GB2336279A

GB2336279A - Echo canceller and method for device having a decoder

Info

Publication number: GB2336279A
Application number: GB9907069A
Authority: GB
Inventors: Tom Hong Li
Original assignee: Motorola Inc
Current assignee: Motorola Solutions Inc
Priority date: 1998-03-30
Filing date: 1999-03-29
Publication date: 1999-10-13
Also published as: KR19990078413A; GB9907069D0

Abstract

An echo canceller (300) is used for digital communication devices where digital speech decoder (106) is employed. The echo canceller (300) receives linear prediction residuals and a linear prediction error at an output (303) of the speech decoder (106). An updating direction for the coefficients of an adaptive filter (110) is formed in two steps in a filter coefficient generating circuit (304). In the first step, a updating direction is formed in linear prediction residual domain from the output of a decoder (106). In the second step, the updating direction used in the coefficients adaptation is obtained by transforming the updating direction from residual domain to speech domain in the filter coefficient generating circuit (304).

Description

2336279 ECHO CANCELLER AND METHOD FOR DEVICE HAVING A DECODER

FIELD OF THE INVENTION

The present invention pertains to echo cancellers, and more particularly to fast converging echo cancellers for digital communication devices.

BACKGROUND OF THE INVENTION

Some bi-directional communication systems have separate paths for signals transmitted in opposite directions. In such systems, signals in one path may be reflected into the other path. These reflected signals, commonly referred to as echo signals, interfere with desired communication signals. Consequently, echo cancellers have been developed to suppress the reflected signals.

Echo cancellers employ adaptive filters to estimate the echo signals reflected into one of the signal paths. The echo estimate is subtracted from the signal path including the echo signal to generate a substantially echo free, or echo suppressed, signal.

Because least mean square (LMS) adaptive filters have relatively simple structures, and are computationally stable and efficient, they are widely used in echo cancellers to estimate echo signals. However, LMS adaptive filters suffer from slow convergence rates when employed in acoustic echo cancellation applications, such as teleconferencing and hands-free cellular communication.

Another disadvantage of LMS adaptive filters is that they often suppress echo signal by no more than 30 dB (decibels). In a typical acoustic application, echo signals are very strong, and may be as strong as the desired communication signals. In such an environment, adaptive filters must converge quickly to simulate the fast-changing echo path and it is highly desirable for echo signals to be suppressed at least 40 dB. If these two requirements can not be satisfied, large echo estimate errors appear, and these large errors can result in severe degradation of the 5 signal quality of the desired communication signals.

Linear prediction methods have been applied to increase the echo canceller rate of convergence. US patent 4,672,665, discloses an echo canceller that estimates the linear prediction filter coefficients of a far-end speech signal and then uses the linear prediction filter coefficients to perform filtering both of the far-end and near-end speech signals. The original far-end and near-end speech signals are replaced by the filtered near-end and far-end signals, respectively, such that adaptation of the echo canceller is done from the filtered far-end and near-end signals. The filtered far-end signal is much closer to a white random process signal than the original signal, and thus the convergence rate is increased.

US patent 4,697,261, discloses an echo canceller for digital communication devices in which linear-prediction based speech coders are employed. The echo canceller adaptive filter uses the available linear prediction filter residual signal of the speech coder. The echo canceller also uses the coefficients of the linear prediction inverse- filter of near-end input signal.

However, these echo cancellers are very sensitive to near-end noise signals. They fail to achieve sufficiently reliable operation to insure that users can communicate clearly throughout a call. Thus, their applications are extremely limited.

The need for a rapidly converging, stable, high suppression, noise robust, and computationally efficient adaptive filter is particularly great in environments such as digital mobile hands-free telephones and digital hands-free teleconferencing devices, where linear-pred iction -based speech coders are employed. The echo path in such an environment is subject to rapid change, strong echo, and loud background noise signal. Accordingly, it is necessary for the echo canceller's adaptive fitter to converge quickly, and suppress echo signals by more than 40 dB. It is also desirable for an echo canceller to be computationally efficient such 2 that it can be implemented in a relatively small circuit having low current drain characteristics. It is also desirable for such an echo canceller take advantage of circuitry that is already present in a communication device.

