. It should be
noted: the SNR calculated in equation (10) is a theoretical maximum. In practice other factors (e.g. power
supply noise) tend to reduce the final SNR. Also, the human voice is considered to have about 40 dB of
dynamic range, however, during most conversations it is typically about 20 dB down from maximum. In other
words, we generally do not shout during normal conversation. Consequently, the average signal to noise ratio
for uniformly quantized speech is about 20 dB less than what would be calculated using equation (10).
Another objective measure of quality which can be derived in a similar manner is dynamic range. Dynamic
range pertains to the resolution of a quantization scheme. It is the ratio of the full scale amplitude to the
smallest quantized amplitude change,
1 2n
DR
2 n 1 . (11)
2
In dB,
DRdB 6.02( n 1) . (12)
2.3 Non-uniform Quantization [1, 7, 9]

To this point the discussion has been based strictly on uniform quantization (linear speech coding). That is,
waveforms are quantized such that all signal levels have the same resolution. An alternative to linear
quantization is to make fine for low level signals and coarse for high level signals, introducing a non-uniform
quantizer characteristic. This type of quantization can result in improved dynamic range for a given number of
bits and effectively raise the SNR for lower level signals. The drawback is lower maximum SNR—equations 9
and 10 are not valid for non-uniform quantization. Figure 3 shows a transfer characteristic for a seven level
non-uniform quantizer. Notice, when compared to the uniform quantizer transfer characteristic shown in
Figure 2, there are three quantization levels for the input s(t) between zero and one, where there is only one
with the uniform quantizer.
q(n)
s(t)
-4 -3 -2 -1 1 2 3 4
-1
-2
-3
Figure 3: Non-uniform quantizer transfer characteristic.
A complete digital communication system, incorporating non-uniform speech coding, compresses the
quantized signal at the transmitter then expands the signal back to linear form at the receiver. The
compressed signal requires fewer bits and therefore consumes less BW. Ideally, the expander transfer
function is the exact inverse of the compressor and is able to reproduce the original uncoded analog signal. In
literature pertaining to telecommunications, the words compressor and expander are generally combined in to
a single word; compandor.
Companding is on of the most popular forms of non-uniform quantization. When compared to uniform
quantization, it allows for bandwidth compression without degradation in dynamic range, at the expense of
peak signal to noise ratio. For example, a signal uniformly quantized to 14 bits, using equations (10) and (12),
would have a peak SNR and DR of approximately 89 dB and 78 dB respectively. If the same signal is non-
uniformly quantized (compressed) using only 8 bits (256 levels), the minimum quantization level can be set
such that DR of 78 dB can be retained, although the peak SNR will be degraded. Degradation in the peak
SNR is due to the course quantization levels used for the large amplitude signals (where fine quantization is
not necessary). Low amplitude signals are quantized at finer resolution (more steps, where it is necessary).
Consequently, the SNR at low signal levels is improved when compared to uniform quantization with the
same number of levels. Essentially, companding strives to make SNR constant over the dynamic range of the
quantizer.
2.3.1 µ-Law Companding [1, 9]

Most forms of non-uniform quantization are derived from a logarithmic transfer function. That is, the output
signal is proportional to the log of the input signal. In North America and Japan, digital telecommunications
networks employ µ-law companding as the standard for PCM encoding (=255 for North America). The
general form of the transfer characteristic is
ln(1 x )
y( x ) sign( x ) . (13)
ln(1 )
Where, 1 x 1
x = input signal
y(x) = compressed output signal
sign(x) = polarity of input signal.
2.3.2 Uniform versus Non-uniform Quantization [7]

An accepted method of objectively measuring quality in signal processing systems is SNR. In the context of
digital coding this would more accurately be called: signal to quantization noise ratio (SQR). Equation (10)
was derived to calculate the SNR for uniform quantization assuming maximum input signal. To find SNR for
input signal amplitudes less than maximum equation (10) must be rewritten
S
SNRdB 4.77 6.02 n 20 log
i . (14)
S MAX
Where, Si = input signal level

