JPH0380370B2 - - Google Patents

Description

Claim 1: Means 116 for controllably adjusting an update gain in response to a tap gain control signal, and in response to a received signal [X, K] representing a first defined characteristic of the received signal. In an adaptive filter subject to external variations, the means 202 for being responsive to the received signal X, K comprises means 201 for generating a first normalized signal; Means 204 for generating a normalized signal and responsive to such first and second characteristics is adapted to generate the tap gain control signal with the greater of the first normalized signal or the second normalized signal. An adaptive filter that selects.

2. In the filter according to claim 1, further the first predetermined characteristic is a first predetermined power estimate, and the second predetermined characteristic is a second predetermined characteristic. An adaptive filter characterized in that it is a predetermined power estimate.

3. In the filter according to claim 2, the first power estimate is a long-term average power estimate, and the second power estimate is a fast attack peak power estimate. An adaptive filter featuring:

4. In the filter according to claim 3, the long-time power estimate (FIG. 3) is the average power of the received signal, and the fast attack power estimate (FIG. 4) is the average power of the received signal. Adaptive filter characterized by maximum power.

5. In the filter according to claim 4, the selection means 204 includes means for selecting the larger of the long-term power estimate and the fast attack power estimate as the effective normalized control signal. An adaptive filter comprising:

6. The adaptive filter according to claim 1, further comprising means (405 in FIG. 4) for scaling the second normalized signal by a predetermined value.

7. In the filter according to claim 6, furthermore, the selection means (204 in FIG. 2) selects the first
normalized signal or the scaled second
An adaptive filter characterized in that the larger one of the normalized signals of is selected as an effective normalized control signal.

8. The filter of claim 1, wherein the received signal includes a sequence of amplitude samples, and the first normalized signal generating means generates an indication of the average power of the magnitude of the predetermined number of received signal samples. means (FIG. 3) for generating a maximum power estimate of the magnitude of such predetermined number of received signal samples; An adaptive filter characterized by comprising:

9. In the filter according to claim 8, the display of the estimated maximum power value is further scaled by a predetermined value (404, 405 in FIG. 4), and the selection means (second Figure 20
4) An adaptive filter comprising means for selecting the greater of the average power (FIG. 3) or the modified maximum power (FIG. 4) as the effective normalized control signal.

Description The present invention relates to adaptive filters, and in particular to updating filter characteristics during reception of a predetermined signal.

The adaptive filter generates an output signal in response to a signal applied to it according to predetermined conditions. Typically, the filter generates a transfer function (impulse response characteristic) according to an algorithm that includes updating the transfer function characteristic in response to an error signal. In this way, the filter characteristics are optimized to produce the desired results.

It is known that it is desirable to normalize the update gain of an adaptive filter. Normalization makes the operation of the filter independent of variations in received signal power. IEEE Newsletter/Communication Magazine (IEEE
Transactions on CommunicationsVol Co −
26, No. 5, May 1978, pp. 647-653, D. L. Duttweiler “A Twelve-Channel Digital Echo Canceller”
Prior art devices use the average of the squared magnitudes of the input signal samples to normalize the gain, as described in the paper entitled ``Channel Digital Echo Canceler''. Using an estimate of the sum of squares, Another gain normalizer is shown in US Pat. No. 3,922,505, issued Nov. 25, 1975.

Although these conventional devices exhibit satisfactory characteristics in certain applications, they suffer from degraded or unstable performance when the received signal contains transient signals, rapidly pulsed signals, or the like. In telephone applications, instability occurs when echo cancellers handle things such as busy signals, pulsed signals such as telephone tones, and data set input and output signals. For transient signals, the average power estimate used in conventional devices is relatively small, so the update gain generated is too large, resulting in degraded performance. Even if the filter does not become unstable, at best it will have poor convergence and the converged characteristic will not be a good approximation of the desired characteristic. This result is undesirable.

According to the invention, the gain coefficient of the adaptive filter subjected to external perturbation is updated under the selective control of first and second normalization signals, the first normalization signal being a long-range of the past gain coefficient. A second normalized signal is generated from an estimate of the average power and a second normalized signal is generated from an estimate of the fast attack peak power of the past gain factor. Further in accordance with a feature of the present invention, the selector modifies the gain factor update accordingly by the greater of the average power or peak power estimates as a single tap gain control signal.

