 |
Description  |
|
|
BACKGROUND OF THE INVENTION
The present invention relates to a type of digital voice switch which is
generally used in voice communication channels to detect speech in the
presence of noise. In particular, the present invention relates to a
digital voice switch which employs a speech detector having a variable
speech threshold level, a noise detector having a variable noise threshold
level, a disabling detector having a fixed maximum threshold level and a
threshold adjustment circuitry which provides rapid adjustment of the
speech and noise threshold levels.
Voice switches are known in the art as devices which distinguish between
vocal sounds and noise carried by a communications channel. Devices of
this nature have a number of known uses. For example, in a communication
system which includes n voice input channels and m voice output channels,
where m<n, voice switches are used to determine when there are vocal
sounds on any of the n input channels. Only those channels carrying vocal
sounds at any instant are connected to an output channel. Clearly, the
acceptable performance of the communication system depends upon the
ability of the voice switches to recognize speech in the presence of noise
and to establish and maintain a communications link between the input and
output channels. A failure to detect speech signals may result in
excessively long clipping of speech utterances and cause user
dissatisfaction. Another important function of voice switches is to
prevent noise signals from activating the communication channel during the
silence intervals in speech so that optimum system loading may be
achieved.
Previously known voice switches use various techniques to distinguish
between noise and speech signals. The earliest and simplest prior art
voice switches employ a detector having a fixed threshold level to compare
digitally encoded samples of a signal on a channel with the fixed
threshold level. If the samples of the signal are above the threshold
level, it is assumed the signal represents voice. If the samples of the
signal are equal to or below the threshold level, it is assumed that the
signal represents noise. Typically, the voice detector detects speech by
detecting a given number of consecutive samples in excess of the threshold
value. Detection of four samples in sucession has been considered
suitable.
Many vocal sounds result in a signal having an amplitude which tapers off
toward the end of the sound. Should the amplitude fall below the threshold
level, the described voice switch would be turned off before the
completion of the sound and result in a clipped speech pattern. To prevent
clipping of the trailing portion of transmitted sounds, the voice switch
would be constructed to operate with a hangover time. For example, when
speech is detected, the voice switch is turned on to pass the detected
samples of the channel signal. Once turned on, the voice switch will
remain on for a hangover period to insure passage of all samples of the
sound. Typically, the prior art voice switches have a hangover time of 150
milliseconds.
Clipping of the front end of the speech segment may also occur because in
certain vocal sounds the amplitude of the leading portion of the signal is
low. To avoid front end clipping, all samples of the signal are delayed a
fixed period of time, say 4 milliseconds, after the samples are received
at the input of the voice switch to permit ample time for the detection of
speech. After the delayed period, the samples are applied to the output of
the voice switch which actually controls the passage of speech samples and
the blockage of noise and other non-speech samples. Consequently, the
voice switch would detect speech prior to the time the leading portion of
the speech signal arrives at the output. Thus, clipping of the front end
of the speech signal is minimized.
The described prior art threshold voice switches have many disadvantages.
For example, because the amplitude of speech signals varies from speaker
to speaker, the prior art voice switches cannot accurately distinguish the
speech of low level talkers from channel noise. Moreover, the prior art
switches may clip speech if the amplitude of the low level speech signals
falls below the fixed threshold. The value of the threshold usually is set
at a level which is a compromise between a high level, yielding minimum
noise triggering, and a low level, yielding maximum speech detection.
Another disadvantage exists because noise on a typical communication
channel also varies over a considerable range and a high noise level could
trigger the voice switch during the silence intervals in speech. The
transmission of noise will use available channel capacity and increase
system loading.
To overcome the shortcomings of the fixed threshold systems, voice switches
having a variable threshold level have been introduced which adjust the
threshold level to the correct level that yields maximum noise immunity
and maximum sensitivity to speech. One such system is disclosed in U.S.
Pat. No. 3,832,491 filed Aug. 27, 1974, issued to Joseph A. Sciulli et al.
and assigned to the assignee of the present application. The invention
discloses a voice switch having a digital adaptive threshold generating
device. The threshold level is varied in accordance with the loudness of
the talker by comparing the number of times the threshold is exceeded over
a given period with a reference number. Maximum and minimum threshold
levels are also provided to prevent the threshold level from rising too
high when there is continuous talking by a loud talker and from falling
too low when there is continuous silence.