SMAX = maximum signal level.
When the quantization characteristic is not uniform (14) is no longer valid. To clarify the performance of
companded quantization, the SNR must be calculated for a full range of input signal levels. For µ-law
companding the signal to noise ratio is given by the following equation [9],

2n

3(2 )
SNR ( dB) 10 log . (15)

[ln(1 )]2 1 1.732 1
Si 2 Si2
It should be noted that as µ 0, equation (15) is equivalent to equation (14), thus µ = 0 implies uniform
quantization. Figure 4 compares SNR versus relative input signal power for 256 level non-uniform
quantization (µ = 255) and 256 level uniform quantization.
60
8 bit linear PCM
8 bit mu-law PCM
50
40
SNR (dB)
30
20
10
0
-60 -50 -40 -30 -20 -10 0
INPUT (dB)
Figure 4: SNR versus input level for 8 bit -law and 8 bit uniform quantization.
2.4 Differential Quantization [4, 8]

Differential quantization is a coding technique where the difference between the present sample and the
predicted value for the next sample is quantized. This type of quantization is advantageous because the
difference signal dynamic range (variance) is less than the unprocessed input signal. Consequently, fewer
quantization levels are required to retain the same Signal-to-Noise Ratio (SNR). Alternatively, the number of
quantization levels can remain the same allowing the difference signal to be encoded using all the available
bandwidth resulting in greater resolution and improved signal quality.
ENCODER DECODER
x(n) d(n) c(n) d QD (n) x QD (n)

+ Q Q -1 +
-
x PD (n) P(z)
Q -1
d Q (n)
x P (n) x Q (n)
P(z) +
LOCAL DECODER
Figure 5: Differential quantization block diagram.
Figure 5 is a block diagram displaying a differential quantization system. All signals are represented in
discrete time notation implying that x(n) is the discrete time version of x(t). Notice the decoder is in the
feedback path of the encoder. Thus, the decoder is performing the inverse of the encoder. The block labeled
Q converts the difference signal d(n) in to a binary representation suitable for transmission and the block
labeled Q-1 does the inverse. In reality, the process of converting d(n) to c(n) and back to dQD(n) is a significant
factor in the non-ideal behavior of differential quantization. Nevertheless, fundamental analysis of the system
is simplified by assuming blocks Q and Q-1 cancel. Employing this assumption, the transfer function of the
encoder in z-domain terms
C( z )
H ENC ( z ) Q( z )[1 P( z )] . (16)
X ( z)
And the decoder transfer function is
X QD ( z ) 1
H DEC ( z ) . (17)
C( z ) Q( z )[1 P( z )]
If DQD(z) D(z), then XQD(z) X(z) and the entire system transfer function can be written as,
H ( z ) H ENC ( z ) H DEC ( z ) 1 . (18)
Although equation (18) is based on several assumptions, it shows a differential quantization system
patterned after the topology displayed in Figure 5 can be made to produce an output signal which
approximates the original input signal. The quality of the approximation is what distinguishes various
differential quantization schemes. Generally, high quality approximation does not come without a price.
A crucial element in differential quantization is the predictor P(z). The output xP(n) of P(z) is a weighted
sum of past input samples. The general form is equivalent to a finite impulse response (FIR) filter
P
x P (n) a k x Q (n k ) . (19)
k 1
Where, P = order of predictor

ak = weighting factor (coefficient)
xP(n) = output of predictor
xQ(n) = dQ(n) + xP(n) (input to predictor).
Applying the z-transform to equation (19) yields the transfer function for the predictor
P
X P (z)
P( z )
XQ (z)
ak z k . (20)
k 0
Equations (19) and (20) show that the predictor output is a linear combination of past inputs giving rise to
the term “linear prediction.” Non-linear predictors (non-linear combination of past input samples) have been
studied, however, due to complexity and stability issues, their popularity is limited.
The coefficients ak are calculated such that P(z) will provide a reasonably accurate model for the behavior
of human speech. In 1966 a paper was published by McDonald [6] suggesting coefficients, based on
normalized autocorrelation of human speech samples, for predictors of order one to ten. Later, in 1972, Noll
[4] published similar data. These two papers are generally referenced in determination of predictor
coefficients. Assuming that P(z) does provide a reasonably accurate model for human speech, xQ(n) x(n) and
d(n) 0. In other words, a good predictor should minimize the difference signal d(n). This is the basis for
differential quantization. In most DM algorithms the predictor order P is set to one.
The potential for instability in differential quantizers exists in the encoder. As mentioned early, equation
(20) is the transfer function of an FIR filter. One of the significant characteristics of FIR filters is, by definition,
they are stable (the transfer function has only zeros). However, when an FIR is placed in a feedback path, as
is the case with the encoder, the zeros become poles—if one of these poles finds its way outside the unit
circle the differential quantizer will be unstable.
3. Delta Modulation [1, 4, 6, 8, 10]

