JPH11154894A - Direction-change echo canceler and method thereof - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an echo canceler, which can be executed in a comparatively small circuit with high calculation efficiency having low drain current characteristic to quickly converge and suppress echo signally more than 40 dB. SOLUTION: An echo canceler 200 converts a remote terminal input voice to its linear predictive remainder. At a coefficient generating circuit 204, an updating direction to the coefficient of the echo canceler is formed in two stages. In the first stage, the updating direction is formed in the region of the linear predictive remainder. On the second stage, the updating direction is converted from the remainder region to an audio region, so that an updating method to be used for adapting the coefficient can be provided.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates generally to echo cancellers, and more particularly to a fast converging echo canceller.

[0002]

2. Description of the Related Art Some two-way communication systems have a separate path for signals transmitting in the opposite direction. In such a system, a signal on one path may be reflected and enter the other path. These reflected signals are generally called echo signals and interfere with desired communication signals. As a result, echo cancellers have been developed to suppress reflected signals.

[0003] An echo canceller uses an adaptive filter to estimate an echo signal that is reflected and enters one of the signal paths. Subtracting this echo estimate from the signal path containing the echo signal produces a signal that is substantially echoless, ie, echo suppressed.

[0004] Mean least square (LMS)
e) Adaptive filters have a relatively simple structure and are computationally stable and efficient, so they are widely used in echo cancellers for estimating echo signals. However, LMS adaptive filters include video conferencing (telec
onference) and hands-free cellular communications (hands-free
When used for acoustic echo cancellation such as cellular communication, there is a disadvantage that the convergence rate is low.

[0005] Another disadvantage of LMS adaptive filters is that they often suppress echo signals below 30 dB (decibel). In a typical acoustic application, the echo signal is very strong and can be as strong as the desired communication signal. In such an environment, the adaptive filter must converge quickly to simulate a rapidly changing echo path, and it is highly desirable to suppress the echo signal by at least 40 dB. If these two requirements cannot be met, large echo estimation errors will occur, which will result in significant degradation in the signal quality of the desired communication signal.

[0006] Repetitive least squares (RLS)
ast square) algorithm Affine projection algorithm (a
Adaptive filters using other algorithms (such as the ffine-projection algorithm) adapt faster than echo cancellers using LMS filters, but they are unstable or computationally expensive. Computationally expensive filters are those that require large circuits or large amounts of processing resources. Unstable adaptive filters generate erratic estimates and therefore cannot achieve sufficiently reliable behavior, ensuring that users can communicate clearly over a call Can not.

[0007]

[Problems to be solved by the invention] Fast convergence, stability,
The need for high suppression and computationally efficient adaptive filters is particularly acute in environments such as mobile hands-free telephones and hands-free video conferencing. The echo path in such an environment can be fast changing, strong echo,
And troublesome background noise is likely to occur. Therefore, the adaptive filter of the echo canceller needs to converge quickly and suppress the echo signal by 40 dB or more. Also,
It is also desirable that the echo canceller be computationally efficient and be implemented in relatively small circuits having low drain current characteristics.

For this reason, in an environment in which an acoustic echo is likely to be generated, it is necessary to keep the filter stable and computationally efficient even when the echo path changes quickly and the echo signal is strong.

[0009]

DETAILED DESCRIPTION OF THE INVENTION An echo canceller converts a far-end speech sample into its linear prediction residual sample. An update direction for the echo canceller coefficients is formed in the linear prediction residual area. Next, the updating direction in the remaining area is converted to the updating direction in the audio area. The coefficient adaptation of the echo canceller is performed using the updated direction converted in the voice domain. By calculating the update direction by converting the linear prediction residual area into the speech area, an optimal update direction can be obtained. This allows the echo canceller to significantly speed up the adaptation to the echo path, enabling a significant increase in echo suppression while maintaining stability, efficiency and robustness.

