 |
Description  |
|
|
FIELD OF THE INVENTION
This invention relates to analog signal processing techniques for enhancing
the quality of signals transmitted through limited bandwidth channels.
More particularly the invention relates to expanded partial response
processing of analog signals.
BACKGROUND OF THE INVENTION
Signal communication privacy is a matter of concern in radiotelephone
systems. Implementation of privacy techniques involving enciphering of
signals and the processing of signals to place them within the
transmission bandwidth of a determined transmission medium, such as a
telephone system voice channel, usually are accompanied by signal
bandwidth expansion. The inconsistent requirements of enlarged bandwidth
due to signal processing and/or enciphering and limited bandwidth of a
transmission channel can result in reduced quality of transmitted signals.
In radiotelephone systems, the need for private transmission is
particularly acute, and the bandwidth conflict is significant because the
radio link in a communication path of such a system is frequently in
series with a relatively narrow band channel of a wire circuit in the
public switched telephone network.
Sample masking is one signal enciphering technique that is experiencing
growing popularity because the enciphered signals lack remanent clear
information energy and so are unintelligible. In sample masking, signal
samples are combined by modulo arithmetic with a sequence of signals
representing randomly, and with substantially equal probability, the
discrete levels of an amplitude range in which the signal samples are
quantized. However, the outlined enciphering process expands the
bandwidth, as noted above, of a voice signal of reasonable quality beyond
the bandwidth of a typical telephone channel. Attempts to meet that
challenge at a given effective bit rate have resulted in arrangements that
either tolerate substantial reduction in the quality of a transmitted
signal, or require special transmission channels of adequate bandwidth, or
are so expensive to implement that they are not practical for commercial
purposes.
S. B. Weinstein teaches a sample masking system in a paper "Sampling-based
Techniques for Voice Scrambling," ICC '80 Conference Record, Vol. 1, pages
16.2.1-16.2.6. Input voice samples are constrained to an amplitude range
which is reduced at each extreme by an amount corresponding to two or
three standard deviations of the ambient noise to make the noise in
recovered speech the same as if no scrambling had been performed. Those
same input signals are also band limited to a 2500 Hertz bandwidth so
that, after the enciphering bandwidth expansion, the signals will have a
bandwidth of about 2933 Hertz that approximately matches that of the
channel to be used.
An R. T. Adams et al. U.S. Pat. No. 4,283,602 teaches a baseband
transmission system of the sample masking type in which enciphered signals
are low pass filtered to the same bandwidth as both the original input
signals and the transmission medium which is to receive the enciphered
signals. Experience has shown that such arrangements, employing only low
pass filtering to a common bandwidth create intolerable intersymbol
interference which renders the signal unintelligible.
In U.S. Pat. No. 4,398,062 to D. D. McRae et al. is shown another system of
the sample masking type and employing low pass filters in the respective
channels of a quadrature modulator to "condition the pulse trains"
provided from the output of a circuit that splits the enciphered signal
into two half-pulse-rate symbol trains. The nature of the conditioning is
not explained.
A paper "An Efficient Technique for Sample-Masked Voice Transmission" by R.
J. Cosentino et al. appeared in the IEEE Journal On Selected Areas in
Communications, Vol SAC-2 No. 3 May 1984 at pages 426-433, and describes
the use of linear filtering of an enciphered and sample masked signal to
modify the enciphered component without processing the signal component. A
so-called spectrum-modifier filter was approximated by truncating a
weighted sin x/x filter to a length of 599 which is expensive to implement
because of the length of the filter.
Another Cosentino et al. paper "Secure Voice-Bandwidth Modem" was presented
at the 1982 Carnahan Conference on Security Technology, Lexington, Ky.,
May 12, 1982; and describes a modem employing single-channel sample
masking with sample rate decimation and interpolation.
In Chapter 7 of Digital Communications: Microwave Applications, by K.
