In the field of telecommunications, such as with speaker phones and in cellular telephony, it is often desirable to allow a user to operate communication equipment without requiring the continued occupation of one or more of the user's hands. This can be an important factor in environments, such as automobiles, where a driver's preoccupation with holding telephone equipment may jeopardize not only his or her safety, but also the safety of others who share the road. Freedom to use one's hands for something other than holding a microphone is useful in other applications as well, such as with internet communication by means of a personal computer, speech recognition by a computer, or with audio-visual presentation systems.

To accommodate these important needs, so-called"hands-free" equipment has been developed, in which microphones and loudspeakers are mounted within the hands-free environment, thereby obviating the need to hold them. For example, in an automobile application, a cellular telephone's microphone might be mounted on the sun visor, while the loudspeaker may be a dash-mounted unit, or may be one that is associated with the car's stereo equipment. With components mounted in this fashion, a cellular telephone user may carry on a conversation without having to hold the cellular unit or its handset. Similarly, personal computers often have microphones and loudspeakers mounted, for example, in a monitor in relatively close proximity to each other.

One problem with a hands-free arrangement is that the microphone tends to pick up sound from the nearby loudspeaker, in addition to the voice of the user of the hands-free equipment (the so-called"near-end user"). This is also a problem in some non-hands-free devices, such as handheld mobile telephones, which are becoming

smaller and smaller. (Because of the small size, a mobile telephone's microphone cannot entirely be shielded from the sound emitted by its loudspeaker.) This sensing by the microphone of sound generated by the loudspeaker can cause problems in many types of applications. For example, in communications equipment, delays introduced by the communications system as a whole can cause the sound from the loudspeaker to be heard by the individual on the other end of the call (the so-called"far-end") as an echo of his or her own voice. Such an echo degrades audio quality and its mitigation is desirable if not an absolute necessity because the presence of a delayed echo seriously impedes the far-end user's ability to carry on a normal conversation. A similar problem can exist, for example, in automated systems that synthesize speech through a loudspeaker, and include voice recognition components for recognizing and responding to spoken commands or other words sensed by the microphone. In such applications, the presence of an echo of synthesized speech in the microphone signal can severely degrade the performance of the speech recognition components. Solutions for ameliorating such echoes include utilizing an adaptive echo cancellation filter, an echo attenuator or both in some kind of combined operation.

As a representative example of hands-free equipment in general, an exemplary"hands-free"mobile telephone, having a conventional echo canceler in the form of an adaptive filter arrangement, is depicted in FIG. 1. A hands-free communications environment may be, for example, an automotive interior in which the mobile telephone is installed. Such an environment can cause effects on an acoustic signal propagating therein, which effects are typically unknown. Henceforth, this type of environment will be referred to throughout this specification as an unknown system H (z). The microphone 105 is intended for detecting a user's voice, but may also have the undesired effect of detecting audio signals emanating from the loudspeaker 109. It is this undesired action that introduces the echo signal into the system.

Circuitry for reducing, if not eliminating, the echo includes an adaptive filter 101, such as an adaptive Finite Impulse Response (FIR) or Infinite Impulse Response (IIR) filter, an adaptation unit 103, such as a least mean square (LMS) cross correlator, and a subtractor 107. If one assumes that the only sound received by the microphone 105 is that originating from the loudspeaker 109 (of course, in practice the microphone will also receive sounds from the intended user), then a digitized echo

signal 126 (also designated"u") is generated. In operation, the adaptive filter 101 generates an echo estimate signal 102, which is commonly referred to as a u signal.

The echo estimate signal 102 is the convolution of the far-end signal 112, and a sequence of m filter weighting coefficients (hi) of the filter 101 (See Equation 1). where: x (n) is the input signal, m is the number of weighting coefficients, and n is the sample number.

When the weighting coefficients are set correctly, the adaptive filter 101 produces an impulse response that is approximately equal to the response produced by the loudspeaker 109 within the unknown system H (z). The echo estimate signal 102 generated by the adaptive filter 101 is subtracted from the incoming digitized echo signal 126, to produce an echo filtered microphone signal, also referred to as an error signal e (n) because its energy is entirely attributable to the error in the echo estimate signal 102 when the only sound reaching the microphone 105 is attributable to the sound emanating from the loudspeaker 109: e (n) =u (n)-u (n) (2) Ideally, any echo response from the unknown system H (z), introduced by the loudspeaker 109, is removed from the digitized echo signal 126 by the subtraction of the echo estimate signal 102. Consequently, when the digitized echo signal 126 also includes signal components attributable to the intended near-end user, the resulting difference signal should ideally include only these near-end user components. Typically, the number of weighting coefficients (henceforth referred to as

"coefficients") required for effectively canceling an echo will depend on the application. For handheld phones, fewer than one hundred coefficients may be adequate. For a hands-free telephone in an automobile, about 200 to 400 coefficients will be required. A large room may require a filter utilizing over 1000 coefficients in order to provide adequate echo cancellation.

