 |
Description  |
|
|
CROSS-REFERENCE TO RELATED AND APPLICATION
The subject matter of this application is related to the currently pending
U.S. patent application of J. B. Allen and D. J. Youtkus entitled
"Background Noise Compensation in a Telephone Network, Ser. No.
08/175,075, filed on even date herewith and assigned to the assignee of
the present invention.
FIELD OF THE INVENTION
The present invention relates generally to the field of telephone sets
connected to a telephone network and specifically to the problem of using
a telephone in a noisy environment.
BACKGROUND OF THE INVENTION
When a person uses a telephone in a noisy environment such as a noisy room,
an airport, a car, a street comer or a restaurant, it can often be
difficult to hear the person speaking at the other end (i.e., the
"far-end") of the connection over the background noise present at the
listener's location (i.e., the "near-end"). In some cases, due to the
variability of human speech, the far-end speaker's voice is sometimes
intelligible over the near-end background noise and sometimes
unintelligible. Moreover, the noise level at the near-end may itself vary
over time, making the far-end speaker's voice level at times adequate and
at times inadequate.
Although some telephones provide for control of the volume level of the
telephone loudspeaker (i.e., the earpiece), such control is often
unavailable. Moreover, manual adjustment of a volume control by the
listener is undesirable since, as the background noise level changes, the
user will want to readjust the manual volume control in an attempt to
maintain a preferred listening level. Generally, it is likely to be
considered more desirable to provide an automatic (i.e., adaptive) control
mechanism, rather than requiring the listener first to determine the
existence of the problem and then to take action by adjusting a manual
volume control. One solution which attempts to address this problem has
been proposed in U.S. Pat. No. 4,829,565, issued on May 9, 1989 to Robert
M. Goldberg, which discloses a telephone with an automated volume control
whose gain is a function of the level of the background noise.
SUMMARY OF THE INVENTION
We have recognized that the use of either conventional manual volume
controls or an automatic mechanism such as that disclosed in the
above-cited U.S. Pat. No. 4,829,565 fails to adequately solve the
background noise problem. In particular, these approaches fail to
recognize the fact that by amplifying the signal which supplies the
handset receiver (i.e., the loudspeaker), the side tone is also amplified.
(The side tone is a well-known feed-through effect in a telephone. A
portion of the input signal from the handset transmitter--i.e., the
microphone--is mixed with the far-end speech signal received from the
network. The resultant, combined signal is then supplied to the handset
loudspeaker.) Since the side tone contains the background noise, itself,
the background noise is, disadvantageously, amplified concurrently with
the far-end speech signal whenever such a volume control (either manual or
automatic) is used to amplify the signal which supplies the handset
receiver. By amplifying both the speech signal and the noise together, the
degrading effect of the noise can actually become worse because of the
properties of the human ear.
In accordance with the present invention, a modified speech signal is
produced from an original speech signal in a telephone set before the side
tone has been combined therewith. Specifically, a gain factor is applied
to the original speech signal to produce the modified speech signal. The
gain factor is a function of a received signal indicative of the
background noise at the given destination at which the telephone set is
located. The side tone may then be combined with the modified speech
signal.
The gain factor may be a function of the level of the background noise, or
it may be a function of both the level of the background noise and the
level of the original (i.e., the tar-end) speech signal. The modified
speech signal may comprise a linear amplification of the original speech
signal or it may comprise an amplified and "compressed" version of the
original speech signal. By "compressed" it is meant that the higher level
portions of the original signal are amplified by a smaller gain factor
than are the lower level portions.
In accordance with one illustrative embodiment, the original speech signal
may be separated into a plurality of subbands, and each resultant subband
signal may be individually modified (e.g., amplified) in accordance with
the technique of the present invention. In particular, these original
subband speech signals may be amplified by a gain factor which is a
function of a corresponding subband-noise-indicative signal. Such
subband-noise-indicative signals may be generated by separating the signal
indicative of the background noise into a corresponding plurality of
subbands. The individual modified subband signals may then be combined to
form the resultant modified speech signal. The modified speech signal may
then be combined with the side tone.
As used herein, the term "telephone set" is intended to include any
apparatus for use by one party in providing a two-way speech communication
linkage between the user and another party, wherein the apparatus has the
effect (intentionally or unintentionally) of combining a side tone with
the received signal. Examples of such devices include conventional
desk-top or other corded telephone sets, cordless and cellular telephones,
and headsets (such as those commonly used by pilots, telephone operators,
air traffic controllers, police dispatchers, etc.). The telephone set may
be connected to the other party by any conventional (or non-conventional)
telephone network means. As used herein, the term "telephone network" is
intended to include conventional terrestrial telephone networks (local or
long distance), wireless (including cellular) communication networks,
radio transmission, satellite transmission, microwave transmission, fiber
optic links, etc., or any combination of any of these transmission
networks.