Feher, Prentice-Hall, Inc., 1981, a paper "Correlative (Partial Response)
Techniques and Applications to Digital Radio Systems" by A. Lender,
includes discussion of the concept of introducing into a data signal a
controlled amount of intersymbol interference so that a somewhat higher
symbol rate can be accommodated in a channel of given bandwidth. Both
binary and non-binary data signals are considered. Modulo addition
precoding is indicated to reduce error propagation. Examples of modulo Q=4
(seven level) multilevel systems are presented. A 31-level system is
mentioned but not disclosed. Bit slicing is used for data detection and is
contemplated for error detection also. In data transmission systems, the
noise effects injected during transmission usually do not exceed the
amplitude span between adjacent levels.
SUMMARY OF THE INVENTION
The foregoing problems of conflict between bandwidth expansion in
enciphering operations and bandwidth limitations on transmission channels
for the enciphered signals are alleviated by applying partial response
processing to enciphered analog signal samples. In one embodiment, the
analog samples are quantized in a sufficient number of levels to
approximate a continuously varying analog signal in which an error of a
small percentage of the total number of levels in the allowable analog
signal range is not readily perceptible to the human ear.
BRIEF DESCRIPTION OF THE DRAWINGS
A more complete understanding of the invention and various features,
objects, and advantages thereof may be obtained from a consideration of
the following Detailed Description in connection with the appended claims
and the attached Drawings in which:
FIG. 1 is a simplified block and line diagram of a communication system
employing the invention;
FIGS. 2-5 are diagrams illustrating operation and advantages of the
invention; and
FIGS. 6 and 7 are further detailed diagrams of the transmitting and
receiving terminals, respectively, of the system of FIG. 1.
DETAILED DESCRIPTION
In FIG. 1, the invention is shown as applied to one direction of
transmission for a communication path in, e.g., a cellular radiotelephone
system. A path for the other direction would be essentially the same. One
example of systems of the type in which the invention is illustratively
useful is described at length in a collection of papers in the January
1979 Bell System Technical Journal. Accordingly, there are omitted from
the drawings herein some details of the system which are not necessary for
an understanding of the invention.
The path illustrated in FIG. 1 includes a mobile facility, or terminal, 10,
such as a vehicular unit which transports radio terminal equipment about
the cells of the system service area. Terminal 10 is coupled through a
radio link 11 to a fixed facility 12 which is in turn advantageously
coupled to a zone office (not shown) of the public switched telephone
network (PSTN). Fixed facility 12 advantageously includes a cell antenna
site 13 with associated channel radio equipment and controls therefor, a
mobile telecommunications switching office (MTSO) 16 and a signal
processing facility receiving terminal 17 corresponding to the mobile
terminal 10.
In the mobile terminal 10 and fixed receiving terminal 17, various
functions are performed; and they can be executed in either analog or
digital implementations using individual function techniques already known
in the art. For example, in one application, they were implemented using
multiple digital signal processors and processor functions of the type
shown in the September 1981 issue of The Bell System Technical Journal,
which issue was subtitled "Digital Signal Processor--A Programmable
Integrated Circuit For Signal Processing." Another suitable circuit of the
type is the TMS-320 signal processor manufactured by Texas Instruments
Incorporated.
Analog speech signals are coupled through an anti-aliasing filter 14 which
limits those signals to about a 300-3000 Hz band. An analog-to-digital
converter 18 receives the filtered signals and schematically represents
the establishment of an amplitude quantization range for the analog signal
information. That range advantageously has a sufficient number Q of levels
to represent speech signal variations with quantizing noise that is low
enough to be of no significant annoyance to parties communicating over the
illustrated path. In a quantizing domain with that many levels, the step
size between levels is usually sufficient in relation to the magnitude of
noise injected in a signal during transmission between terminals 10 and 17
that the noise can cause level errors of plural levels n; but n is such a
small percent of the total number Q of levels that the error occurrence is
not readily perceived by the typical human hearing function, hereinafter
characterized simply as the human ear. In other words, the quantizing step
size between levels is so related to the extent of noise that might be
injected during transmission that transmission-noise-induced errors of a
small number, at least one, of levels are in a range that is compatible
with the error tolerating characteristic of the human ear. It has been
found further that the ear can also tolerate errors of five percent or
more in the number of levels in the quantizing range without introducing a
significant degradation in perceived speech quality.