A conventional Least Mean Square (LMS) echo canceller 100 is shown in device 101 of FIG. Device 101 includes a microphone 104 and a speaker 102 for hands-free operation. The speaker path is a digital-to-analog (D / A) converter 1 connected to the decoder 106 and the receiver output of the transceiver 120.
08. The microphone path includes an analog-to-digital (A / D) converter 114, a combiner 112, and an encoder 116 coupled to the transmitter of the transceiver 120.
including. Therefore, the signals x (n), y (n) and e
(N) is a digital signal. Transceiver 120 is coupled to antenna 122 in the wireless telephone.

Device 101 can be a hands-free accessory connected to a mobile radiotelephone,
Alternatively, it may be a wireless telephone or a hands-free video conferencing device. Echo Canceller 10
One skilled in the art will recognize that 0 can be implemented in a two-way communication device having digital and / or analog circuitry. Thus, decoder 106 and encoder 116 are for illustrative purposes. Decoder 106 may provide A / A if transceiver 120 is in an analog cellular horn or analog landline horn, or if device 101 is connected to the analog output of a digital cellular horn or digital landline horn.
It can be a D converter. Alternatively, the decoder 10
6 may be a digital decoder in a digital device such as a digital cellular horn. Encoder 116 may be connected when transceiver 120 is in an analog cellular horn or analog landline horn, or when device 101 is connected to the analog output of a digital device such as a digital cellular horn or digital landline horn. If so, it can be a D / A converter. Alternatively, the encoder 116
In a digital device such as a digital cellular horn, it can be a digital encoder.

The echo canceller 100 will now be described with reference to FIGS. Let n be the current sampling point, the far-end audio sample x (n) is output to the speaker 102, and the near-end audio signal y
(N) is received from microphone 104, and for the purposes of this description, signals x (n) and y (n) are assumed to be synchronized. That is, the digital-analog converter 10
8 and the analog-to-digital converter 114 mean using the same clock. Near-end audio signal y
(N) consists of near-end speech s (n), echo t (n), and near-end noise N (n). The echo t (n) is a portion where the signal x (n) output from the speaker 102 is reflected and returns to the microphone 104. The noise N (n) is, for example, ambient noise in a vehicle cabin.

In the following description, it is assumed that near-end speech s (n) is zero (ie, near-end speech s (n) does not exist) while performing the adaptation of the coefficients of the adaptive filter. . If both near-end and far-end audio are present,
That is, adaptation must be stopped in a state commonly referred to as double-talk. Double talk detectors for detecting this situation and inhibiting adaptation during this state are known in the art and will not be described in detail below for brevity. Therefore, when cited here, it is assumed that the near-end audio signal y (n) consists only of the echo t (n) and the noise N (n).

The echo canceller 100 includes a coefficient generation circuit 109 and an LM for modeling an echo path.
An S adaptive filter 110 is included. LMS adaptive filter 11
The estimated echo value z (n) at which 0 occurs is determined by the combiner 11
The signal is combined with the signal y (n) in 2 to remove or suppress the echo signal. The echo estimation error e (n) is the difference between the echo estimate z (n) and the near-end signal y (n). The coefficient of the LMS adaptive filter 110 is updated by the coefficient generation circuit 109 based on the received far-end voice sample and the echo estimation error.

At time n, the LMS adaptive filter 11
Coefficient of 0, W (n) = [w 0 (n) w 1 (n). . .
w _L-1 (n)] ^T. L is the filter length, and the superscript T (ie, [] ^T ) means transposition of the vector. The L far-end received far-end speech samples are X (n) =
[X (n), x (n-1),. . . , X (n−L +
1)] ^T. Echo estimate z of LMS adaptive filter
(N) is represented by the following equation. z (n) = X (n) ^T W (n) (1) Echo estimation error e for LMS adaptive filter 110
(N) is represented by the following equation.

E (n) = y (n) −z (n) (2) The coefficient W (n) of the LMS adaptive filter 110 is updated according to the following equation.

[0017]

(Equation 3) Where μ is the adaptation step size, and
X (n) || ² = X (n) ^T X (n).