DM is a differential quantization scheme that uses two level quantization (i.e. one bit quantizer). By using a
single bit to represent each sample, the sample rate and the bit rate are equivalent. Consequently, sample
rate is directly related to signal quality (SNR). Also, BW of the input signal and band-limiting of the output are
significant factors in determining signal quality.
The first DM algorithm was conceived in 1946. During the last 50 years, two algorithms; Linear DM (LDM)
and CVSD, received a significant amount of attention from those interested in efficient methods to digitize
speech. LDM is the most fundamental, least complex form of DM. Hence, this tutorial will attempt to provide
an understanding of LDM as a foundation for understanding more complex DM algorithms. The CVSD
algorithm, an accepted standard in the tactical community and finding a growing number of applications in
“point-to-point” wireless applications, will also be covered.
3.1 Linear DM [1, 4, 6, 8]

The predictor P(z) in LDM is first order and the quantizer Q is two level. Figure 6 is a flow diagram describing
the LDM encoder and decoder algorithms.
ENCODER DECODER
x(n) d(n) c(n) d QD (n) x QD (n) x QDF (n)

+ Q Q -1 + LPF
-
Q -1 az -1
x PD (n)
d Q (n)
INTEGRATOR
x P (n) x Q (n)
az -1 +
INTEGRATOR
Figure 6: LDM block diagram.
The predictor is shown as a single tap FIR filter with transfer function P(z) = az-1. In the encoder, the
transfer function between dQ(n) and xP(n) can be expressed in the z-domain terms as
X P (z) az 1
(21)
DQ ( z ) 1 az 1
In the decoder, the transfer function between dQD(n) and xQD(n) is
X QD ( z ) 1
(22)
DQD ( z ) 1 az 1
Equations (21) and (22) represent discrete time integrators (if a = 1). If a < 1 they become damped or
lossy integrators. McDonald [6] and Noll [4] both suggest a = 1 for optimum prediction gain. A pure integrator
(a = 1) will cause bit errors to propagate longer than if a < 1. In practice, a value of a< 1 is preferred.
The quantizer Q in figure 6 functions like a comparator. When the input d(n) exceeds zero it outputs a
logic one, when the input is less then zero the output is a logic zero. Hence, the output of Q is a single bit
indicating the sign of the magnitude of d(n). The inverse quantizer Q-1 converts logic levels to delta dQ(n) as
shown in the table below.
sign of d(n) c(n) dQ(n)

- 0
+ 1
Table 1: Sign of d(n) with corresponding logic levels and deltas.
The value of plays an important role in the performance of LDM. If is relatively small, tracking of
slowly changing, low amplitude signals is quite good at the expense of poor tracking for fast, abruptly
changing signals. When DM is not able to keep up with the input signal a phenomena called slope overload is
exhibited, as shown in Figure 7.
x Q (n)
x(t)
SLOPE OVERLOAD
SLOPE OVERLOAD
Figure 7: Slope overload in LDM.
Increasing the value of can lessen the effects of slope overload but creates a new problem; granular
noise. With to large, low amplitude signals will not be quantized at fine enough levels and they appear as
idle channel noise , see Figure 8. The idle channel pattern is simply an alternating one-zero sequence
indicating the input signal amplitude is not changing. Since an alternating one-zero bit pattern has a mean
value of zero, the signal out of the decoder will integrate to zero.
x Q (n)
x(t)
GRANULAR NOISE
Figure 8: Granular noise in LDM.
3.2 CVSD [2, 4, 10]

