 |
Description  |
|
|
TECHNICAL FIELD
This invention relates to signal classifiers and, more particularly, to an
arrangement for classifying an incoming signal among one of a plurality of
classifications.
BACKGROUND OF THE INVENTION
In recent times bit rate reduction techniques have been employed to
increase transmission capacity over digital transmission facilities. One
such technique is adaptive differential pulse code modulation (ADPCM).
ADPCM is employed to increase capacity over voiceband digital transmission
facilities. Use of 32 kilobit/sec ADPCM is increasing and, normally,
doubles the capacity of T carrier facilities. Greater transmission
capacity may be realized by judiciously transmitting the voiceband signals
at still lower bit rates than the 32 kilobit/sec rate.
The 32 kilobit/sec rate ADPCM, however, presents a problem when
transmitting certain non-voice signals. Typically, non-voice signals, for
example, voiceband data signals, are transmitted at the 32 kilobit/sec
rate ADPCM. That is, no bits are allowed to be dropped to lower the
transmission bit rate. When transmitting "higher" bit rate voiceband data
signals, for example, those generated by a 9600 bit/sec or higher rate
modem, the use of the 32 kilobit/sec so-called fixed rate ADPCM results in
unacceptable bit error rates. Consequently, the data must be retransmitted
thereby resulting in unacceptable transmission throughput. In order to
minimize this problem it is desirable to transmit the 9600 bit/sec and
higher rate voiceband data signals at an ADPCM transmission bit rate or
other PCM transmission bit rates higher than the present fixed ADPCM bit
rate of 32 kilobit/sec. Additionally, it may be acceptable and desirable
to transmit voiceband data signals having "lower" bit rates at a bit rate
less than the 32 kilobit ADPCM. In order to effect transmission of the
voiceband data signals as bit rates higher or lower than the 32
kilobit/sec ADPCM rate, they must be classified as to their respective
baud rates.
Heretofore, attempts at classifying voiceband data signals have used a
so-called ordinary autocorrelation of the signal. A problem with the use
of the ordinary autocorrelation is that the results are modulated by the
carrier frequency of the data signal. Consequently, the results of such a
classifying arrangement do not accurately reflect the baud rates of the
voiceband data signals.
SUMMARY OF THE INVENTION
Classification of an incoming signal is realized, in accordance with an
aspect of the invention, by employing a classification arrangement which
is based on the autocorrelation of a complex low-pass version of an
incoming signal, i.e., the complex autocorrelation.
In acccordance with a particular aspect of the invention the magnitude of
the autocorrelation of the complex low-pass version of the incoming signal
is uniquely employed to classify incoming voiceband data signals.
More specifically, the normalized magnitude of the complex autocorrelation
function determined for a specific delay interval, i.e., lag, is related
to the spectral width of the incoming signal independent of carrier
frequency and, in turn, the spectral width is related to the baud rate for
the voiceband data signals. The normalized magnitude is compared to
predetermined threshold values to classify the incoming signal as one of a
plurality of classifications, for example, one of a plurality of baud
rates.
BRIEF DESCRIPTION OF THE DRAWING
The invention will be more fully appreciated from the following detailed
description when considered in conjunction with the accompanying figures,
in which:
FIG. 1 shows in simplified block diagram form a signal classification
arrangement including an embodiment of the invention; and
FIGS. 2 and 3 when combined A--A and B--B form a flow chart illustrating
operation of a classification arrangement in accordance with aspects of
the invention.
DETAILED DESCRIPTION
FIG. 1 shows in simplified block diagram form an arrangement for
classifying voice band signals in accordance with aspects of the
invention. Accordingly, shown is incoming digital signal d(n), being
supplied to multipliers 10 and 11. In this example, signal d(n) is in
linear PCM form with a sampling rate of 8 kHz. Thus, a sample interval is
125.mu. seconds. A signal representative of cos (.pi.n/2) is supplied from
cos (.pi.n/2) generator 12 to multiplier 10. In turn, multiplier 10 yields
a(n)=d(n) cos (.pi.n/2). Similarly, a signal representative of sin
(.pi.n/2) is supplied from sin (.pi.n/2) generator 13 to multiplier 11. In
turn, multiplier 11 yields b(n)=d(n) sin (.pi.n/2). Signal a(n) is
supplied to low pass filter 14 which yields a low pass version thereof,
namely, u(n). Similarly, signal b(n) is supplied to low pass filter 15
which also yields a low pass version thereof, namely, v(n). In this
example, low pass filters 14 and 15 are each a second order recursive
filter with a cutoff frequency at 2 kHZ. Both u(n) and v(n) are supplied
to complex signal generator 16 which yields .gamma.(n)=u(n)-jv(n).
.gamma.(n) is a complex low pass version of d(n). It is noted that the
complex low-pass version, .gamma.(n), may be generated by other
arrangements; one example being a Hilbert filter. Signal .gamma.(n) is
supplied to multiplier 17, complex conjugate generator 18 and magnitude
generator 19. The complex conjugate .gamma.*(n) of the complex low pass
version signal .gamma.(n) is supplied from complex conjugate generator 18
to delay unit 20. In turn, delay unit 20 delays each sample representation
of .gamma.*(n) a predetermined number, k, of sample intervals. In this
example, a delay k, i.e., lag, of two (2) sample intervals is
advantageously used. The delayed complex conjugate .gamma.*(n-k) is
supplied to multiplier 17 where it is combined via multiplication with
.gamma.(n) to yield .gamma.(n).gamma.*(n-k). In turn, the combined signal
.gamma.(n).gamma.*(n-k) is supplied to averaging filter 21 which yields
the complex autocorrelation of .gamma.(n), namely,
##EQU1##
where N is a number of samples, i.e., window size, used to generate a
so-called estimate of R(k). In one example, N=1024 for classifying voice
band data signals and N=256 for classifying between speech and voice band
data. Averaging filter 21 generates the complex autocorrelation
R(k)=R(k)+.gamma.(n).gamma.*(n-k)/N, i.e., the present estimate, R(k) is
the previous estimate of R(k) plus an averaged update portion
.gamma.(n).gamma.*(n-k)/N. It is important to note that the magnitude of
the complex autocorrelation R(k) of digital signal .gamma.(n) is
independent of the carrier frequency of the voice band data signal d(n).
Consequently, the results of the classifying arrangement of the invention
are not modulated by the voice band data signal carrier frequency and,
accurately, reflex the baud rates of the voiceband data signals. The
complex autocorrelation R(k) is supplied to normalized magnitude unit 22
and normalized real part unit 23.
Normalized magnitude unit 22 generates
##EQU2##
normalized by R(O), because the signal level of d(n) may vary. R(O) is
representative of the power of incoming signal d(n). In this example, the
value of C(k) employed is, as indicated above, at delay k=2 and the
normalization factor is R(k) at delay k=0. The output C(k), or in this
example C(2), from normalized magnitude unit 22 is supplied to threshold
detectors unit 24. Threshold detectors unit 24 includes a plurality of
threshold detectors (not shown) which discriminate between the baud rates
of the voice band data signals. The particular threshold levels are
obtined by minimizing the probability of false detection under the
assumption that C(k), at a given delay k, i.e., lag, if Gaussian
distributed over many experimental results. The delay value k=2 was
selected in this example because it yields the best overall results.