In one application of the invention, Q=256 levels were found to satisfy the
above requirements. The exact number is not critical. It has been found
that 256 levels, which can be represented by 8-bit binary coded digital
words, are convenient for digital signal processing, e.g., functions such
as limiting, enciphering, and partial response encoding of signals, to
achieve an advantageous compromise between maximizing information
throughput rate and transmitted signal error rate.
However, it is convenient for other signal processing functions such as
scaling, filtering, and modulation to use 16-bit arithmetic to handle
arithmetic operation carries and truncate the results to the 8-bit level
for the previously indicated functions.
Given such a value, i.e. 256, for Q, it has been further found that each
change of one bit in the number of bits used to represent digitally the
analog signal amplitudes causes about a six decibel (dB) change in
quantizing noise--fewer bits meaning more noise and vice versa. It has
been found that a quantizing range of 64 levels (represented by 6 bits) is
about as small as can be employed and still generally get user reactions
of at least good signal quality. In a number of instances, signals were
quantized and digitally represented in a quantizing domain of many more
levels than 256; but only the 256 levels (the 8 most significant bits of
the digital representations) were used for certain of the illustrated
functions, as previously noted, as an advantageous compromise between a
desire for low quantizing noise and a need for a practical word bit length
for those digital signal processing functions.
Output of converter 18 is applied to a preliminary speech processing
function block 19 which provides functions such as limiting, scaling,
splitting, filtering, and decimating. These will be described in more
detail in connection with FIG. 6, and they are provided for protection
against noise for certain modulo arithmetic functions to be described and
for developing in-phase and quadrature-phase versions of the signal to
facilitate the filtering and other functions to be described. The
resulting in-phase and quadrature versions of digitally represented analog
sample amplitudes are applied through a sample masking and partial
response encoding function block 20. Each of the masking and encoding
functions involves a different modulo arithmetic operation. Yet another
function block 21 provides one part, i.e. the square root, of a partial
response filter function and a quadrature modulation function to translate
the processed baseband in-phase and quadrature signals up to a suitable
passband range, e.g., 300-3000 Hz, Which is characteristic of many
circuits in the PSTN. The digitally represented masked sample amplitudes
are thereafter converted to analog form by a digital-to-analog converter
22.
As will be further described subsequently, at least the sample masking and
the partial response functions take place in the same signal amplitude
domain, that is, in the same signal quantizing range or at least in
different compatible ranges. Compatible domains, or ranges, are those in
which the respective quantizing ranges each meets the foregoing
specifications, as to low quantizing noise and as to compatibility with
the error tolerating capability of the human ear, and signals can be
translated between them by appropriate scaling.
In addition, the cumulative partial response filtering (product of the
square roots of transmit and receive parts of the filtering) is of the
type which injects a predetermined controlled amount of intersymbol
interference. Such filtering has heretofore been employed only for data
transmission systems as to data symbols of a relatively few number of
levels. Usually that number is at least an order of magnitude less than
the number of levels that are needed for adequate representation of speech
signal samples as above outlined. Also the step size with that smaller
number of levels is usually so large that noise is unlikely to cause
errors of even one level. More importantly, however, the intersymbol
interference contemplated in those data systems is between data symbols
whereas it is here contemplated in a sense which might be characterized as
"intersample interference" where each sample is an analog value that is
digitally represented, in the partial response filtering operation, by a
plurality of discrete data symbols for convenience of processing. In other
words, the transfer function of the partial response filtering function is
determined largely by a frequency parameter represented by the frequency
of occurrence of analog sample values which have a discrete
representation.
In the FIG. 2 amplitude versus time diagram are shown analog wave
representations of two time-adjacent impulses, such as data bits in a data
system or speech samples in a voice system. These impulses are of the type
that may have appeared at the input of the converter 22 in FIG. 1 if
conventional Nyquist type filtering had been employed instead of the
indicated partial response filtering. In FIG. 2, the solid line wave is
the current impulse and the dashed wave is the one that preceded it in
point of time. This type of filtering is designed theoretically to shape
the impulses so that they each exhibit maximum amplitude in the center of
a bit time and low amplitude so-called sidelobes on either time side which
have zero crossings at the center points of adjacent bit intervals. That
approach is taken to minimize intersymbol interference; but, as noted in
the aforementioned Lender paper, the theoretical result is physically
unrealizable. Consequently, practical applications have typically reduced
the data transmission bit rate to achieve a reasonably low error rate.