R (n) = E {X (n) X (n) ^T } =
{R _ij (n) | i, j = 0,1,. . . , L-1} is
6 is an autocorrelation matrix of far-end speech samples. E
｛ ^* ｝ Is the statistical average. Far-end audio signal is non-stationary
(nonstationary), R (n) is a time variation Lx
L matrix, L time-varying positive eigenvalues ｛λ _i
(N) i = 0, 1,. . . , L-1} and L time variation eigenvectors _{{v i (n) i =} 0,1 ,. . . , L
-1) has｝. S (n) = E {y (n) X (n)}
= [S ₀ (n) s ₁ (n). . . S _L-1 (n)] ^T =
[E ｛y (n) x (n)｝ E ｛y (n) x (n−
1) II. . . E {y (n) X (n-L + 1)}] ^T is
It is an autocorrelation vector of a far-end audio sample and a near-end audio signal.

The performance of the adaptive filter depends on the values of the L time-varying eigenvalues and the eigenvectors. It is known that μ is fixed and must be chosen to obey the following rules:

0 <μ <2 / λ _max (n) (4) where λ _max (n) is the maximum eigenvalue of R (n).
Since the far-end speech is non-stationary, λ _max (n) has a large dynamic range. However, in order to maintain the stability of the adaptive filter, μ must be very small so that it always falls within the range of equation (4). A small μ to maintain this stability is LM
This is one of the causes of the slow convergence of the S adaptive filter.

The optimum coefficient of the LMS adaptive filter 110 is
W ₀ = [w ₀ w ₁ ,. . . , W _L-1 ] ^T. Normal,
W ₀ changes very slowly over time. Therefore, here, this is regarded as a constant vector. LMS
The echo estimation error generated by the adaptive filter using the current coefficient W (n) is given as follows.

Ε (n) = y (n) −z (n) = y (n) −X (n) ^T W (n) (5) where y (n) is a near-end echo signal. The situation where y (n) is composed of echo and noise is discussed below. The following analysis assumes that there is no near-end noise. If W (n) is not correlated with y (n) and X (n) (this assumption is approximately correct if W (n) converges and W (n) is considered a constant vector) ), The mean square error of the echo estimate of the LMS adaptive echo canceller is defined by:

Ξ = E ｛ε ² (n)｝ = E ｛[y (n) −X (n) ^T W (n)] ² ＝= E ｛y (n) ² ｝ + E ｛W (n) ^T X (n) X (n) T W (n)} - 2E {y (n) X (n) W (n) T} = E {y (n) 2} + W (n) T E {X (n ) X (n) ^T {W (n) -2E} y (n) X (n)} W (n) ^T = E {y (n) ² } + W (n) ^TR (n) W (n) −2S (n) W (n) ^T (6) Since W ₀ is the optimal solution of W (n) and ξ is a function of W (n), the slope of ξ at W ₀ is zero. This satisfies the following equation:

[0024] ∂ξ / ∂W (n) = 0 => 2R (n) W 0 -2S (n) = 0 => W 0 = R (n) -1 S (n) (7) (7) equation Into the equation (6), the minimum mean square error of the echo estimation value can be obtained using the following equation.

Ξ _min = E ｛y (n) ² ｝ −S (n) ^T W ₀ (8) Although ξ _min fluctuates with time, it changes very slowly with time, so it is regarded as a constant here. By substituting Equations (8) and (7) into Equation (6), the following relationship can be proven.

Ξ = ξ _min + [W (n) −W ₀ ] ^TR (n) [W (n) −W ₀ ] (9) The LMS adaptive filter converts W (n) to its true value W ₀ . Try to get closer. That is, ξ is made to approach ξ _min .
The convergence behavior of the adaptive filter is the case where there are two coefficients,
It is possible to understand the case where L = 2 and W (n) = [w ₀ (n) w ₁ (n)] and W ₀ = [w ₀ w ₁ ]. Therefore, the convergence behavior of the adaptive filter will be described with reference to the equal mean square curves shown in FIGS. These curves set ξ to different constant values, w ₀ (n) and w ₁
This is plotted by using (n) as an axis. Curve is w ₀ (n) oval with -w ₁ (n) plane, the eigenvectors of R (n) is defined the major axis of the ellipse, the eigenvalues defines a sudden Hod (steepness) of the error surface.
Thus, the larger the eigenvalue, the longer the axis, and the eigenvalue of the matrix R (n) defines the length of the axis.