Another type of prior art voice switches having a variable threshold is
taught in the U.S. Patent application Ser. No. 606,828, filed Aug. 21,
1975, filed by Raymond H. Lanier and assigned to the assignee of the
present invention. In the application of Lanier the threshold is shifted
in response to changes in the noise level itself. This invention is based
upon the recognition that over a given interval of time "T" speech will
appear as random talk spurts separated by periods of silence, while noise
(generally Gaussian distributed) will be continuous. This difference
between speech and noise makes it possible to detect the noise level with
respect to the voice switch threshold. To detect noise, a time interval T
is divided in equal subintervals .tau.. The number of samples that exceed
the threshold in each subinterval is then counted. If the values of
samples tend to be non-uniform over the interval T, then it is assumed
that active speech is present. If, on the other hand, the values of
samples tend to be uniform over the time interval T, then it is assumed
that noise is present. In the latter case, when the number of samples
accumulated during .tau. is large, the threshold level would be raised,
whereas when the number of samples accumulated is small, the threshold
level would be lowered. To maintain the threshold level just above the
noise level, a threshold zone is provided wherein the zone is varied to
cause the peak of the noise level to be above a minimum level of the zone
but below a maximum level of the zone.
In the prior art variable threshold voice switches described above, the
adjustment time initially required to increase or decrease the threshold
level, and subsequently to vary the threshold level in response to a
change in noise level, is relatively slow. The delay in system response
resulting from these adjustments results in unsatisfactory switch
performance. Another problem with the described systems is that the voice
threshold level, when adjusted to uniform noise samples, is positioned too
close to the noise level. Consequently, high noise pulses which are
present in normal telephone line noise, quite often exceed the voice
threshold level and cause false triggering of the voice switch.
SUMMARY OF THE INVENTION
The present invention relates to a variable threshold digital voice switch
which detects speech signals in the presence of noise in communications
channels. The present invention is designed to overcome the disadvantages
of previously known voice switches by providing:
a greater immunity to false detection of noise;
a faster threshold adjustment in response to varying noise levels;
a simplification in design; and
a minimization of speech clipping.
The voice switch of the present invention employs three threshold detectors
and a threshold adjustment circuitry. In particular, the voice switch
provides a speech threshold detector having a high speech threshold level
T.sub.H to detect the presence of speech, a noise threshold detector
having a low noise threshold level T.sub.L to detect the presence of
noise, a threshold adjustment circuitry operating in conjunction with the
noise threshold detector to detect the noise level and to position T.sub.H
and T.sub.L according to the noise level, and a disabling threshold
detector having the maximum threshold level T.sub.M to disable the
threshold adjustment circuitry when speech is present. The threshold
levels of T.sub.H and T.sub.L are variable while the threshold level of
T.sub.M is fixed. The threshold adjustment circuitry operates at a high
speed and is capable of performing rapid adjustment of T.sub.H and T.sub.L
in response to varying noise levels.
The voice switch of the present invention is designed to operate in a
digital communications system which transmits voice signals in digital
form. The voice signals are first sampled and encoded into digital form
before they are applied to the input of the voice switch. The input
samples are applied to a delay device which delays the application of the
samples to the output of the voice switch for a fixed period of time. This
delay provides a buffer against clipping of the front end of the speech
burst and allows ample time for detection of speech.
The speech threshold detector having T.sub.H as the speech threshold level
is provided to detect the presence of speech and operates as follows. The
input samples, which are applied to the delay device, are also applied to
the input of the speech detector and the magnitude of the samples is
compared with the speech threshold level T.sub.H. When three consecutive
samples are detected to be greater in magnitude than T.sub.H, speech is
determined to be present. The three consecutive sample period, instead of
the conventional four consecutive period, is utilized as the basic
decision interval for detecting speech signals because experimentation has
revealed that on any given speech waveform the speech threshold level for
three consecutive sample detection would be positioned further above the
noise level than the level for four consecutive sample detection without
sacrificing any speech detection capability. This means that the present
invention having a higher threshold level T.sub.H than the conventional
systems would yield greater noise immunity. Upon detecting speech, the
speech detector applies an output signal to the output of the voice switch
and causes it to be turned on. When the voice switch is turned on, it will
permit the passage of the speech samples which are delayed by the delay
device. Once the voice switch is in the "on" state, it will remain on for
a hangover period, which is set at a fixed period of time, approximately
170 milliseconds, to minimize clipping of the trailing portion of the
speech burst. The hangover period is set only after the detection of the
last three consecutive speech samples in a speech burst. Of course, for a
long speech burst, the voice switch will remain on without interruption
for so long as consecutive speech samples are detected in the speech
detector.
The noise threshold detector having T.sub.L as the noise threshold level is
provided to detect the presence of noise. The input samples, which are
applied to the delay device and the input of the speech threshold
detector, are also applied to the input of the noise detector. The
magnitude of the samples is compared with the noise threshold level
T.sub.L. Each time the magnitude of a sample exceeds T.sub.L, the noise
detector produces an output signal representing the presence of noise. The
threshold adjustment circuitry operates in conjunction with the noise
detector to detect the noise level and to simultaneously adjust the speech
and noise threshold levels according to the noise level. To accomplish the
threshold adjustment, the output signals from the noise detector are
accumulated over a given interval of time i. During the period of time i,
the number of signals (Ni) is accumulated. If the accumulation Ni is
greater than a first predetermined percentage x of the total number of
samples, which indicates that T.sub.L is below the noise level, both
T.sub.H and T.sub.L are increased by a fixed increment. T.sub.H is
separated by a fixed distance .DELTA. above T.sub.L. If the accumulation
Ni is less than a second predetermined percentage y of the samples, which
indicates that T.sub.L is above the noise level, T.sub.H and T.sub.L are
decreased by the same increment. In this manner the threshold levels
T.sub.H and T.sub.L are adjusted until Ni is within a desired range which
is between x% and y% of the total number of samples during the sampling
period of i. For example, a range between 3.3% and 5% is found to be
suitable. At this range, T.sub.L is positioned near the noise level and
T.sub.H is positioned just slightly above the noise level. At this
position, the speech threshold level T.sub.H is far enough above the noise
level to screen out most of the noise signals, yet low enough to detect
low-level speech signals.
Since the noise level changes from time to time, the positions of T.sub.H
and T.sub.L are constantly adjusted according to the changes in the noise
level. Because the input samples are continuously applied to the input of
the noise detector, the level of noise is periodically measured by
accumulating over time i, the number of signals (Ni) which exceed the
noise threshold level T.sub.L. The positions of T.sub.H and T.sub.L are
then adjusted accordingly until Ni is within the desired range. At this
range, T.sub.L and T.sub.H are again properly adjusted with respect to the
new noise level.
The adjustment time required by the voice switch of the present invention
for the initial adjustment when an idle channel becomes active or for the
threshold levels to react to a change in noise is only dependent upon the
time needed to detect the noise level and the time required to adjust
T.sub.L and T.sub.H until T.sub.L is positioned near the noise level.
Compared with the prior art variable threshold noise detectors, the
adjustment circuitry of the present invention operates at a much faster
rate and thus provides a better switching performance than the previously
known detectors.
It is known that in a typical communications channel the noise appears
punctuated by spurts of speech. During active speech, the speech samples
that are applied to the input of the noise detector will greatly increase
Ni and will cause the thresholds to be misadjusted to high levels. To
overcome the incorrect adjustments during the presence of speech, the
disabling threshold detector having T.sub.M as the disabling threshold
level is employed to disable the threshold adjustments of the T.sub.H and
T.sub.L while speech is present. T.sub.M is fixed at a level which is high
enough so that it will not be exceeded by typical noise level and yet is
low enough so that it will be easily exceeded at least once during a
speech burst. When T.sub.M is exceeded and the hangover is placed in an ON
state due to detection by the speech threshold that three consecutive
samples have exceeded T.sub.H, all threshold level adjustments are
disabled and will remain disabled for the entire duration of the hangover
period.
