JP2005318571A - Echo canceler - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To solve a problem that it is necessary to identify a double talk and an echo path variation using a predetermined identification parameter in order to apply an echo canceler to a system wherein the double talk and the echo path variation occur, but an overlap that occurs in distribution given by the double talk and the echo path variation is too great in a conventional identification parameter, so that identification is delayed, the securing of an echo erasure quantity is delayed and operation of the echo canceler becomes unstable. <P>SOLUTION: In an adaptation algorithm of an echo canceler, a product of an error signal and an input signal to each of coefficient taps within a predetermined search range of an FIR filter 212 is calculated, and a representative tap is selected out of the coefficient taps within the search range based on a first value determined based on the product. It is then determined whether or not a double-talk state is generated based on a second value determined based on a product of a difference signal and an input signal to the representative tap, and a third value determined based on a power value of the difference signal. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention reduces the amount of echo cancellation even when double talk (a state where a near-end speaker and a far-end speaker speak simultaneously), which is known as one of the factors that degrade the performance (echo cancellation amount) of the echo canceller. The present invention relates to an echo canceller that maintains a predetermined value or more and quickly recovers an echo cancellation amount when the echo path changes.

One of the specific application fields to which the identification technique of double talk and echo path variation is applied is a hands-free communication device. This is an apparatus for making a call using a microphone and a speaker, and has an advantage that a call can be made without a handset (handset). FIG. 2 shows an example of a circuit that realizes the hands-free call by introducing the acoustic echo canceller 210. In FIG. 2, the adaptive filter 212 synthesizes a pseudo echo u _j approximated to an echo d _j that circulates from the speaker 201 to the microphone 200 using the far-end speaker voice x _j . Also, the filter coefficient

Is the output of the subtractor 211

Is updated to be minimal. In the above equation, I is the number of taps of the adaptive filter, and n _j is the near-end speaker voice. That is, the echo _dj is canceled as a result of the coefficient update, and the amount transmitted to the far-end speaker side is suppressed, thereby enabling a smooth conversation.

The problem is that the near-end speaker's voice n _j becomes the microphone output y _j

It is to mix as. The adaptive filter uses the echo d _j as a target and synthesizes a pseudo echo u _j approximated thereto. At that time, the near-end speaker's voice n _j acts as a disturbance to the synthesis. The effect is already the adaptive filter coefficient update algorithm, the most common learning identification method

It has been revealed in the case. Where ‖x _j ‖ ² is the square norm

And μ is a constant called step size.

The range is selected. In this learning identification method, the amount of echo cancellation is an estimation error, that is, the impulse response of the acoustic system from the speaker 201 to the microphone 200

Between the coefficient and the adaptive filter coefficient

And the magnitude of the estimation error Δ _j is expressed as its root mean square σ _d ²

It is formulated as As can be seen from this, the speech of the near-end speaker increases σ _n ^2, and as a result, the echo cancellation performance decreases due to the increased estimation error. Here, σ _x ² is the average power of the far-end speaker speech, and σ _n ² is the average power of the near-end speaker speech.

  In response to this decrease, equation (9) indicates that observation of the presence / absence of double talk and control to stop updating the coefficient of the adaptive filter or to reduce the step size when it is observed are effective. . The problem with that control is that the echo path, i.e., when the h changes, on the contrary, requires continued coefficient updating. As described above, the echo canceller requires reverse control with respect to the echo path fluctuation and the occurrence of double talk. Therefore, for stable operation of the hands-free communication device, it is indispensable to identify both without error and perform an operation suitable for each. The problem is how to identify it.

The existence of this double talk problem became known, and the first method examined was not the identification method but the double talk detection method (see, for example, Non-Patent Document 1). The reason is that it was simply considered that the coefficient update should be stopped only in the case of double talk with the coefficient update in a normal state. However, when the goal is to secure an echo cancellation amount of, for example, 30 dB, it is necessary to secure σ _d ² = 0.001 as an estimation error. This difficulty in securing is, for example, that the voice power of the near-end speaker to be detected is about 1/1000 of the voice power of the far-end talker when the step size μ is set to 1 as is apparent from Equation (9). It can also be seen from the fact that This fact means that the utterance of the near-end speaker buried in the echo should be detected. However, it is clear that the detection of this buried voice is actually difficult.

Therefore, a method for the generation of double talk then with an increase in the subtractor output e _j has been proposed (e.g., see Non-Patent Document 2). In this case, it has been subtracted echoes in a subtractor output e _j, therefore, the near-end speaker's voice buried in the echo can be detected. However, the subtractor output e _j increases like the double talk even echo path changes. Accordingly, in the method using the subtractor output e _j, which requires identification of the echo path change and double-talk occurs. That is, in order to maintain a high amount of echo cancellation in the double-talk is not forced to see the change in the subtractor output e _j from the need to detect the near-end speaker's speech buried in echo, further double the change In order to use for talk detection, an accurate discrimination method of echo path fluctuation and double talk occurrence is required.

This identification is equivalent to knowing the proportion of components constituting the microphone output y _j or the subtractor output e _j . For example, when (d _j −u _j ) increases in equation (2), it can be determined that the echo path changes, and when n _j increases, it is determined as double talk. The problem is how to correctly determine which of the two has increased.

Use of correlation can be considered as a method of distinguishing and knowing the increase. Furthermore, there are a far-end speaker voice x _j and a pseudo echo u _j as signals that can be used for the calculation of the correlation. For example, in the case of using the latter pseudo echo u _j , it can be assumed that the pseudo echo u _j and the near-end speaker voice n _j have a low correlation, and therefore, there is no fluctuation in the echo path, and the echo is sufficiently Offset

If the above relationship holds, the expected value for the occurrence of double talk

Becomes zero. That is, in double talk, the correlation e _j u _j gives a distribution centered on zero. here,

Represents an expected value operation. Conversely, since the unerased occurs for echo d _j when the echo path changes, the expected value

Is no longer zero. Thus, for example, considering a case where complete echo path change the pseudo echo u _j and echo d _j has no correlation at all occurs, a negative value of the average power of the correlation e _j u _j is pseudo echo in this case

Gives a distribution centered on. That is, if this principle is used, for example,

Than this,

When the value is large, it can be determined as double talk, and when it is small, it can be determined as echo path fluctuation.