It can be seen that the effectiveness of the echo canceler is directly related to how well the adaptive filter 101 is able to replicate the impulse response of the unknown system H (z). This, in turn, is directly related to the set of coefficients, hi, maintained by the filter 101.

It is advantageous to provide a mechanism for dynamically altering the coefficients, hi, to allow the adaptive filter 101 to adapt to changes in the unknown system H (z). In a car having a hands-free cellular arrangement, such changes may occur when a window or car door is opened or closed. A well-known coefficient adaptation scheme is the Least Mean Square (LMS) process, which was first introduced by Widrow and Hoff in 1960, and is frequently used because of its efficiency and robust behavior. As applied to the echo cancellation problem, the LMS process is a stochastic gradient step method which uses a rough (noisy) estimate of the gradient, g (n) = e (n)-x (n-1), to make an incremental step toward minimizing the energy (and the mean square error) of the error signal, e (n), where x (n) is in vector notation corresponding to an expressionx (n) = [x (n) x (n-1) x (n-2) x (n-m+1)]. The update information produced by the LMS process e (n) K (n) is used to determine the value of a coefficient in a next sample. The expression for calculating a next coefficient value hi (n + 1) is given by: hj (n+l) =hj (n) +e ()'jc(-/),!o..-lr) where x (n) is an nth sample of the digitized input signal 112, hi (n) is a filter weighting coefficient, i designates a particular coefficient, m is the number of coefficients,

n is the sample number, and , u is a step or update gain parameter.

The LMS method is very common in real-time applications due to its efficiency and simplicity. It produces information in incremental portions each of which portions may have a positive or a negative value. The information produced by the LMS process can be provided to the adaptive filter 101 to update the filter's coefficients.

Referring back to FIG. 1, the conventional echo cancellation circuit includes a filter adaptation unit 103 in the form of an LMS cross correlator for providing coefficient update information 104 to the filter 101. In this arrangement, the filter adaptation unit 103 monitors the error signal e (n) that represents the digitized echo signal 126 minus the echo estimate signal 102 generated by the filter 101. The echo estimate signal 102 is generated, as described above, with the use of update information 104 provided to the adaptive filter 101 by the filter adaptation unit 103.

The coefficients, hi, of the adaptive filter 101 accumulate the update information 104 as shown by Eq. (3).

Despite the presence of echo cancellation circuitry, such as that described above, the signals generated for further processing (e. g., for transmission to the far-end user in a communications system, or for near-end speech recognition or for controlling the operation of the echo cancellation filter 101) may very often still include echo components. This may occur, for example, because the adaptive filter has not yet converged to a fully adapted state, or even after such convergence whenever the unknown environment H (z) changes, thereby requiring the adaptation process to be repeated. The presence of strong echo signal components in the signal can cause degraded or even faulty operation of the down-stream processing components, since these echo signal components may be mistaken for near-end speech. For these reasons, particular applications may set requirements for echo return loss enhancement (ERLE).

In the field of cellular radio communications, the typical (linear) echo cancellation arrangement described above has problems fulfilling the ERLE requirements, which are quite high for delayed systems such as the Global System for Mobile communications (GSM) and the Digital Advanced Mobile Phone System

(DAMPS). The required ERLE figure can be as high as 40-50 dB in these cases. With good audio components it might be possible to achieve 30-35 dB ERLE by means of the linear echo canceler alone. To substantially eliminate the remainder, referred to as the residual echo, some extra processing is needed. This is called residual echo suppression.

Residual echo suppression is especially difficult because one is dealing with the non-linear components of the echo. This means that linear methods cannot be applied (at least not in a straightforward manner). Instead, non-linear methods tend to be better at this task.

A technique that has been applied to the problem of residual echo suppression is the use of a center clipper, which"clips"out a slice of the signal around the middle or zero level. An example of this technique is illustrated in FIGS. 2a, 2b and 3. In FIG. 2a, an exemplary signal 201 is plotted as a function of time prior to application of center clipping. For center clipping, respective positive and negative threshold levels 203,205 are defined. The transfer function for center clipping, showing the relationship between an input signal and a corresponding output signal, is depicted in FIG. 3. It can be seen that when the magnitude of the input signal lies within the range defined by the positive and negative threshold levels 203,205, then the output signal has a magnitude of zero. When the magnitude of the input signal is greater than the positive threshold 203, then the output signal is given by: output = (input)- (positive threshold).

