JP4642711B2 - Echo canceller - Google Patents

Description

  The present invention relates to an echo canceller, and more particularly to an echo canceller that performs stable operation by discriminating echo path fluctuations and double talk.

  Along with changes in call formats and hardware, we have a greater chance of using microphones. Our auditory attention can be focused on the target sound, so extraneous noise and reverberation aren't bothered. However, since the microphone acquires not only the necessary sound but also unwanted noise and reverberation sound, it becomes the output of the microphone as it is. In recent years, speakers such as hands-free phones and video conferencing systems often talk at a position away from a microphone. For this reason, the level of the required voice is lowered, and the influence of noise and reverberant sound is relatively increased. One device that reduces this effect is an acoustic echo canceller. The echo is a sneak signal to the microphone of the sound output from the speaker or the indirect wave reflected by the wall surface in the room. A system for canceling this echo is an acoustic echo canceller. Acoustic echo cancellers are composed of adaptive digital filters to automatically adapt quickly to changes in the room environment, such as changing the location of microphones and speakers, as the speaker moves around the room, or changes in the echo path .

  One of the most important characteristics of this echo canceller is stability against double talk. Double talk is a near-end speaker speaking. The error signal used for updating the adaptive filter is a residual signal after the echo is canceled by a pseudo echo. When the near-end speaker speaks, since the near-end speaker's voice remains in the residual signal, the risk of divergence increases if the adaptive filter is updated in this state. In order to detect this double talk, it is necessary to distinguish between double talk and echo path fluctuation. This is because both double talk and echo path variations increase the residual signal. Many methods have been proposed so far regarding this identification processing [see, for example, Patent Document 1, Non-Patent Document 1, Non-Patent Document 2, Non-Patent Document 3, Non-Patent Document 4, and Non-Patent Document 5]. Naturally, the risk of howling will increase if the identification is wrong or it takes time. Further, if the identification is made accurate, there arises a problem that the determination is further delayed with respect to a small echo path variation (small estimation error).

Thus, one of the causes of the increase in the estimation error in the acoustic echo canceller is the fluctuation of the echo path, which can be solved by identifying the double talk and the echo path fluctuation. The present invention presents an echo canceller that employs a new identification method for identifying double talk and echo path variations.
JP 2005-318571 A S. Minami and T. Kawasaki, “A double talk detection for an echo canceller,” ICC'85 Rec. , Pp. 1492-1497, 1985. Kensaku Fujii and Toshiro Oga, “Study on double talk and echo path fluctuation detection for acoustic echo canceller,” IEICE Tech., EA-91-82, Jan. 1991. Hau Ye and Bo-Xiu wu, “A new double-talk detection algorithm based on the orthogonality theorem,” IEEE tranc. Commun. , Vol. COM-39, no. 11, pp. 1542-1545, Nov. 1991. Kensaku Fujii, Toshiro Oga, “Distinction of Double Talk and Echo Path Variation,” IEICE Tech., EA-94-15, May 1994. Kensaku Fujii, Toshiro Oga, “Double-talk detection method combined with detection of echo path fluctuations,” IEICE (A), vol. J78-A, no. 3, pp. 314-322, March 1995.

  The object of the present invention is to accurately distinguish between echo path fluctuations and double talk, to accelerate the convergence of coefficients for echo path fluctuations, and to enable stable echo cancellation for double talk. It is in the realization of a canceller.

In order to solve the above-described 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 a 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 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, and a ratio value between the first value and the second value it is determined whether double talk state is generated based on, first value, suspected of the FIR filter A value determined based on the power value of the echo signal, the second value, Ri value der determined based on the power value of the error signal, the calculation of the ratio value is repeated at predetermined time intervals, the current Based on the magnitude of the difference between the calculated ratio value and the previously calculated ratio value, it is determined whether or not a double talk state has occurred.