However, for lower transmission rates, e.g., 1200 and 300 FSK, a delay of
k=3 seems to produce better results. In this example, if
0.ltoreq.C(2).ltoreq.0.646, then the voice band data signal has a baud
rate of 2400/sec which relates to a 9600 or higher bit/sec voice band data
signal; if 0.646<C(2).ltoreq.0.785, the voice band data signal has a baud
rate of 1600/sec which relates to a 4800 bit/sec voice band data signal,
if 0.785<C(2).ltoreq.0.878, then the voice band data signal has a baud
rate of 1200/sec which relates to a 2400 bit/sec voice band data signal;
and if 0.878<C(2).ltoreq.1, then the voice band data signals has a baud
rate of .ltoreq.600/sec which relates to voice band data signals having
bit rates less than 1200 bit/sec. The results from threshold detectors
unit 24 are supplied to utilization means 32 for use as desired. For
example, the results are advantageously used to adjust the number of bits
used in an ADPCM coder for improving the quality and efficiency of
transmitting voice band data signals.
Normalized real part unit 23 generates R.sub.d (k)=-Real[R(k)]/R(O) which
is related to the phase of the complex autocorrelation of .gamma.(n). The
real part of the complex autocorrelation R(k) is normalized by the
autocorrelation value at k=0 to compensate for level changes in d(n).
Again, the best overall results are obtained at a delay lag k=2. Thus, if
R.sub.d (2)>0 the complex autocorrelation has a first phase, for example,
a phase in the second and third quadrants and if R.sub.d (2).ltoreq.0 the
autocorrelation has a second phase, for example, a phase in the first and
fourth quadrants. It has been determined that if R.sub.d (2).ltoreq.0 that
d(n) is a voice band data signal and if R.sub.d (2)>0 the signal is a
speech signal. The R.sub.d (2) signal is supplied to an input of two
dimensional threshold detector 25. Threshold detector 25 is jointly
responsive to R.sub.d (k) and signal .eta. from ratio -1 unit 29 to yield
a final determination of whether d(n) is a speech or voice band data
signal. As is explained hereinafter .eta.=m.sub.2 /m.sub.1.sup.2 -1 where
m.sub.1 is the first order absolute moment of the low pass version
.gamma.(n) of d(n), namely,
##EQU3##
or m.sub.1 =m.sub.1 +.vertline..gamma.(n).vertline./N and m.sub.2 is the
second order absolute moment of the low pass version .gamma.(n) of d(n),
namely,
##EQU4##
or m.sub.2 =m.sub.2 +.vertline..gamma.(n).vertline..sup.2 /N. In this
example, N is 256 for speech detection and 1024 for voice band data
detection. Threshold detector 25, in this example, yields a signal
representative that d(n) is a speech signal when R.sub.d (2)>0 or
.eta.>0.3, otherwise it yields a signal representative that d(n) is a
voice band data signal. Such a threshold detector would include two
separate detectors having their outputs ORed. The output from threshold
detector 25 is supplied to utilization means 32 for use as desired.
Although both the so-called phase R.sub.d (2) and the normalized variance
.eta. are used to distinguish between speech and voice band data, it will
be apparent that either one may be used individually to make such a
determination.
It has also been determined that it is desirable and important to detect
the type of modulation scheme used in the voice band data signal in order
to accurately distinguish between certain of the voice band data signals.
For example, use of the complex autocorrelation related parameter C(k)
described above does not accurately distinguish a 1200 FSK signal from a
2400 bit/sec or 4800 bit/sec signal. It has been determined that a
predetermined relationship between a first order absolute moment and a
second order absolute moment of the complex low pass version .gamma.(n) of
d(n) adequately distinguishes as to whether the modulation type is FSK,
PSK, and QAM. By definition, the moment of order P of a signal x(n) is the
average of x.sup.P (n) and the absolute moment of order P of a signal x(n)
is the average of .vertline.x(n).vertline..sup.P.