The present invention will be described below with reference to the accompanying drawings.

FIG. 1 is a simplified block diagram of an adaptive filter that is an embodiment of the present invention; FIG. 2 is a diagram showing simplified details of an embodiment of the power estimator used in FIG. 1; Figure 4 is a detailed diagram in the form of a simplified block diagram of the long-time estimator used in Figure 2; Figure 4 is a simplified block diagram of the fast attack estimator used in the power estimator of Figure 2; FIG.

An adaptive filter 100 that includes one embodiment of the invention is shown in simplified detailed diagram form in FIG. In general, adaptive filter 100 is similar to the adaptive filters used in echo cancellers shown in US Pat. Nos. 3,499,999 and 3,500,000. Briefly, adaptive filter 100 includes an adjustable signal processor with a closed loop error control system that is self-adaptive to automatically track variations in the signal in the output path. Specifically, filter 100 employs a system output estimator 101 that includes an adaptive transversal filter arrangement to synthesize a linear approximation of any desired system represented by block 102.

For this purpose, input signals X, K from the far end
is fed from the far end signal source through a first transmission path, eg, lead 103, to a first input of filter 100, and therein to the input of power estimator 104 and the input of arbitrary system output estimator 101.
The far end signals X, K may be, for example, digitally sampled audio signals, where K is an integer indicating the sampling period. far end signals X,K
is also provided to optional system 102 via lead 105, possibly a conversion circuit such as a digital-to-analog converter not shown. In an echo canceller application, optional system 102 includes a hybrid 106, a matched impedance 108, and a bidirectional transmission line 107. lead 105
It is usually desirable that the input signal from the listener to the hybrid 106 be provided to the near-end listener via a bidirectional transmission line 107. However, because of the impedance mismatch typically caused by the inability of balanced impedance 108 in hybrid 106 to precisely match the impedance of bidirectional transmission line 107, a portion of the hybrid's input signal appears on output lead 109. It is reflected back to the signal source at the far end as an echo signal. Similarly, any system 102 will produce a system output signal different from that produced by estimator 101 until the adaptive filter converges to the arbitrary system characteristics. Therefore, the output of arbitrary system 102 is equivalent to the echo signal of an echo canceller application. The output of optional system 102 is provided via lead 109 to the other input of filter 100 and therein to a first input of combinational network 110. Lead 109 may also include a conversion device, such as an analog-to-digital converter, not shown. The second input of combinational network 110 is the signal estimate of any system output signal produced by estimator 101. The output estimate of the arbitrary system is provided from the output of estimator 101 via lead 111 to a second input of combinational network 110. Combinatorial network 110 combines arbitrary system output estimates from estimator 101 and arbitrary system 102
The error signal E, corresponding to the algebraic difference of the outputs from
Generates K. Error signals E and K are on the second transmission line,
For example, it is supplied to the far end signal source and estimator 101 via lead 112.

Estimator 101 includes a so-called tapped delay line containing delay units 115-1 to 115-N-1 (shift registers) for achieving the desired delay in taps corresponding to a convenient Nyquist interval. Therefore, the delayed copies X,K-1 to X,K- of the far-end signals X,K entering the corresponding taps
N+1 occurs. The signals X, K-1 through X, K-N-1 and X, K at each tap position are adjusted in response to the error signals E, K. Specifically, signals X,K through X,K-N+1 are individually weighted in response to E,K via corresponding ones of conditioning circuitry 116-0 through 116-N-1, respectively. Adjustment circuit networks 116-0 to 116-N
-1 each includes multipliers 117 and 118 and a feedback loop 119. Feedback loop 119 adjusts the tap weights to the desired values in a manner well known to those skilled in the art and described in the references cited above. In this adjustment, it is important to properly adjust the individual loop gains to yield a suitable system. This is achieved by normalizing the loop gain by dividing the loop gain G by an estimate of a predetermined characteristic of the input signals X,K. In this example, input signal power estimates P, X are used as normalized control signals and are generated by power estimator 104. The normalized control signals P,X are provided to each controllable gain unit 121 of the adjustment network 116. Regarding this, the above-mentioned "12
See the paper entitled "Channel Digital Echo Canceller". Coordination Network 116-
The weighted copies of X,K from 0 to 116-N-1 are summed by summing circuitry 120 to generate an echo estimate signal that approximates the output from arbitrary system 102 or the echo to be canceled. Ru. The output estimate for any system is lead 11.
1 to the second input of the combinational network 110.