Further, in order to solve the above problems, another echo canceller according to the present invention, the echo signals received signal is supplied to a speaker is produced by sneaking from the speaker to the first microphone, the first microphone a echo canceller for removing from the output signal, and the coefficient updatable FIR filter for generating an echo replica signal based on the received signal, the pseudo echo signal which the FIR filter is generated from the output signal of the first microphone A subtractor that generates an error signal by subtraction, and 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, it is determined whether double talk state has occurred based on the ratio value between the first value and the second value, said Value is a value determined based on the power value of the pseudo echo signal of the FIR filter, the second value, the power value of the error signal, and the power value of the ambient noise signal around the first microphone value der determined based on is, the ambient noise signal is acquired by the second microphone that is placed in the vicinity of the first microphone, the calculation of the ratio value is repeated at predetermined time intervals, the calculated time Based on the difference between the ratio value and the previously calculated ratio value, it is determined whether or not a double talk state has occurred.

  In the echo canceller, the ratio value is a ratio of the second value to the first value, and a difference between the ratio calculated this time and the ratio calculated last time is equal to or greater than a predetermined threshold value. In such a case, it may be determined that a double talk state has occurred.

  According to the echo canceller of the present application, double talk and echo path fluctuation are accurately identified, and stable operation is guaranteed.

Hereinafter, an echo canceller according to an embodiment of the present invention will be described with reference to the drawings.
(1. Introduction)
In the embodiment of the present invention, a block execution learning identification method is used as an adaptive algorithm, and the block length is controlled by the voice power of the far-end speaker and the step size is controlled by the power of the residual signal. According to this combined control method, since it is not necessary to distinguish from double talk for small echo path fluctuations where the estimation error is -20 dB or less, large echo path fluctuations and double talk with an estimation error exceeding -20 dB are eliminated. A method of identification is proposed.
(2. Adaptive algorithm and control method)
[2.1 Block execution learning identification method]
The basic structure of the acoustic echo canceller is shown in FIGS. 1A and 1B. FIG. 1A shows the path from the speaker 201 to the microphone 200 in FIG. 1B using the transfer function h, and the structures of FIG. 1A and FIG. 1B are equivalent. Further, FIG. 1C is a diagram showing a structure equivalent to FIG. 1B, but the structure of the adaptive FIR filter 212 in FIG. 1B is shown in more detail. In these figures, the echo canceller is applied in a telephone device.

  Referring to FIG. 1C, 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 FIR filter 212a and the coefficient updating unit 212b constitute an adaptive FIR filter 212.

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 g _j generated by this wraparound. Note that n _j is the near-end speaker's voice and ambient noise input to the microphone 200.

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. The FIR filter 212a outputs a signal G _j as an output signal (pseudo echo signal).

The subtractor 211 receives the signal y _j and the signal G _j . The subtractor 211 subtracts the signal G _j from the signal y _j . Subtraction result as a signal (error signal) e _j 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.

By such a system, the generated echo g _j is canceled by the pseudo echo G _j synthesized by the adaptive FIR filter 212, and the residual signal is obtained.

Is transmitted to the far end. Here, n _j is the disturbance input to the microphone, that is, the near-end speaker's voice and ambient noise.

  In this embodiment, a block execution type learning identification method is used as an adaptive algorithm for updating the coefficient of the adaptive filter.

Is used. Here, J is the block length and ρ is the step size.

[2.2 Block length control]
In the learning identification method, the mean square of the estimation error when the adaptive filter coefficient converges is

It is estimated. Here, μ is the step size, σ _N ² is the average power of the disturbance, and σ _x ² is the average power of the reference signal. Regarding this estimation error, between the learning identification method and the block execution type learning identification method,

There is a relationship. For example, μ = 1 in the learning identification method corresponds to ρ = J. By substituting this relationship into Equation (3), the mean square of the estimation error obtained by the block execution learning identification method is

It is estimated. In order to simplify the relationship between the mean square of the estimation error and the block length J and facilitate control, formula (5) can be summarized for the case where 2J >> ρ can be approximated.

