JP5016551B2 - Echo canceller - Google Patents

Description

  The present invention relates to an echo canceller using an adaptive filter, and more particularly to an echo canceller that can converge filter coefficients of an adaptive filter at high speed without discriminating between echo path fluctuations and double talk.

There are two factors that cause the acoustic echo reduction amount to fluctuate in the acoustic echo canceller. One is power fluctuation of the far-end speaker voice that works as a reference signal, and the other is superposition of the near-end talker voice that is a disturbance (called double talk). Among these, the former problem is solved by applying block length control to the adaptive algorithm [Non-Patent Document 1]. By this application, the coefficient of the adaptive filter is updated even in a section where there is no sound, and acoustic echo can be reduced quickly. The problem lies in the latter, and the difficulty in solving the problem is that it is difficult to distinguish between the near-end speaker input to the microphone and the acoustic echo, and the opposite control is required for both inputs. In general, both inputs are identified using a difference in correlation strength with the far-end speaker voice [Non-Patent Document 2]. The problem is that since the strength of the correlation appears as a probability phenomenon, identification errors are inevitable, and accurate identification takes time. Therefore, as a method for quickly and accurately performing this identification, a method of arranging identification parameters in two dimensions has been proposed [Non-patent Document 3]. However, the procedure is too complicated for identification accuracy, and there is a problem that it takes a long time to identify small echo path variations. For this difficulty of identification, a control method that does not require identification can be one of the solutions. However, [Non-Patent Document 4] and [Non-Patent Document 5] have a problem in that the response to a large echo path variation is delayed, and the identification has not been made completely unnecessary [Non-Patent Document 5]. . Furthermore, as a technique for not performing this identification, a method using an adaptive filter with a coefficient semi-fixed has been studied [Non-Patent Document 6]. This is to monitor the magnitude of the residual echo obtained by the adaptive filter that constantly updates the coefficient and the filter whose coefficient is semi-fixed, and when it is judged that the former residual echo is small, the coefficient is transferred to the semi-fixed filter. This is a method for ensuring a stable echo cancellation amount. The problem is that the coefficients are not updated during double talk, and therefore convergence is delayed. In an acoustic echo canceller, double talk is normal, and it is desirable that the coefficient of the adaptive filter is updated even during double talk.

[Acoustic echo canceller using 16-bit fixed-point DSP] [
[Non-patent document 7], and [Non-patent document 8] shows [Convergence characteristics of a tri-quantized X-NLMS algorithm], and [Normalized dolly mean EE 'algorithm and its convergence condition]. There is [Non-patent document 9], and [Non-patent document 10] shows [block type execution of adaptive digital filter], and [derivation and analysis of convergence condition of EE 'method]
There is [Non-patent Document 11].
Kensaku Fujii, Toshiro Oga, “Continued update method of adaptive filter coefficients in silent speech interval, useful for acoustic echo canceller,” IEICE Transactions (A), vol. J78-A, no. 11, pp. 1403 -1409, Nov. (1995). Kensaku Fujii and Toshiro Oga, “Double-talk detection method using echo path variation detection,” IEICE Transactions (A), vol. J78-A, no. 3, pp. 314-322 (1995). Zhang Wakao, Kensaku Fujii, Meiji Muneyasu, “A Study on Double Talk and Echo Path Change Detection,” IEICE Technical Report, EA2003-146, Jan. (2004). Kensaku Fujii, Toshiro Oga, “Reduction of convergence time in block length control in additive normalization LMS method,” IEICE Transactions (A), vol. J80-A, no. 1, pp. 27-35, Jan (1997). Kensaku Fujii, Toshiro Oga, “A method to keep the estimation error for acoustic echo canceller to the required value,” IEICE Transactions (A), vol. J83-A, no. 2, pp. 141-151, Feb (2000). Yoichi Haneda, Shoji Makino, Masafumi Tanaka, Suetsu Shimauchi, Junji Kojima, “Duo Filter Control System for Acoustic Echo Canceller,” 1995 IEICE General Conference, A-350, March (1995). J. Sakaguchi, T. Hoshino, K. Fujii and J. Ohga, “Acoustic echo canceller system materialized with a 16-bit fixed point processing type DSP,” IEICE Trans. Fundamentals, vol. E82-A, no. 12, pp 2818-2821, Dec. 1999. K. Fujii and Y. Tanaka, “Convergence property of triquantized-x NLMS algorithm,” IEICE Trans. Fundamentals, vol. E83-A, no. 12, pp. 2739-2742, Dec. 2000. K. Fujii, M. Muneyasu, T. Hinamoto, and Y. Tanaka, “Normalizedleast mean EE 'algorithm and its convergnec condition,” IEICE Trans. Fundamentals, vol. E84-A, no. 4, pp. 984-990, April 2001. GA Clark, SK Mitra, and SR Parker, “Block implementation of adaptive digital filter,” IEEE Trans. Circuits & Syst., Vol. CAS-28, no. 6, pp. 584-592, June 1981. Yusuke Iwamatsu, Kensaku Fujii, Meiji Muneyasu, “Derivation and analysis of convergence condition of EE method,” IEICE Technical Report, EA2006-99, Jan. (2007).