The idea of using the correlation with this pseudo echo u _j is

The same applies to the above. That is, the correlation between the microphone output y _j and the output u _j of the adaptive filter

On the other hand, when the pseudo echo u _j closely simulates the echo d _j

Is the average power of the pseudo echo

Since the correlation between the pseudo echo u _j and the echo d _j is low when the echo path changes,

Is distributed around zero. Therefore,

Is the threshold

It can be determined that the echo path is changed when exceeding the value, and double talk when the value is below. Further, the correlations y _j u _j and e _j u _j may be values normalized by the power of u _j and y _j as parameters used for the determination. A detailed description of this identification method can also be found in conventional literature (see, for example, Non-Patent Document 3).

As an example showing that this correlation is effective for identification, FIG. 3 shows the correlation between the subtractor output e _j and the pseudo echo u _j , which is said to have a relatively high separation performance in Non-Patent Document 3, as the microphone output y. Normalize with _j power

Shows the distribution calculated for echo path variation and double talk. However, the following conditions are set for this calculation.
(1) M = 128
(2) Sampling frequency: 8 kHz
(3) The voice shown in FIG. 4 is used as the far-end speaker voice and the near-end talker voice.
(4) The coefficient of the adaptive filter is fixed when the estimation error becomes -30 dB to generate double talk.
(5) The impulse response of the unknown system is changed to a value at which the estimation error becomes 0 dB, and the echo path variation is set.
(6) The tap length of the adaptive filter and the impulse response length of the unknown system are equal 512.
(7) The impulse response of an unknown system is given as a normal random number that exponentially decays.
(8) Average power ratio of far-end speaker voice to ambient noise and near-end talker voice to ambient noise: 30 dB
From the results shown in FIG. 3, it can be seen that there is a large overlap in the distribution given by the double talk and the echo path variation. The overlap means that an identification error occurs. Therefore, in order to reduce this error, it is necessary to reduce the variance of each distribution. For this purpose, it is necessary to increase the number of samples (M). However, this increases the identification delay, which delays securing the amount of echo cancellation and destabilizes the operation of the echo canceller.

Another problem in the method using the correlation between the subtracter output e _j and the pseudo echo u _{j is that} the correlation between y _j and u _j becomes strong when the fluctuation of the echo path is small, and the distribution and echo path given by double talk. This is because the distribution given by the fluctuations approaches, making identification more difficult. In order to accurately determine this small echo path variation, a larger number of samples is required. That means that the response to echo path variations is further delayed. In addition, the correlation between the pseudo echo u _j and the near-end speaker's voice n _j is assumed to be zero. However, even if different speakers utter, the phonemes may coincide, so there is actually a correlation. In this case, even if M is increased, the distribution overlap is not easily reduced.

As another method of using the correlation with the subtracter output e _j , there is a method using the correlation with the far-end speaker voice x _j (see, for example, Non-Patent Document 4), and the effectiveness is referred to white noise. It has been confirmed that it is a signal. However, after all the correlation between the previous subtractor output e _j and the pseudo echo u _j utilizes the correlation between the subtractor output e _j and the far-end speaker's speech x _j. Therefore, also in this method, there is a similar problem to the method of using the correlation of the subtractor output e _j and the pseudo echo u _j.

  Then, next, a method that can deal with echo path fluctuation and double talk without distinguishing them was examined (for example, see Non-Patent Document 5). In this method, an adaptive filter coefficient update algorithm is used as a block-execution learning identification method.

And constantly updating the coefficient by absorbing the power fluctuation of the far-end speaker voice

To calculate the step size ρ _n

As described above, a method of making it variable in units of blocks has been proposed. Here, a block execution type adaptive algorithm will be described. The block execution type adaptive algorithm is an update algorithm that cumulatively adds a numerator vector of an update vector and a normalized denominator in a time block. In other words, in equation (4), the second term on the right-hand side is the part of the update vector, but the numerator (numerator vector) and denominator (normalized denominator) of the second term are cumulatively added in the time block. Equation (14) is modified to the execution type. Where J is the block length,

Is a constant previously calculated using

It is extended until it becomes above, and the coefficient of the adaptive filter is updated when P _n ≧ P ₀ . Furthermore, ρ ₀ is the maximum step size set in this control, C ₀ is an estimation error corresponding to the amount of echo cancellation to be secured, and Q ₀ is the average power of steady ambient noise. That is, when double talk occurs, Q _n increases, the step size decreases according to equation (16), and the estimation error is maintained at a predetermined size.

  Here, Expression (16) is a control expression when the far-end speaker voice is white noise. When using speech with strong autocorrelation as a reference signal

Then, the operation becomes more stable when a = 10 and a small step size is given (for example, see Non-Patent Document 6).

According to this step size control, it has been confirmed that it is not necessary to distinguish from double talk for small echo path fluctuations in which the estimation error increases to about −20 dB (see Non-Patent Document 5). That is, if this control is adopted, it is only necessary to find a method for speeding up the identification with respect to echo path fluctuations in which the estimation error increases beyond -20 dB.
Seiji Kuruyama, Junzo Tamura, Seiichi Yamamoto, Hikoichi Ishigami, “Multiple Echo Canceller with Common Adaptive Control Unit”, IEICE Technical Report, CS78-23 (1978-05) Furuya Seiji, Fukushi Yuzo, Ito Eiki, Tanabe Junji, Sugawara Yukio, "A Method and Experimental Results of Superimposed Call Detection in Adaptive Echo Canceller," 1984 IEICE General Conference, 2343 (1984-03) Kensaku Fujii, Toshiro Oga, “Identification of Double Talk and Echo Path Fluctuation”, IEICE Technical Report, EA94-15 (1994-05) Kensaku Fujii, Toshiro Oga, “Reduction of convergence time by block length control in additive normalization LMS method”, Japanese Institute of Electronics, Information and Communication Engineers (A), vol. J80-A, no. 1, pp. 27-35 (1997-01) Kensaku Fujii, Toshiro Oga, “Method of keeping the estimation error for acoustic echo canceller to the required value”, IEICE Japanese Journal (A), vol. J83-A, no. 02, pp. 141-151 (2000-02) Zhang Wakao, Kensaku Fujii, Meiji Muneyasu, “A Study on Double Talk and Echo Path Change Detection”, 18th Digital Signal Processing Symposium, B3-6 (2003-11)

  It is an object of the present invention to quickly and accurately discriminate between echo path variations and double talk where this estimation error increases beyond -20 dB, speed up the convergence of the coefficients for echo path variations, Is to realize a device capable of maintaining a stable echo cancellation amount.