Since the audio signal is non-stationary, the equal mean square error curve has a time-varying ellipses.
Becomes The LMS adaptive filter uses the update direction shown in FIG.
3 and 13 in FIG. The tilt direction is the only simple and reliable update direction. As can be seen, the direction 11 fluctuates significantly from the optimal direction 12 shown in FIG. 2 and the update direction 13 is almost always the same as FIG. 3 where the matrix R (n) includes these two examples with time-varying eigenvalues. Greatly fluctuates from the optimal direction shown in FIG.

The shape and size of the mean-square error curve change over time. Thus, LMS adaptive filters often use update directions that have a large divergence from the optimal direction. This is another reason why the convergence of the LMS adaptive filter is so slow that it is not possible to suppress the echo signal by more than 40 dB in applications such as acoustic echo cancellation.

The system architecture of the improved echo canceller 200 is shown in FIG. Echo canceller 200 may be implemented using a microprocessor, digital signal processor, microcomputer, computer, or any other suitable circuit.
It is feasible. The echo canceller 200 includes a linear prediction circuit 202, and a far-end audio signal x as an input signal.
(N), and outputs a linear prediction coefficient and a residual signal d (n) to an output 203. The linear prediction circuit can be implemented using any suitable linear prediction filter circuit. The linear prediction error filter is advantageous because many algorithms are available with a fast processing speed to estimate its coefficients. Preferably, a fast algorithm that is computationally efficient and stable is used. Choosing a linear prediction error filter with these characteristics ensures that the echo canceller 200 is stable and efficient.

Linear prediction analysis can provide very accurate estimates of speech parameters. Linear prediction systems are excited by an impulse train for voiced speech or a random noise sequence for unvoiced speech. The parameters of a linear prediction system vary very slowly over time. The relationship between the audio sample s (n) and the excitation u (n) is represented by the following equation.

[0031]

(Equation 4) Where p is the order of the time-varying digital filter, G is the gain factor, and ｛α _k | k = 1,
2,. . . , P} are the coefficients of the time-varying digital filter.

Linear prediction of speech samples generates an estimate of the current speech sample from a linear combination of past samples. Assuming that the s (n) hat is a linear prediction value of s (n), the s (n) hat has prediction filter coefficients {α _k | k = 1, 2,. . . , P} using a linear predictor.

[0033]

(Equation 5) The prediction error of the linear predictor of Expression (11) is given by the following expression.

[0034]

(Equation 6) From equation (12), it can be seen that the prediction error sequence is the output of the linear prediction error filter, and its transfer function is as follows.

[0035]

(Equation 7) Linear prediction error filter coefficient ｛α _k | k = 1,
2,. . . , P} are known. The linear prediction error filter is a whitening filter because it generates a residual signal having white noise characteristics. In the present invention, Levinson-D
The urbin iterative algorithm is particularly advantageous. This filter has a number of desirable properties. It is popular for speech prediction and is therefore very well understood and developed. This is computationally efficient, so
It can be implemented in relatively small circuits. Another important property of this algorithm is that the linear prediction error filter employing this filter is stable.

The echo canceller 200 operates on a frame basis of a received audio sample. The received audio samples refer to signal samples from both the far-end signal output to the speaker 102 and the near-end signal input from the microphone 104. Audio samples from both the far-end and near-end signals are synchronized. The frame size K is usually 50 to 20
0 samples. Once the frames of K new audio samples have been received from both the far end and the near end, the operation of the algorithm begins. The K received far-end speech samples are x (n), x (n-1),. . . x (n
−K + 1), and the K received near-end signal samples are y
(N), y (n-1),. . . y (n−K + 1). Here, n is the current sample time.

