JP4094522B2 - Echo canceling apparatus, method, echo canceling program, and recording medium recording the program - Google Patents

Description

  The present invention relates to an erasing apparatus and method for erasing an echo that circulates from a speaker to a microphone, an echo elimination program, and a recording medium on which the program is recorded.

  An echo canceling apparatus that cancels echo that circulates from a speaker to a microphone is configured as shown in FIG. Conventionally, in the echo canceling apparatus 100, a simulated echo path 101 that retains a simulation characteristic h ′ (k) of a vector h of length L in the time domain having an impulse response of the echo path between the speaker 1 and the microphone 2 as an element. Have Here, k is the number of steps indicating discrete time at a predetermined interval. The reproduced signal x (k) generates a simulated reverberation signal y ′ (k) by convolution with the simulated characteristics h ′ (k), and the simulated error calculation means 102 collects the microphone 2 including the actual reverberation signal. By subtracting from the sound signal y (k), an error signal e (k) that is also an output signal of the echo canceller is calculated. Further, the echo path estimation means 103 estimates the characteristic of the characteristic h ′ (k) as needed using the reproduced signal x (k) and the error signal e (k).

The estimation of the simulation characteristic h ′ (k) in the echo path estimation unit 103 is performed using the method described in (Non-patent Document 1).
h ′ (k) = [h ₁ (k), h ₂ (k),..., h _L (k)] Each element h _i (k) represented by ^T (i = 1, 2,...・ About L)

と分解した場合の（非特許文献１）に基づく反響消去装置の構成である。ここで、sgn(x(k-i))は、x(k-i)／|x(k-i)|と等しく、x(k-i)の符号を表す。
図２の反響路推定手段１０３における再生信号列正規化手段１１１は、（式２）における

の計算に相当し、誤差ベクトル化手段１１２は、（式２）におけるe(k-1)／|x(k-i)|をｉ＝１，２，・・・，Ｌに対して計算し、Ｌ個の絶対値除算誤差e(k-1)／|x(k-i)|を各要素に持つ絶対値除算誤差ベクトルとして出力する。また、特性差分更新手段１１３は、再生信号列正規化手段１１１、誤差ベクトル化手段１１２の結果を基に、（式２）に基づき、ｉ＝１，２，・・・，Ｌに対する要素を計算し、１ステップ過去に推定された模擬特性h'(k-1)をステップｋで用いるh'(k)に更新する手段である。
J.Nagumo,A.noda“A learning method for system identification”IEEE Trans.Automatic control, Vol.AC-12,No.3, June 1967 FIG. 2 shows (Equation 1) by the inventor in order to easily compare the difference from the configuration according to the present invention described later.

This is a configuration of an echo canceling device based on (Non-Patent Document 1). Here, sgn (x (ki)) is equal to x (ki) / | x (ki) | and represents the sign of x (ki).
The reproduction signal sequence normalization means 111 in the echo path estimation means 103 in FIG.

The error vectorization means 112 calculates e (k−1) / | x (ki) | in (Equation 2) for i = 1, 2,... The absolute value division error e (k-1) / | x (ki) | is output as an absolute value division error vector having each element. Further, the characteristic difference update unit 113 calculates elements for i = 1, 2,..., L based on (Expression 2) based on the results of the reproduction signal sequence normalization unit 111 and the error vectorization unit 112. The simulation characteristic h ′ (k−1) estimated one step in the past is updated to h ′ (k) used in step k.
J. Nagumo, A. noda “A learning method for system identification” IEEE Trans. Automatic control, Vol. AC-12, No. 3, June 1967

は、全てのｉ＝１，２，・・・，Ｌに対して、|x(k-i)|≦||x(k-1)||が成り立つから、常に１を越えることはない。一方、誤差ベクトル化手段１１２の出力である絶対値除算誤差ベクトルの各要素e(k-1)／|x(k-i)|は、その大きさが一定ではない。特に、大きさの小さい再生信号x(k-i)に対し、誤差信号e(k-1)が大きい場合には、e(k-1)／|x(k-i)|の値は非常に大きくなる。大きくなる原因が、本来の更新に意味を持つものであれば問題は無いが、多くの場合、もともとの収音信号y(k)に反響信号以外の外乱が多く含まれることに起因する。このため、不要に大きな模擬特性h'(k)の更新を実行してしまい、反響消去性能は、大きく乱れる。
本発明の課題は、収音信号y(k)に反響信号以外の外乱が多く含まれることに起因する
e(k-1)／|x(k-i)|（ｉ＝１，２，・・・，Ｌ）の大きさの増大のために、模擬特性h'(k)の更新が不適切となり、反響消去性能が劣化することを防ぐことである。 The problem in the structure of FIG. 2 demonstrated by (background art) is shown.
Each element of the normalized reproduction signal sequence which is the output of the reproduction signal sequence normalization means 111 inside the echo path estimation means 103