  In order to solve the above problems, an echo canceller according to the present invention is an echo canceller that removes an echo signal generated when a reception signal supplied to a speaker wraps around from the speaker to the microphone from the output signal of the microphone. A coefficient updateable FIR filter that generates a pseudo echo signal based on the received signal, and a subtractor that generates an error signal by subtracting the pseudo echo signal generated by the FIR filter from the output signal of the microphone; A coefficient updating unit that updates a coefficient of the FIR filter by an adaptive algorithm so that the error signal is minimized based on the received signal and the error signal, and the adaptive algorithm includes the error signal and the error signal Calculate the product of the input signal to each coefficient tap within the predetermined search range of the FIR filter A representative tap is selected from coefficient taps within the search range based on a first value determined based on the product, and a first value determined based on a product of the error signal and an input signal to the representative tap. Based on the binary value and a third value determined based on the power value of the error signal, it is determined whether or not a double talk state has occurred.

  In the echo canceller, the first value is a product of the error signal and the input signal to each coefficient tap, and the second value is a product of the error signal and the input signal to the representative tap. Yes, the third value may be the power value of the error signal.

  In the echo canceller, the adaptive algorithm is a block execution type adaptive algorithm, and the first value is a cumulative sum in a time block of a product of the error signal and an input signal to each coefficient tap, The second value is a cumulative sum in the time block of the product of the error signal and the input signal to the representative tap, and the third value is a cumulative sum in the time block of the power value of the error signal. Also good.

  In the echo canceller, it may be determined that a double talk state has occurred when the absolute value of the ratio of the second value to the third value is less than a predetermined threshold.

  Further, in the echo canceller, the absolute value of the ratio of the second value to the third value is equal to or greater than the predetermined threshold value, and the second consecutive second for the representative tap selected this time or before. When the signs of the values are different, it may be determined that a double talk state has occurred.

  In the echo canceller, the absolute value of the ratio of the second value to the third value is equal to or greater than the predetermined threshold value, and the current input signal to the representative tap selected this time and the current error signal When the sign of the second value determined based on the product and the sign of the second value determined based on the product of the previous input signal to the representative tap selected this time and the previous error signal are different, It may be determined that a talk state has occurred.

  In the echo canceller, the absolute value of the ratio of the second value to the third value is equal to or greater than the predetermined threshold value, and the current input signal to the representative tap selected last time and the current error signal When the sign of the second value determined based on the product is different from the sign of the second value determined based on the product of the previous input signal to the representative tap selected last time and the previous error signal, It may be determined that a talk state has occurred.

  In the echo canceller, when the third value is equal to or less than a predetermined value, the predetermined threshold value may be corrected to a larger value.

  In order to solve the above problem, another echo canceller according to the present invention provides an echo signal generated when a reception signal supplied to a speaker circulates from the speaker to a microphone from an output signal of the microphone. An echo canceller that eliminates the error signal by subtracting the FIR filter capable of updating the coefficient based on the received signal and the coefficient-updatable FIR filter that generates the pseudo echo signal from the output signal of the microphone. A subtractor for generating, and a coefficient updating unit for updating a coefficient of the FIR filter by an adaptive algorithm so that the error signal is minimized based on the received signal and the error signal, the adaptive algorithm comprising: The error signal and the input signal to each coefficient tap within the predetermined search range of the FIR filter And a representative tap is selected from among the coefficient taps within the search range based on a first value determined based on the product, and continuous for the representative taps selected this time or before When the two values of the second value have different signs, it is determined that a double talk state has occurred, and the second value is determined based on the product of the error signal and the input signal to the representative tap. It may be a value.

  In the echo canceller, the sign of the second value determined based on the product of the current input signal and the current error signal to the representative tap selected this time and the previous input to the representative tap selected this time When the sign of the second value determined based on the product of the signal and the previous error signal is different, it may be determined that the double talk state has occurred.

  In the echo canceller, the sign of the second value determined based on the product of the current input signal and the current error signal to the representative tap selected last time, and the previous input to the representative tap selected last time When the sign of the second value determined based on the product of the signal and the previous error signal is different, it may be determined that the double talk state has occurred.

  In the echo canceller, the first value is a product of the error signal and the input signal to each coefficient tap, and the second value is a product of the error signal and the input signal to the representative tap. There may be.

  In the echo canceller, the adaptive algorithm is a block execution type adaptive algorithm, and the first value is a cumulative sum in a time block of a product of the error signal and an input signal to each coefficient tap, The second value may be a cumulative sum in a time block of a product of the error signal and an input signal to the representative tap.

  In the echo canceller, a coefficient tap having the largest absolute value of the first value may be selected as a representative tap among coefficient taps within the search range.

  In the echo canceller, the search range may be all coefficient taps of the FIR filter.

  In the echo canceller, the search range may be a part of coefficient taps of the FIR filter.

  In the echo canceller, the search range may be a continuous coefficient tap from the first stage to the intermediate stage of the FIR filter.

  In the echo canceller, when the adaptive algorithm determines that the double talk state has occurred, the update amount of the coefficient of the FIR filter is reduced or the update of the coefficient is stopped. Good.

  In the echo canceller, the adaptive algorithm may determine that an echo path variation state has occurred when it is not determined that a double talk state has occurred for a predetermined number of times. Also good.

  In the echo canceller, the adaptive algorithm may increase the coefficient update amount of the FIR filter when it is determined that an echo path fluctuation state has occurred.

  According to the echo canceller of the present application, double talk and echo path fluctuation are identified with a small delay, and a stable operation is guaranteed for the system.

  Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

  FIG. 1A is a diagram showing a schematic configuration of the present invention (echo canceller) according to this embodiment applied to a communication device, and FIG. 1B shows each function of the present invention (echo canceller) according to this embodiment. FIG.

  Referring to FIG. 1A, the call device has a microphone 200 and a speaker 201. The echo canceller mainly includes an FIR filter 212a capable of updating coefficients, a coefficient updating unit 212b, and a subtractor 211. The adaptive filter 212 is configured by the FIR filter 212a and the coefficient updating unit 212b.

A signal x _j (received signal) received by the call device is supplied to the speaker 201. Sound radiated from the speaker 201 goes around the microphone 200. The output signal y _j of the microphone 200 includes an echo signal generated by this wraparound.

The signal x _j is also input to the FIR filter 212a. The FIR filter 212a includes I coefficient taps, and H (1), H (2), H (3),..., H (I) are set as coefficients for each coefficient tap. Yes. From the FIR filter 212a, a signal u _j is output as an output signal.

The subtractor 211 receives the signal y _j and the signal u _j . The subtractor 211 subtracts the signal u _j from the signal y _j . Signal e _j as the subtraction result is sent from the communication device as a transmission signal.

The coefficient updating unit 212b, the signal e _j is inputted as an error signal. A signal x _j is also input to the coefficient updating unit 212b. In the coefficient update unit 212b, an adaptive algorithm (coefficient update algorithm) is operating. Adaptive algorithm of the coefficient update unit 212b (the coefficient update algorithm), based on the signal x _j and the signal e _j, coefficients of the coefficient taps of the FIR filter so that the signal e _j is minimized H (1), H (2 ), H (3),..., H (I) are updated.

  In this embodiment as well, the block execution type learning identification method is used as the coefficient update algorithm, and the above step size control is applied to the echo path variation in which the increase in the estimation error is -20 dB or less. In the present embodiment, an echo path variation in which an increase in estimation error exceeds −20 dB is identified as the occurrence of double talk, and a large step size is provided to quickly secure an echo cancellation amount.

In order to simplify the description, the equation (14) is calculated for each element of m = 1,..., I of the coefficient H _n of the adaptive filter.

Change to The numerator of the second term on the right side of the above equation (20) is the cumulative sum in the time block of the product of the error signal e _j and the input signal x _j (m) to the m-th coefficient tap of the FIR filter 212a. It is. In this embodiment, the molecule of the second term

Pay attention to. And the estimation error in the nth block is

If the maximum value of A _n (m) given in equation (21) is obtained at m = k, the k-th coefficient tap is selected as the representative tap. Its maximum value A _n (k) is

When the echo path changes,

Is obtained as a mean distribution. Furthermore, during double talk

In this case, A _n (k) gives a distribution centered on zero. Accordingly, the if A _n (k) of a parameter is expected to be able to identify double-talk and echo path change the ^{_{Δ n (k) Jσ x 2}} /2 as a threshold value. However, the search range of A _n (k) is k = 1 to I / 16 for the reason described later. This search range may be all or a part of the I coefficient taps of the FIR filter, but a part of the search range is used here. Here, the search range is a continuous coefficient tap from the first stage (1 stage) to the intermediate stage ((I / 16) stage) of the FIR filter.

FIG. 5 shows that A _n (k) is a function of the power of the far-end speaker voice.

Divided by

(However, maxA _n shown in the figure is A _n (k) in the text. Obviously, the distribution of a _n (k) is expressed as Δ _n (k) when the echo path changes. average value, when the double-talk has given distribution to average zero. However, similar to the distribution also Figure 3 of FIG. 5, the overlap occurs in the distribution. Therefore, the a _n (k) It can be seen that it is difficult to identify a simple distribution by 2 minutes, where the distribution shown in FIG.

This is a result of calculating a _n (k) by extending the block length J until. Further, in the simulation of FIG. 5, in addition to the conditions of FIG. 3, (11) average power of stationary ambient noise: Q ₀ = 0.001
(12) Minimum block length: J ₀ = 128
(13) Target estimation error: −30 dB, that is, C ₀ = 0.001
(14) a = 10.0
(15) ρ ₀ = 12.8
(16) P ₀ = 64I
Is set.

In contrast to the method using a _n (k) as an identification parameter, the simplest method for reducing the variance is to increase the block length J. However, in that case, there arises a problem that the identification delay increases. Moreover, the extension also causes a delay in convergence in the block execution type learning identification method in which the reference signal is a speech having a strong autocorrelation. That is, even in the method using A _n (k), a countermeasure for reducing the overlap without increasing the block length J is required.

Therefore, in double talk, a _n (k) becomes large and overlaps with the distribution of echo path fluctuation when the variance of A _n (k) becomes large, and further, the variance becomes large in the near end. Note that this is when the power of the speaker voice increases. In this case, from equation (15)

As can be seen, Q _n also increases simultaneously. On the other hand, Q _n increases due to echo path variation when a large echo path variation occurs, and at that time, A _n (k) also increases simultaneously. That is, when the distribution of the distribution at the time of double talk becomes large and the distribution of FIG. 5 spreads in the right direction, Q _n becomes large. At that time, the average value of the distribution of echo path fluctuations also becomes large. The distribution moves to the right. Therefore, if the distribution of echo path fluctuations is moved to the right in accordance with this Q _n , it can be expected that the overlap when the double distribution occurs and the distribution spreads will be reduced. This is equivalent to expressing the distribution of A _n (k) as a two-dimensional distribution by adding Q _n to the parameter.

FIG. 6 is a distribution obtained when the echo path changes, with the horizontal axis representing JQ _n and the vertical axis representing A _n (k). However, the echo path variation is given here as a variation in which the estimation error is 0 dB. In this case, the expected value of JQ _{n on} the horizontal axis is

It becomes. From this, JQ _n has a magnitude proportional to the sum of the product of the mean square σ _d ² of the estimation error and the average power σ _x ² of the far-end speaker voice and the average power σ _n ² of the near-end talker voice. Understand. Therefore, if this was the horizontal axis, when the increase in e _j is happening only echo path change is

Can be approximated with

Is

The straight line with the slope is distributed as an average. This can also be confirmed from the distribution shown in FIG. However, in FIG. 6, since A _n (k) has no meaning in the positive and negative, A _n (k) is displayed as an absolute value. Further, the conditions for calculating this distribution are the same as those in FIG. 5, and only the display method is different.