In operation, a linear prediction analysis of the K received far-end speech samples, ie, the input signal samples, is first performed in the linear prediction circuit 202. Linear prediction error filter coefficients {α _k | k = 1, 2,. . . , P} can be estimated by various algorithms such as the Levinson-Durbin algorithm. The order p of the linear prediction error filter used in the linear prediction circuit 202 may be, for example, in the range of 5 to 15. The linear prediction error filter for the linear prediction circuit 202 has a transfer function A (z) represented by the following equation.

[0038]

(Equation 8) The coefficients of the adaptive filter are updated in a coefficient generating circuit 204, which is a coefficient updating circuit. LMS adaptive filter 1
Using the updated coefficients at 10, adaptive echo cancellation is performed on the K received audio samples, ie, the input signal samples. The time index i is within the time range of the K new received audio samples. That is, n ≦ i ≦ n−K + 1.
_{W (i) = [w 0} (i) w 1 (i). . . w _L-1
(I)] ^T is the coefficient of the adaptive echo canceller at time point i, L is the length of the adaptive echo canceller, 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.

For each of the K received samples, i.e., i = n-K + 1, nk + 2,. . . , N
In response, the echo canceller 200 executes the echo cancellation. Echo canceller 200
Performs echo echo cancellation using the coefficient W (n-K + 1) at i = n-K + 1, and updates the coefficient W (n-K + 1) to the coefficient W (n-K + 2).

At time i = n−K + 2, the coefficient (n−
The echo cancellation is performed using (K + 2), and the coefficient (n−K + 2) is updated to the coefficient (n−K + 3).

Further, at the last sampling time point i = n, echo cancellation is performed using the coefficient W (n), and each successive echo is canceled until the coefficient W (n) is updated to the coefficient W (n + 1). This operation is repeated at the sample time, waiting for the next frame, and this sequence is repeated for that frame.

At the specified time point i, the echo estimation error of the adaptive echo canceller is as follows.

[0043] Here e (i) = y (i ) -X (i) T W (i) (15), y (i) is the near-end audio sample at time i.

At time i, the coefficient update direction of the adaptive echo canceller is constructed in the linear prediction residual area. The L linear prediction residual samples d (i), d (i-1),. . . d (i-L + 1)
Is generated by filtering the received far-end speech sample at time point i using a linear prediction error filter A (z).

[0045]

(Equation 9) The coefficient update direction of the adaptive echo canceller in the linear prediction residual region is Q (i), and is expressed as follows.

[0046]

(Equation 10) Here, D (i) = [d (i), d (i-1),. . .
d (i−L + 1)] ^T , and Q (i) = [q ₀
(I), q ₁ (i),. . . q _L-1 (i)] ^T and μ are the steps of adaptation, and || D (i) || ² = D (i) ^T
D (i).

After obtaining the coefficient change direction of the adaptive echo canceller in the linear prediction residual area, the coefficient generation circuit 204
Then, the update method in the audio domain is represented by A (z) and Q
It can be calculated based on (i). The coefficient update direction of the adaptive echo canceller in the voice domain is G
(I) = [g ₀ (i) g ₁ (i). . . g _L-1 (i)]
^T. G (i) is calculated by the following equation.
This is performed for each component of (i).

[0048]

[Equation 11] Here, g _j (i) = 1 (j = −1, −2,.
p) was assumed to calculate the above equation.

Equations (18) and (16) represent filters with different operating directions. Equation (16) is used for forward linear prediction filter processing (forward linear prediction filter processing).
ring), and converts the far-end speech sample from the speech area into a linear prediction residual area. Equation (18) performs backward linear prediction and converts the update direction from the linear prediction residual area to the speech area.

After obtaining G (i), the coefficients of the adaptive echo canceller can be updated directly as follows.

W (i + 1) = W (i) + G (i) (19) Here, W (i + 1) is a coefficient updated at time point i + 1 for use in the next sample.