Since | x (ki) | ≦ || x (k−1) || holds for all i = 1, 2,..., L, it does not always exceed 1. On the other hand, the magnitude of each element e (k−1) / | x (ki) | of the absolute value division error vector which is the output of the error vectorization means 112 is not constant. In particular, when the error signal e (k−1) is large with respect to the small reproduction signal x (ki), the value of e (k−1) / | x (ki) | becomes very large. There is no problem as long as the cause of the increase is significant in the original update, but in many cases, the original sound collection signal y (k) includes many disturbances other than the echo signal. For this reason, the large simulation characteristic h ′ (k) is updated unnecessarily, and the echo cancellation performance is greatly disturbed.
The problem of the present invention is that the collected sound signal y (k) includes many disturbances other than the echo signal.
Due to the increase in the size of e (k-1) / | x (ki) | (i = 1, 2,..., L), the update of the simulation characteristic h ′ (k) becomes inappropriate and the response This is to prevent the erasing performance from deteriorating.

The echo canceling apparatus and method of the present invention for solving the above problems will be described with reference to FIG.
3 has a simulated echo path 101 and simulated error calculation means 102 as in FIG. 2, and differs from the configuration of FIG. 2 in that the output of the error vectorization means 112 is provided inside the echo path estimation means 103. Error vector nonlinear conversion means 114 for inputting each element e (k-1) / | x (ki) | (i = 1, 2,..., L) of the absolute value division error vector. The error vector nonlinear conversion means 114 has a large size according to the size of each element of the input e (k−1) / | x (ki) | (i = 1, 2,..., L). A value smaller than the actual size of e (k-1) / | x (ki) | That is, by applying an arbitrary function ψ (a) (herein, ψ (a) is referred to as a limiter function) having a characteristic of suppressing the input value a as the input value a increases, the error vector nonlinear conversion means 114 is ψ (e (k−1) / | x (ki) |) (i = 1, 2,..., L) is output.

  The echo canceling apparatus and method according to the present invention provide each element e (k−1) / | x () of the absolute value division error vector resulting from the fact that the collected sound signal y (k) contains many disturbances other than the echo signal. ki) | (i = 1, 2,..., L) prevents the improper update of the simulation characteristic h ′ (k) by providing an error vector nonlinear conversion means that prevents an increase in size. Even if the collected sound signal y (k) includes a disturbance, the echo canceling performance is not deteriorated. In particular, when the echo canceling apparatus is applied to a communication conference, the disturbance is ambient noise generated in a room in which speakers and microphones are installed or voice of a speaker in the room. When the upper limit of the gain of the echo path between the speaker and the microphone can be assumed, and when the influence of the disturbance is small, each element e (k-1) / | x (ki ) | (I = 1, 2,..., L) also has an absolute value division error proportional to the upper limit value of the gain of the echo path, and does not increase beyond this. On the other hand, when the voice of the speaker in the room becomes a disturbance and the voice of the communication partner is simultaneously spoken, each element e (k-1) / | x (ki) | (i = 1) of the absolute value division error vector , 2,..., L) can be larger than the absolute value division error upper limit value assumed when the influence of the above-described disturbance is small. Therefore, by performing error vector nonlinear conversion means according to the present invention so that each element of the absolute value division error vector does not exceed the absolute value division error upper limit value, echo cancellation can be performed even in simultaneous speech with the communication partner. Performance degradation can be prevented.