  As described above, the problem with the echo canceller of the type that distinguishes between echo path fluctuations and double talk is that it is difficult to quickly and reliably identify, and that the coefficient is not updated at the time of double talk. The convergence is delayed. In an acoustic echo canceller, double talk is normal, and it is desirable that the coefficient of the adaptive filter is updated even during double talk.

  The object of the present invention is to solve such problems.

In order to solve the above-mentioned problems, an echo canceller according to the present invention is an echo canceller that removes an echo signal generated by a supply signal supplied to a speaker from the speaker to the microphone from the output signal of the microphone. An FIR type main adaptive filter that inputs the supply signal as a reference signal and generates a pseudo echo signal as an output signal and that can update filter coefficients, and a pseudo echo generated by the main adaptive filter from the output signal of the microphone A main subtractor that generates a main error signal by subtracting the signal, a FIR type sub-adaptive filter that can update the filter coefficient that inputs the supplied signal as a reference signal and generates an output signal, and a sub error signal are generated A sub-subtractor, wherein the main adaptive filter is coupled with the supply signal by an adaptive algorithm. Based on the main error signal, the coefficient of the filter is updated so that the power of the main error signal is minimized, and the sub-adaptive filter is based on the supply signal and the sub-error signal by an adaptive algorithm. The coefficient of the filter is updated so that the power of the sub error signal is minimized, and the number of filter taps I ₁ of the sub adaptive filter is smaller than the number of filter taps I of the main adaptive filter, and the sub subtractor Generates the sub error signal by subtracting the added value of the second half tap output value of the main adaptive filter and the tap output value of the sub adaptive filter from the output signal of the microphone, and generates the second half tap of the main adaptive filter. In the main adaptive filter, the output value is obtained by multiplying each signal subsequent to I ₁ on the delay line that sequentially delays the input reference signal by the corresponding filter tap coefficient. The tap output value of the sub-adaptive filter is obtained by shifting, and the tap output value of the sub-adaptive filter is a filter tap coefficient corresponding to each signal on the delay line that sequentially delays the input reference signal in the sub-adaptive filter. The main adaptive filter updates the step size in updating the filter coefficient based on the sub error signal. In this echo canceller, for example, the number of filter taps I ₁ of the sub adaptive filter 128 may be a 512 filter taps I of the main adaptive filter. In this case, in order to obtain the latter half tap output value of the main adaptive filter, first, 384 multiplications are performed by multiplying each of the 129th to 512th delay signals in the main adaptive filter by the corresponding filter tap coefficient. Get the value. Then, the latter half tap output value of the main adaptive filter can be obtained by summing these 384 multiplied values. In order to obtain the tap output value of the sub-adaptive filter, first, 128 delayed values are obtained by multiplying the first to 128th delay signals in the sub-adaptive filter by the corresponding filter tap coefficients. . The tap output value of the sub-adaptive filter can be obtained by summing these 128 multiplication values.

  In the echo canceller, the adaptation algorithm of the main adaptive filter and the adaptation algorithm of the sub adaptive filter may both be learning identification methods.