Since the coefficients of the linear prediction error filter for the speech signal change slowly with time, the coefficients of the linear prediction error filter within each frame are substantially 1
It does not change between frames. Therefore, within one frame of the audio sample, this is a linear time-invariant FI
It can be regarded as an R filter (linear time-invariant FIR filter). The coefficient for one frame of the audio sample is estimated by the Levinson-Durbin algorithm. Also, the linear prediction error filter is a whitening filter, and the linear prediction residual is effectively white noise. Therefore, the autocorrelation matrix of the linear prediction residual is that of white noise.

P (i) + E {D (i) D (i) ^T } =
{P _ij (n) | i, j = 0,1,. . . , L-1} is
It is an autocorrelation matrix of a linear prediction residual. When the update direction Q (i) in the linear prediction residual region is used as the update direction in the adaptation, the mean square error β (i) of the LMS echo estimation value at the time point i is given by the following equations (5) to (9). It is obtained from the following equation.

[0054] _{β (i) = β min +} [V (i) -V 0] T P (i) [V (i) -V 0] (20) Here, β _min is the smallest of the β (i) Value. As before, this can be considered a constant since it changes slowly over time. V (i) is a coefficient of the echo canceller at time point i, and is a convolution of W (i) by (1 / A (z)). V ₀ is the optimal solution of the previously defined coefficient V (i), which changes slowly with time and is therefore considered a constant vector. V (i)
Is the convolution of W ₀ and (1 / A (z)).

FIG. 5 shows an equal mean square error curve for the linear prediction residual region. P (i) has L equal eigenvalues and the equal mean square error curve is a circle. As a result, the tilt direction 51 is the only reliable update direction obtained for the LMS filter, and the tilt direction 51 is the optimal direction at any point on the circle. V (i) is W (i)
Converges to V ₀ much faster than converges to W ₀ .
Note that V (i) is not equal to W (i), and in fact, (1)
/ A (z)) is the convolution of W (i). The optimal update direction can be obtained as Q (i) in the remaining area.

The optimal update direction in the voice domain for W (i) can be obtained from the information obtained from Q (i) and A (z). Consider the following equation for the adaptation of V (i):

V (i + 1) = V (i) + Q (i) (21) This is the same as the LMS adaptive filter of FIG. 1, except that the update direction in the linear prediction residual region is used.
After convergence, the coefficients are not equivalent to the true coefficients W (i), but are equivalent to convolution with (1 / A (z)). Since the linear prediction error filter A (z) is time-invariant within one frame of the received far-end speech sample, W
(I) is the convolution of V (i) and A (z). The convolution according to A (z) in equation (21) is performed as follows.

[0058]

(Equation 12) Here, x in a circle represents convolution of two filters. Therefore, the above adaptive equation derives the update direction by convolution of the update direction in the residual area with A (z), and it is the conversion update direction that is used in the echo canceller 200.

The conversion update direction indicates the optimum direction in the audio area. This is demonstrated by: (1) Since Q (i) is the optimal direction for the adaptation of V (i), there exists a scalar η such that V (i) + ηQ (i) −V ₀ = 0. (2) If the convolution of A (z) and Q (i) is not the optimal update direction for W (i),

[0060]

(Equation 13) Does not exist. (3) A (z) is a linear time invariant filter,

[0061]

[Equation 14] Therefore, the above equation can be rewritten as follows.

[0062]

(Equation 15) (4) Clearly, at least one that satisfies the above equation
There are two scalars α = η. Therefore, the convolution of A (z) and Q (i) produces an optimal update direction for W (i).

Similarly, the root-mean-square error of the echo estimated value of the echo canceller 200 can be expressed as follows according to equations (5) to (9).