It becomes. From this equation, the power of the reference signal required to obtain the required estimation error C ₀ with respect to the average power Q ₀ when the disturbance is ambient noise is

It is obtained. Furthermore, the power σ _x ² of this reference signal is the addition norm

If the number of taps of the adaptive filter is I, the average

Has the size of From this, control the block length J

If the coefficient is updated at this time, the estimation error is expected to decrease to C ₀ regardless of the power fluctuation of the reference signal. Further, in order to perform the identification of double talk and echo path change stably in this embodiment, defines the shortest block length, put the J _0.

[2.3 Step size control]
Step size control requires calculation of disturbance power. In the acoustic echo canceller, only the output of the microphone can be used for the calculation. However, the echo g _j is mixed in the output. Therefore, in the present embodiment, the residual signal when the echo g _j is sufficiently canceled by the pseudo echo G _{j in} the calculation of the power of the disturbance n _j.

Is approximated by the disturbance n _j , the mean square of the residual signal

Is approximated to the disturbance power. Using the addition norm Pn calculated simultaneously with the estimated value Qn of the disturbance power

If the step size is controlled in units of blocks, the estimation error is maintained at the required value C ₀ even if the coefficient of the adaptive filter is updated in the double talk section. However, ρ ₀ = 16 corresponding to μ = 0.25 of the learning identification method is set as the upper limit of the step size ρ _n in consideration of the variation of Pn and the color of the speech.
(3. Discrimination method of double talk and echo path fluctuation)
[3.1 Identification parameters]
In the present embodiment, a method is proposed in which the power average Gn of the pseudo echo and the power average Qn of the residual signal are two-dimensionally arranged and identified.

For example, in equation (15), Gn can be obtained by accumulating power values G _j ² of the pseudo echo signal G _j in a time block (time range) having a time width J and dividing by J. It is shown. In equation (16), has been shown to Qn is obtained from the error signal e _j in the same manner.

The ratio of the Gn and Qn formula (15), which approximates the ratio of the average power of G _j and e _j from (16). Here, in the case of echo path fluctuation, when the near-end speaker's voice is not contained, n _j can be ignored in equation (1), so the ratio between Gn and Qn is approximated to the estimation error. Since the coefficient is now updated slowly, the ratio between Gn and Qn becomes almost constant. On the other hand, in the case of double talk, when the near-end speaker's voice is _input , the influence of n _j becomes large in equation (1), so the ratio between Gn and Qn is not constant. FIG. 2 (a) shows the distribution of the same ratio when an echo path variation with an estimation error of 0 dB, −10 dB, and −20 dB occurs, and FIG. 2 (b) shows the occurrence of double talk in a state where the echo is sufficiently reduced. It is the distribution of the same ratio. Here, in FIG. 2, ρ ₀ = 16, J ₀ = 64, C ₀ = 0.001, I = 512, the power ratio between the far-end speaker and the ambient noise is 30 dB, and the power of the near-end speaker and the ambient noise. The simulation is performed with a ratio of 30 dB. The same conditions apply in the following simulation.

  As can be seen from FIG. 2, the echo path variation is linear, and double talk gives an irregular distribution. That is, it is expected from this difference in distribution that it is possible to discriminate echo path fluctuations when there is little change in slope in the adjacent time, and double talks when there is a large change. The slope (ratio of Qn to Gn) is

Is calculated.

In the above equation (equation (17)), the Gn denominator first value, the Qn molecules when the second value, r _n can be thought of as the ratio of the second value to the first value.