Further, in the echo canceller, the main adaptive filter has a step size in updating the filter coefficient, a multiplication value of a predetermined value and an addition norm Pn of the supply signal vector, an average power of the sub error signal, and the main adaptation. You may make it update to the value (rho) n divided by the multiplication value with the filter tap number I of a filter. Here, the predetermined value may be determined as appropriate. However, when the required estimation error C ₀ is determined, a value (2C ₀ ) obtained by doubling the value (C ₀ ) may be used as the predetermined value. .

  In the echo canceller, both the adaptive algorithm of the main adaptive filter and the adaptive algorithm of the sub adaptive filter may be a block execution type learning identification method.

Further, in the echo canceller, the main adaptive filter has a step size in updating the filter coefficient, a multiplication value of a predetermined value and an addition norm Pn of the supply signal vector, an average power of the sub error signal, and the main adaptation. It is updated to a value ρn divided by the product of the filter tap number I of the filter, and the addition norm is an addition norm in the block length J of the block execution type learning identification method, and the average power is The average power in the block length J may be used. Here, the predetermined value may be determined as appropriate. However, when the required estimation error C ₀ is determined, a value (2C ₀ ) obtained by doubling the value (C ₀ ) may be used as the predetermined value. .

In the above echo canceller filter taps I ₁ of the sub adaptive filter may be substantially one quarter of the number of filter taps I of the main adaptive filter. For example, the number of filter taps I ₁ of the sub adaptive filter 128 may be a 512 filter taps I of the main adaptive filter.

  According to the echo canceller of the present application, by separately preparing a sub-adaptive filter and controlling the step size of the main adaptive filter with the power of the residual echo calculated using the filter output, echo path variation and double talk are reduced. The filter coefficients of the main adaptive filter can be converged at high speed without identification.

Hereinafter, an echo canceller according to an embodiment of the present invention will be described with reference to the drawings.
(1. Overview)
Here, an echo canceller that does not require discrimination between the echo path variation and the double talk is realized by adding the following modifications to the methods shown in [Non-Patent Document 6] and [Non-Patent Document 5]. In this echo canceller, the main adaptive filter proposed in [Non-patent Document 5] that controls the step size by regarding the residual echo as a disturbance, and the step size is fixed to a value that provides the fastest convergence, and the number of taps is small. Prepare two sub-adaptive filters. Further, the step size is controlled by the residual echo obtained by the sub adaptive filter. As a result, the step size is controlled by a residual echo that rapidly decreases, and the former adaptive operation is speeded up. On the other hand, in the double talk, the residual echo rapidly increases in the latter, thereby setting the step size small and preventing the disturbance of the coefficient.
(2. Conventional control method)
Figure 1 shows the structure of an acoustic echo canceller to which the control method proposed in [Non-Patent Document 5] is applied. This study is based on this control method. In this structure, the coefficient of the adaptive filter is calculated using the block execution type learning identification method [Non-Patent Document 10].

And updated. Where Hn is the coefficient vector of the adaptive filter updated in the nth block, ρn is the step size, j is the time (sample time index), J is the block length, ej is the residual echo, and Xj is the reference It is a signal vector. By controlling the block length J and the step size ρn for this coefficient updating algorithm, the estimation error decreases toward the required value C ₀ even if the coefficient is constantly updated regardless of the power fluctuation of the reference signal and double talk. [Non-Patent Document 5].
(2.1 Block length control)
[Non-Patent Document 5] controls the block length J as follows. That is, the constant P ₀ is determined, and the addition norm

But

The coefficient of the adaptive filter is updated after extending the block length J until. The principle by which this control maintains the estimation error at the required value C ₀ is an expression that prescribes the estimation error, which is derived in the learning identification method.

It is in. This equation shows that if the step size is set to μ ₀ when the power ratio of the reference signal to the disturbance is σ _X ² / Q ₀ , the estimation error is reduced to C ₀ calculated from the above equation. . This also indicates that even if the average power Q _{0 of the} disturbance is constant, if the average power σ _X ² of the reference signal varies, the estimation error also varies inversely.

In [Non-Patent Document 5], in order to find a control method that keeps the estimation error constant even if the power of the reference signal fluctuates,

And approximate. Even with this approximation, the step size is usually chosen.