平均値を瞬時値に置き換えるとこの関係は成り立たない場合があるが、それは、特に再生信号列の要素x(k-i)（ｉ＝１，２，・・・，Ｌ）の中で小さい値を持つ要素に対してである。このため仮に、収音信号y(k)に含まれる外乱が小さい場合でも、e(k-1)／|x(k-i)|（ｉ＝１，２，・・・，Ｌ）の上限値をACLと制限することの影響は小さい。従って、適切なリミッタ関数ψ(a)の例として、

が挙げられる。ここで、βは、反響路の上限値として想定されたACLの値の信頼度に応じ、調整するための係数で、ACLが上限値として信頼できるのであれば、β＝１が最も理に適った値といえる。 In the present invention, it is important to select the characteristic of the aforementioned limiter function ψ (a) used in the error vector nonlinear transformation means 114 in FIG. Here, assuming that the upper limit of the gain of the echo path is ACL,

If the average value is replaced with an instantaneous value, this relationship may not hold, but it has a small value among the elements x (ki) (i = 1, 2,..., L) of the reproduction signal sequence. Is against the element. Therefore, even if the disturbance included in the collected sound signal y (k) is small, the upper limit value of e (k−1) / | x (ki) | (i = 1, 2,..., L) is set. The impact of restricting with ACLs is small. Therefore, as an example of a suitable limiter function ψ (a),

Is mentioned. Here, β is a coefficient for adjustment according to the reliability of the ACL value assumed as the upper limit value of the echo path. If ACL is reliable as the upper limit value, β = 1 is most reasonable. It can be said that.

As an embodiment of the present invention, an outline of the characteristic of a function that can be applied other than the example of the limiter function ψ (a) is shown in FIG.
When the influence of disturbance is small, the norm division error does not exceed the norm division error upper limit value. In a sufficiently reliable implementation environment, for a norm division error of a predetermined magnitude or more as in alternative 5, A characteristic having a decreasing tendency may be given.

The configuration shown in FIG. 5 is an example of another embodiment of the device of the present invention.
The signal conversion units 121 to 129 and the signal conversion units 131 to 133 do not all function at the same time, but realize some different processing modes by functioning some of them. Signal conversions realized by them include frequency conversion and inverse frequency conversion. Furthermore, frequency conversion includes (fast) Fourier transform, (fast) discrete cosine transform, subband band division, etc., and inverse frequency transform includes inverse (fast) Fourier transform, inverse (fast) discrete cosine transform, sub For example, band band synthesis.
For example, (1) when the signal conversion units 121 and 124 in FIG. 5 perform frequency conversion and the signal conversion unit 126 performs inverse frequency conversion, the simulated echo path 101, the simulated error calculation means 102, and the echo path estimation means 103 All function for each frequency with respect to the signal after frequency conversion. (2) Alternatively, frequency conversion may be performed in the signal conversion units 131 and 132 provided outside the echo canceling apparatus 100 instead of the signal conversion units 121, 124, and 126, and inverse frequency conversion may be performed in the signal conversion unit 133. It is the same. However, in this case, the signal input / output by the echo canceller is a frequency conversion signal. (3) As another form, there is a form in which the signal conversion unit 121 performs frequency conversion, the signal conversion unit 123 performs inverse frequency conversion, and the signal conversion unit 128 performs frequency conversion. In this case, the simulated echo path 101 and the echo path estimation means 103 process the frequency conversion signal, but the simulation error calculation means 102 performs the simulation error calculation while keeping the time signal. (4) Further, as another form, there is a form in which the signal conversion units 127 and 128 perform frequency conversion and the signal conversion unit 129 performs reverse frequency conversion. In this case, only the echo path estimation means 103 is executed for the frequency conversion signal, and the simulated echo path 101 and the simulation error calculation means 102 execute the processing while keeping the time signal. In the inverse frequency conversion in the signal conversion unit 129, the conversion from the frequency domain characteristic to the time domain characteristic is performed on the estimated value of the simulated echo path characteristic, not on the signal.

  In the above example, it is assumed that signal converters not mentioned are not functioning. Whether or not the signal converters 127 and 128 are included in the echo path estimation unit 103 is not the essence of the present invention, and the same effect can be obtained in any case.