Molecule on the right side of the above equation (29) corresponds to the cumulative sum at time block of the product of the input signal x _j to representative taps and the error signal e _j. Further, the denominator of the right side of the equation (29), the square value of the error signal e _j, i.e., corresponding to the accumulated sum at the time the block of the power value of the error signal e _j (see equation (15)).

Here, the effect of JQ _n on the distribution of FIG. 6 will be examined in detail. First, when P ₀ and the maximum step size ρ ₀ are set so that the block length is the shortest J ₀ when the power of the reference signal is maximum and a predetermined estimation error is ensured, P _n exceeds P ₀ . Every time, the coefficient is updated with P _n = P ₀ . That is, in the case of an echo path variation where the estimation error is 0 dB, σ _d ² = 1 (in this simulation, the power gain of the unknown system is set to 1), so JQ _n is

It will be concentrated in the vicinity. However, when the shortest block length is selected in this way, the frequency with which the block length becomes J ₀ or more increases, and the convergence of the coefficients is delayed when speech with strong autocorrelation is used as a reference signal. If the reference signal has a strong autocorrelation, the maximum value of the step size ρ ₀ is set small so that a predetermined estimation error is secured even with a small P ₀ , that is, a short block, and the update frequency is increased. You should do it. In the simulation of FIG. 6, taking this point into consideration, the maximum step size ρ ₀ is selected to be small so that the coefficient update frequency is high, and P ₀ = 64I is set as previously shown as the simulation condition (16). . In this case, when the power of the reference signal is minimum

Becomes the minimum 64, and when the power of the reference signal is larger than this, JQ _n > 64. FIG. 7 shows changes in P _n that occur in the distribution calculation of FIG. Apparently, the minimum value of P _n shown in FIG. 7 is P ₀ , and the frequency of P ₀ is low. The distribution of FIG. 6 is a distribution reflecting the result of FIG.

  On the other hand, when double talk occurs alone, in equation (21)

Therefore, A _n (k) has a distribution centered on zero. Further, because the dispersion is proportional to the power of the near-end talker speech, a distribution as shown in FIG. 8 spread with increasing JQ _n. From the distributions shown in FIG. 8 and FIG. 6, it can be seen that the separation is well performed in the occurrence of strong double talk in which the power of the near-end speaker voice increases. Further, as compared with the distribution of FIG. 5, it can be seen that in FIG. 5, the disadvantage that the distribution greatly overlaps in strong double talk is separated and solved by using a two-dimensional distribution. FIG. 9 shows the distribution of FIG. 8 and FIG. 6 together.

From FIG. 6, it can be seen that the event of r _n <0.6 does not occur in the echo path fluctuation, and therefore it can be determined that double talk is possible when r _n <0.6. Of course, it cannot be said that there is no possibility of echo path fluctuations where r _n <0.6. However, even if it is determined that there is a double talk, there is no problem because there is a high probability that it is determined that the echo path changes in the next block. The problem is that, as can be seen from FIG. 8, even in the case of double talk, there are cases where r _n ≧ 0.6, and when r _n = 0.6 is determined as a threshold, double talk may be mistaken for echo path fluctuation. It is. When this error occurs, the step size is set large, and the estimation error increases rapidly.

FIG. 10 shows the result of calculating the relationship between JQ _n and the probability that the double talk is r _n ≧ 0.6 in the distribution of FIG. Obviously, the probability that the determination error occurs is higher as JQ _n is smaller, and lower as JQ _n is larger. This can be predicted from the distribution of FIG. However, a determination error in which the echo path variation is double talk does not occur. That is, according to this identification method, the identification error is reduced by a large double talk that is difficult to separate from the echo path fluctuation in the one-dimensional distribution of FIG.

Further, in this result, when JQ _n = 0 to 50, double talk of 80% or more is determined as an echo path fluctuation. The reason why the rate of occurrence of determination errors in this way is high is that the average of the minimum values of JQ _n is 64 in the echo path variation where the estimation error, which is the simulation condition, is 0 dB. In this case, it is very unlikely that JQ _n <50 in the echo path variation. This can also be seen from FIG. That is, most of them are double talk when JQ _n <50. In any case, with a simple method using r _n = 0.6 as a determination threshold, there is a high possibility that double talk will be mistaken for echo path fluctuation.

As a measure to reduce this error, in this embodiment, the reverberation of the room is first concentrated (Shinji Makino, Nobuo Koizumi, “Improvement of adaptive characteristics of echo canceller in room sound field”, Electronic Information Communication) The search range of A _n (k) is set to k = 1 to I / 16 in consideration of the Japanese academic journal (A), vol. J71-A, no. 12, pp. 2212-2214 (1988-012)). And As a result, a reduction in the search processing amount can be obtained at the same time. That is, in the echo path fluctuation, the probability that the maximum is within this range is high, and in double talk, it appears equally in the entire range. Therefore, if the search is limited to this range, the possibility that A _n (k) becomes small in double talk increases, and the probability of r _n > 0.6 decreases. 11, the double-talk is the relationship probability and JQ _n is determined that echo path change when the narrowed thus search range. It turns out that it has decreased compared with FIG.

Next, attention is paid to the sign change of A _n (k). In other words, A _n (k) is

As can be seen from the distribution as an average, it is highly possible that A _n (k) has the same sign in adjacent blocks in the echo path variation. Conversely, in double talk, A _n (k) is expressed by equation (23).

As can be seen from the result of applying, the correlation between adjacent blocks is low, and there is a high possibility that codes are different in consecutive blocks. Therefore, if the code is different, it can be determined as double talk.

In the above, the sign of A _n of elected representatives taps in the time block of the current minute (k) depending on whether different or not between this date and the previous minute, an example of determining double talk. However, the determination may be made based on whether or not the sign of the representative tap An _n (k) selected in the previous time block is different between the current time and the previous time. Alternatively, the determination may be made based on whether or not the sign of the representative tap An _n (k) selected before that differs between the current time and the previous time. Furthermore, double talk may be determined based on whether or not the past two consecutive times of A _n (k) have different signs for representative taps selected this time or before.