To this end, magnitude unit 19 generates
.vertline..gamma.(n).vertline.=.sqroot.u.sup.2 (n)+v.sup.2 (n). Then the
first order moment of .vertline..gamma.(n).vertline. can be evaluated as
m.sub.1 =m.sub.1 +.vertline..gamma.(n).vertline./N; and the second order
moment of .vertline..gamma.(n).vertline. can be evaluated as m.sub.2
=m.sub.2 +.vertline..gamma.(n).vertline..sup.2 /N. Again, in this example,
for detecting speech N=256 and for detecting voice band data N=1024. Thus,
the first order moment m.sub.1 of .vertline..gamma.(n).vertline. is
generated by averaging filter 26 which yields m.sub.1 =m.sub.1
+.vertline..gamma.(n).vertline./N. Then, squarer unit 28 yields
m.sub.1.sup.2 which, in turn, is supplied to ratio -1 unit 29. Similarly,
the second order moment m.sub.2 of .vertline..gamma.(n).vertline. is
generated by supplying .vertline..gamma.(n).vertline. to squarer unit 27
to yield .vertline..gamma.(n).vertline..sup.2 and then averaging filter 30
yields m.sub.2 =m.sub.2 +.vertline..gamma.(n).vertline..sup.2 /N. Then,
m.sub.2 is supplied to ratio -1 unit 29 which, in turn, yields a so-called
normalized variance .eta. of .vertline..gamma.(n).vertline., namely,
.eta.=m.sub.2 /m.sub.1.sup.2 -1.
As indicated above the normalized variance .eta. is supplied to two
dimensional threshold detector 25 for use in distinguishing between speech
and voice band data signals. The normalized variance .eta. is also
supplied to threshold detectors 31 for distinguishing between several
types of voice band data modulation. In this example, the modulation types
being distinguished are frequency shift keying (FSK), pulse shift keying
(PSK) and quadrature amplitude modulation (QAM). In this example, it has
been determined that if 0<.eta..ltoreq.0.021, then the modulation type is
FSK; if 0.021<.eta..ltoreq.0.122 then the modulation type is PSK; and if
0.122<.eta. then the modulation type is QAM. The results from threshold
detectors 31 are supplied to utilization means 32 where they are used for
determining the particular voice band data signal being received.
Thus, it is seen that use of .eta. allows to discriminate between FSK, PSK
and QAM voice band data signals, while C(2) can be used to discriminate
among 2400 baud/sec, 1600 baud/sec, 1200 baud/sec and 600 baud/sec or
lower baud signals. These latter signals are related to 9600 bit/sec, 4800
bit/sec, 2400 bit/sec and 1200 bit/sec or lower bit rate signals. If
desired C(k) at delay k=3, i.e., C(3), can be generated as described above
for C(2) and used to discriminate between 1200 bit/sec and 300 bit/sec
voice band data signals.
In situations where it is desired only to discriminate 9600 bit/sec voice
band data signals from all others and can tolerate assigning to the 4800
QAM voice band data signal a higher speed classification, then use of the
normalized variance .eta. for N.gtoreq.512 is sufficient.
Preferably, the above described classification arrangements are to be
implemented on a very large scale integrated (VLSI) circuit. However, the
classification arrangements are also readily implemented via use of a
processor, for example, an array processor. To this end, FIGS. 2 and 3
when combined A--A and B--B form a flow chart illustrating the steps for
implementing the classification of incoming digital signals, in accordance
with aspects of the invention. Accordingly, the program routine is entered
via initialized step 201. Conditional branch point 202 tests to determine
if input energy is present. If the test result is YES, energy is present
and operational block 203 causes and N to be set to N=256. As noted above
N=256 is the number of samples used to detect whether the incoming signal
d(n) is speech or voice band data. Operational block 204 causes n, R(k),
m.sub.1 and m.sub.2 to be set to n=1, R(k)=0, m.sub.1 =0 and m.sub.2 =0.
Operational block 205 causes the computation of a(n)=d(n) cos (.pi. n/2).
and b(n)=d(n) sin (.pi.n/2). Operational block 206 causes generation of
the complex low pass version .gamma.(n) of incoming signal d(n) by low
pass filtering by the filter function g(n) the results of step 205, namely
.gamma.(n)=[a(n)-jb(n)] g(n), where indicates the convolution function.