Accordingly, a need remains for a stable, fast-converging, noise robust, and computationally efficient adaptive filter for the environments such as those subject to acoustic echoes, where echo paths experience rapid changes, echo signals are strong, and near-end noise signals are present, and taking full advantage of circuitry present in some environments. 1 BRIEF DESCRIPTION OF THE INVENTION

FIG. 1 is a circuit schematic illustrating an echo canceller according to the prior art.

FIG. 2 is a circuit schematic illustrating an echo canceller according to the prior art.

FIG. 3 is a circuit schematic illustrating an echo canceller.

FIG. 4 is a circuit schematic illustrating the echo canceller according to FIG. 3.

FIG. 5 is a flow chart illustrating operation of the echo canceller according to FIGs. 3 and 4.

DETAILED DESCRIPTION OF THE DRAWINGS

An echo canceller is implemented in a communication device including a speech decoder in a receive path. The echo canceller receives linear prediction residual samples and a linear prediction error signal from the speech decoder. The updating direction for the coefficients of the echo canceller are formed in the linear prediction residual domain. The updating direction in the residual domain is transformed to an updating direction in speech domain. Coefficient adaptation of the echo canceller is executed using the transformed updating direction in the speech domain. By calculating the updating direction through the transformation from the linear prediction residual domain to the speech domain, the optimal updating direction can be 3 obtained. This echo canceller adapts much more quickly to the echo path, suppresses echo signals to a greater degree while remaining stable, efficient, and noise robust.

A conventional least means square (LMS) echo canceller 100 is illustrated in FIG. 1. The echo canceller is implemented in a communication device 101 including a microphone 104 and a speaker 102 for hands-free operation. The receive path, which is connected to speaker 102, includes a decoder 106 and a digital-to-analog (D/A) converter 108 connected to the receiver output of transceiver 120. The transmit path, which is connected to microphone 104, includes an analog-to-digital (A/D) converter 114, a combiner 112, and an encoder 116 coupled to the transmitter input of a transceiver 120. The signals x(n), y(n) and e(n) are digital signals. The transceiver 120 is coupled to an antenna 122 when implemented in a wireless radiotelephone, which may for example be a cellular radiotelephone, a satellite telephone, a cordless telephone, or the like.

The communication device 101 may be a radiotelephone, as mentioned above, a hands-free teleconferencing device, or any other suitable communication device. Those skilled in the art will recognize that the echo canceller 100 can be implemented in bi-directional communication devices having digital circuitry. The decoder 106 is a digital decoder in a digital device such as a digital cellular radiotelephone. The encoder 116 is a digital encoder in a digital device such as a digital cellular radiotelephone.

An echo canceller 100 will now be described with reference to FIG. 1. The current sampling instant is n, a far-end speech sample x(n) is the output to a speaker 102, a near-end audio signal y(n) is received from a microphone 104, and for purposes of this description, signals x(n) and y(n) are synchronized, meaning that a digital-to-analog converter 108 and an analog-to-digital converter 114 use the same clock. A near end audio signal y(n) comprises a near-end speech signal s(n), an echo signal t(n), and a near-end noise signal N(n). An echo signal t(n) is the portion of a signal x(n) output by a speaker 102 that is reflected back to a microphone 104. A noise N(n) may, for example, be ambient noise in the compartment of a vehicle.

4 In the following discussion, near-end speech s(n) is assumed to be zero (i.e., a near-end speech s(n) does not exist) while the adaptation of the coefficients of an adaptive filter is executed. When both near-end and far-end speech are present, commonly called a double-talk condition, adaptation has to stop. Double talk detectors for detecting this situation, and preventing adaptation during this condition, are known in the art, and will not be described in greater detail hereinafter for brevity. Thus, as referred to herein, a near-end audio signal y(n) comprises an echo signal t(n) and a noise signal N(n) only.

An echo canceller 100 includes a coefficient generating circuit 109 and a FIR filter 110 to model an echo path. The echo estimate z(n) generated by a FIR filter 110 is combined with a signal y(n) in combiner 112 to remove, or suppress, an echo signal. An echo estimate error e(n) is the difference between an echo estimate z(n) and a near-end signal y(n). The coefficients of a FIR filter 110 are updated by a coefficient generating circuit 109 based on far-end speech samples and an echo estimate error.