BRIEF DESCRIPTION OF THE DRAWINGS
The specific nature of the invention, as well as other objects, aspects,
uses, and advantages thereof, will clearly appear from the following
description and from the accompanying drawing, in which:
FIG. 1 is a graphical representation showing the positions of the speech
threshold level T.sub.H, the noise threshold level T.sub.L and the
disabling threshold level T.sub.M with respect to the noise and speech
levels.
FIG. 2 is a block diagram of the preferred embodiment of the present
invention.
DESCRIPTION OF THE PREFERRED EMBODIMENT
The effectiveness of a voice switch is dependent upon the placement of a
speech threshold level with respect to the speech and noise levels.
Ideally, the speech threshold level should be positioned just above the
noise level to maximize sensitivity to speech signals and remain immune to
false triggering caused by high level noise signals. Since noise on a
typical communication channel varies over a considerable range of levels,
it also is critical to adjust the speech threshold level according to
changes in noise level.
The voice switch of the present invention utilizes a speech detector having
a variable speech threshold level T.sub.H to detect the presence of
speech, a noise detector having a variable noise threshold level T.sub.L
to detect the presence of noise, a threshold adjustment circuitry
operating in conjunction with the noise detector to measure the noise
level and to adjust the threshold levels T.sub.H and T.sub.L and a
disabling detector having a fixed disabling threshold level T.sub.M to
disable the adjustment circuitry when speech is present. An illustration
of the positions of the speech threshold level T.sub.H, the noise
threshold level T.sub.L and the disabling threshold level T.sub.M with
respect to the speech and noise levels is shown in FIG. 1. To position the
level T.sub.H just above the noise level, it is necessary to periodically
measure the noise level and correspondingly adjust T.sub.H. As illustrated
in FIG. 1, the speech threshold level T.sub.H is maintained at a fixed
distance .DELTA. above the noise threshold level T.sub.L, where T.sub.H =
T.sub. L + .DELTA.. (A preferred value for .DELTA. for a particular code
is given below; for example, for the code contemplated in the example
described herein, a delta value corresponding to seven binary steps may be
utilized.) To measure the noise level, the noise detector and the
threshold adjustment circuitry are employed, wherein the number of samples
Ni, which exceed the variable noise threshold level T.sub.L, is
accumulated over a given interval of time i. A time interval of 150
milliseconds is determined to be sufficient. If Ni is greater than say 5%
of the total number of samples in the time interval, both T.sub.L and
T.sub.H are increased by a step increment so that the number of samples
above T.sub.L will be reduced. If Ni is less than say 3.3% of the samples,
the levels of T.sub.L and T.sub.H are similarly reduced thus causing an
increase in the number of noise samples above T.sub.L. The threshold
levels are adjusted until Ni falls within the range between 3.3% and 5% of
the total number of samples or is approximately equal to 4% of the
samples. When Ni is approximately equal to 4% of the total number of the
samples, the speech threshold T.sub.H is thus properly adjusted to the
optimum position which is slightly above the noise level and yet low
enough to detect low level speech signals.
The disabling threshold T.sub.M is also employed in the present invention
to disable the threshold adjustment circuitry while speech is present. As
shown in FIG. 1, T.sub.M is set to a fixed level, say -23dBmO, which is
considerably above a typical line noise level and yet low enough to be
exceeded at least once during a speech burst.
The preferred embodiment of the digital voice switch which accomplishes the
foregoing results is illustrated in FIG. 2. As is conventional in a
digital communications channel which transmits voice information in
digital format, the analog voice information is applied to a conventional
encoder wherein the analog signals are sampled, typically, at an 8-KHz
rate, and subsequently encoded into an 8-bit digital sample. As well known
in the art, the 8-bit samples comprising 7 amplitude bits and 1 sign bit
are applied to the input of the digital voice switch. The 8-bit samples,
indicated as SIGN, B.sub.1, B.sub.2, . . . , B.sub.7, are applied in
parallel by the input lines shown generally at 1. The switching portion of
the digital voice switch comprises 8 parallel front end delay units, shown
generally at 3, which consist of serial shift registers clocked at the
sampling frequency of 8kHz.