Similarly, when the magnitude of the input signal is less than the negative threshold 205, then the output signal is given by: output = (input)- (negative threshold).

When this transfer function is applied to the exemplary signal 201, a clipped signal 207 results, as depicted in FIG. 2b.

Using a center clipper on speech signals has to be done with great care because the clipping process removes low level signals in a very crude manner. If the center clip threshold levels are not held as low as possible, the distortion that is introduced by the clipping process rapidly dominates the resulting signal.

U. S. Patent No. 5,475,731, which is hereby incorporated herein by reference, discloses an arrangement that uses the echo estimate signal 102 (M) to

control a center clipper. A block diagram of this arrangement is shown in FIG. 4. The echo estimate signal 102 (û) is generated as described above, and is subtracted from the digitized echo signal 126 (u) to generate an error signal e (n) that includes the residual echo components. Even though the echo estimate signal 102 does not completely cancel the echo (i. e., some residual echo energy is present in the error signal e (n)), the echo estimate signal 102 does provide an accurate indication of when and at what magnitude the echo appears. This dynamically changing information is supplied to a threshold calculator 401, which generates therefrom positive and negative thresholds 203,205. The threshold calculator 401 may be for example, an envelope, or level detector. The threshold values may be determined by: Positive and Negative Threshold Values (n) = ABS (e (n)) x K, where K is a constant that may be in the range from 0.1 to 0.5.

These generated positive and negative threshold values are then supplied to a control input of a center clipper 403 which generates a center-clipped output signal from the supplied error signal e (n).

In alternative embodiments, the center clipper 403 may be replaced by an attenuator, whose degree of attenuation is dynamically controlled by information derived from the echo estimate signal 102.

While the purpose of the center clipping (or attenuation) is to remove the echo residuals, this approach suffers from a drawback in that it introduces quite a bit of distortion into the signal. This is due to the fact that center clipping cuts out not only the residual echo, but also some wanted parts of the signal. Background noise sounds are also especially distorted after being subjected to center clipping. The use of attenuation instead of center clipping does not really improve matters, since it also degrades the wanted parts of the signal. Consequently, attenuation should also be avoided.

SUMMARY It is therefore an object of the present invention to provide methods and apparatuses that substantially reduce residual echo signals in an echo cancellation system.

It is another object of the present invention to perform residual echo cancellation without introducing distortion into a desired signal.

The foregoing and other objects are achieved in methods and apparatuses for reducing an echo in an electrical audio signal. In accordance with one aspect of the invention, reducing an echo in an electrical audio signal, includes converting a first acoustic signal into the electrical audio signal; generating an echo estimate signal from a loudspeaker signal; and generating an error signal that is a difference between the electrical audio signal and the echo estimate signal. A residual echo contained in the echo estimate signal is then reduced by decorrelating the error signal.

In another aspect of the invention, decorrelation is applied to the error signal only when a magnitude of the error signal is in a range defined by positive and negative threshold levels.

In yet another aspect of the invention, the positive and negative threshold levels are determined as a function of the echo estimate signal. For example, the positive and negative threshold levels may be determined by detecting an envelope of the echo estimate signal.

In still another aspect of the invention, the positive and negative threshold levels are determined periodically, thereby providing dynamic adjustment.

In yet another aspect of the invention, decorrelation is accomplished by differentiating the error signal to reduce the residual echo contained in the echo estimate signal.

BRIEF DESCRIPTION OF THE DRAWINGS The objects and advantages of the invention will be understood by reading the following detailed description in conjunction with the drawings in which: FIG. 1 is a block diagram of a conventional hands-free transceiver that includes an acoustic echo canceler in the form of an adaptive filter arrangement; FIG. 2a is a graph of an exemplary signal plotted as a function of time prior to application of center clipping; FIG. 2b is a graph depicting the resulting signal after application of center clipping in accordance with conventional techniques;

FIG. 3 is a graph depicting the transfer function of a center clipper in accordance with conventional techniques; FIG. 4 is a block diagram of an arrangement that uses an echo estimate signal (û) to control a center clipper for the purpose of reducing a residual echo in accordance with conventional techniques; FIG. 5 is a block diagram of an apparatus that applies decorrelation to reduce a residual echo in accordance with one aspect of the invention; FIG. 6 is a block diagram of an exemplary embodiment of a decorrelator for use in the invention; FIG. 7 is a graph depicting the transfer function of a selective decorrelator, in accordance with one aspect of the invention; FIG. 8a is a graph of an exemplary time-domain signal prior to application of decorrelation; FIG. 8b is a graph depicting the resulting signal after application of selective decorrelation in accordance with one aspect of the invention; and FIG. 9 is a block diagram of an exemplary embodiment of a selective decorrelator in accordance with one aspect of the invention.