In the above, an example in which double talk is determined based on whether or not the two consecutive A _n (k) codes are different when r _n is 0.6 or more is shown. However, regardless of what value r _n takes, double talk may be determined only by whether the signs of two consecutive A _n (k) are different.

In the above description, it has been described that if the code of A _n (k) is a continuous block and different code, it can be determined as double talk. However, it cannot be denied that there is a possibility of different signs due to the influence of the second term of equation (23) even if the echo path changes. However, even if the echo path variation is mistaken as double talk by the determination, there is a high possibility that the next block is determined as the echo path variation. In fact, FIG. 12 shows that the tap k _n-1 given the maximum value A _n-1 (k _n-1 ) observed in the previous block is the same as A _n (k _n-1 ) in the current block. This is a result of examining whether the code is given or a different code is given in the case of echo path fluctuation and double talk. The upper part of FIG. 12 shows an echo path fluctuation where the estimation error is 0 dB, the middle stage is an echo path fluctuation where the error is −10 dB, and the lower part is a double talk, where 1 indicates a code match and −1 indicates a code mismatch. ing. In FIG. 12, the probability that the codes do not match due to the echo path variation where the estimation error is 0 dB is 1.5%, and the probability that the estimation error is −10 dB is 5%. Further, as can be seen from this result, it is unlikely that two blocks continue to have different codes when the echo path changes, and in many cases the same code is used in the next block, and it is determined that the echo path changes. In addition, even if the echo path variation is mistaken as double talk, coefficient update as double talk is executed, and if it is determined that the echo path variation is correct in the next block, the processing as echo path variation is one block. Just a minute delay will not cause any major problems.

  On the other hand, FIG. 12 also shows that the probability of different signs increases as the increase in estimation error decreases. Therefore, it is also an option not to use a sign change for an echo path variation in which the increase in estimation error is less than −10 dB. The determination of whether or not the increase in the estimation error is less than −10 dB can be known from the step size. That is, the equation (19)

And substituting C ₀ = 0.001 and a = 10.0

Is obtained. However, since the case of an echo path fluctuation in which the estimation error is −10 dB is considered, σ _d ² = 0.1, and Q _n is σ _n ² << σ _x ² σ because the residual echo is dominant. _d ^2, the power sigma _x ² of the far-end speaker's speech is assumed to give in terms of the block length in the shortest J ₀ = 128. From this equation (32), it can be seen that when the estimation error is less than −10 dB, the step size ρ is larger than 0.256. Here, if the change in the sign is used for the determination with respect to the echo path variation in which the estimation error is −10 dB or more, the sign determination is performed when ρ ≦ 0.256. That is,
(1) The search for A _n (k) is limited to k = 1 to I / 16, and if r _n <0.6, it is determined as double talk.
(2) If ρ ≦ 0.256, the code is determined, and when A _n (k _n−1 ) of the current block is different from the maximum tap k _n−1 in the previous block, double Judge as talk.
(3) If ρ> 0.256, no sign determination is performed.

  FIG. 13 shows the probability that an error (an error that determines double talk as an echo path variation) occurs when these two methods are used together. It can be seen that the probability further decreases.

Next, consider that to observe this r _n for consecutive K blocks. That is,
(4) The above procedure is repeated K times, and when all are not determined as double talk, it is determined as echo path fluctuation.

The probability of occurrence of an error in determining double talk as an echo path variation when this block alone is determined is as shown in FIG. Now, if the occurrence probability is p, the probability that the error occurs K times consecutively is p ^k . Therefore, if the probability of occurrence is p ₀ , the number of blocks K to be continuously observed is

Than

Is calculated.

FIG. 14 shows the number of blocks K calculated using this relationship when the error occurrence probability is p ₀ = 10 ⁻⁴ . From this result, it can be seen that for JQ _n ≧ 50, the echo path variation is detected with an error rate of 10 ⁻⁴ until 5 consecutive times. The detection delay is 128 × 5/8 = 80 ms in time. On the other hand, FIG. 14 shows that JQ _n <50 requires 10 times continuous observation. However, this JQ _n <50 occurs when the power of the reference signal is minimum as described above, and the frequency at which the minimum power is generated is very low as can be seen from FIG. 7, and therefore JQ _n <50. Is unlikely to occur continuously many times, and in fact, JQ _n > 50 is likely to be detected with a small number of consecutive times. In addition, even if it can be detected that there is an echo path fluctuation when the power of the reference signal is minimum, the estimation error is reduced little, and the echo transmitted to the far end side also has a small power. There is little need for quick detection. In this case, the determination delay is large, and there are few problems even if it is processed as double talk.

Next, FIG. 15 and FIG. 16 show the distribution of _An (k) when the echo path variation has an estimation error of −10 dB and −20 dB. This magnitude of the echo path changes, as can be seen in the formula (24) from equation (29), the molecular tilt r _n 1 square of the estimation error, since the denominator is given by the square, a small echo of estimation error inclination of about path changes r _n increases. For example, the echo path fluctuation in which the estimation error in FIG. 15 is increased by −10 dB is r _n > 1.8, and at −20 dB in FIG. 16, r _n > 5.33. That is, in all cases, the echo path variation does not satisfy r _n <0.6. Therefore, if r _n <0.6, it is possible to determine double talk regardless of the magnitude of the echo path fluctuation.

The problem is that the distribution in the case of double talk has nothing to do with the magnitude of the echo path variation. Further, JQ _n is concentrated to 50 or less when the echo path variation is such that the estimation error is −10 dB and −20 dB. Therefore, if the determination threshold is 0.6, the number of blocks necessary for determination is 10. Become. In an actual system, it is expected that the frequency of occurrence of echo path fluctuations with estimation errors of −10 dB and −20 dB is expected to be high in this way. Therefore, a device for reducing the determination delay in this case is necessary.