As indicated above, in this example, a low pass filter function g(n) is
employed that is a second order recursive filter with a cutoff frequency
at 2 kHz. Operational block 207 causes estimates of R(k), m.sub.1 and
m.sub.2 to be updated. As indicated above, R(k) is the autocorrelation of
incoming complex digital signal .gamma.(n) and the updated value is
R(k)=R(k)+.gamma.(n).gamma.*(n-k)/N where * indicates the complex
conjugate. In this example, a delay, i.e., lag of k=2 sample intervals is
used. Again m.sub.1 is the first order moment of
.vertline..gamma.(n).vertline. and its updated value is m.sub.1 =m.sub.1
+.vertline..gamma.(n).vertline./N and m.sub.2 is the second order moment
of .vertline..gamma.(n).vertline. and its updated value is m.sub.2
=m.sub.2 +.vertline..gamma.(n).vertline..sup.2 /N. Operational block 208
causes the setting of n=n+1. Conditional branch point 209 tests whether
n.ltoreq.N. If the test result is YES control is returned to operational
block 205 and steps 205-209 are iterated until the test result in step 209
is NO. This indicates that the 256 samples window has occurred over which
the values of R(k), m.sub.1 and m.sub.2 are being estimated. Then,
operational block 210 causes the following calculations to be performed:
the normalized magnitude C(k) of the complex autocorrelation of
.gamma.(n), namely C(k)=.vertline.R(k).vertline./R(0) where R(0) is the
complex autocorrelation of .gamma.(n) at delay k=0; the normalized real
part R.sub.d (2) of the complex autocorrelation at delay k=2, namely,
R.sub.d (2)=-Real[R(2)]/R(0); and the normalized variance .eta. of the
magnitude of the complex low pass version .gamma.(n) of incoming signal
d(n), namely, .eta.=m.sub.2 /m.sub.1.sup.2 -1, where m.sub.1 is the first
order moment of .vertline..gamma.(n).vertline. and m.sub.2 is the second
order moment of .vertline..gamma.(n).vertline. from step 207. Conditional
branch point 211 tests to determine if the incoming signal is speech or
voice band data by determining, in this example, if R.sub.d (2)>0 or
.eta.>0.3. If the test result in step 211 is YES, operational block 212
sets an indicator that the incoming signal is speech. Thereafter, the
process is stopped via 213. If the test result in step 211 is NO,
operational block 214 sets an indicator that the incoming signal is voice
band data. Conditional branch point 215 tests to determine if N=256. If
the test result is YES, operational block 216 sets N=1024 and n=1, and
control is returned to operational block 204. As indicated above, in this
example a window of 1024 samples is used to generate the estimates of
R(k), m.sub.1 and m.sub.2 for voice band data signals. Thereafter, steps
204 through 211, 214 and 215 are iterated. Since N=1024 the test result in
step 215 is NO. Thereafter, operational block 217 determines the voice
band data signal parameters in this example, as follows: if
0.ltoreq.C(2).ltoreq.0.646 then the incoming signal baud rate is 2400/sec;
if 0.646<C(2).ltoreq.0.785 then the incoming signal baud rate is 1600/sec;
if 0.785<C(2).ltoreq.0.878 then the incoming signal baud rate is 1200/sec;
if 0.878<C(2).ltoreq.1 then the incoming signal baud rate is equal to or
less than 600/sec; if 0<.eta..ltoreq.0.021 then the modulation type for
the incoming signal is FSK; if 0.021<.eta..ltoreq.0.122 then the
modulation type for the incoming signal is PSK; and if 0.122<.eta. then
the modulation type for the incoming signal is QAM. Thereafter, the
process is stopped via 218.
* * * * *
|
|
 |
|
 |
Description |
|