The shift-registers of the front end delay 3 have a sufficient number of
stages to provide a 4 millisecond delay to allow ample time for speech
detection which will be explained below and thus provide a buffer against
clipping of the leading portion of speech signals. The outputs of the
delay units 3 are fed directly to output AND gates shown generally at 5.
The output AND gates are turned on to pass voice samples when speech
signals are present in the communication channel. The output gates are
turned off to block the passage of non-voice or noise samples when
non-voice signals are present in the channel.
The magnitude bits, B.sub.1, B.sub.2, . . . , B.sub.7, of the input samples
of lines 1 also are applied to a speech threshold detector 7. A digital
representation, TH1 - TH7, of the threshold level, also is applied to the
detector 7 by lines 6. Lines 6 are connected to and fed back from a
portion of the threshold adjustment circuit which will be explained below.
Since the threshold level will always be positive, it is not necessary to
provide a sign bit for the digital threshold value. The speech threshold
detector may consist of a conventional comparator constructed in a well
known manner as an operational amplifier. The comparator digitally
compares the magnitude of the sample represented by the signals in lines 1
with the magnitude of the speech threshold level represented by the
signals in lines 6 (TH1 - TH7). The comparator in the speech detector
generates a binary 1 output if the magnitude of the sample exceeds the
threshold level and a binary 0 output if the magnitude of the sample is
equal to or less than the threshold level. The binary outputs from the
threshold detector 7 are clocked by an 8 kHz clock into a 3-bit shift
serial register 9. When the shift register 9 is completely filled with
three binary 1 bits indicating that three consecutive samples exceed the
threshold level, the outputs of the shift register will be all binary 1
and will energize AND gate 11. Thereupon, the AND gate 11 applies a binary
1 output to the triggering input of a one-shot multivibrator 13. If the
shift register 9 is not filled with all binary 1 bits, the AND gate 11
will not be energized indicating that speech is not present or is no
longer present in the communication channel.
The one-shot 13 is a conventional retriggerable device having a fixed time
pulse width which provides a hangover time. The hangover time may be set
at a time period typically between 150 and 180 milliseconds. Thus, the
output of the one-shot 13 will rise to its active level upon triggering
and will drop to its non-active level say 170 milliseconds after the last
received trigger. The active output of the one-shot device 13 energizes
the output AND gates 5 to pass the delayed speech samples to the output
terminal.
If the AND gate 11 is not energized because the speech detector fails to
detect three consecutive samples exceeding the threshold level, the
one-shot 13 will not be triggered to its active level and the output AND
gates 5 will not be turned on. Consequently, the AND gates 5 will block
the passage of the delayed non-voice samples.
If a long and high amplitude speech burst is present in the communication
channel, all of the samples of the speech signal probably will exceed the
speech threshold level and only consecutive binary 1 outputs will be
generated by the speech detector 7. Thus, the shift register 9 will be
continuously filled with binary 1 bits and the one-shot 13 will be in the
active state for as long as speech is detected to be present in the
channel. The output AND gates 5 will be turned on to pass the entire
speech burst without any interruption and will remain on for the period of
the hangover time after the detection of last three consecutive speech
samples.
Except for the introduction of the 3-bit shift register 9 in place of the
conventional 4-bit shift register, the voice switch described thus far is
conventional. The major improvement provided by the subject invention is
in the apparatus for adjusting the speech threshold level according to the
changes in the noise level and in the device for disabling the threshold
adjustment circuitry when speech is present.