At time n, the coefficient vector of a FIR filter 110 is W(n) = [wo(n) Wl(n)... WL-l(n)IT, where L is the length of a FIR filter 110 and superscript T (i.e., [IT) means the transpose of a vector or a matrix. The vector that is formed by L received far-end speech samples is X(n) = [x(n), x(n- 1),..., x(n-L+1)IT. The echo estimate z(n) of a FIR filter 110 filter is:

z(n) = X(n)T W(n).

The echo estimate error e(n) of a FIR filter 110 is:

e(n) = y(n) - z(n).

(1) (2) The coefficient vector W(n) of a FIR filter 110 is updated according to the following equation:

W(n+l) = W(n) + m e (n) X (n) (3) IIX(n)11' wherep is an adaptation step size and IIX(n)11' = X(n)' X(n). A linear prediction echo canceller 200 is illustrated in FIG. 2. Echo canceller 200 is in a digital communication device having a linearprediction-based speech decoder 106 according to the prior art. 5 Decoder 106 generates linear prediction filter coefficients {a k 1 k

1, 2,..., p}which are used by the decoder in a convention manner and are also supplied to a linear prediction error filter (A(z)) 205 and a preemphasis filter (P(z)) 206 at output 203, where p is the order of the linear prediction error filter.

Decoder 106 also generates a linear prediction residual signal d(n) which is provided at output 202. The linear prediction residual signal d(n) is used internally to the decoder, and is also communicated to the filter coefficient updating circuit 204. The filter coefficient updating circuit is also coupled to receive echo residual signals e(n) at output 208 via a pre-emphasis filter 206 and an inverse linear prediction error filter (l/A(z)) 205. The Filter coefficient updating circuit generates coefficients (W(n)) input to a finite impulse response filter (FIR) filter 110. The FIR filter 110 generates an echo estimate z(n).

The echo estimate z(n) generated by a FIR filter 110 is combined with a signal y(n) in combiner 112 to remove, or suppress, an echo signal. An echo estimate error e(n) is the difference between an echo estimate z(n) and a near-end signal y(n). The coefficient vector of a FIR filter 110 is updated by a coefficient generating circuit 204 based on linear prediction residual signals d(n) and the filtered echo cancellation error signals at output 207. The coefficient generating circuit 204 is the same as above LMS echo canceller except a far-end signal and an echo cancellation error signals are replaced by their filtered ones.

The echo canceller according to FIG. 2 is are very sensitive to near-end noise signals due to inverse linear prediction filtering in filter 205 of the near-end signals y(n). Inverse linear prediction filtering of the near-end signal y(n) significantly amplifies the power of near-end noise signal because linear prediction filters typically have very large gain. Additionally, the performance of adaptive echo cancellers depends upon the power of near-end noise signals. The higher the power of near-end 6 noise signals, the higher the error of adaptive filter coefficients, and the slower the convergence rate of echo cancellers. Thus, inverse filtering of the near end signal detrimentally impacts on the ability of the echo canceller to converge quickly and accurately.

The system architecture of an improved echo canceller 300 is illustrated in FIG. 3. The echo canceller operates with a speech decoder 106. The speech decoder 106 can be implemented using any suitable decoder based on linear speech decoding, which are well known to those skilled in the art, and accordingly are not described in greater detail herein. Such The decoders are used in digital cellular systems, such as those operating according to the Global System for Mobile Communications (GSM) standard.

Echo canceller 300 can be implemented using a microprocessor, a digital signal processor, a microcomputer, a computer, or any other suitable circuitry. An echo canceller 300 includes outputting linear prediction error filter coefficients {a k 1 k = 1, 2,..., p} and residual signals d(n) at output 303, where p is the order of a linear prediction error filter. The transfer function of a linear prediction error filter is A(z)=11+ -k.

lakZ (4) k=l Because speech coders in digital communication devices normally operate on a frame-by-frame basis, echo canceller 300 operates on the received audio samples on a frame-by-frame basis. The received audio samples include the signal samples from the far-end which are output to speaker 102, and the near-end input from microphone 104. The audio sqmples from both the far-end and near-end are synchronized. The frame size K for an echo canceller may or may not be the same as the frame size of speech coders. Once a frame of K new audio samples from both the far-end and near-end are received, the operation of the new echo canceller starts. K received far-end speech samples are x(n), x(n-1), x(n-K+1) and K received near-end signal samples are y(n), y(n-1), y(n-K+1), where n is the current sampling instant.