FIG. 3 (a) shows the time variation of the slope obtained with respect to the echo path fluctuation from above and the voice of the far-end speaker, and FIG. 3 (b) shows the time change of the slope obtained with respect to double talk from above. The voice of the near-end speaker (upper) and the voice of the far-end speaker (lower). However, in the figure, the upper limit of inclination is displayed as 3 for easy viewing. As can be seen from the graph, in the case of double talk, there is a significant change except for the section where the near-end speaker is silent. It can be seen that the echo path variation continues with a substantially constant slope. However, the slope is large in the section where the far-end speaker is silent. This is because the influence of ambient noise is strong in the silent section. Therefore, the slope (ratio of (Qn−Q ₀ ) to Gn) is

And calculate. Here, Q ₀ is the ambient noise power.

In the above formula (formula (18)), the denominator term is the first value, the numerator term is the second value, and R _n can be considered as the ratio of the second value to the first value. The numerator term is a subtraction result value obtained by subtracting the ambient noise power from Qn. This ambient noise signal can be obtained from an output signal from the microphone 200 (see FIGS. 1B and 1C) when both the near-end speaker and the far-end speaker are silent. Further, the ambient noise signal may be obtained from the output signal of another microphone installed in the vicinity of the microphone 200.

Further, Q ₀ in the above equation (Equation (18)) is regarded as a predetermined value that does not change with time (that is, Q ₀ is set as a constant), and the predetermined value (Q ₀ ) is subtracted from Qn. A molecular term as a binary value may be obtained. Such a method is effective when the noise level around the microphone 200 is known and the change in the noise level is small. This eliminates the need to detect ambient noise signals from the microphone, thus simplifying the system.

  The result of calculation of equation (18) is shown in FIG. Compared to FIG. 3A, it can be seen that the influence of ambient noise is smaller in the echo path variation.

Therefore, than r _n of formula (17), who was using Rn in the formula (18), it can be expected to be able to perform identification of double talk and echo path change more accurately. In the following, an example in which double talk and echo path variation are identified using this slope Rn (Rn in equation (18)) will be described. In this (Equation (18)), the absolute value is taken to prevent the slope value from becoming negative.

  If it is possible to discriminate double talk and echo path variation from the magnitude of the change in the slope, the identification is performed and the step size of the adaptive FIR filter is changed or the filter coefficient is updated according to the result. It is only necessary to control the operation of the adaptive FIR filter, such as stopping or restarting. More specifically, when it is determined that the double talk state has occurred, the step size of the adaptive FIR filter is reduced to reduce the update amount of the filter coefficient, or the update of the filter coefficient is stopped. It may be. Further, when it is determined that the echo path variation state has occurred, the updating of the filter coefficient once stopped may be resumed, or the update amount of the filter coefficient may be increased.

  FIG. 5 shows the result of calculating the magnitude of the change in order to investigate the possibility of discriminating double talk and echo path fluctuations from the magnitude of the change in slope. (A) shows the magnitude of the change in tilt due to the echo path variation and the voice of the far-end speaker, and (b) shows the magnitude of the change in tilt and the voice of the near-end talker during double talk. As can be seen from this result, the change in the slope is large when the far-end speaker's voice is silent in the echo path variation, and the slope change is small in the double-talk when the near-end speaker is silent. I understand that. That is, it can be seen that identification errors are likely to occur in a section where both are silent.

[3.2 Sections that do not require identification]
In order to solve this problem, the state of control when there is no sound at the far end side and near end side simultaneously is examined. In this case, the maximum step size ρ ₀ = 16 is set as single talk, and the coefficient is updated so that the estimation error converges to C ₀ by block length control. When the far-end speaker is silent and the near-end speaker utters, the power of the reference signal is much smaller, in which case step size control is sufficient,
1. When the far-end speaker is determined to be silent, identification is not necessary. Similarly, the section where the near-end speaker is silent is a single talk state regardless of whether or not the far-end speaker is generated, and step size control is sufficient. That is, 2. When the near-end speaker is determined to be silent, identification is not necessary.

These two judgments can be judged by observing Pn and Qn and when these are equal to or higher than Q ₀ . However, considering the observation error

When is true, identification is not necessary.