  In the second embodiment, when the processing of the echo path estimation unit 103 is executed in the frequency domain, the reproduction signal vector x (k) and the error signal e (k) corresponding to each frequency are included. Become. Here, in the case of performing frequency conversion with time decimation, the length of the reproduction signal vector x (k) corresponding to each frequency does not necessarily match the length L given in the time domain. It's okay. That is, according to the time thinning interval, for example, when the frequency is converted by thinning the time interval to 1 / R, the length is set to L / R. For example, x (k) Can handle. Also, the ACL value assumed as the upper limit value of the gain in the echo path may be given as a different value for each frequency.

となり、ｉ＝１，２，・・・，Ｌの中で最大値をとる要素x(k-i)と誤差信号e(k-1)との比のみ評価される。また、このとき、図３または図５における再生信号列正規化手段１１１は、（式５）における、

の計算を実行することになる。 According to the present invention, the estimation of each element of the simulated characteristic h ′ (k) is written in contrast to (Equation 2).

Thus, only the ratio between the element x (ki) having the maximum value in i = 1, 2,..., L and the error signal e (k−1) is evaluated. At this time, the reproduction signal string normalizing means 111 in FIG. 3 or FIG.

Will be executed.

  The embodiment shown in FIG. 6 is an embodiment that can be applied to the case where the ACL that is the assumed upper limit value of the gain of the echo path used in (Equation 3) is unknown or the accuracy of the assumed value is to be improved. Means 115 are provided. The signal conversion units 121 to 129 and the signal conversion units 131 to 133 illustrated in FIG. 6 can perform frequency conversion and inverse frequency conversion processing as necessary, as shown in (Example 2). This includes the case where none of the signal conversion units 121 to 129 and the signal conversion units 131 to 133 function (bypass). The acoustic coupling amount evaluation means 115 measures the level ratio of the reproduced signal and the collected sound signal or the reproduced signal and the error signal and gives it as an ACL, as shown in (Example 2). When the path estimation unit 103 functions, the acoustic coupling amount evaluation unit 115 includes a function of calculating an ACL value for each frequency.

The echo canceling apparatus of the present invention can be composed of a computer having a CPU, a memory, and the like, a user terminal, and a recording medium such as a CD-ROM, a magnetic disk device, and a semiconductor memory.
The echo canceling program recorded on the recording medium or the echo canceling program transmitted via the line is read by the computer, realizes each component described above on the computer, and executes each process.

  It can be applied to hands-free calling and hands-free speech recognition.

The figure which shows the basic composition of an echo cancellation apparatus. The figure which shows the structure of the echo cancellation apparatus which has the conventional echo path estimation means. The figure which shows the structure of the echo cancellation apparatus which has the echo path estimation means of this invention. The figure which shows the alternative example of a limiter function. The figure which shows the structure of the Example of the echo cancellation apparatus which has a signal conversion part of this invention. The figure which shows the structure of the Example of the echo cancellation apparatus which has a signal conversion part and acoustic coupling amount measurement means of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Speaker (acoustic signal reproduction | regeneration means), 2 ... Microphone (acoustic signal sound collection means)
DESCRIPTION OF SYMBOLS 100 ... Echo cancellation apparatus 101 ... Simulated echo path, 102 ... Simulated error calculation means 103 ... Echo path estimation means, 111 ... Reproduction signal sequence normalization means, 112 ... Error vectorization Means 113, characteristic difference updating means 114, error vector nonlinear conversion means 121-129, 131-133, signal conversion unit

Claims (12)