If it is limited to this range, only a maximum of 1/2 (3dB) will occur. In [Non-patent document 5], between the step size of the learning identification method and the block length J ₀

Leading to a relationship. If this relationship is substituted into equation (5),

Is obtained. However, in the above equation, σ _X ² J ₀ I is the average of the addition norms Pn when the block length in equation (1) is J ₀ . I is the number of taps of the adaptive filter. This formula is
Even if the power of the reference signal is smaller than the assumed σ _X ² , the block length J ₀ is extended, and the addition norm Pn is

If the coefficient is updated when

Or

In other words, it is maintained at a constant C ₀ .

Here, the following cautions are necessary when applying this block length control to a real system. That is, when the sound is collected in a quiet environment, the power of the reference signal becomes small in a section where there is no sound, and the echo may not reach the microphone. If the coefficient is updated in this state, the coefficient of the adaptive filter is disturbed [Non-patent document 7] and [Non-patent document 8]. One method for preventing this disturbance is to check the size of each sample value of the reference signal, and to perform control that is not used for coefficient update when the value is below a certain value [Non-Patent Document 7], [Non-Patent Document 8]. ]. Another method is to determine the longest block length, and when the above relationship is not satisfied even in the longest block length, it is determined that the echo does not reach the microphone and the coefficient is not updated. When applying this block length control to an actual system, it is necessary to apply any one of these methods.
(2.2 Step size control)
Next, a control method at the time of occurrence of double talk, which is another factor that varies the amount of acoustic echo reduced by the acoustic echo canceller, will be described. In this study, focusing on equation (9),

And deformed. Here, it is assumed that double talk occurs and the power of disturbance increases from Q ₀ to Q n. At this time, in order to maintain the estimation error at C ₀ , the step size is

It is necessary to make it smaller. That is, if the step size is controlled in this way, the required value C ₀ is maintained as an estimation error. Furthermore, the addition norm does not always coincide with P ₀ in practice. Actually, the addition norm when executing coefficient update is Pn, so the step size is

It is necessary to calculate.

  The problem is that it is necessary to know the disturbance power Qn for this step size control. [Non-Patent Document 5] shows the power of the disturbance.

And approximate. Figure 1 shows the structure for calculating the step size according to this approximation. However, this approximation results in including the unerased echo power in Qn, giving a step size smaller than necessary in the echo path variation, and greatly delaying the convergence of the coefficients. In this study, we propose one way to solve this problem.

In addition, as with block length control, there are points to be noted when applying this step size control. That is, when the average power ratio of the disturbance to the reference signal becomes smaller than the required estimation error C _{0, the} step size calculated from the above equation becomes large. For example, if the coefficient is updated when the addition norm is exactly P ₀ , the step size is

Is calculated. Among, Qn / σ _X ² is the power ratio of the disturbance versus the reference signal. Therefore, if the disturbance power Qn is smaller than expected, C ₀ σ _X ² / Qn may become extremely large. For example, when the required estimation error is designed as C ₀ = 0.001 (−30 dB) and the average power ratio of the disturbance to the reference signal is reduced to −40 dB, the step size calculated from the above equation is 20 J ₀ . On the other hand, the upper limit of the step size set in the learning identification method is 1, and in the block execution type learning identification method, the upper limit is J ₀ . That is, the coefficients of the adaptive filter diverge when ρ> J ₀ . In this control, it is necessary to set an upper limit for the set step size.
(3. Proposed control method)
In order to increase the step size and speed up the convergence when changing the echo path, it is necessary to quickly discharge the power of the echo remaining from Qn that approximates the power of the disturbance. Here, in order to perform the discharge at high speed, a separate adaptive filter is prepared as a sub-adaptive filter with the step size fixed to the value that maximizes the convergence, and the residual echo generated by the sub-adaptive filter is generated.

The power of

We propose a method to control the step size using. The structure is shown in Fig.2. However, in this structure, the tap of the main adaptive filter is divided into the first half and the second half, and the number of taps of the sub-adaptive filter is given as the first tap number I ₁ of the first half of the main adaptive filter so that the convergence is further accelerated. Moreover, the residual echo is obtained by subtracting the tap output u _j ^b in the latter half of the main adaptive filter in addition to the output u _j ^C of the sub adaptive filter.