7 The coefficient vector of a FIR filter is updated in a coefficient vector generating circuit 304, which is a coefficient updating circuit. Echo cancellation is performed on K received audio samples, input signal samples, using the updated coefficients in a FIR filter 110. A time index i is inside the time range of K new received audio samples, i.e., n: i:, nK+1. W(i) = [wo(i) wl(i)... WL_1 (iff is the coefficient vector of a FIR filter 110 at time i, L is the length of a FIR filter 110, and X(i)=[X(i) x(i-1)... x(i-L+1 T is a vector of L received far-end speech samples at time i. L may be smaller or larger than K.

On a sample-by-sample basis for K received samples, i.e., i=n- K+t n-K+2,..., n, echo canceller 300 performs echo cancellation. Echo canceller 300: at i=n-lK+l, performs echo cancellation using a coefficient vector W(n-K+1), and updates a coefficient vector W(n-K+1) to a coefficient vector W(n-K+2); at time i=n-lK+2, performs echo cancellation using a coefficient vector W(n-K+2), and updates a coefficient vector W(n-K+2) to a coefficient vector W(n-lK+3); continues this operation at each successive sampling instant until a last sampling instant, i=n, at which echo cancellation is performed using a coefficient vector W(n), and a coefficient vector W(n) is updated to a coefficient vector W(n+l); and waits for the next frame to repeat this the sequence for that f rame. At a specified time i, an echo estimate error of an adaptive echo canceller is:

e(i) = y(i) _ X(j)T W(i) (5) where y(i) is a near-end audio sample at time i.

At time i, a coefficient vector updating direction of an adaptive echo canceller is constructed in linear prediction residual domain in updating direction circuit 402 of coefficient vector generating circuit 304. L linear prediction residual samples d(i), d(i-1),..., d(i-L+1) are obtained 35 from a speech decoder. A coefficient vector updating direction of an 8 adaptive echo canceller in linear prediction residual domain is Q(i), wherein:

Q0) = m e (i) D (i) lID(n)11' (6) where D(i) = [d(i), d(i-1),... d(i-L+1) f and Q(i) = [qo(i), ql(i),.-- qL- 1 (i)]T, p is an adaptation step size and IID(i)il' = D(j)T D(i).

After obtaining a coefficient vector updating direction in updating direction circuit 402, the updating direction is transformed to the speech domain in transform circuft 404. The updating direction in the speech domain is calculated based on Affl and Q(i). The updating direction and transformation are described in copending patent application serial number 08/923,574, entitled DIRECTION TRANSFORM ECHO CANCELLER AND METHOD, filed 4 September 1997, in the name of Tom Hong Li, the disclosure of which is incorporated herein by reference.

The coefficient vector updating direction of an adaptive echo canceller in the speech domain is G(i) = [go(i) gi(i)... 9LA (i)]T. G(i) is calculated for each component of G(i) through the following equation:

gj(i) q (i) + O kqj-k(i) j = 0, k1 1,., L-1 (7) where gj(i) = 0 for j=-1, -2...... p are assumed in order to perform the above equation.

After G(i) is obtained, a coefficient vector of an adaptive echo canceller can be updated diredly in update coefficient circuit 406 as follows:

W(i+l) = W(i) + G(i) (8) where W(i+l) is the updated coefficient vector for use in the next sample at time 1+1.

9 The operation of echo canceller 300 will now be summarized with reference to FIGs. 4 and 5. The echo canceller generates current coefficients for the adaptive filter in circuit 304 as indicated in block 502.