In the FIG. 3 amplitude versus time diagram are shown analog wave
representations of two time-adjacent samples such as are applied to the
input of the converter 22 in FIG. 1 when partial response filtering is
employed. The solid line wave is the current one, and the dashed wave is
the one that preceded it in point of time. The dotted wave is the sum of
the current and prior waves and represents the type of waveshape that
would be recovered at the output of block 30 (to be described) in fixed
facility 12. It can be seen from that figure that the central, or main
lobe, portions of the current and preceding impulses overlap by a known
amount that is fixed for partial response functions by choosing partial
response filtering passband in relation to the signal symbol rate (sample
rate in the case of FIG. 1). Consequently, the partial response processing
introduces a known and controlled amount of intersymbol interference
(ISI); and, because it is known, it can be conveniently removed at a
receiving terminal. Since the ISI is removable, it is not necessary to
reduce the transmission rate of impulses. It will be hereinafter shown
that this controlled ISI feature, which has been known for use in data bit
impulse transmission systems is, when appropriately applied, also
advantageously useful in connection with speech sample impulses.
FIG. 4 depicts a set of superimposed signal frequency bandwidth diagrams to
illustrate the advantage of partial response processing in a bandwidth
sense as newly applied to the speech transmission problem. For the purpose
of these diagrams, any arbitrarily given uniform amplitude in a passband,
impulse transmission rate, impulse error rate, and transmission channel
bandwidth are assumed. It is also assumed that, in the illustrative
application to sample masking, the baseband input signals before sample
masking are band limited to a band such that after any intervening
processing the resulting signal produced by digital-to-analog converter 22
will have a bandwidth approximately the same as that of the transmission
channel, e.g., the allowable bandwidth of the wire and radio channel
including radio link 11.
The portion of the FIG. 4 diagram indicated by the bracket A represents a
maximum limited band that must be imposed on input signals to prior
systems employing Nyquist type filtering of the sample masked impulses
that are to be applied to the radio link 11. The wider bandwidth portion
indicated by the bracket B represents the wider limited band that must be
imposed on input signals to systems of the type in FIG. 1 and employing
partial response type filtering of the sample masked impulses to be
applied to the same radio link 11. Portions A and B are typically passband
signals here assumed to have a common lowest frequency such as 300 Hz.
Portions A and B are double crosshatched to distinguish visually the
premasking part of the diagram from the remainder.
FIG. 4 portions indicated by brackets C and D represent the proportionately
similar bandwidth expansion that normally results from sample masking for
the input signals represented by portions A and B, respectively. Parts of
C and D extending beyond the upper frequency end of B are single cross
hatched. Finally, the FIG. 4 portion indicated by bracket E represents the
ultimate bandwidth achieved for application of both the Nyquist and the
Partial Response filtered signals to the common channel 11; and the part
of E extending beyond D is uncrosshatched. It should be understood that
although for convenience the portion E has been shown as having the same
lowest frequency as the portions A through D, E is typically modulated up
to be a passband signal of, e.g., 300-3000 Hz, for either A or B input in
order to match typical telephone channel bandwidths in the PSTN.
It should be apparent from the foregoing that the bandwidth expansion
associated necessarily with the Nyquist filtered signal, i.e., from the
upper frequency of C to the upper frequency of E, is much greater than the
expansion associated with the partial response filtered signal, i.e., from
the upper frequency of D to the upper frequency of E. It can be seen in
FIG. 4 that the D and E bandwidths are approximately the same.
Consequently, given a system with a transmission channel bandwidth
corresponding to E of FIG. 4, a system employing the heretofore typical
Nyquist filtering must start with a band limited input signal of
significantly narrower bandwidth, than is the case for a system using
partial response filtering, in order to allow for the bandwidth expansions
of both the sample masking operation and that of the Nyquist filtering to
be done. It has been found that this difference in starting bandwidth is
the margin between a receiving terminal signal that has sufficient
bandwidth to represent an adequate frequency range of both male and female
voices to be easy to listen to for extended periods and a receiving
terminal signal that has insufficient bandwidth to represent those same
voices without significant distortion, which renders the signals hard to
listen to without undue fatigue for extended periods of time.
FIG. 5 illustrates the discovery that the relation between quantizing step
size and transmission noise required for data transmission is not
necessary for analog speech sample transmission because of the
aforementioned quantizing level detection error tolerance capability of
the human hearing function, i.e., the so-called human ear. In FIG. 5,
there is shown a randomly varying analog signal wave segment 37
representing transmission noise typically injected into a transmitted
signal during transmission through a medium such as the FIG. 1 radio link
11. A pair of solid horizontal lines 38 represent a minimum binary coded
data signal quantizing step size required to enable transmitted signal
reception with a reasonably low error rate. By contrast, the dashed lines
39 represent a minimum, though not critical, analog speech sample signal
quantizing step size for enabling transmitted signal reception for good
quality speech recovery. Thus, quantizing step size for speech can be made
small enough to reduce quantizing noise to reasonable levels with
sufficient resolution for good speech reproduction even though that small
step size is much less than the expected transmission noise amplitude.
That is, the human ear does not readily perceive level errors such as can
be produced by noise of the indicated relative magnitude.
Returning to FIG. 1, an equalization training signal generator 15 provides
a signal for training an equalization function in terminal 17. Output of
generator 15 is switchably substituted for output of block 21 at
appropriate times in the course of system operation, as is well known.
A timing unit 23 in FIG. 1 produces various clock signals employed for the
function blocks in the remainder of the terminal 10. In addition the unit
provides clock to a generator 26 of tones for, e.g., pilot tones employed
in a single sideband system of the sort contemplated in connection with
the single sideband receiver of the U.S. Pat. No. 4,403,348 of K. W.
Leland and N. R. Sollenberger. Such tones can also be utilized for
enciphering function countdown handshake between terminals and for other
synchronizing functions, as is known in the art. The output of generator
26 is, in a digitally implemented embodiment, injected into the signal
path, e.g., by arithmetic adding of time coincident sample values, at the
input to converter 22. This location is employed because the tones are
advantageously generated digitally, e.g., by known digital signal
processing operations; and the resulting tone sample values can be here
injected by adding digitally represented sample amplitudes. The resulting
digitally represented speech and tone sample sums are converted to analog
by converter 22 and applied via a sample smoothing filter 24 and
appropriate transmitter radio frequency circuits, i.e. cellular audio
processing and modulation circuit 25, to the radio link 11.
Radio signals received at the cell site 13 are restored to their passband
position, e.g., 300-3000 Hz. Those signals are then passed on a voice
signal trunk used for the particular radio channel employed to the MTSO 16
where they are switched to a wire path, including terminal 17 for recovery
of clear speech signals to be forwarded to the PSTN zone office. Terminal
17 is advantageously located on the zone office side of the MTSO so one
need not deal with enciphering key changes when performing handoff.
It can be seen from FIG. 1 that the receiving terminal 17 functions are in
many respects essentially the inverse of those in the mobile terminal 10.
Received analog samples are coupled through an anti-aliasing filter 34 and
again converted to digital format by an analog-to-digital converter 27.
Tone samples are extracted for detection by digital filtering functions in
tone detectors 28. Output of those detectors is applied to a timing unit
29 which produces the various clock signals required for the functions of
terminal 17. Output of converter 27 is also applied to a quadrature
demodulator function block 30, which develops in-phase and quadrature
phase versions of the signal samples, applies the remaining square root
portion of the partial response filtering, and performs decimation prior
to application of the signals to an equalizer function block 31. That
block is operative for equalization training at appropriate times in
system operation, e.g., when the terminal is started up for the incoming
message. Otherwise, block 31 equalizes signals which are passed on to a
partial response decoding and sample unmasking function block 32.
It is advantageous in some applications, as here, to split the partial
response filtering function between two square root portions in terminal
10 block 21 and terminal 17 block 30. However, that filtering function can
be split in other ways which are well known in the art.
Equalized outputs of block 31 are in block 32 unmasked by subtraction of
the pseudorandom sample sequence used at terminal 10 and partial response
decoded, both being modulo-Q operations. The deciphered speech samples are
passed on to a speech deprocessing function block 33 where limiting,
inverse scaling, interpolation, and filtering take place. Finally, the
clear speech samples in digital form are restored to analog form by
another digital-to-analog converter 36 prior to being coupled through a
filter 40 for transmission to the local zone office in the public switched
telephone networks (PSTN). Filter 40 is a low pass filter with cutoff
frequency between the upper edge of the voice band and the sample rate for
integrating the samples to restore the speech signal information.
It has been found that the outlined technique for sample masking and
partial response processing of analog sample amplitude values allows use
of the desired speech bandwidth and results in relatively minor
user-perceived reduction in signal quality, in the range of 0-3 dB,
compared to the signal quality of a standard cellular radiotelephone
channel.
FIG. 6 shows additional detail of the operations of the various functional
blocks in the transmitter terminal 10 depicted in FIG. 1. Both analog and
digital techniques are known in the art for implementing each of the
circuit blocks shown in FIG. 6 in isolation, and such block details
comprise no part of the present invention. The input to the speech
processing function 19 in FIG. 6 is the digital form of analog speech
samples provided from converter 18 in FIG. 1. Signal energy in the
300-2800 Hz baseband, corresponding to the band B in FIG. 4, is presented
illustratively in an 8 kiloHertz, Q-level quantized sample stream. The
sample amplitudes are in a range .+-.y volts. Signal energy is split into
two paths in which it is applied to modulators 41 and 42 for multiplying
with sin .omega..sub.1 t and cos .omega..sub.1 t signals, illustratively
at 1550 Hz, to create the usual in-phase and quadrature-phase signals for
convenience of further processing. Local oscillation sources for the
modulators 41 and 42, and for other modulators and pulse sources to be
mentioned, are advantageously a common oscillator in timing unit 23, with
outputs properly phased, to assure coordinated operation of such sources
within the transmitter. Operation of that oscillator is further
coordinated with corresponding circuit elements in the receiving terminal
by virtue of transmission of outputs of tone generator 26, as previously
mentioned. Since further processing in these two in-phase and
quadrature-phase paths is identical up to the point of later quadrature
modulation to place the signals in a proper passband, only one path will
be described.
Output products from modulator 41 are low pass filtered in a filter 43 to
confine further processing to the baseband product in the illustrative
band 0-1250 Hz.
From a Nyquist sampling standpoint, a minimum sampling rate of 2500 Hz is
required. Accordingly, the output of filter 43 is applied to a decimating
function circuit 49 wherein only every third sample in the normal signal
sequence is selected for use in the processing operations which follow.
Decimating the 8 kHz sampled signal by three is easy to do and yields a
2667 Hz sampling rate. That new low rate also reduces the complexity of
the digital signal processing to follow, and it makes the subsequent
problem of transmitting masked sample information over a channel of
limited bandwidth tractable.
Output of decimating circuit 49 is applied through a multiplier 51, to be
described, to one input of an amplitude scaling multiplier 46 which brings
the received sampled signals into a range for efficient use of the
available part of the quantizing range. The received sample amplitude
digital representations have already been quantized in a range of 0-255
levels, i.e., Q=256, wherein each level number is the number of a voltage
at the lowest voltage of a step of uniform size throughout the range. It
is assumed that the maximum anticipated normal signal excursion
substantially fills the level range. Modulo arithmetic operations are to
be performed in the signal processing path as has been mentioned; and it
is assumed illustratively that they are all in the same amplitude domain
of 256 levels of the same step size.
It is permissible for normal signal processing operations, e.g., addition
and subtraction, to cause modulo wrap-around because it is well known that
a modulo subtractor following a modulo adder automatically provides
compensating "unwrapping" in the receiving terminal. However, it is not
good if transmission noise injected between terminals should cause
wrap-around because that would cause large errors in the speech signal
restoration result. For example, a 2-level transmission noise added to a
255-level speech sample equals a 1-level, modulo-256, result; and the
receiving terminal operations will perceive it as a small signal instead
of a large one because no inverse function is in the receiver terminal to
take the transmission noise out. Thus, a small amount of transmission
noise can create a large error in the recovered waveform. Therefore,
output signals from the decimating circuit 49 are scaled so their maximum
excursion fills a range less than that of the full Q=256 levels.
To that end, the total range of levels is reduced by some number, X, of
levels at each of the low and high ends of the range to leave an effective
signal range of Q-2X levels. For example, if transmission noise is
expected normally to be a maximum of 10 levels, X is chosen to be about 28
levels to be sure that no noise-induced wrap-around can occur. Multiplier
46 performs the necessary scaling to bring the .+-.y volt signal range
into the Q-2X level range and receives both the output of multiplier 51
and a multiplier factor (Q-2X)/2Y.
An adder 47 biases the scaled bipolar signals by Q/2 levels to shift the
center of the signal range to the center of the Q=256 level range in order
to facilitate modulo arithmetic. The centered signals are then amplitude
limited to the range from X to Q-X levels in a limiter 48 to remove any
extraordinary excursions that may have been present in the quantized
received signals and persisted through the mentioned scaling and level
shifting functions.
Output from the limiting circuit 48 in the speech processing function 19 is
applied to one input of an adder 50 for sample masking the speech samples.
Adder 50 is operated for modulo operation, as is known in the art for
sample masking, and here the operation is modulo Q. To each successive
signal sample Sn is added a random number Rn drawn from an enciphering
sequence having values which are mutually independent and uniformly
distributed over the desired coding interval Q, i.e., illustratively the
same amplitude domain as that of the samples being masked. The sum An is
taken modulo-Q so that
An=Sn+Rn (Mod-Q).
The result An of this process is a sequence of random variables, the
elements of which are uniformly distributed over the range Q, provided
Q>Maximum Magnitude of Sn, and have amplitudes statistically independent
of all other element values. Forming the enciphered sequences in the
foregoing manner from the sampled version of the signal permits a secure
waveform to be produced for transmission. The random number sequence used
in the masking process is advantageously derived through use of the well
known DES algorithm approved by the National Bureau of Standards. Although
the same symbol Rn is employed for the random sequences of both the
in-phase and quadrature-phase branches in FIG. 6, the sequences used in
the two branches are advantageously different, mutually independent
sequences. The same is true of the sequences Rn in the receiver of FIG. 7.
An interesting property of sample masking is that when low level
unenciphered speech signals, e.g., from cross talk originating in the
PSTN, are superimposed on enciphered speech signals in the receiving
terminal, the low level speech will appear after deciphering as
intelligible cross talk superimposed on the desired deciphered component.
Cross talk of this nature will tend to give the false impression to the
user of the enciphered channel that the channel is not secure. In
applications where such crosstalk is a problem, this residual
intelligibility of crosstalk is advantageously eliminated by multiplying
speech signal samples prior to sample masking by a pseudorandom sequence
Br of positive and negative unity gain functions in the multiplier 51.
These effects are removed in the receiver after the deciphering process.
Since the unenciphered crosstalk will not contain this pseudorandom
multiplication, the removal of these factors from the desired signal in
the receiver will also apply them to the crosstalk. This operation will
render the crosstalk unintelligible.
The masked samples are next applied to a partial response precoding circuit
52. There the samples are coupled through another modulo-Q adder wherein
there is subtracted an impulse from the most recent prior sample time as
represented by the T-delay circuit 56 coupled between the output and
second input of adder 53. T is the duration of a sample interval in the
decimated sample sequence and is the reciprocal of the decimated sequence
sample rate. It is, however, emphasized here that this precoding operation
operates, as indicated by the T-delay, on the basis of the decimated
sequence sample interval or rate rather than on the bit interval of the
digital data bits used to represent the sample magnitude in the
illustrative digital processing implementation of the sample masking,
partial response, and associated signal processing. The indicated
precoding operation permits independent interpretation of each sample
amplitude at the receiving terminal 17 in spite of the correlation
properties between successive samples, which will be incorporated in the
partial response filtering to be mentioned. Consequently, there can be no
propagation of errors. A subtractor 55 subtracts a value corresponding to
Q/2 levels from the output of adder 53 to remove the offset imposed by
adder 47 and restore the bipolar characteristic of the signal.