DETAILED DESCRIPTION The various features of the invention will now be described with respect to the figures, in which like parts are identified with the same reference characters.

In accordance with one aspect of the invention, decorrelation is applied to the error signal, e (n), in order to decorrelate the echo residuals, that is, to decrease the autocorrelation characteristic of the echo residuals. This will cause the residual echoes, which are essentially non-linearly distorted speech, to sound like noise, thereby avoiding the detrimental affect that they would otherwise have on the far-end user.

In accordance with another aspect of the invention, decorrelation of the desired near-end speech signal (which is combined with the residual echo signal) is avoided by selectively applying decorrelation to substantially only those parts of the error signal, e (n), where the echo residuals are estimated to reside.

In yet another aspect of the invention, a dynamic determination is made of where the echo residuals are estimated to reside, so that the residual echo cancellation will dynamically adapt to changing conditions.

These and other aspects of the invention will now be described with respect to an exemplary embodiment depicted in the block diagram of FIG. 5. The variously depicted components can be embodied in any number of ways, as will be apparent to one of ordinary skill in the art. Of course, hard-wired circuits (either custom logic or even analog components) could be constructed to perform the various functions described herein, and these could be connected together substantially as illustrated. Alternatively, the invention could be embodied, in whole or in part, as a stored program executed within any of a number of commercially available digital signal processors (DSPs), microcontrollers or other processors. In this respect, the invention could be considered to be the DSP, microcontroller or other processor in combination with a stored program. Alternatively, the invention could also be embodied as a computer-readable storage medium having stored therein the program instructions for effecting the invention. Computer readable storage media include, but are not limited to, volatile and non-volatile Random Access Memories (RAM), magnetic storage media (e. g., magnetic disk, diskette or tape) and optical storage media (e. g., compact disk read only memory (CD ROM)).

As can be seen from FIG. 5, the primary echo cancellation functions are performed by an adaptive filter 101, LMS cross correlator 103 (or other type of filter adaptation device, such as a recursive least squares or normalized LMS filter) and subtractor 107 arranged to operate as in the prior art circuits depicted in FIGS. 1 and 4. As explained earlier, the resultant error signal, e (n) generated by the conventional arrangement includes unacceptable levels of residual echo components. To eliminate the disturbing effect that these residual echoes can have on the far-end user, the error signal e (n) is supplied to a decorrelator 503. As mentioned above, decorrelation causes speech signals to sound like noise, which does not have the disturbing effect associated with echo signals. The decorrelator 503 can be implemented in any of a number of ways, including but not limited to implementation as a module that takes the derivative of the input signal (i. e., the output of the decorrelator 503 can be a signal that is proportional to ()). dn

A simple embodiment of the decorrelator 503 is illustrated by the block diagram of FIG. 6. The error signal e (n) is supplied directly to an input of a subtractor 603 and also to an input of a delay device 601. The output of delay device 601 is a delayed error signal 605, which is a delayed version of its input. The delayed error signal 605 is then supplied by the delay device 601 as the subtrahend to the subtractor 603. The output of the subtractor 603 is a signal that is proportional to the derivative of the error signal e (n).

Returning now to FIG. 5, operation of the residual echo canceler can be further improved by accounting for the fact that the simple decorrelator illustrated in FIG. 6 will decorrelate the wanted components contained in the error signal e (n) as well as the residual echo components. To solve this problem, the strategy applied is to determine a center slice in the error signal, e (n), that will be decorrelated. Those portions of the error signal, e (n), whose magnitude falls outside of the center slice are left unaffected. In yet another aspect of the invention, the threshold levels that define the center slice can be dynamically changed as a function of the echo estimate signal 102.

The transfer function of this selective decorrelation is depicted in FIG.