FIG. 2 shows, in accordance with one feature of the invention, first and second representations representing estimates of predetermined characteristics of the received signals X, K are generated and normalized according to predetermined conditions. One embodiment of a power estimator 104 that selects one of the estimates as control signals P, X is shown in simplified block diagram form. This example does not limit the scope of the invention, but
The predetermined characteristics of the signals X,K are their long-term power estimates PL,X and fast attack power estimates.
PS,X.

As mentioned above, under certain received signal conditions, the use of a long-term, or average, estimate of the received signal PL,X can produce undesirable results. Similarly, under certain received signal conditions, use of the fast attack estimate of received signal power PS,X may produce undesirable results. Specifically, when using approximation methods to generate fast attack power estimates, the crest or peak factor used is selected for the particular type of received signal, whether speech, sinusoidal, or noise. Using a crest or peak factor for voice means that this will not be the correct value when a sine wave or single frequency tone or noise is received. For example, the crest or peak factor of the voice is too large for a sine wave, and the normalized control signals P, X are too small, making the normalized update gain too large. If the update gain is large,
Filter performance tends to deteriorate. According to a feature of the invention, both the long-term power estimate and the fast attack power estimate are used, and one of the two is normalized according to predetermined conditions by the control signal P,
Choosing as X solves this problem. Specifically, the larger of the corrected long-term and fast attack estimates is P,
Selected as X. Therefore, the update gain G is always divided by the larger of the power estimates, which results in a smaller update gain and a more stable adaptive filter with higher quality convergence. Therefore, stability with respect to transient signals is improved while maintaining high quality of the other signals.

Accordingly, FIG. 2 shows a long-time estimator 20
1. Fast attack estimator 202, divider 20
3 and selector 204 are shown. As mentioned above, the predetermined input signal characteristic estimates used in this embodiment are the long-term and fast attack power estimates, PL,
PS,X. Long-term power estimate PL of X, K,
X is generated by long-term estimator 201 and fast attack power estimate PS,X is generated by fast attack estimator 202. Estimator 2
Details of the embodiments 01 and 202 are provided in the third section, respectively.
4 and described below.
The divider 203 calculates the estimated fast attack value PS,X
is corrected, that is, scaled by a predetermined value α, to obtain a modified version of PS,X, that is, PS′,
Used to generate X. The modification of PS,X ensures that high quality long-term estimates are used most of the time except when transient signal conditions exist. In other words, the fast attack estimate is corrected due to sudden changes in the received signals X and K.
The fast attack estimate is used only when PS',X exceeds the long-term power estimate PL,X. In this example, PS _,
X] by subtracting log ₂ α from it to produce PS′,X, where α is 6 dB. As is well known, subtraction of logarithms is the same as division, so a simple anti-logarithm operation yields PS',X.
The selector 204 selects the long-term estimated value PL, as the update gain normalization control signal according to predetermined conditions.
X or modified fast attack estimate
Used to select either PS or X.
In this example, the predetermined condition is to modify the fast attack estimate PS, It is to choose the one of value. The selection in this example is log ₂ PL,
X and log ₂ PS',X, and is realized in a well-known manner. The normalization control signal from selector 204 normalizes the update gain, i.e., G/P,
X is provided to each of the conditioning circuitry 116 to In one example, log _{2 P} ,
Produces P,X.

FIG. 3 shows in simplified form a first predetermined characteristic of
2 shows a simplified block diagram of 01. The input signals X, K are samples of the signal in digital form, each representing a μ-law quantized amplitude sample. Typically the sampling frequency is
At 8kHz, N=128. Each digital sample representation includes a sign bit, 3 segment bits and 4 step bits.