Therefore, since σ _d ² becomes 1/10 when the echo path variation has an estimation error of −10 dB, attention is paid to JQ _n <200 or less. This is also confirmed from FIG. Further, if JQ _n ≧ 200 or more, the number of necessary continuous blocks is three even if the threshold value is 0.6. Therefore, here, JQ _n <200 or less may be considered as a countermeasure for reducing the number of blocks required when an echo path variation with an estimation error of −10 dB or less occurs. Of course, in this case as well, there is a possibility that an echo path variation with an estimation error of 0 dB has occurred. However, in that case, the above measures are sufficient, and no other measures are required.

Therefore, in order to find the countermeasure, the number of continuous blocks required when the threshold value is set to r _n = 1.8 for an echo path variation with an estimation error of −10 dB is calculated. The result is shown in FIG. From this result, the number of continuous blocks required for 100 ≦ JQ _n is zero. This indicates that the probability of occurrence of double talk _where 100 ≦ JQ _n <200 and r _n > 1.8 is 10 ⁻⁴ or less (note that double talk is not related to the magnitude of echo path fluctuations). ). This can also be confirmed from FIG. That is,
(5) If JQ _n ≧ 100 when r _n > 1.8 is calculated, it can be immediately determined as double talk.

Similarly, from FIG. 17, the number of consecutive blocks required when (6) r _n > 1.8 and 50 ≦ JQ _n <100 is three.
(7) The number of consecutive blocks required when r _n > 1.8 and 10 ≦ JQ _n <50 is 5.
(8) The number of continuous blocks required when r _n > 1.8 and JQ _n <10 is 8.
Can be identified. That is, it is possible to reduce the number of consecutive blocks required for determination by raising the threshold to 1.8 when exceeding r _n of 1.8 in the current block by JQ _n <200.

Similarly, in the case of JQ _n <20, the threshold value can be increased to 5.33 from FIG. However, in the echo path variation where the estimation error is −20 dB, it is not necessary to distinguish between double talk and echo path variation by the step size control, and it is not necessary to change the threshold value substantially.

22A and 22B are control flows when the present invention is applied to a practical apparatus. Here, k ₁ to k ₆ are parameters for counting the number of continuous blocks, and are set to 0 when r _n <0.6 and at the initial stage of control. A _n (m) is the same as A _k in the text, and m in this control flow is the same as k in the text. What should be noted in this control flow is that, for example, when counting the number of consecutive blocks, r _n ≦ 1.8 and 200 ≦ JQ _n <700 include r _n ≦ 1.8 and JQ _n <200. Is a point. Therefore, when r _n ≦ 1.8 and 200 ≦ JQ _n <700, and then r _n ≦ 1.8 and JQ _n <200, the number of consecutive blocks k _{2 in} this case is 2 It is counted. In addition, when JQ _n <10 and r _n > 5.33, it indicates that the estimation error is −20 dB or less, and thereafter it is no longer necessary to control the echo path variation and double talk separately. Represents. Therefore, in this flow, control for setting the counting parameter to 0 is incorporated.

  FIG. 18 shows a decrease characteristic of ERLE obtained by detecting the occurrence of echo path fluctuation with an estimation error of 0 dB according to the present invention. From this result, it can be seen that the echo rapidly decreases after the occurrence of the echo path fluctuation according to the present invention. FIG. 19 shows the result of an ideal state in which the echo path variation is detected with a delay of zero. From this comparison, the effectiveness of the present invention can be confirmed. Furthermore, FIG. 20 shows a reduction characteristic of the estimation error. Here, the broken line is the result of an ideal detection with a detection delay of zero. Although the influence of the detection delay of 75 ms according to the present invention remains, the estimation error decreases at the same speed as in the ideal case. It is confirmed that On the other hand, FIG. 21 shows ERLE given by the present invention in the case of double talk. Thus, it can be seen that ERLE is stably maintained in double talk.

  As described above, according to the present invention, double talk and echo path fluctuation are identified with a small delay, and a stable operation is guaranteed for the system. Here, the description of the present invention has been made using the adaptive algorithm as a block execution type learning identification method. However, it should be noted that the present invention can be applied to adaptive algorithms other than the block execution type learning identification method.

  For example, a learning identification method that is not a block execution type can be used, or an adaptive algorithm other than the learning identification method can be used. Furthermore, an adaptive algorithm other than the learning identification method, which is a block execution type, can also be used. If the time block of the block execution type adaptive algorithm is shortened, the result is finally the same as that of the non-block execution type adaptive algorithm. Therefore, it can be expected that the same effect as the effect of the present invention described above will also occur in an adaptive algorithm that is not a block execution type.

  According to the echo canceller of the present application, double talk and echo path fluctuation can be identified with a small delay, and can be used, for example, in the field of electroacoustics such as a hands-free call device.

It is a figure which shows schematic structure of this invention (echo canceller) applied to the telephone apparatus. It is a block diagram which shows each function of this invention (echo canceller). It is a block diagram of a hands-free communication device using an acoustic echo canceller. It is a frequency distribution diagram of the identification parameter R _Ey (n). It is a figure which shows the audio | voice used for simulation. It is a histogram of a _n (k). It is a distribution map of _Ak obtained by the echo path fluctuation | variation in which an estimation error increases to 0 dB. It is a figure which shows the change of _Pn . It is a distribution diagram of the A _k at the time of double-talk. Estimation error is a distribution diagram of A _k when the echo path change and double-talk increases to 0 dB. is a diagram illustrating the probability of mistaking the echo path change doubletalk when the r _n = 0.6 and the decision threshold. It is a figure which shows an error generation probability when a search range is set to 1-16. It is a figure which shows the frequency with which the code | symbol of A _n (k) of an adjacent block corresponds or does not correspond. It is a figure which shows the probability that a double talk is mistaken for an echo path | route fluctuation | variation in the combined method of code | symbol determination. It is a figure which shows the number of continuous blocks required in order that the probability that a double talk will be mistaken for echo path | route fluctuation | variation will be 10 <^-4> or less. It is a distribution map of _An (k) at the time of the echo path fluctuation | variation in which an estimation error is -10dB, and double talk. It is a distribution map of _An (k) at the time of the echo path fluctuation | variation in which an estimation error is -20dB, and double talk. It is a figure which shows the number of blocks required in order to detect the echo path | pass fluctuation | variation in which an estimation error is -10dB. FIG. 5 is an ERLE characteristic diagram obtained by detecting an echo path variation according to the present invention. FIG. 5 is an ERLE characteristic diagram when echo path variation can be detected with a delay of zero. It is a comparison figure of the reduction characteristic of an estimation error. It is an ERLE characteristic diagram given by the present invention at the time of double talk. It is a control flowchart in the case of applying this invention to a practical apparatus. It is a control flowchart in the case of applying this invention to a practical apparatus.