Prevents saturation at an early stage.

  In this structure, if the coefficient of the sub-adaptive filter with a small number of taps is updated with the step size that provides the fastest convergence, the residual echo of the sub-adaptive filter will be more than the residual echo ej generated by the main adaptive filter.

Will decrease faster. Utilizing this characteristic, the residual echo

 From

Calculate the step size

Can be expected to speed up convergence.
(3.1 Comparison of convergence characteristics)
Figure 3 compares the convergence characteristics. However, the following conditions are set in this simulation.
(1) Sample value of impulse response of unknown system: normal random number with exponential decay
(2) Number of samples: 512.
(3) Number of taps of main adaptive filter: I = 512
(4) Number of taps of the sub-adaptive filter: I ₁ = 128
(5) Reference signal and disturbance: white noise
(6) Power ratio of reference signal to disturbance: -40dB
(7) Coefficient update algorithm: Block execution type learning identification method
(8) Block length: J ₀ = 16 (fixed)
(9) Sub-adaptive filter coefficient update step size: ρ ₀ = 16 (fixed)
(10) Upper limit of coefficient update step size of main adaptive filter: ρ ₀ = 16
(11) Required estimation error: C ₀ = 0.001 (−30 dB)
FIG. 3 (a) shows the convergence characteristics obtained by the method of [Non-Patent Document 5] without using the sub-adaptive filter. Naturally, in (a), since the residual echo is large in the early stage of convergence, the step size is set small, and the convergence of the coefficient is the slowest.

Figure (b) does not use u _j ^b, but uses only the output u _j ^C of the sub-adaptive filter to generate residual echo.

And the convergence characteristics when the step size is controlled by the power. in this case,

Decreases rapidly and a large step size is set, and convergence is accelerated in the early stage of convergence. However, since the number of taps is set low after that

Saturates quickly and the step size is fixed to a small value. Therefore, the convergence speed is fixed as shown in Fig. (B). However, since the step size is fixed to a small value, the estimation error is greatly reduced.

  Figure (c) shows this

In order to prevent the decrease of the signal from being saturated early, the tap output u _j ^b in the latter half of the main adaptive filter is added to the output u _j ^C of the sub adaptive filter to further reduce the residual echo. . This configuration causes residual echo

Can be reduced and convergence is further accelerated. From this result, the convergence time is about 1/7.
It can be seen that the number has decreased.

  Finally, Figure (d) shows the residual echo power Qn and

These are convergence characteristics when the step size is controlled using the smaller one (Fig. 2). The convergence characteristics at this time coincide with (c). That is the power of residual echo

The step size of the main adaptive filter is always smaller than 1, so that the convergence of the main adaptive filter is delayed, and until the convergence

This is because ej is not selected for this purpose.
(3.2 Required estimation error and reference signal to disturbance power ratio)
In this simulation, the required estimation error is −30 dB, and the power ratio of the reference signal to the disturbance is −40 dB. Since the block length is 16, the step size calculated when the estimation error is reduced to −30 dB reaches the upper limit of 16. Accordingly, since the step size is fixed at 16 thereafter, the estimation error is saturated at the reference signal to disturbance power ratio of -40 dB. Further, the residual echo ej given by the main adaptive filter and the residual echo given by the sub-adaptive filter by the process of fixing the step size to the upper limit of 16

Match, Qn and

The same step size as in (iii) is calculated even in (iv) in which the smaller one is selected. As a result, both convergence characteristics coincide.

On the other hand, FIG. 4 shows the convergence characteristics when the required estimation error is C ₀ = 0.0001 (−40 dB). Thus, if the required estimation error is reduced to C ₀ = 0.0001, the step size calculated from the equation (9) is further reduced, and the convergence is delayed from the result of FIG. On the other hand, there is a difference between (iii) and (iv) in the magnitude of the estimation error after convergence of the coefficients. This is due to the dispersion of the power of the residual echo. That is, the step size of the sub-adaptive filter is fixed at 1, and the step size of the main adaptive filter is controlled. Further, since the required estimation error is set to C ₀ = 0.0001 and the power ratio of the reference signal to the disturbance is set to −40 dB, the step size is controlled with an average of 16. On the other hand, since the upper limit of the step size is set to 16, the average of the controlled step sizes is smaller than 16. As a result, the residual echo ej is smaller, and the calculated step size is

And ej are larger when selecting the smaller one. Therefore, the estimation error is not reduced,

Since the step size is given small when controlling only, the estimation error is further reduced. In this respect, (iii) is a more desirable control method than (iv).
(4. Characteristics during double talk)
In finding the optimum step size control method, the amount of echo cancellation secured by the control method during double talk is one of the judgment materials. In ITU-T G.167, the echo cancellation amount can be reduced to 15 dB in double talk. In this case, the required estimation error is C ₀ =
0.0316. FIG. 5 shows C ₀ = 0.0316, C ₀ = 0.01 (echo cancellation amount 20 dB).
), C ₀ = 0.001 (corresponding to the echo cancellation amount of 30 dB), and the convergence characteristic obtained by the control method (iv) is compared with the convergence characteristic obtained by fixing the step size to 16. is there. This convergence characteristic gives the ideal convergence speed. Here, the reason why the simulation was performed using (iv) instead of (iii), which made the control method desirable as described above, is that when the required estimation error is set to be equal to or greater than the power ratio of the disturbance versus the reference signal, both converge. This is because the same result can be obtained by using either of them because the characteristics match.

According to the results shown in FIG. 5, the time required for convergence is the case of C ₀ = 0.001 in the proposed method (iv), and the ideal characteristic with the step size fixed at 16 is about 3.5, C ₀ = It is about 1.5 times at 0.01 and about 1.2 times at C ₀ = 0.0316. From this result, it can be seen that the proposed method (iv) gives a characteristic very close to the ideal characteristic with respect to the echo path variation according to the standard of ITU-T G.167. Here, in the case of C ₀ = 0.001 which is the slowest convergence, a further increase in the convergence will be examined.

The problem is the operation during double talk. FIG. 6 shows the convergence characteristics when the step size is fixed at 16 and the coefficient is continuously updated even when the reference signal to disturbance ratio temporarily increases to −10 dB when the estimation error decreases to −20 dB. Naturally, the estimation error increases to -10 dB because the coefficient is continuously updated. In order to prevent this increase, the method of fixing the step size to 16 discriminates the echo path fluctuation and the double talk at the time of the occurrence of the double talk, and if double talk, the coefficient update is stopped or the disturbance included in the residual echo. It is necessary to control the step size according to the ratio. However, if the ratio is found, it is not necessary to distinguish between echo path variations and double talk. The proposal here is a proposal of a control method that makes the identification unnecessary.

In contrast, Fig. 7 shows the convergence characteristics obtained by applying the proposed method (iv). However, the required estimation error is C ₀ = 0.001 (corresponding to an echo cancellation amount of 30 dB). However, the disturbance is increased so that the reference signal to disturbance ratio is temporarily -10 dB when the estimation error is reduced to -20 dB as in FIG. Obviously, according to the proposed method (iv), the estimation error continues to decrease even in the double talk section, and is saturated with an estimation error of -30 dB. This is because the required estimation error is set to -30 dB. After that, with the end of the double talk, the estimation error quickly decreases again to a power ratio of disturbance to reference signal of −40 dB. From this, it can be seen that the proposed method can flexibly cope with the case where double talk occurs during the convergence of the coefficients. In real systems there is no guarantee that double talk will only occur after convergence of the coefficients. Thus, it can be said that the proposed method that can automatically cope with the double talk that occurs during the convergence of the coefficients is more practical in this respect.

FIG. 8 shows the results of a simulation performed for the proposed method (iii) under the same conditions as in FIG. When double talk occurs in this way, the coefficient of the sub-adaptive filter is greatly disturbed, and the disturbance of the coefficient is larger than ej.

Increases significantly. In the proposed method (iv)

Since the step size is controlled by selecting the smaller one of ej and ej, the step size becomes smaller in (iii) than in (iv), and the convergence speed in the double talk section becomes lower than in (iv). With this ej

This difference also occurs in double talk after the convergence of coefficients. That is, when double talk occurs after the coefficient converges,

Is larger than ej, and the step size is set smaller. As a result, the estimation error is further reduced in (iii). At first glance, this seems like a desirable property. However, the amount of echo cancellation required in double talk is designed as the required estimation error C ₀ , and there is no need to reduce the echo further. For an actual system, it is more useful to improve the convergence speed than the amount of echo cancellation in double talk is more than the design value. Therefore, here we consider (iv) as the proposed method.
(5. Verification by voice)
Here, the effectiveness of the proposed method (iv) is verified using speech as a reference signal. In this way, when the reference signal is speech, it is necessary to increase the frequency of coefficient update by setting the shortest block length short in order to suppress the influence of autocorrelation. On the other hand, in order to accurately calculate the disturbance power Qn required for step size control, it is necessary to select a long block length. In this verification, the calculation accuracy of the disturbance power Qn is given priority.
(1) Shortest block length: J ₀ = 16.

Next, when the adaptive algorithm is a learning identification method, it is desirable that the step size μ is as close to 1 as possible. However, when speech is used as a reference signal, the influence of the autocorrelation cannot be ignored. In particular, when the learning identification method is a block execution type, it is necessary to set a small step size. According to [Non-Patent Document 3], the step size μ needs to be about 0.1. Here we show the effectiveness of the proposed method
(2) Step Size Limit: verifying (corresponding to _{ρ 0 = μ 0 J 0 =} 2.0) μ 0 = 1/8.

The remaining constants that need to be set for this control are the required estimation error C ₀ , the average power ratio of the reference signal to the disturbance, and the threshold value P ₀ that determines the coefficient update. In this verification,
(3) Required estimation error: C ₀ = 0.001 (equivalent to an echo cancellation amount of 30 dB)
(4) Reference signal: Audio read from CD-ROM + white noise (corresponding to -30 as the power ratio of audio to audio)
(5) Disturbance: white noise
(6) Average power ratio of reference signal to disturbance: 30dB
And Next, these are transformed into the equation (10)

If you assign to. From the above condition (6), Q ₀ = 0.001σ _x ²

And obtained. That is,
(7) Update decision threshold P ₀ = σ _x ² I
And Other conditions are the same as in FIG.

Figure 9 compares the ideal convergence characteristic with the step size fixed at the fastest ρ ₀ = 2 and the convergence characteristic obtained by the proposed method (iv). The upper part of Fig. 10 shows the speech waveform used as the reference signal. According to this result, the proposed method has a slightly slower convergence characteristic than the ideal characteristic. The reason is that when the reference signal is speech as shown in [Non-Patent Document 1], the echo cancellation amount increases quickly even if the estimation error is delayed. That is, as can be seen from Fig. 11, the residual echo

Decreases rapidly in the case of speech, which increases the calculated step size.

FIG. 12 shows the step size calculated from this residual echo. From this result, it can be seen that the step size is set small in the speech interval at the early stage of convergence, but rapidly increases. In addition, the residual echo is reduced in the non-voice interval and the step size is set to 2 which is the upper limit, which is effective for speeding up the convergence. This also has an advantage in updating the coefficient even in a section where there is no speech by block length control.

FIG. 13 shows the results of a simulation confirming that the proposed method (iv) can stably reduce the estimation error even for double talk. Obviously, in the proposed method, the estimation error decreases stably even in the case of double talk. The lower waveform in Fig. 10 is the waveform of the near-end speaker voice added as a double talk.
(6. Summary)
Here, a sub-adaptive filter with a small number of taps is prepared separately, and the step size is controlled by the power of the residual echo calculated using the filter output, thereby speeding up the convergence faster than the conventional method [Non-Patent Document 5]. I showed that I can do it. In addition, when using speech as a reference signal, it was confirmed by simulation that almost the same convergence characteristic as that obtained when a step size for maximizing convergence was given. This solves the problem of discriminating between echo path fluctuation and double talk, which has been an obstacle to the realization of an acoustic echo canceller.

In the future, we plan to actually construct the system and verify the effectiveness of the proposed method. In addition, an adaptive algorithm [Non-Patent Document 9] proposed for further speeding up the convergence will be applied and its effect will be confirmed.

  According to the echo canceller of the present application, the filter coefficients of the main adaptive filter can be converged at high speed without discriminating between the echo path fluctuation and the double talk, and can be used in the technical field of telecommunication devices.

It is a structural diagram of an acoustic echo canceller to which a conventional control method is applied. 1 is a structural diagram of a system according to the present invention. It is a figure which shows the comparison of a convergence characteristic. It is a figure which shows the comparison of a convergence characteristic. It is a figure which shows the comparison with the convergence characteristic when a step size is fixed at 16. It is a figure which shows the influence which double talk has on a convergence characteristic. It is a figure which shows the influence which double talk has on a convergence characteristic. It is a figure which shows the influence which double talk has on a convergence characteristic. It is a figure which shows the comparison with this invention when using an audio | voice as a reference signal. It is a wave form diagram of a far-end speaker voice and a near-end talker voice. It is a wave form diagram of an acoustic echo and a residual echo. It is a figure which shows transition of step size (rho). It is a figure which shows the convergence characteristic at the time of double talk.

Claims (6)

An echo canceller that removes an echo signal generated by a supply signal supplied to a speaker from the speaker to the microphone from the output signal of the microphone,
An FIR type main adaptive filter capable of updating the filter coefficient for inputting the supply signal as a reference signal and generating a pseudo echo signal as an output signal;
A main subtractor for generating a main error signal by subtracting the pseudo echo signal generated by the main adaptive filter from the output signal of the microphone;
An FIR type sub-adaptive filter capable of updating a filter coefficient for inputting the supply signal as a reference signal and generating an output signal;
A subtractor for generating a sub error signal,
The main adaptive filter updates the coefficient of the filter based on the supply signal and the main error signal by an adaptive algorithm so that the power of the main error signal is minimized,
The sub-adaptive filter updates the coefficient of the filter by an adaptive algorithm based on the supply signal and the sub-error signal so that the power of the sub-error signal is minimized,
The number of filter taps I ₁ of the sub-adaptive filter is smaller than the number of filter taps I of the main adaptive filter,
The sub-subtracter generates the sub-error signal by subtracting the added value of the latter half tap output value of the main adaptive filter and the tap output value of the sub-adaptive filter from the output signal of the microphone,
The latter half tap output value of the main adaptive filter is obtained by multiplying each signal subsequent to I ₁ on the delay line for sequentially delaying the input reference signal by the corresponding filter tap coefficient in the main adaptive filter. The sum of the multiplication values obtained,
The tap output value of the sub-adaptive filter is a sum of multiplication values obtained by multiplying each signal on the delay line that sequentially delays the input reference signal by the corresponding filter tap coefficient in the sub-adaptive filter. ,
The main adaptive filter is an echo canceller that updates a step size in updating the filter coefficient based on the sub error signal.   The echo canceller according to claim 1, wherein both the adaptation algorithm of the main adaptive filter and the adaptation algorithm of the sub-adaptive filter are learning identification methods.   The main adaptive filter includes a step size in updating the filter coefficient, a multiplication value of a predetermined value and an addition norm of the supply signal vector, an average power of the sub error signal, and a filter tap number I of the main adaptive filter. The echo canceller according to claim 1, wherein the echo canceller is updated to a value divided by the multiplication value.   The echo canceller according to claim 1, wherein both of the adaptive algorithm of the main adaptive filter and the adaptive algorithm of the sub adaptive filter are block execution type learning identification methods. The main adaptive filter includes a step size in updating the filter coefficient, a multiplication value of a predetermined value and an addition norm of the supply signal vector, an average power of the sub error signal, and a filter tap number I of the main adaptive filter. Update to the value divided by the multiplication value,
The addition norm is an addition norm in the block length J of the block execution type learning identification method,
The echo canceller according to claim 4, wherein the average power is an average power in the block length J. Number of filter taps I ₁ of the sub adaptive filter, is approximately one quarter of the number of filter taps I of the main adaptive filter echo canceller according to any one of claims 1 to 5.

Images