Output from adder 55 is also applied to an interpolating circuit 57 wherein
each masked and precoded sample is interleaved with two zero-amplitude
samples to accomplish a one-to-three interpolation, the inverse of the
decimation in circuit 49, and thereby restore the sample rate to that at
the input of speech processing circuit 19. The interpolated sample
sequence is then coupled through a square root partial response filter 58
to an input of a modulator 59. Filter 58 has a square root of cosine type
of amplitude characteristic as indicated by the small diagram of its
frequency response, or transfer function, in FIG. 6. The filter has a
cutoff frequency substantially less than the sample rate in the precoding
circuit 52. Illustratively, that cutoff is at half of the decimated sample
recurrence frequency, i.e. illustratively 1333 Hz, as indicated by 1/2T in
the diagram, where T is the period of the decimated recurrence frequency,
i.e. illustratively 2667 Hz. The result of such filtering is to
incorporate a fixed, i.e. controlled, amount of intersymbol interference,
again wherein the symbols under consideration are the sample amplitudes
rather than the individual bits of a digitally represented value of that
amplitude. Filtered signals in the in-phase and quadrature-phase paths of
terminal 10 are applied to an input of each of the modulator 59 and
another modulator 60. Those modulators also receive local oscillator
signals sin .omega..sub.2 t and cos .omega..sub.2 t for shifting the
signals in the two paths to the passband position, i.e. 300-3000 Hz, of
the channel to be used for transmission. Outputs of the modulators are
combined in an adder 54 to form the sample masked analog sample stream for
application to the converter 22 in FIG. 1.
In FIG. 7 is shown additional detail for the receiving terminal 17 at a
level similar to that of the terminal 10. In FIG. 7 again the in-phase and
quadrature-phase paths after demodulation are the same so only one is
described. The quantized, sample masked, digitally encoded signals from
analog-to-digital converter 27 in FIG. 1 are quadrature demodulated in
modulators 61 and 62 in accordance with local oscillator signals at the
same frequency, i.e. 1750 Hz, as was employed at modulators 59 and 60 in
FIG. 6 mobile terminal 10. The resulting bandpass product at 0-1333 Hz is
selected by a square root cosine filter 63 which also completes the
intersymbol interference insertion effect of filter 58. Samples in the
output of filter 63 have amplitudes in a range .+-.Q. A three-to-one
decimation is performed in a decimating circuit 66 since the full 8 kHz
sample rate is not needed for the decoding and unmasking operations to be
performed. The training equalizer 31 function of FIG. 1 is not shown in
FIG. 7 for clarity. The decimated 2667 Hz sample rate signal is applied to
a partial response decoding circuit 67. In an adder 64 the value Q is
added to all samples to convert them from a bipolar to a single sided
form. In the latter form the signal is taken modulo-Q in circuit 65. This
removes the transmitter partial response coding effect as well as the
intentionally inserted intersymbol interference and leaves discrete,
masked samples.
Masking is removed in a modulo-Q subtractor 70 by subtracting the same
pseudorandom pulse sequence Rn used in terminal 10. In the speech
deprocessing circuit 33, the clear samples are limited (after both
modulo-Q arithmetic operations) by a limiter 72 to the same {X, Q-X} range
employed in limiter 48, restored to their normal ground reference by
subtractor 73 removing the Q/2 bias, and scaled back to the originally
received sample amplitude range by multiplication in 76 by 2Y/(Q-2X).
Likewise, the unmasked samples are, in a multiplier 71, multiplied by the
reciprocal pseudorandom, plus and minus, unity gain sequence 1/Br to
remove the crosstalk protection effect. An interpolating circuit 77
interleaves each sample with two zero-amplitude samples, in a manner that
is the inverse of the decimation in circuit 66, to restore the original 8
kHz sample rate; and a low pass filter | | |