Reproducing the reproduction signal in response to the input of the reproduction signal to be reproduced from the acoustic reproduction means while retaining the simulated characteristics of the echo path that is the acoustic transmission path between the acoustic reproduction means and the acoustic sound collection means existing in the same space A simulated reverberation path that outputs a simulated reverberation signal that simulates a reverberation signal obtained by collecting the signal by the acoustic sound collection means, a collected sound signal that is all signals collected by the acoustic sound collection means, and the simulation Simulated error calculation means for inputting an echo signal and outputting an error signal that is the difference between them, the reproduction signal and the error signal are inputted, the characteristics of the echo path are estimated, and the simulated echo path is held Reverberation path estimating means for updating the simulated characteristics, and in the echo canceling apparatus for canceling the echo signal included in the collected sound signal,
The echo path estimation means includes
The reproduction signal is input as a finite-length time sequence, and the result obtained by dividing the absolute value of each element of the reproduction signal sequence by the norm of the reproduction signal sequence itself is raised to the power of each element, and the result is the reproduction signal sequence. Reproduction signal sequence normalizing means for multiplying the sign of each element and outputting as a normalized reproduction signal sequence,
Input the error signal and the absolute value of each element of the reproduction signal string or the reproduction signal string, divide the error signal by the absolute value of each element of the reproduction signal string, and the number of elements of the reproduction signal string An error vectorization means for outputting an absolute value division error vector as a division result equal to
An error vector for inputting the absolute value division error vector and outputting a nonlinear transformation absolute value division error vector having a value smaller than the size of the element as the size of each element of the absolute value division error vector is larger Non-linear conversion means;
The normalized reproduction signal sequence and the nonlinear transformation absolute value division error vector are input, the difference between the characteristics of the echo path and the simulated echo path is estimated, and the estimated difference characteristics and the characteristics of the simulated echo path are Characteristic difference updating means for outputting the estimated characteristics of the new echo path,
An echo canceling apparatus characterized by comprising: The echo canceling device according to claim 1,
An echo canceling apparatus, wherein the nonlinear transformation in the error vector nonlinear transformation means is realized by a limiter function that suppresses a large value exceeding a predetermined magnitude (upper limit value). In the echo cancellation apparatus according to claim 2,
An acoustic coupling amount measurement unit is provided that calculates a ratio between the reproduced signal and the collected sound signal, or the reproduced signal and the error signal, and determines an upper limit value used in the limiter function based on the ratio. Echo canceling device. Reproducing the reproduction signal in response to the input of the reproduction signal to be reproduced from the acoustic reproduction means while maintaining the simulation characteristics of the echo path that is an acoustic transmission path between the acoustic reproduction means and the sound collection means existing in the same space A collected sound signal that is a signal generated by the acoustic sound collecting means to generate a simulated echo signal that simulates an echo signal obtained by the acoustic sound collecting means and constitutes a simulated echo path, and is a signal that is collected by the acoustic sound collecting means. And the simulated echo signal are input, an error signal which is a difference between them is obtained, the reproduction signal and the error signal are input, the characteristics of the echo path are estimated, and the simulated echo path holds In the echo canceling method for canceling the echo signal included in the collected sound signal by updating the characteristics,
Estimating the characteristics of the echo path is
The reproduction signal is input as a finite-length time sequence, and the result obtained by dividing the absolute value of each element of the reproduction signal sequence by the norm of the reproduction signal sequence itself is raised to the power of each element, and the result is the reproduction signal sequence. Is multiplied by the sign of each element to obtain a normalized reproduction signal sequence,
Input the error signal and the absolute value of each element of the reproduction signal string or the reproduction signal string, divide the error signal by the absolute value of each element of the reproduction signal string, and the number of elements of the reproduction signal string Obtain the absolute value division error vector as the division result equal to
The absolute value division error vector is input, and the larger the size of each element of the absolute value division error vector is, the larger the magnitude of each element is, the non-linear transformation absolute value division error vector whose value is smaller than the element is generated.
The normalized reproduction signal sequence and the nonlinear transformation absolute value division error vector are input, a difference between the echo path and the simulated echo path characteristic is estimated, and the estimated difference characteristic and the simulated echo path characteristic And a new echo path estimation characteristic is generated. In the echo cancellation method according to claim 4,
An echo canceling method characterized in that non-linear transformation for each element of the absolute value division error vector is realized by a limiter function that suppresses a large value exceeding a predetermined magnitude (upper limit value). In the echo cancellation method according to claim 5,