7. A center slice is defined as an region in which the magnitude of an input signal lies between a positive threshold 701 and a negative threshold 703. Outside of this region, the signal is permitted to pass unchanged. However, when the input signal falls within the center slice region, it is subjected to decorrelation as described above. This process is further illustrated in FIGS. 8a and 8b. In FIG. 8a, an exemplary signal 801 is graphed in the time domain. The result of selective decorrelation of the signal 801 is shown in FIG. 8b. It can be seen that the selectively decorrelated signal 803 is identical to the exemplary signal 801 for those regions where the exemplary signal 801 either exceeds a positive threshold 701 or falls below a negative threshold 703. Of course, the slope of the lines representing transfer function in these regions need not be 1: 1, as illustrated. In other embodiments, the slope could be greater than 1 (amplification) or less than 1 (attenuation), as required by the particular application.

For those regions where the exemplary signal 801 lies between the positive and negative thresholds 701,703, the output signal is a decorrelated version of the exemplary signal 801, as depicted by the shaded region 805 of the graph.

The positive and negative thresholds 701,703 could be static. In this case, appropriate values should be selected at the time that the residual echo canceler is designed. Suitable threshold values will permit the wanted voice signal to pass substantially unaffected by decorrelation, while applying decorrelation to signal energy levels at which the residual echo most likely appear.

An improvement can be achieved, however, by dynamically adjusting the positive and negative thresholds 701,703 so that they will more closely track the actual energy levels of the residual echo. To accomplish this, the exemplary residual echo canceler depicted in FIG. 5 further includes a decorrelator threshold calculator 501. The decorrelator threshold calculator 501 may be for example, an envelope, or level detector. The threshold values may be determined by: Positive and Negative Threshold Values (n) = ABS (e (n)) x K, where K is a constant that may be in the range from 0 to 1.

For this embodiment, the decorrelator 503 needs to include means for making use of the threshold information. An exemplary embodiment of a decorrelator 503 that operates in this fashion is shown in FIG. 9. An error signal, e (n), is supplied as an input to the decorrelator 503. This signal is supplied directly to an input of a subtractor 903 and also to an input of a delay device 901. The output of delay device 901 is a delayed error signal 907, which is a delayed version of its input. The delayed error signal 907 is then supplied by the delay device 901 as the subtrahend to the subtractor 903. The output of the subtractor 903 is a signal that is proportional to the derivative of the error signal e (n). This signal is supplied to a first input of a multiplexor (MUX) 909.

The error signal e (n) is also supplied directly to a second input of the MUX 909. The purpose of the MUX is to selectively output either the unchanged error signal e (n), or the signal that is proportional to the derivative of the error signal e (n). A control signal for the MUX 909 is generated by a comparison circuit 911, which receives as inputs: the error signal e (n), the positive threshold 701 and the negative threshold 703. In this embodiment, the output of the comparison circuit 911 is asserted whenever the magnitude of the error signal e (n) is in the range falling between the positive and negative thresholds 701,703. This signal, when asserted,

selects the signal that is proportional to the derivative of the error signal e (n) to appear at the output of the MUX 909. When the output signal from the comparison circuit 911 is not asserted, the output of the MUX 909 is the unchanged error signal e (n). In this way, the output of the MUX 909 (which is also the output from the decorrelator 503) selectively supplies either an unchanged error signal e (n) or a decorrelated signal as a function of whether the error signal e (n) is outside the center slice defined by the positive and negative thresholds 701,703.

The above-described invention provides improved methods and apparatus for reducing residual echos. Benefits of the invention include improved background noise quality compared to the use of a center clipper, because where a center clipper cuts out the signal entirely, the decorrelator only decorrelates.

Decorrelated noise sounds like noise with a small change in character. The background noise quality is one of the most important quality parameters in contemporary full- duplex hands-free designs.

Another benefit of the invention is that the residual echoes, which are essentially non-linearly distorted speech signals, will also sound like noise after decorrelation.

The inventive methods and apparatuses described herein can be applied in any hands-free microphone-loudspeaker arrangement to reduce the residual echoes of the loudspeaker signal present in the microphone signal. Such arrangements include, but are not limited to, hands-free telephones (including desk and mobile telephones such as vehicular hands-free telephones) and computer-based hands-free arrangements.

The inventive methods and apparatuses described herein can also be applied advantageously in non-hands-free arrangements that nonetheless suffer from acoustic echo problems, such as hand-held mobile telephones.

The invention has been described with reference to particular embodiments. However, it will be readily apparent to those skilled in the art that it is possible to embody the invention in specific forms other than those of the preferred embodiments described above. This may be done without departing from the spirit of the invention. The preferred embodiments are merely illustrative and should not be considered restrictive in any way. The scope of the invention is given by the appended claims, rather than the preceding description, and all variations and equivalents which fall within the range of the claims are intended to be embraced therein.