Here, it is desirable to generate a mean square estimate of the power of X and K, that is, PL, X=1/N _N-1 〓 ⁱ⁼⁰ X ² (K-i) (1). Here in this example N
=128. However, for this application, the geometric mean square estimate has been found to be sufficiently close to the desired long-term power estimate. i.e. becomes. Furthermore, addition is known to be easier to implement than multiplication or squaring. Furthermore, in μ-law digital sample representation, the segment bits approximate the base 2 logarithm of the corresponding sample size;
U.S. Patent No. 4,189,715 to Duttweiler, February 19, 1980
Please refer to

Therefore, as shown in FIG. 3, received signal samples X, K are provided to a segment extractor 301 to obtain segment bits of the received samples. This means that only 3 segment bits are connected to delay element 3.
02-1, then delay elements 302-2 to 302-
This is easily accomplished by using a corresponding plurality of gates that are energized to pass through (N-1). The delay element 302 has segment bits XL,K
may be a shift register for storing N delayed copies of XL(K-1) to XL(K-N+1). As mentioned above, segment bits
XL, K to XL(K-N+1) approximately represent the logarithm of the corresponding sample size.
Signals XL, K to XL (K-N+1) are sent to the adder 30
3. XL, K to XL (K-N+1)
Since is the base 2 logarithm of the magnitude of the corresponding sample, adder 303 generates a signal representing the logarithm of the product of magnitudes, i.e. log _2N-1 π ⁱ⁼⁰ X(K-i) (3) occurs. The output of adder 303 is then fed to an amplifier with a gain of 1/N, thereby giving 1/Nlog _2N-1 π ⁱ⁼⁰ X(K-i) or (4) log ₂ [ _N-1 π ⁱ⁼⁰ X(K-i)] ^1/N (5) is generated. Next, the output from the amplifier 304 is given to the amplifier 305 as follows: 2log ₂ [ _N-1 π ⁱ⁼⁰ X(K-i)] ^1/N (6) That is, the logarithm of PL, ₂ PL,X (7) is generated. In practice amplifiers 304 and 305 can be combined. log ₂ PL of amplifier 305,
The output of X is provided to selector 204 (FIG. 2).

In the above-mentioned paper entitled “12 Channel Digital Echo Canceller”, the long-term power estimate PL,
A similar device for generating X is described.

FIG. 4 shows a simplified embodiment of a fast attack estimator 202 for generating a second predetermined characteristic estimate of X, K, ie a so-called fast attack power estimate PS, The block diagram is shown below. The received signals X and K are each μ
contains samples of the signal in digital form representing quantized amplitude samples of the law. Each digital sample representation includes a sign bit, 3 segment bits and 4 step bits.

It is known that the peak power and the estimated mean square power are related by a so-called crest or peak factor γ. That is, P, XP, X, MAX/γ ² (8). It is also known that subtraction is easier than multiplication. Therefore, if the logarithms of P, X, MAX and γ ² are available, the desired mean square fast attack power estimate PS, As mentioned above, the segment bits of a μ-law digital sample are approximate representations of the base 2 logarithm of the corresponding sample magnitude.

Therefore, as shown in FIG. 4, the received signal samples X, K are provided to a segment extractor 401 to obtain the segment bits of the received samples. This means that only the 3 segment bits are sent to delay element 402-1, and then to delay element 402-1.
This is easily accomplished by using a corresponding plurality of gates that operate to provide signals 2 through 402-N. Delay element 402 is a segment bit.
N delayed copies of XS,K, i.e., XS,
A shift register that stores K to XS, (K-N+1) may be used. As mentioned above, segment bits
XS, K to XS, (K-N+1) are approximate expressions of the logarithm of the corresponding sample size.
The largest of the signals XS,K to XS,(K-N+1) is obtained by a maximum selector 403 in a well-known manner and is then fed via an amplifier 404 to the summing input of an adder 405. Ru. Amplifier 404 has a gain of 2, for example. Therefore, the output of amplifier 404 is 2log ₂ X,K,MAX or log ₂ P,X,
It represents MAX. Signal β is added to adder 40
5 subtraction input. The signal β is chosen to be 2log ₂ γ ² . Here γ is the crest or peak factor. In this example, the factor is chosen for the audio signal. Such crest or peak factor values are easily obtained by computer simulation in a well-known manner. Therefore, the output of adder 405 approximately represents log ₂ PS,X. where PS,X is the desired fast attack power estimate. The log ₂ PS,X signal is applied to one input of divider 203 (FIG. 2).