Explanation of symbols

200 microphone
201 Speaker
212 Adaptive filter

Claims (20)

An echo canceller that removes an echo signal generated by a reception signal supplied to a speaker from the speaker to the microphone, from the output signal of the microphone,
A coefficient updatable FIR filter that generates a pseudo echo signal based on the received signal;
A subtractor that generates an error signal by subtracting a pseudo echo signal generated by the FIR filter from an output signal of the microphone;
A coefficient updating unit that updates the coefficient of the FIR filter by an adaptive algorithm so that the error signal is minimized based on the received signal and the error signal;
The adaptive algorithm calculates a product of the error signal and an input signal to each coefficient tap within a predetermined search range of the FIR filter, and based on a first value determined based on the product, the search A representative tap is selected from the coefficient taps within the range,
Based on the second value determined based on the product of the error signal and the input signal to the representative tap and the third value determined based on the power value of the error signal, a double talk state occurs. Echo canceller to determine whether or not. The first value is the product of the error signal and the input signal to each coefficient tap;
The second value is the product of the error signal and the input signal to the representative tap;
The echo canceller according to claim 1, wherein the third value is a power value of the error signal. The adaptive algorithm is a block execution type adaptive algorithm;
The first value is a cumulative sum in a time block of a product of the error signal and the input signal to each coefficient tap;
The second value is a cumulative sum in a time block of a product of the error signal and an input signal to the representative tap;
The echo canceller according to claim 1, wherein the third value is a cumulative sum in a time block of the power value of the error signal.   The echo according to any one of claims 1 to 3, wherein when the absolute value of the ratio of the second value to the third value is less than a predetermined threshold, it is determined that a double talk state has occurred. Canceller. The absolute value of the ratio of the second value to the third value is greater than or equal to the predetermined threshold, and
The echo canceller according to claim 4, wherein the double-talk state is determined to occur when the signs of the second values for two consecutive times of representative taps selected this time or before are different. The absolute value of the ratio of the second value to the third value is greater than or equal to the predetermined threshold, and
The sign of the second value determined based on the product of the current input signal to the representative tap selected this time and the current error signal, the previous input signal to the representative tap selected this time, and the previous error signal, The echo canceller according to claim 4, wherein the double talk state is determined to occur when the sign of the second value determined based on the product of is different. The absolute value of the ratio of the second value to the third value is greater than or equal to the predetermined threshold, and
The sign of the second value determined based on the product of the current input signal to the representative tap selected last time and the current error signal, the previous input signal to the representative tap selected last time, and the previous error signal, The echo canceller according to claim 4, wherein a double talk state is determined to occur when a sign of the second value determined based on the product of is different.   The echo canceller according to any one of claims 4 to 7, wherein when the third value is equal to or smaller than a predetermined value, the predetermined threshold is corrected to a larger value. An echo canceller that removes an echo signal generated by a reception signal supplied to a speaker from the speaker to the microphone, from the output signal of the microphone,
A coefficient updatable FIR filter that generates a pseudo echo signal based on the received signal;
A subtractor that generates an error signal by subtracting a pseudo echo signal generated by the FIR filter from an output signal of the microphone;
A coefficient updating unit that updates the coefficient of the FIR filter by an adaptive algorithm so that the error signal is minimized based on the received signal and the error signal;
The adaptive algorithm calculates a product of the error signal and an input signal to each coefficient tap within a predetermined search range of the FIR filter, and based on a first value determined based on the product, the search A representative tap is selected from the coefficient taps within the range,
When the sign of the second value for two consecutive times for the representative tap selected this time or before is different, it is determined that a double talk state has occurred,
An echo canceller, wherein the second value is a value determined based on a product of the error signal and an input signal to the representative tap.   The sign of the second value determined based on the product of the current input signal to the representative tap selected this time and the current error signal, the previous input signal to the representative tap selected this time, and the previous error signal, The echo canceller according to claim 9, wherein a double talk state is determined to occur when a sign of the second value determined based on the product of is different.   The sign of the second value determined based on the product of the current input signal to the representative tap selected last time and the current error signal, the previous input signal to the representative tap selected last time, and the previous error signal The echo canceller according to claim 9, wherein a double talk state is determined to occur when a sign of the second value determined based on the product of is different. The first value is the product of the error signal and the input signal to each coefficient tap;
The echo canceller according to any one of claims 9 to 11, wherein the second value is a product of an error signal and an input signal to the representative tap. The adaptive algorithm is a block execution type adaptive algorithm;
The first value is a cumulative sum in a time block of a product of the error signal and the input signal to each coefficient tap;
The echo canceller according to any one of claims 9 to 11, wherein the second value is a cumulative sum in a time block of a product of the error signal and an input signal to the representative tap.   The echo canceller according to any one of claims 1 to 13, wherein, among coefficient taps within the search range, a coefficient tap having the largest absolute value of the first value is selected as a representative tap.   The echo canceller according to any one of claims 1 to 14, wherein the search range is a full coefficient tap of the FIR filter.   The echo canceller according to any one of claims 1 to 14, wherein the search range is a partial coefficient tap of the FIR filter.   The echo canceller according to any one of claims 1 to 14, wherein the search range is a continuous coefficient tap from an initial stage to an intermediate stage of the FIR filter.   The adaptive algorithm reduces the coefficient update amount of the FIR filter or stops updating the coefficient when it is determined that a double talk state has occurred. The echo canceller described in the item.   The adaptive algorithm determines that an echo path variation state has occurred when it is not determined that a double talk state has occurred continuously for a predetermined number of times. The echo canceller described in the section. The echo canceller according to claim 19, wherein the adaptive algorithm increases the coefficient update amount of the FIR filter when it is determined that an echo path fluctuation state has occurred.

Images