Issues with granularity and slope overload can be drastically reduced by making dynamic adjustments to the
quantizer step size . Adaptive DM (ADM) algorithms attempt to do this by making small for slowly changing
signals and large for rapidly changing signals. The most publicized ADM algorithm is known as CVSD. It was
first proposed by Greefkes and Riemens in 1970 [2]. In their CVSD algorithm, adaptive changes in are
based on the past three or four sample outputs (i.e. c(n), c(n-1), c(n-2), c(n-3)). Figure 9 and Figure 10 show
flow diagrams of the algorithms for the encoder and decoder respectively. Notice only the last three samples
of c(n) are used here.
x(n) d(n) c(n)

+ Q
-
x P (n)
d Q (n) ±1
x L
I1
∆
z -1 z -1
+ x
I2
∆ MIN ∆ MAX
Q -1
Figure 9: CVSD encoder block diagram.
c(n) ±1 d QD (n) x QD (n) x QDF (n)

L x LPF
I1
∆
z -1 z -1
x +
I2
∆ MAX ∆ MIN
Q -1
Figure 10: CVSD decoder block diagram.
The block labeled Q-1 , previously shown in Figure 6, has been replaced by two cascaded one sample
delays (z-1) and logic to determine when three consecutive ones or zeros have occurred. The minimum and
maximum step height is set by MIN. and MAX, respectively. When strings of three consecutive zeros or ones
have not occurred for a period of time long enough for the output of the integrator I2 to decay to near zero, the
algorithm is equivalent to LDM (section 3.1). The time constants for the integrator I2 and integrator I1 are
typically 4 ms and 1 ms respectively. In CVSD literature integrator I1 is referred to as the principle integrator
and integrator I2 as the syllabic integrator. The so called syllabic integrator derives its name from the length of
syllable. Actually a syllable is about 100 ms in duration, however, pitch changes are on the order of 10 ms.
Consequently, 4 ms seems to work best for the CVSD syllabic (pitch) time constant. The block labeled L
performs simple level conversion (i.e. c(n) =1, L outputs 1; c(n) = 0, L outputs -1). This CVSD algorithm is also
known as “Digitally Controlled DM” [10].
Figure 11 displays the output of the integrator I1 (xP(n) in the encoder and xPD(n) in the decoder). When
compared to Figure 7 and Figure 8 slope overload and granular noise are reduced.
x P (n)
x(t)
Figure 11: CVSD quantization noise.
3.3 Performance Measurement of DM [1,4]

The maximum SNR for DM speech coders is [1]
f3
SNRdB 10 log
S 2 14.0 . (23)
f BW f
Where, fS = data rate (sample rate)
fBW = signal BW (cutoff frequency of LPF)
f = signal frequency.
Most literature on DM coding agrees on the terms involving fS, fBW, and f in equation (23) but most do not
agree on -14 dB. This term varies from -17.0 dB to -11.7 dB depending on a variety of different input signals
and assumptions about environment. Nevertheless, equation (23) is considered a maximum because it
assumes no slope overload and minimum granular noise, i.e. optimum step size.
The adaptive behavior of CVSD results in a SNR versus input level characteristic similar to non-uniform
quantization (mentioned in section 2.3). This non-linear SNR characteristic is due to companding, where the
quantization level (slope) is adjusted to a larger or smaller value according to past pitch changes of the input
signal. The number of past samples (bits) used to make a prediction is normally three or four. Four bit
companding has proven to be most effective for data rates greater than 32 Kb/s, three bit companding seems
to work better for data rates less than 32 Kb/s.
Figure 12 shows a typical SNR versus input level characteristic for 32 Kb/s CVSD (four bit companding
algorithm, input signal = 820 Hz sinewave). Equation (23) predicts a maximum SNR of 27.6 dB which is nearly
met in Figure 12 for and input signal level of -15 dbm0 (0 dBm0 = 489 mVrms). Above and below the -15
dBm0 input level granular and slope overload noise start cause distortion. Finally, It should be noted that
CVSD does not perform well with continuous tone or data input signals, particularly at data rates below 32
Kb/s.
30
25
20
SNR (dB)
15
10
0
-40 -35 -30 -25 -20 -15 -10 -5 0 5 10
INPUT (dBm0)
Figure 12: Measured SNR versus input level for 32 Kb/s CVSD (input = 820 Hz sinewave).
SNR is the most used method to objectively quantify performance of speech coding algorithms. However,
it does not always correspond with perceived quality, particularly for differential and adaptive algorithms using
actual voice as the input. In addition, it is difficult to make reliable SNR measurements in the presence of
random bit errors. In an effort to quantify the perceived quality of a speech coding algorithm, Mean Opinion
Score (MOS) testing was developed [4]. Table 2 summarizes the five point scale used to judge quality and
impairment.
Number Quality Impairment