If a section that does not require this identification is removed, the result is as shown in FIG. However, for the sake of easy understanding, only Gn and Qn are displayed for portions below 0.01. Compared with FIG. 6 before removing unnecessary sections, it can be confirmed that when Gn and Qn have small values, they are removed as unnecessary identification. Further, when the magnitude of the change in the inclination is examined with 0 as the section where identification is not required, it is as shown in FIG.
(4. Simulation)
Check if double talk and echo path variation are identified by simulation. The simulation conditions are the same as in FIG.

[4.1 Discrimination results of double talk and echo path fluctuation]
As an identification condition, if the magnitude of the change in inclination is less than 0.1, the echo path is changed. That is, 0.1 is used as a threshold value (predetermined threshold value) to determine whether or not a double talk state has occurred. Since this condition is provisionally determined in the present embodiment, further study is necessary. Identification is performed under these conditions, and 1 is output when step size control is performed as double talk, ₀ is output when identification is not required, and -1 is output when ρ = ρ ₀ is given as echo path variation. Here, the estimation error is set to 0 dB when the echo path is changed, and the estimation error is set to −30 dB when the echo is double talk. FIG. 9 shows the identification result at this time. As can be seen from FIG. 5A, all the sections other than the section where identification is not necessary are judged to be echo path fluctuations, so that the identification is successful. Also, looking at the vicinity of 10,000 samples where the near-end speaker is silent in Fig. 5 (b), it is correctly determined that the echo path changes when the far-end speaker is speaking, and that identification is unnecessary when there is no speech. You can see that. Further, as seen in the vicinity of 9,000 samples in FIG. 9B, it is determined that the echo path changes despite the double talk. This is considered to be because the change in inclination becomes small when the power of the voices of both speakers does not change greatly. To reduce this error, when the magnitude of changes in Gn and Qn is small, ie

Then, the condition that the previous process is continued is added. In the present embodiment, it is assumed that α = 0.001. At this time, the identification result is as shown in FIG. Obviously, it can be confirmed that the judgment is correct.

[4.2 Convergence characteristics when double talk and echo path change]
Next, the convergence characteristics when the identification is performed using the above-described identification conditions are confirmed. However, ρ ₀ = 6.4 was used, and FIG. 11 was used as the far-end speaker and FIG. 12 was used as the near-end speaker. Moreover, it is the second term of the formula (2) for stabilization.

It is necessary to decide the upper limit. In the present embodiment, the upper limit is set to A (m) = 0.1 / I. The convergence characteristics are shown in FIG. From this, it can be confirmed that the estimation progresses in the case of single talk and the estimation does not advance in the case of double talk. Further, FIG. 14 shows 50,000 samples, and FIG. 15 shows that the value of the unknown filter h _j that is an echo path is changed with 150,000 samples. Then, the value of the estimation error increased. FIG. 16 and FIG. 17 are enlarged in order to confirm the cause. From this, it can be seen that at the point _where the value of ρ _n is changed, the maximum value is ρ ₀ = 6.4. Since the value of ρ _n is small in the case of double talk, it can be confirmed that this is an echo path fluctuation.

[4.3 Conclusion]
As described above, a simulation was performed to confirm whether identification was operating effectively. From this result, it was confirmed that the identification was performed effectively. Further, as a future problem, the upper limit of A (m) will be examined and other parameters will be examined.
(5. System outline)
FIG. 18 is a block diagram showing a schematic configuration and each function of the present invention (echo canceller) according to this embodiment applied to a communication device.

  Referring to FIG. 18, the call device has a microphone 200 and a speaker 201. The echo canceller has an adaptive FIR filter 212. The configuration of the adaptive FIR filter 212 is the same as that of the adaptive FIR filter 212 shown in FIG. 1C.

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 g _j generated by this wraparound. Note that n _j is the near-end speaker's voice and ambient noise input to the microphone 200.