An echo canceling method characterized in that an upper limit value used in the limiter function is determined based on a ratio between the reproduction signal and the sound pickup signal or a ratio between the reproduction signal and the error signal. Reproducing the reproduction signal in response to the input of the reproduction signal to be reproduced from the acoustic reproduction means while maintaining the simulation characteristics of the echo path that is an acoustic transmission path between the acoustic reproduction means and the sound collection means existing in the same space A process of generating a simulated echo signal that simulates an echo signal obtained by collecting the signal by the acoustic sound pickup means to construct a simulated echo path, and a collection of all signals collected by the acoustic sound pickup means. A process of inputting a sound signal and the simulated echo signal, obtaining an error signal that is a difference between them, a process of inputting the reproduction signal and the error signal, estimating a characteristic of the echo path, and the simulated echo In the echo canceling program for canceling the echo signal included in the collected sound signal, for causing the computer to execute the process of updating the simulation characteristics held by the road,
The process of estimating the characteristics of the echo path is as follows:
The reproduction signal is input as a finite-length time sequence, and the result obtained by dividing the absolute value of each element of the reproduction signal sequence by the norm of the reproduction signal sequence itself is raised to the power of each element, and the result is the reproduction signal sequence. A process of obtaining a normalized reproduction signal sequence by multiplying the sign of each element of
Input the error signal and the absolute value of each element of the reproduction signal string or the reproduction signal string, divide the error signal by the absolute value of each element of the reproduction signal string, and the number of elements of the reproduction signal string A process of obtaining an absolute value division error vector as a division result equal to
A process of inputting the absolute value division error vector, and generating a nonlinear transformation absolute value division error vector having a value smaller than the size of the element as the size of each element of the absolute value division error vector is larger ,
The normalized reproduction signal sequence and the nonlinear transformation absolute value division error vector are input, a difference between the echo path and the simulated echo path characteristic is estimated, and the estimated difference characteristic and the simulated echo path characteristic And a process for generating an estimated characteristic of a new echo path, and an echo canceling program. In the echo cancellation program according to claim 7,
An echo canceling program that implements nonlinear transformation for each element of the absolute value division error vector in the process of generating the nonlinear transformation absolute value division error vector by a limiter function that suppresses a large value exceeding a predetermined size (upper limit value). . In the echo cancellation program according to claim 8,
An echo canceling program having processing for calculating an upper limit value used in the limiter function by calculating a ratio between the reproduction signal and the collected sound signal or a ratio between the reproduction signal and the error signal and based on the ratio. Reproducing the reproduction signal in response to the input of the reproduction signal to be reproduced from the acoustic reproduction means while maintaining the simulation characteristics of the echo path that is an acoustic transmission path between the acoustic reproduction means and the sound collection means existing in the same space A process of generating a simulated echo signal that simulates an echo signal obtained by collecting the signal by the acoustic sound pickup means to construct a simulated echo path, and a collection of all signals collected by the acoustic sound pickup means. A process of inputting a sound signal and the simulated echo signal, obtaining an error signal that is a difference between them, a process of inputting the reproduction signal and the error signal, estimating a characteristic of the echo path, and the simulated echo In a recording medium recorded with an echo canceling program for canceling an echo signal included in the collected sound signal for causing a computer to execute a process of updating a simulation characteristic held by a road,
The process of estimating the characteristics of the echo path is as follows:
The reproduction signal is input as a finite-length time sequence, and the result obtained by dividing the absolute value of each element of the reproduction signal sequence by the norm of the reproduction signal sequence itself is raised to the power of each element, and the result is the reproduction signal sequence. A process of obtaining a normalized reproduction signal sequence by multiplying the sign of each element of
Input the error signal and the absolute value of each element of the reproduction signal string or the reproduction signal string, divide the error signal by the absolute value of each element of the reproduction signal string, and the number of elements of the reproduction signal string A process of obtaining an absolute value division error vector as a division result equal to
A process of inputting the absolute value division error vector and generating a nonlinear transformation absolute value division error vector having a value smaller than the size of the element as the size of each element of the absolute value division error vector is larger ,
The normalized reproduction signal sequence and the nonlinear transformation absolute value division error vector are input, a difference between the echo path and the simulated echo path characteristic is estimated, and the estimated difference characteristic and the simulated echo path characteristic And a process for generating a new reverberation path estimation characteristic. In the recording medium which recorded the echo cancellation program of Claim 10,
An echo canceling program that implements nonlinear transformation for each element of the absolute value division error vector in the process of generating the nonlinear transformation absolute value division error vector by a limiter function that suppresses a large value exceeding a predetermined size (upper limit value). A recording medium on which is recorded. In the recording medium which recorded the echo cancellation program of Claim 11,
An echo cancellation program having a process of calculating the ratio between the reproduced signal and the collected sound signal or the reproduced signal and the error signal and determining the upper limit value used in the limiter function based on the ratio is recorded. recoding media.

Images