[0064] ε (n) = y (n ) -z (n) = y (n) -X (n) T W (n) (23) where, y (n) is the near-end signal. W (n) = [w ₀
(N) w ₁ (n). . . w _L-1 (n)] ^T
This is the coefficient of the canceller 200. X (n) = [x
(N), x (n-1),. . . , X (n−L + 1)] ^T
Is the received far-end speech sample. The root mean square error of the echo estimate of this new adaptive echo canceller is defined by:

[0065] ^{_{ξ N = E {ε 2 (}} n)} = E {[y (n) -X (n) T W (n)] 2} = ξ min + [W (n) -W 0] T R (n) [W (n) -W 0] _{where, ξ min = E {y (} n) 2} -S (n) T W 0
Is the minimum mean square error by the echo canceller 200. w ₀ = [w ₀ w ₁ ,. . . , WL _-1 ] ^T are the optimal coefficients of the echo canceller 200. R (n) =
{X (n) X (n) ^T } is defined as before.
The equal mean square error curve is identical to that of the LMS echo canceller of FIG. However, as shown in FIG. 6, the update direction has been improved.

Two coefficients, L = 2, W (i) =
FIG. 6 shows the root mean square error curve of the new adaptive filter for [w ₀ (i) w ₁ (i)] and W ₀ = [w ₀ w ₁ ]. The new adaptive filter does not employ the tilt direction 61 as its update direction as in the case of the LMS adaptive filter of FIG. The coefficient generation circuit 204
The coefficient updating direction is determined in stages. First, an update direction is obtained in the linear prediction residual area of the far-end speech. Second, the update direction is converted into a voice domain to obtain an optimal update direction 62. Since the coefficient generation circuit 204 always uses the optimal update direction, it avoids the detrimental effects of changes in eigenvectors and eigenvalues of far-end speech that occur in prior art LMS filters.

The echo canceller 200 (FIG. 2) converges much faster than the conventional LMS adaptive echo canceller. In addition, the echo canceller 200
B or more acoustic echo can be suppressed. This is particularly useful when used for acoustic echo cancellation applications.

Therefore, the echo canceller 200
2 shows that the conventional LMS adaptive filter 110 is used. However, an improved update direction is used by the coefficient generation circuit 204. The update direction is derived from the output of the linear prediction error filter, which calculates linear prediction coefficients from the input signal samples, as shown in block 700 (FIG. 7). The algorithm for calculating the linear prediction error filter coefficients is easily implemented, fast, and stable. Levinson-Durbi
The n algorithm is known, computationally efficient,
Since the echo canceller circuit can be mounted without excessively increasing the size of the echo canceller circuit, it can be advantageously employed. Other known low order finite impulse response (FIR)
There is a filter that can be used to calculate linear prediction information.

The update direction in the linear prediction residual region is derived from the linear prediction residual, as shown in block 701. This is done after converting the far-end speech samples into linear prediction residuals. This involves FIR filtering of the input signal samples. The order of the FIR filter is, for example,
It can be about 10. Therefore, the power required for calculation is small, and stability is guaranteed. Therefore,
The update direction is initially in the linear prediction residual region (the signal processed by the linear prediction filter is in the linear prediction region).

Next, as shown in block 702, the update direction is converted from the linear prediction residual area to the speech area. The audio area is a digitized audio. The domain conversion can be performed using an FIR filter. The update direction in the voice domain is used by the coefficient generation circuit 204 to update the adaptive filter coefficients.

It is known that the LMS echo canceller is simple, has high robustness, and is computationally efficient. The echo canceller 200 is an LMS
It is preferable to take advantage of these characteristics of the LMS echo canceller while employing filters and increasing the convergence speed of the adaptation by using an improved update direction. It can be concluded that the improved echo canceller is easy to implement, stable, robust and computationally efficient. Simulations have shown that the echo canceller 200 has high robustness even in noisy environments. The conversion of the update direction significantly increased the convergence speed, and the convergence speed increased the echo suppression. One skilled in the art will also recognize that the benefits of the transformed adaptation direction can be employed in non-LMS echo cancellers. Therefore, as used herein, "adaptive filter" and "echo canceller" are not limited to LMS adaptive filters and LMS echo cancellers.