The updating direction is then calculated in the linear prediction residual domain in circuit 402, as indicated in block 504. The updating direction is transformed to the speech domain in circuit 404, as indicated in block 506. Updated coefficients are then generated in updating coefficient generating circuit 406, as indicated in block 508. The updated coefficients are input to finite impulse response (FIR) filter 110, which generates an echo cancellation signal. Those skilled in the art will recognize that the circuits 402, 404 and 406 can be implemented in the digital signal processor (DSP), microprocessor, microcontroller or computer implementing the coefficient vector updating circuit 304, or in separate programmable logic circuits that together implement the coefficient vector updating circuit 304.

Even though an LMS echo canceller has a slow convergence rate, it is simple, robust, and computationally efficient. Echo canceller 300 preferably employs a FIR filter to take advantage of these characteristics of an LMS echo canceller while improving the speed of adaptation convergence by using an improved updating direction. Echo canceller is very sensitive to near-end noise signals due to filtering of both far end and near-end signals using a linear prediction error filter. However, echo canceller 300 does not require linear prediction filtering of near end signals due to the conversion of the updating direction. Accordingly, it is much more robust than echo canceller 200. The improved echo canceller is readily implemented, stable, robust, and computationally efficient.

Simulations show that echo canceller 300 is robust in noisy environments. An updating direction transformation provides a much faster convergence rate, and a much higher echo suppression. Those skilled in the art will also recognize that the advantages of the transformed adaptation direction can be employed in non LMS echo cancellers. Thus "adaptive filter" and "echo canceller" as used herein is not limited to LMS adaptive filters and LMS echo cancellers. It is also envisioned that the echo canceller will be implemented with a double talk detector to disable adaptation when near-end and far-end signals are both present.

Claims

1. A method of generating an updating direction in an echo canceller for a digital communication device, the echo canceller including an adaptive filter and the digital communication device including a speech decoder in a receive signal path to decode receive signals and generate output signals to drive a speaker, the method comprising the steps of:

outputting linear prediction signals from the speech decoder; calculating in the adaptive filter an updating direction in a linear prediction residual domain using the linear prediction signals; and transforming the updating direction from the linear prediction residual domain to a speech domain using the linear prediction signals to generate a speech domain updating direction for use in updating the coefficients of an adaptive filter.

2. The method as defined in claim 1, wherein the step of outputting linear prediction signals includes outputting linear prediction residuals and said step of calculating in the adaptive filter an updating direction in the linear prediction residual domain uses the linear prediction residuals.

3. The method as defined in claim 1, wherein the step of outputting linear prediction signals includes outputting linear prediction coefficients and wherein said step of transforming the updating direction from linear prediction residual domain to the speech domain uses the linear prediction coefficients.

4. The method as defined in claim 3, wherein the step of transforming the updating direction from linear prediction residual domain to the speech domain includes filtering the updating direction by a finite impulse response filter to obtain the updating direction in speech domain.

5. The method as defined in claim 3, wherein the step of transforming the updating direction from linear prediction residual domain to the speech domain includes transforming the updating direction to speech domain using a linear prediction error fitter.

6. The method as defined in claim 5, where the step of transforming the updating direction from linear prediction residual domain to the speech domain includes calculating the updating direction g,(i) in speech domain according to:

p g,(i) = q, (i) + X_.' X kqn(i-k) k=l where L is the length of an adaptive filter, p is an order of a linear prediction error filter, Q(i) = [qo(i), ql(i),.... qL-l(i)] T is a updating direction in linear prediction residual domain, G(i) = [go(i), gi(i), W-1 (i)]T is a speech domain updating direction, and (aJ are coefficients of a linear prediction error filter.

7. The method as defined in claim 1, wherein the communication device includes an input coupled to a microphone for inputting near-end signals and a linear prediction based speech coder which coder encodes an echocancelled signal into information bits for transmission, and further comprising the steps of: obtaining linear prediction filter coefficients from the decoder; obtaining linear prediction residual signals from the decoder; wherein the step of calculating includes the step of forming an updating direction for an echo canceller in a linear prediction residual domain from the linear prediction residual signals; wherein the step of transforming the updating direction from the linear prediction residual domain to a speech domain in a 12 linear prediction error filter uses the linear prediction filter coefficients; and further including the step of generating updated coefficients for an adaptive filter based on current coefficients and the updating direction in the speech domain.

8. The method as defined in any one of claims 1-7, wherein the device is a hands-free speaker phone.

13