The signal x _j is also input to the adaptive FIR filter 212, and the adaptive FIR filter 212 outputs a signal G _j as an output signal (pseudo echo signal).

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

The calculation unit 213, a pseudo echo signal G _j of adaptive FIR filter 212 has generated, is input and the error signal e _j, or slope (ratio) r _n, of the formula (17) Based on these equations (18 ) Is calculated by the calculation unit 213. Then, the calculated value (r _n or Rn) is sent to the identification unit 214, where it is identified (determined) whether a double talk state has occurred or an echo path fluctuation state has occurred. . The identification unit 214 outputs the identification (judgment) result as an identification signal. Based on the identification signal, a control unit (not shown) controls the operation of the adaptive FIR filter 212, such as changing the step size of the adaptive FIR filter 212, stopping or restarting the filter coefficient update.
(6. Summary)
As described above, according to the present invention, double talk and echo path fluctuation are accurately identified, and stable operation is guaranteed for the system. In the description of the present invention, the adaptive algorithm is used 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 more accurately identified, and can be used in the field of electroacoustics such as hands-free communication devices.

It is a principle diagram of an acoustic echo canceller. It is a principle diagram of an acoustic echo canceller. It is a principle diagram of an acoustic echo canceller. It is a figure which shows an identification parameter. It is a figure which shows inclination _R0 . It is a figure which shows inclination Rn at the time of echo path | route fluctuation | variation. It is a figure which shows the change of inclination. It is a figure which shows an identification parameter. It is a figure which shows an identification parameter when an identification unnecessary part is removed. It is a figure which shows the change of the inclination except the unnecessary part of identification. It is a figure which shows an identification result. It is a figure which shows the identification result of a double talk. It is a figure which shows the voice of a far-end speaker. It is a figure which shows the voice of a near-end speaker. It is a figure which shows a convergence characteristic. It is a figure which shows a convergence characteristic. It is a figure which shows a convergence characteristic. It is an enlarged view of convergence characteristics. It is an enlarged view of convergence characteristics. It is a block diagram which shows schematic structure and each function of the echo canceller applied to the telephone apparatus.

Explanation of symbols

200 microphone
201 Speaker
211 Subtraction part
212 Adaptive FIR filter
213 Calculator
214 Identification part

Claims (3)

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 for updating 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;
Determining whether or not a double talk state has occurred based on a ratio value between the first value and the second value;
The first value is a value determined based on a power value of a pseudo echo signal of the FIR filter;
Second value, Ri value der determined based on the power value of the error signal,
The calculation of the ratio value is repeated at predetermined time intervals,
An echo canceller that determines whether or not a double talk state has occurred based on a magnitude of a difference between the ratio value calculated this time and the ratio value calculated last time . The echo signals received signal is supplied to a speaker is produced by sneaking from the speaker to the first microphone, a echo canceller for removing from the output signal of the first microphone,
A coefficient updatable FIR filter that generates a pseudo echo signal based on the received signal;
A subtracter for generating an error signal by subtracting the pseudo echo signal which the FIR filter is generated from the output signal of the first microphone,
A coefficient updating unit for updating 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;
Determining whether or not a double talk state has occurred based on a ratio value between the first value and the second value;
The first value is a value determined based on a power value of a pseudo echo signal of the FIR filter;
Second value, Ri value der determined based the power value of the error signal, to the power value of the ambient noise signal around the first microphone,
The ambient noise signal is acquired by a second microphone installed in the vicinity of the first microphone;
The calculation of the ratio value is repeated at predetermined time intervals,
An echo canceller that determines whether or not a double talk state has occurred based on a magnitude of a difference between the ratio value calculated this time and the ratio value calculated last time . The ratio value is a ratio of the second value to the first value;
The echo canceller according to claim 1 , wherein when the magnitude of the difference between the ratio calculated this time and the ratio calculated last time is equal to or greater than a predetermined threshold, it is determined that a double talk state has occurred.

Images