5 excellent imperceptible
4 good perceptible but not annoying
3 fair slightly annoying
2 poor annoying
1 bad very annoying
Table 2: Mean Opinion Score testing guidelines.
An MOS rating of 4 to 4.5 is considered Toll Quality (equivalent to commercial telephony). Where as
Communication Quality MOS ratings are 3 to 4 (barely perceptible distortion, but no degradation in
intelligibility).
Figure 13 compares MOS ratings for -law PCM (the standard for Toll Quality), CVSD and ADPCM.
Notice CVSD performs as well or better than both -law PCM and ADPCM in the presence of bit errors.
Specifically, CVSD retains quite good MOS ratings at bit error rates exceeding 1%, and at 10% has an MOS
rating of 3 (Communication Quality). It is this robustness to bit errors (channel noise) that makes CVSD an
ideal solution for many wireless speech communication applications.
5
48 kb/s CVSD
64 kb/s u-law PCM
32 kb/s ADPCM
4
MEAN OPINION SCORE

3
0
1e-05 0.0001 0.001 0.01 0.1
BIT ERROR RATE
Figure 13: MOS versus bit error rate for ADPCM, -law PCM, and CVSD from [4].
4. Summary
CVSD has several attributes that make it well suited for digital coding of speech. One bit words eliminate the
need for complex framing schemes. Robust performance in the presence of bit errors make error detection
and correction hardware unecessary. Other speech coding schemes may require a digital signal processing
engine and external analog to digital/digital to analog converters to convert the analog signal in to a form that
can be processed digitally—the entire CVSD codec algorithm, including input and output filters, can be
integrated on a single silicon substrate. Despite this simplicity, CVSD has enough flexibility to allow digital
encryption for secure applications. Finally, CVSD can operate over a wide range of data rates—it has been
successfully used from 9.6 kB/s to 64Kb/s. At 9.6Kb/s audio quality is not particularly good, however, it is
intelligible. At data rates of 24 Kb/s to 48 Kb/s it is judged as quite acceptable. And above 48 Kb/s it is
comparable to toll quality. All of these attributes make CVSD attractive to wireless telecommunication systems
(e.g. digital cordless telephones, digital Land Mobile Radio). The defence industry has been using CVSD for
decades in wireline and wireless systems as specified in Mil-Std-188-113. More recently, Federal Standard
1023 proposed CVSD for 25 Khz channel radios operating above 30 Mhz. Figure 14 is a block diagram
showing a CVSD Codec in a digital mobile radio system.
MIC CVSD CODEC CVSD CODEC MIC

ENC DATA
DIGITAL RF DIGITAL RF ENC DATA
ENC IN TRANSMITTER TRANSMITTER ENC IN
ENC CLK ENC CLK
ENC OUT ENC OUT

DEC CLK DIGITAL RF DIGITAL RF DEC CLK
DEC DATA RECEIVER RECEIVER DEC DATA
SPKR SPKR
Figure 14: Digital wireless system incorporating CVSD Codec.
This tutorial has attempted to shed some light on the fundamental aspects of CVSD. It has shown that CVSD
is a differential adaptive quantization algorithm with one bit coding and first order prediction (one bit ADPCM).
In addition, objective and subjective methods of measuring signal quality showing that CVSD performs quite
well in the presence of bit errors (noisy channel) have been presented.
5. References
[1]J. C. Bellamy, Digital Telephony, Wiley and Sons, New York, 1982.
[2] J. A. Greefkes and K. Riemens, “Code Modulation with Digitally Controlled Companding for Speech
Transmission,” Philips Tech. Rev., pp. 335-353, 1970.
[3] A. Gersho, "Principals of Quantization," IEEE Transactions on Circuits and Systems, pp. 427-436, July
1978.
[4] N. S. Jayant and P. Noll, Digital Coding of Waveforms: Principles and Applications to Speech and Video,
Prentice-Hall, Englewood Cliffs, N. J., 1984.
[5] A. B. Jerri, "The Shannon Sampling Theorem—Its Various Extensions and applications: A Tutorial
Review,"Proceedings of the IEEE, pp. 1565-1596, November 1977.
[6] R. A. McDonald, "Signal-to-Noise and Idle Channel Performance of Differential Pulse Code Modulation
Systems-Particular Applications to Voice Signals," Bell System Technical Journal, pp. 1123-1155,
Sept. 1966.
[7] P. Noll, "A Comparative Study of Various Quantization Schemes for Speech Encoding," Bell System
Technical Journal, pp. 1597-1614, November 1975.
[8] L. R. Rabiner and R. W. Schafer, Digital Processing of Speech Signals, Prentice-Hall, Englewood Cliffs, N.
J., 1978.
[9] M. Schwartz, Information, Transmission, Modulation, and Noise, McGraw Hill, New York, 1980.
[10] R. Steele, Delta Modulation Systems, Pentech Press, London, England, 1975.

CVSD - A Tutorial

Uploaded by

Copyright:

Available Formats

CVSD - A Tutorial

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

CVSD - A Tutorial

Uploaded by

Copyright:

Available Formats

APPLICATION

Continuously Variable Slope Delta Modulation:

2. Waveform Quantization [1, 3, 5, 9]

2.1 Sampling [5, 9]

Where, FS ( ) sampled signal spectrum

Where, fS = sampling frequency in Hz

2.2 Uniform Quantization [3, 4, 9]

Figure 2: Uniform quantizer transfer characteristic.

Where, A = full scale amplitude

Where, E = expected value operator

Where, n = number of bits in quantizer.

SNRdB 4.77  6.02 n . (10)

DRdB 6.02( n 1) . (12)

2.3 Non-uniform Quantization [1, 7, 9]

Figure 3: Non-uniform quantizer transfer characteristic.

2.3.1 µ-Law Companding [1, 9]

2.3.2 Uniform versus Non-uniform Quantization [7]

Where, Si = input signal level

2.4 Differential Quantization [4, 8]

x(n) d(n) c(n) d QD (n) x QD (n)

Figure 5: Differential quantization block diagram.

And the decoder transfer function is

H ( z ) H ENC ( z ) H DEC ( z ) 1 . (18)

Where, P = order of predictor

3. Delta Modulation [1, 4, 6, 8, 10]

3.1 Linear DM [1, 4, 6, 8]

x(n) d(n) c(n) d QD (n) x QD (n) x QDF (n)

Figure 6: LDM block diagram.

In the decoder, the transfer function between dQD(n) and xQD(n) is

sign of d(n) c(n) dQ(n)

Table 1: Sign of d(n) with corresponding logic levels and deltas.

Figure 7: Slope overload in LDM.

Figure 8: Granular noise in LDM.

3.2 CVSD [2, 4, 10]

x(n) d(n) c(n)

Figure 9: CVSD encoder block diagram.

c(n) ±1 d QD (n) x QD (n) x QDF (n)

Figure 10: CVSD decoder block diagram.

Figure 11: CVSD quantization noise.

3.3 Performance Measurement of DM [1,4]

Number Quality Impairment

Table 2: Mean Opinion Score testing guidelines.

MEAN OPINION SCORE

BIT ERROR RATE

MIC CVSD CODEC CVSD CODEC MIC

ENC OUT ENC OUT

Figure 14: Digital wireless system incorporating CVSD Codec.

You might also like

SNRdB 4.77 6.02 n . (10)