To adjust the level of the speech threshold detector 7 according to the
noise level in the input channels, the subject invention employs a noise
threshold detector 15 and a threshold adjustment circuitry 16. As shown in
FIG. 2, the magnitude bits, B.sub.1, B.sub.2, . . . , B.sub.7, of the
input samples in lines 1 are simultaneously fed to the noise threshold
detector 15 as well as to the speech threshold detector 7. The noise
threshold detector 15 may consist of a conventional comparator constructed
in a well known manner as an operational amplifier. The comparator
compares the magnitude of the input samples in lines 1 with a noise
threshold level indicated as TL1 - TL7 in lines 14, which are connected to
and fed back from a portion of the threshold adjustment circuitry 16 which
will be explained below. The comparator provides a binary 1 at its output
if the input sample exceeds the threshold level and a binary 0 if the
input sample is equal to or less than the threshold level. The threshold
adjustment circuitry 16 is comprised of an accumulator 17, comparators 19
and 21, a counter 25 and an adder 27. The outputs from the noise threshold
detector 15 are applied to the input terminal of the accumulator 17, which
may be a conventional counter or shift register. The accumulator 17 counts
the number of binary 1 outputs received from the noise detector 15 during
a given period of time, say 150 milliseconds. The accumulator is reset to
zero every 150 milliseconds by a 6.67 Hz clock signal. The output of the
accumulator 17 is applied to the inputs of two comparators 19 and 21.
Comparators 19 and 21 are conventional devices which compare the state of
the accumulator 17 with preset numbers. In the specific example described,
comparator 19 compares the accumulated number with a fixed number, 60,
which represents 5% of the total number of samples in the 150-millisecond
interval. If the accumulated number is greater than 60, the comparator
output provides a binary 1 to one of two inputs of an AND gate 23. The
other input to the AND gate 23 is connected to a latch 33 which performs
the disabling function and will be explained below. When both inputs of
the AND gate 23 receive binary 1 inputs, gate 23 is enabled and passes a
binary 1 output to the count-up input of the up-down counter 25.
Similarly, comparator 21 compares the accumulated number with a fixed
number 40, which represents 3.3% of the total samples in the
150-millisecond interval. If the accumulated number is less than 40,
comparator 21 provides a binary 1 output to one of two inputs of an AND
gate 24. The other input to the AND gate 24 is connected to the latch 33
which will be explained below. When both inputs of the gate 24 receive
binary 1 inputs, gate 24 is enabled and passes a binary 1 output to the
count down input of the up-down counter 25. If neither of the two
conditions is met or when the accumulation is .gtoreq. than 40 and
.ltoreq. than 60, then gates 23 and 24 will not be enabled. When the
latter condition occurs, it represents that the noise threshold level as
indicated by signals in lines 14 is properly positioned with respect to
the noise level and no adjustment is needed.
It will be appreciated from the foregoing that the count in the accumulator
17 is the number Ni of the samples which exceed the noise threshold level,
as indicated by signals in lines 14, in the time interval i. Although the
time interval i may be any desirable period of time, a time interval i of
150 milliseconds is used as an example in explaining the preferred
embodiment of the present invention. Comparators 19 and 21 determine
whether the accumulation Ni is in one of the three following ranges:
1st range: Ni > 60
2nd range: Ni < 40
3rd range: 40 .ltoreq. N.ltoreq. 60
The first two ranges indicate that the noise threshold level is positioned
either too low or too high, respectively, whereas the third range
indicates that the threshold level is properly positioned with respect to
the noise level.
After determining the relative position of the noise threshold level,
appropriate adjustment to the noise threshold level in the noise detector
15 and speech threshold level in the speech detector 7 is carried out. If
the count up or count down input of the up-down counter 25 is active
during the 6.67 Hz clock pulse, which indicates that the accumulation Ni
is greater than 60 or less than 40, then the value of the noise threshold
level, TL1 - TL7, applied at the input of the counter 25 is increased or
decreased, respectively, by one quantization step in the binary form. The
output of the up-down counter 25, which now contains the adjusted value,
TL'1 - TL'7, of the noise threshold level, is applied to the input of the
counter 25, to the input of the noise threshold detector 15 and to the
input of an adder 27 by lines 14. As mentioned in the foregoing, the
speech threshold level of the detector 7 is maintained at a fixed distance
.DELTA. above the noise threshold level and is adjusted simultaneously
with the noise threshold level, the adder 27 is employed to carry out the
aforementioned adjustment function. When the adjusted value of noise
level, TL'1 - TL'7, is applied to the adder 27 | | |