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1 shows a telephone set which includes a noise compensation system in
accordance with an illustrative embodiment of the present invention.
FIG. 2 shows a system-level diagram of a broadband-based illustrative
embodiment of a noise compensation system in accordance with the present
invention.
FIG. 3 shows an illustrative implementation of the noise level estimation
unit of the system of FIG. 2.
FIG. 4 shows an illustrative implementation of the gain computation unit of
the system of FIG. 2.
FIG. 5 is a graph which shows a compressor gain which may be applied to the
original speech signal by the signal boost unit of the system of FIG. 2
applying compressed amplification.
FIG. 6 is a graph of the corresponding transfer function for the
illustrative signal boost unit which results from applying the gain shown
in FIG. 5.
FIG. 7 shows an illustrative implementation of the signal boost unit of an
embodiment of the system of FIG. 2 applying a compressed amplification as
shown in the graphs of FIGS. 5 and 6.
FIG. 8 shows an alternative illustrative implementation of the gain
computation unit of FIG. 2 for use in an embodiment applying compressed
amplification in an alternative manner.
FIG. 9 shows a system-level diagram of a multiband-based illustrative
embodiment of the present invention in which noise compensation is
performed in individual subbands.
DETAILED DESCRIPTION
Introduction
The present invention improves the signal-to-noise ratio (SNR) of a far-end
speaker's speech in the near-end listener's ear when the near-end listener
is using a telephone in a noisy environment. The level of the noise in the
ear of the near-end listener can be estimated from the signal levels
picked up by the transmitter (microphone) in the near-end listener's
handset. Based on these levels, the original speech signal generated by
the far-end speaker may be modified within the telephone by being
amplified by a variable gain factor so as to provide a more intelligible
signal to the listener. This modification may advantageously also be a
function of the level of the original speech signal itself. For example,
the speech power level (i.e., a "long-term" average level of the original
speech signal) may be incorporated into the determination of the gain
factor. In this manner, relatively quiet signals may be boosted (i.e.,
amplified) by a larger gain factor than relatively loud signals.
Moreover, the modification of the speech signal may comprise either a
linear amplification or a non-linear, (illustratively) compressed,
amplification. Compressed amplification, in particular, boosts loud
portions of the original speech signal by a lesser amount (i.e., with a
smaller gain factor) than quiet portions. Thus, it is possible in this
manner to, on a short-term basis, boost the signals which fall below the
background noise level without boosting the signals which are already
significantly above the background noise level. Simple linear
amplification, by contrast, boosts all signal levels by an equal amount.
When used to boost low-level signals above the background noise, linear
amplification can in some circumstances result in distortion, since the
higher level signals (already above the noise) could receive excessive
amplification.
FIG. 1 shows telephone set 20 which includes a noise compensation system
embodying the principles of the present invention. Specifically, telephone
set 20 of FIG. 1 comprises a deskset 18 and a handset 13h (having
microphone 13m and loudspeaker 13s). Included within deskset 18 is noise
compensation system 14. For illustrative convenience, the cord which
connects handset 13h to deskset 18 is shown as emanating from the right
side of deskset 18, even though such cords most commonly emanate from the
left side of a deskset.
Deskset 18 further includes ideal hybrid 19 and side tone adder 21. Ideal
hybrid 19 converts between standard two-wire and four-wire telephone
lines. It is ideal in that it substantially isolates the incoming signal
from the outgoing signal on its four-wire side. In this manner, there is
essentially no side tone component to the original speech signal which
emanates from the four-wire side of the hybrid. Side tone adder 21
combines the signal from microphone 13m with the incoming (far-end) speech
signal before it is provided to loudspeaker 13s. In particular, a reduced
level of the signal from microphone 13m is mixed with the speech signal.
Thus, a side tone is provided to the loudspeaker. Ideal hybrids and side
tone adders are conventional components which may be found, for example,
in certain electronic telephone sets. Other conventional components of a
telephone set which are not relevant to the present invention are not
shown.
Noise compensation system 14 receives a noise-indicative signal from
microphone 13m which is representative of the background noise (as
illustrated as emanating from loudspeaker 17) as well as any speech
provided to microphone 13m by the telephone user. Noise compensation
system 14 also receives the original speech signal from a far-end speaker
(whose telephone and the network connection thereto are not shown). Noise
compensation system 14 determines the level of background noise from the
noise-indicative signal, and boosts the original speech signal by a gain
factor based on the background noise level to produce a modified speech
signal. The side tone is then added to the modified speech signal by side
tone adder 21, and the resultant signal is then provided to loudspeaker
13s in handset 13h. Note that the original speech signal as provided to
noise compensation system 14 is substantially free of the background
noise, since the side tone has not yet been added to the original
(far-end) speech signal. Thus, the far-end speech is boosted without any
simultaneous boost of the background noise in the side tone.
An Illustrative Broadband Implementation with Linear Amplification
FIG. 2 shows a system-level diagram of a broadband-based illustrative
embodiment of noise compensation system 14. Inputs to the system include
the original speech signal and the noise-indicative signal, which may
further include speech provided by the near-end listener. The system
produces a modified speech signal for improved intelligibility as output.
All of the signals described with reference to the illustrative embodiment
present herein are presumed to be in digital form.
Based on the noise-indicative signal, noise level estimation 22 determines
the "noise floor" and outputs a signal representing that value. In
particular, this signal represents the noise level over a first
predetermined period of time. By setting this first predetermined period
to a relatively short value (e.g., 250 milliseconds or less), the
determined noise floor will substantially follow changing levels of
background noise in the near-end environment. Specifically, the noise
floor signal represents a short-term (e.g., 250 milliseconds) minimum
value of an "exponentionally mapped past average" signal, and can be
generated using known techniques. An illustrative implementation of noise
level estimation 22 is shown in FIG. 3 and described below.
Gain computation 24 produces a gain signal, GAIN, whose value is
proportional to the noise floor signal and inversely proportional to an
average speech power level signal. This gain signal represents a gain
factor (i.e., a multiplicative factor) by which the original speech signal
may be amplified. The average speech power level signal is generated by
speech power estimation 23, and represents the average level of the
original speech signal over a second predetermined period of time. That
is, the average speech power level measures the "energy" level of the
speech signal. Providing such a gain dependence on the far-end speech
level allows relatively quiet calls to receive a sufficient boost for a
given background noise level, while preventing loud calls from being
over-boosted. By setting the second predetermined period to a relatively
long value (e.g., one second), it can more readily be determined whether
the current far-end speech comprises a loud or soft segment of the call.
Thus, the average speech power level signal represents a long-term average
level. Speech power estimation 23 may be implemented by conventional
signal energy estimation techniques. An illustrative implementation of
gain computation 24 is shown in FIG. 4 and described below.
The gain signal and the original speech signal are provided to signal boost
25 which produces the modified speech signal. Where only linear
amplification is desired, signal boost 25 may comprise a conventional
amplifier (i.e., a multiplier). In this case, the original speech signal
is amplified by a gain factor equal to the value of the gain signal, GAIN.
Where, on the other hand, compressed amplification is desired, signal
boost 25 may comprise circuitry (or procedural code) which amplifies the
original speech signal by a gain factor less than or equal to the value of
the gain signal, wherein the gain factor further depends on the level of
the original speech signal itself. That is, the gain signal, GAIN,
represents the maximum gain which will be applied by the "compressor." An
illustrative implementation of signal boost 25 providing compression is
shown in FIG. 7 and described below.
FIG. 3 shows an illustrative implementation of noise level estimation 22 of
the system of FIG. 2. First, high pass filter (HPF) 31 removes DC from the
input signal. It may be conventionally implemented as a first order
recursive digital filter having a cutoff frequency of, for example, 20 Hz,
and may be based on a standard telephony sampling frequency of 8 kHz.
Absolute value block (ABS) 32 computes the magnitude of the sample and is
also of conventional design. Low pass filter (LPF) 33 computes the
exponentially mapped past average (EMP). As described above, the
exponentially mapped past average comprises an exponentially weighted
average value of the noise level. Low pass filter 33 is also of
conventional design and may illustratively be implemented as a first order
recursive digital filter having the transfer function y(n)=(1-.beta.)
x(n)+.beta.y(n-1), where .beta.=e.sup.-T/.tau., with T a sampling period
and .tau. a time constant. Illustratively, T=0.125 ms and .tau.=16 ms.
Minimum sample latch (MIN) 34 stores the minimum value of EMP over the
first predetermined time period (e.g., 250 milliseconds). The output
signal of latch 34, MEMP, therefore represents the short-term minimum of
the exponentionally mapped past average, and thus represents the
short-term minimum value of the averaged noise-indicative signal. This
signal is subsequently used to represent the noise floor over which
far-end speech should be boosted. In a corresponding manner, maximum
sample latch (MAX) 35 stores the maximum value of EMP over the same
predetermined period. The output signal of latch 35, PEMP, therefore
represents the short-term peak of the exponentionally mapped past average,
and thus represents the short-term peak value of the averaged
noise-indicative signal. Latches 34 and 35 may be implemented by
conventional digital comparators, selectors and storage devices, with the
storage devices reset at the start of each cycle of the predetermined time
period.
Speech detector and noise floor estimator 36 generates the noise floor
signal output based on signals MEMP and PEMP. Specifically, it performs
two functions. First, it is determined whether the noise-indicative signal
presently includes only noise or whether it presently includes speech as
well. This question may be resolved by conventional techniques, such as
those used in the implementation of conventional speakerphones. For
example, the quotient of PEMP (representing the short-term peak value of
the noise-indicative signal) divided by MEMP (representing the short-term
minimum value of the noise-indicative signal) may be compared with a
predetermined threshold. The larger this quotient, the larger the
variability in the level of the input signal. If the level of the input
signal is sufficiently variable within the first predetermined time
period, it is presumed that speech is present. (Note that the variation in
signal level of speech typically exceeds that of background noise.)
Second, speech detector and noise floor estimator 36 sets the output noise
floor signal to a value which represents the estimated level of the noise
floor. If it is determined that speech is not present, the noise floor
signal is set to MEMP, the short-term minimum value of the
noise-indicative signal. Otherwise, the noise floor signal remains
unchanged--that is, the previous value is maintained. In this manner, when
the presence of speech makes it difficult to determine the actual present
level of background noise, it is presumed that the noise level has not
changed since the previous period.
In one alternative embodiment, the value of PEMP alone may be compared with
a predetermined threshold (rather than using the quotient of PEMP divided
by MEMP), since speech is generally of a significantly higher intensity
than is background noise. And in a second alternative embodiment, speech
detection may be bypassed altogether, on the assumption that the far-end
speaker will not be speaking at the same time that the near-end listener
is speaking. In other words, we may not care what the "noise floor" is
determined to be during periods when the near-end listener is speaking. In
this second alternative embodiment, maximum sample latch 35 and speech
detector and noise floor estimator 36 may be removed from noise level
estimation 22 of FIG. 3, and the output of minimum sample latch 34 (i.e,
signal MEMP) may be used directly as the noise floor signal output of
noise level estimation 22.
FIG. 4 shows an illustrative implementation of gain computation 24 of the
system of FIG. 2. The gain signal is generated based on the noise floor
signal from noise level estimation 22 and on the average speech power
level signal from speech power estimation 23. Specifically, the computed
gain is advantageously proportional to the noise floor and inversely
proportional to the average speech power level. Moreover, the gain is
never less than one (i.e., the original speech signal is never attenuated)
nor is it ever more than a maximum specified value.
First, amplifier 41 multiplies the noise floor by a noise scale factor.
This noise scale factor is set to an appropriate value so that the output
signal of amplifier 41, which is representative of a gain factor, is of
the appropriate magnitude. In particular, the noise scale factor acts as a
"sensitivity" control--a smaller scale factor will result in more gain
being applied for a given level of background noise. The magnitude of this
signal may be advantageously set to that gain factor which will boost the
lowest far-end speech levels by an appropriate amount to overcome the
noise level. For example, the noise scale factor may illustratively be set
to a fractional value between zero and one, such as 0.4.
Next, minimizer (MIN) 42 compares the gain factor output by amplifier 41 to
a maximum permitted gain factor to ensure that the system does not attempt
to apply an excessive gain factor to the original speech signal. For
example, the maximum permitted gain factor may illustratively be set to
5.6 (i.e., 15 dB). Maximizer (MAX) 43 then ensures that the resultant gain
factor is in no case less than one, so that the original speech signal is
never attenuated.
Divider 44 and minimizer (MIN) 45 determine an additional multiplicative
factor to be incorporated in the gain computation so that the resultant
gain will be inversely proportional to the average speech power level as
provided by speech power estimation 23. Divider 44 computes the quotient
of a minimum far-end speech level divided by the average speech power
level for use as this additional multiplicative factor. The minimum speech
level represents the minimum level which is to be considered actual
far-end speech, as distinguished from mere background noise during a
period of silence by the far-end speaker. For example, the minimum speech
level may illustratively be set to a value representing -30 dBm. Minimizer
45 then ensures that | | |