[Brief description of the drawings]

FIG. 1 is a circuit diagram showing an echo canceller according to a conventional technique.

FIG. 2 shows that when only two coefficients are used, that is, when L = 2, LM
5 is a graph showing an example of an equal mean square error curve of an echo estimate generated by an incorrect estimate of the coefficient of the S adaptive echo canceller.

FIG. 3 shows that when only two coefficients are used, that is, when L = 2, LM
9 is a graph illustrating another example of an equal mean square error curve of an echo estimate generated by an incorrect estimate of the coefficient of the S adaptive echo canceller.

FIG. 4 is a circuit diagram showing an echo canceller.

FIG. 5 shows an example of an echo estimation value generated by an inaccurate estimation value of a coefficient of an adaptive echo canceller when only two coefficients are used, that is, L = 2, and an update direction in a linear prediction residual region is used. 7 is a graph showing an example of a mean square error curve.

FIG. 6 shows an inaccurate adaptive echo canceller coefficient when using only two coefficients, ie, L = 2, and using the update direction in the voice domain, which is transformed from the update direction in the linear prediction residual area. The figure which shows the example of the equal mean square error curve of the echo estimation value generated by an estimation value.

FIG. 7 is a flow chart showing the occurrence of an update direction.

[Explanation of symbols]

 REFERENCE SIGNS LIST 100 Mean least square (LMS) echo canceller 101 Device 102 Speaker 104 Microphone 106 Decoder 108 Digital-to-analog (D / A) converter 112 Combiner 114 Analog-to-digital (A / D) converter 116 Encoder 120 Transceiver 200 Echo Canceller 202 Linear prediction circuit 203 Output 204 Coefficient generation circuit

Claims (10)

[Claims] 1. A method for generating an update direction in an echo canceller including an adaptive filter, comprising: calculating an update direction in a linear prediction residual area; and converting the update direction from the linear prediction residual area to a speech domain. Generating an audio domain update direction for use in updating the coefficients of the adaptive filter. 2. The method according to claim 1, further comprising generating a linear prediction residual signal using a Levinson-Durbin algorithm, and calculating the update direction in the linear prediction residual region. 3. The method according to claim 1, wherein the step of calculating the update direction in the linear prediction residual area includes a step of deriving a linear prediction residual signal calculated using a linear prediction coefficient. Method. 4. The method of claim 3, wherein said linear prediction residual signal is generated using a linear prediction error filter having a plurality of linear prediction coefficients. 5. The linear prediction residual signal is d (n). Where p is the order of the filter, x (n) and x (nk) are the current and past samples of the input signal, respectively, and {α _k } is the linear prediction coefficient of the linear prediction error filter. 5. The method according to claim 4, wherein
The described method. 6. The step of converting includes performing a filtering process in the updating direction in a finite impulse response filter,
4. The method of claim 3 including obtaining the update direction in the audio domain. 7. The method of claim 3, wherein the step of converting includes the step of converting the update direction to the audio domain using the linear prediction coefficients. 8. The step of transforming the update direction comprises updating the update direction g _n in the audio domain according to the following equation:
Calculating (i), Here, L is the length of the adaptive filter, p is the order of the linear prediction error filter, and Q (i) = [q ₀ (i), q ₁
(I),. . . , Q _L−1 (i)] ^T is the update direction in the linear prediction residual region, and G (i) = [g ₀ (i), g ₁
(I),. . . g _L-1 (i)] ^T is the voice region update direction, and {α _k } is a coefficient of the linear prediction error filter. 9. Estimating an echo estimate using the current coefficients of the adaptive filter and far-end speech samples; generating a linear prediction filter coefficient based on the far-end speech samples; Generating a linear prediction residual signal from the linear prediction residual signal; calculating the update direction for the echo canceller in the linear prediction residual region from the linear prediction residual signal; and, based on the current coefficient and the voice region update direction, 2. The method of claim 1, further comprising: generating update coefficients for the adaptive filter. 10. An echo canceller operating according to the method defined in claim 1. Description: