JP4425114B2 - Echo canceling method, echo canceling apparatus, echo canceling program, and recording medium recording the same - Google Patents

Description

  The present invention relates to an echo canceling method, an echo canceling apparatus, a program, and a recording medium for canceling echo that circulates from a speaker to a microphone.

An echo canceling device that cancels the echo that circulates from the speaker as the sound reproducing means to the microphone as the sound collecting means is connected as shown in FIG. Conventionally, in echo canceller 10, simulated characteristic h of vectors h _* length L having an impulse response of the echo path between the acoustic reproducing means 1 and sound pickup means 2 as elements _(hereinafter * _* denotes a vector) It has a simulated echo path 11 holding ′ _* (k). Here, k is the number of steps indicating discrete time at a predetermined interval. The reproduction signal x (k) generates a simulated echo signal d ′ (k) by convolution with the simulation characteristic h ′ _* (k), and the simulated error calculation means 12 converts the actual echo signal d (k) to By subtracting the simulated echo signal d ′ (k) from the collected sound signal y (k) of the included sound pickup means 2, an error signal e (k) that is also an output signal of the echo canceller 10 is calculated. The echo path estimation means 13 estimates the characteristic of the characteristic h ′ _* (k) as needed using the reproduction signal x (k) and the error signal e (k).

と表せる誤差信号ベクトルである。 The estimation of the simulation characteristic h ′ _* (k) in the echo path estimation means 13 uses the affine projection algorithm described in (Non-Patent Document 1).
h ′ _* (k + 1 ) = h ′ _* ( k) + μ (k) X _* (k) [X _* ^T (k) X _* (k)] ⁻¹ e (k) (Formula 1)
It becomes. Here, x _* (k) = [x (k), x (k−1),..., X (k−L + 1)] ^T , X _* (k) = [x _* (k), x _* (k −1),..., X _* (k−p + 1)] ^T and T are transposes of vectors or matrices, μ (k) is a fixed or time-varying update adjustment coefficient, and takes a value between 0 and 2. Also,

It is an error signal vector that can be expressed as

Now, in the case where the influence of the disturbance included in the collected sound signal y (k) can be ignored and can be regarded as y (k) = ˜d (k), e _* (k) = ˜X _* ^T (k) [h _* −h ′ _* (k)]. In addition, A = -B represents that A and B approximate. At this time, for the update vector of the second term on the right side of (Equation 1),
X _* (k) [X _* ^T (k) X _* (k)] ⁻¹ e _* (k) = ˜X _* (k) [X _* ^T (k) X _* (k)] ⁻¹ X _* ^T (k) [h _* −h ′ _* (k)]
= P _{p *} (k) [h _* −h ′ _* (k)] (Formula 3)
Is considered to hold. Here, P _{p *} (k) = X _* (k) [X _* ^T (k) X _* (k)] ⁻¹ X _* ^T (k) is a reproduction signal vector x _* (k), x _* ( k−1),..., x _* (k−p + 1) is a projection matrix onto a p-dimensional subspace. Therefore, the affine projection algorithm uses the orthogonal projection vector of the error characteristics h _* −h ′ _* (k) of the echo path and the simulated echo path to the subspace characterized by the projection matrix P _{p *} (k). It can be said that this algorithm is ideal for use in road updating.

FIG. 3 shows a configuration of a conventional echo canceling apparatus 100 based on the affine projection algorithm. That is, the reproduction signal vectorization means 101 for accumulating a finite number of samples of the reproduction signal x (k), vectorizing it and outputting the reproduction signal vector x _* (k), and the echo characteristic h ′ _* (k) of the echo path, A simulated echo path 102 that inputs a signal vector x _* (k) and outputs a simulated echo signal d ′ (k) that simulates the echo signal, and a simulated echo signal d ′ (k) from the collected sound signal y (k). Subtraction, output error calculation means 103 for outputting an output error signal e (k), and a reproduction signal for accumulating the reproduction signal vector x (k) over a predetermined number of steps in order to generate a reproduction signal matrix X _* (k) Vector delay accumulating means 104 and error signal vectorizing means 105 for generating an error signal vector e _* (k) by accumulating the error signal e (k) over a predetermined number of steps by applying the weighting shown in (Equation 2). If, based on the error signal vector _e * information (k), the reproduced signal vector _x * (k), to the vector subspace formed by the delay storage means 104 to the stored reproduced signal vector _x * (k) The projection component P _{p *} (k) [h _* −h ′ _* (k)] of the echo path and the error characteristic h _* −h ′ _* (k) of the echo path is a simulated error projection vector X _* (k) [X _* ^T (k) X _* (k)] ⁻¹ e _* Equation unit 106 estimated by (k) and a simulated error projection vector as an update vector, multiplied by an update adjustment coefficient μ (k), In addition to the simulated characteristic h ′ _* (k), it is constituted by an echo path updating means 107 that updates the characteristics of the simulated echo path 102.
Kazuhiko Ozeki, Tetsuo Umeda: "Adaptive filter algorithm using direct projection to affine subspace and its properties" IEICE Transactions (A), J67-A, pp126-132, 1984.

According to the configuration of FIG. 3, if the collected sound signal y actually includes a disturbance, (Equation 3) may not be established accurately, and the error characteristic h _* −h ′ between the echo path and the simulated echo path 102. _* The orthogonal projection vector of (k) is erroneously estimated including the disturbance.
The object of the present invention is to reduce the estimation error of the orthogonal projection vector of the error characteristics h _* −h ′ _* (k) of the echo path and the simulated echo path 102 even when the collected sound signal y includes disturbance. Another object of the present invention is to propose an echo canceling method and an echo canceling apparatus that can realize echo canceling with high accuracy.

ここで、関数ψ（ａ）は、リミッタ関数と呼ぶこととし、入力する値ａが大きくなるほどその値を抑圧する特性を持つ任意の関数である。 In order to solve the above-described problem, the echo canceling method and echo canceling apparatus according to the present invention use the following update formula for estimating the simulation characteristic h ′ _* (k).

Here, the function ψ (a) is called a limiter function, and is an arbitrary function having a characteristic of suppressing the value as the input value a increases.

が満足されるべきである。但し、‖ｈ_＊‖はベクトルｈ_＊のノルムである。これより、通信会議システム等に反響消去方法を適用する場合、遠端の発話者のみが発話し、再生信号として再生され、収音信号において、再生信号に起因する反響信号が支配的である、シングルトークの状態においては、

が成り立つ。 If (Equation 3) is satisfied, the simulated error projection vector

Should be satisfied. However, ‖h _* ‖ is the norm of the vector h _* . From this, when applying the echo cancellation method to a communication conference system or the like, only the far-end speaker speaks and is reproduced as a reproduction signal, and in the collected sound signal, the echo signal due to the reproduction signal is dominant. In the single talk state,

Holds.

なる関係があり、さらに、一般に、ある行列の固有値の総和がその行列の対角成分の総和（トレース）に等しいという関係から、

が成り立つ。これより、通信会議において、収音信号に遠端の発話音声に起因する反響信号と近端の発話音声とが混合されている、ダブルトーク状態において、収音信号に含まれる外乱成分に起因する成分が支配的な場合は、

が成り立つ。 On the other hand, the error signal vector _e * (k), the error characteristics _h * -h _'rather than * (k), the component _n * (k) is dominant due to the disturbance component included in the collected sound signal, the When the error vector at the time can be expressed as e _* (k) = ˜n _* (k),

Furthermore, in general, from the relationship that the sum of the eigenvalues of a matrix is equal to the sum of the diagonal components of the matrix (trace),

Holds. As a result, in a communication conference, the echo signal resulting from the far-end utterance voice and the near-end utterance voice are mixed with the collected voice signal in the double talk state. If the ingredient is dominant,

Holds.

Therefore, the magnitude of the simulated error projection vector √ [e _* ^T (k) [X _* ^T (k) X _* (k)] ⁻¹ e _* (k)] takes values in single talk and double talk. It can be seen that the range is different. That is, in the single talk state, there is an upper limit for the value of √ [e _* ^T (k) [X _* ^T (k) X _* (k)] ⁻¹ e _* (k)], whereas double In the talk state, there is no upper limit.
Therefore, in (Equation 4) according to the present invention, a limiter function is introduced, and the value of √ [e _* ^T (k) [X _* ^T (k) X _* (k)] ⁻¹ e _* (k)] is By limiting so as not to exceed the upper limit value assumed in the single talk state or a value equivalent thereto, the estimation of the simulation characteristics h ′ _* (k) may cause a disturbance in the collected sound signal that does not satisfy (Equation 3). Even if it is included, it is not disturbed greatly.

In the echo canceling apparatus according to the present invention, the magnitude of the simulated error projection vector √ [e _* ^T (k) [X _* ^T (k) due to the fact that the collected sound signal y (k) contains many disturbances other than the echo signal. ) X _* (k)] ⁻¹ e _* (k)] is introduced to prevent the improper update of the simulation characteristics h ′ _* (k), and the sound pickup signal y Even if a disturbance is included in (k), 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 a microphone are installed or voice of a speaker in the room. If the upper limit of the gain of the echo path between the speaker and the microphone can be assumed, or if the influence of disturbance is small, the upper limit of the magnitude of the simulated error projection vector is the upper limit of the gain of the echo path. According to the value.

  On the other hand, if the voice of the speaker in the room becomes a disturbance and the other party's voice is uttered at the same time, the size of the simulated error projection vector exceeds the upper limit assumed when the influence of the disturbance is small. Can be bigger. Therefore, according to the present invention, by performing processing so that the size of the simulated error projection vector does not become larger than expected, deterioration of echo cancellation performance can be prevented even in simultaneous speech with a communication partner.

を得る正規化手段２０２と、ノルム√［ｅ_＊ ^Ｔ（ｋ）［Ｘ_＊ ^Ｔ（ｋ）Ｘ_＊（ｋ）］^-1ｅ_＊（ｋ）］の値が所定の大きさよりも小さい傾向にあるとき、その値と同等の値を出力し、ノルムの値が所定の大きさより大きい傾向にあるとき、その値を小さく抑圧して出力する特性を有する非線形処理を前記ノルムに与え、非線形ノルムψ｛√［ｅ_＊ ^Ｔ（ｋ）［Ｘ_＊ ^Ｔ（ｋ）Ｘ_＊（ｋ）］^-1ｅ_＊（ｋ）］｝を出力するノルム非線形処理手段２０３と、を本発明固有に備え、図３に示した反響路更新手段１０７の代わりに、非線形ノルムを正規化模擬誤差射影ベクトルに乗じ、前記模擬反響路１０２の更新ベクトルを得、前記更新ベクトルを、前記模擬反響路１０２に、そのまま加えるか、あるいは、固定または時変の更新調整係数μ（ｋ）を乗じて加え、（式４）に示すように、模擬反響路１０２の模擬特性ｈ′_＊（ｋ）を更新する模擬反響路更新手段２０４を構築する。 The echo canceling apparatus according to the present invention can be configured by hardware, but the embodiment in which the echo canceling program proposed in the present invention is installed in a computer and the computer functions as the echo canceling apparatus according to the present invention is the best. .
In order for the computer to function as the echo canceling apparatus according to the present invention, as shown in FIG. 1, the reproduction signal x (k) is accumulated in the computer in the same manner as described in FIG. _* A reproduction signal vectorization means 101 for outputting (k) and a simulation characteristic h ' _* (k) of an echo path, a reproduction signal vector x _* (k) and a simulation echo signal for simulating the echo signal a simulated echo path 102 that outputs d ′ (k), and an output error calculation means 103 that subtracts the simulated echo signal d ′ (k) from the collected sound signal y (k) and outputs an output error signal e (k). , Reproduction signal vector delay accumulation means 104 for accumulating the reproduction signal vector x _* (k) for the predetermined number of steps in order to generate the reproduction signal matrix X _* (k), and an error signal for the predetermined number of steps. the error signal vector means 105 for generating the e (k) a _* accumulates by performing weighting as shown in (Equation 2) the error signal vector e _(k), based on the error signal vector e _* information _(k), reproduced signal vector _x * and (k), the error characteristics _h * -h of the echo path and the echo path into constructed vector subspace by the delay storage means 104 to the stored reproduced signal vector _x * (k) _'* The projection component P _{p *} (k) [h _* −h ′ _* (k)] of (k) is simulated error projection vector X _* (k) [X _* ^T (k) X _* (k)] ⁻¹ e _* In addition to the echo path error projection estimation means 106 estimated by (k), the norm of the echo path simulation error projection vector √ [e _* ^T (k) [X _* ^T (k) X _* (k)] ⁻¹ e _* (k)] and the norm calculating unit 201 for calculating ^{_{a, √ [e * T (k}} ) ^{_{X * T (k) X *}} (k)] -1 e * (k)] by the simulated error projection vector ^{_{X * (k) [X *}} T (k) X * (k)] -1 e * (k ), And normalization simulation error projection vector

And the norm √ [e _* ^T (k) [X _* ^T (k) X _* (k)] ⁻¹ e _* (k)] tend to be smaller than a predetermined size. When the norm value tends to be larger than a predetermined value, a non-linear process having a characteristic of suppressing the value to be small and outputting is given to the norm, and the non-linear norm ψ { Norm nonlinear processing means 203 that outputs √ [e _* ^T (k) [X _* ^T (k) X _* (k)] ⁻¹ e _* (k)]}, which is unique to the present invention, is shown in FIG. Instead of the echo path updating means 107 shown, the normalized simulation error projection vector is multiplied by a nonlinear norm to obtain an update vector of the simulated echo path 102, and the update vector is added to the simulated echo path 102 as it is. Alternatively, a fixed or time-varying update adjustment factor μ ( ) The multiplied addition, (as shown in equation 4), to build a simulated echo path updating means 204 for updating the simulated characteristic _h of simulated echo path 102 _'* (k).

  By adding the norm calculation means 201, normalization means 202, nonlinear processing means 203, and simulated echo path updating means 204 specific to the present invention, processing is performed so that the size of the simulated error projection vector does not become larger than expected. In addition, even in simultaneous speech with the other party, an effect unique to the present invention that can prevent deterioration of the echo canceling performance can be achieved.

The selection of the characteristic of the limiter function ψ (a) used in the norm nonlinear processing means 203 in FIG. 1 is important. Here, a case is considered where the frequency characteristic of the echo path h _* is substantially flat and the average gain at each frequency can be approximated by ‖h _* ‖. At this time, with respect to the average value | x (k) ⁻ | of the amplitude | x (k) | of the reproduction signal and the average value | e (k) ⁻ | of the amplitude | e (k) | When the disturbance included in y (k) is small, a relationship of | e (k) ⁻ | / | x (k) ⁻ | <‖h _* ‖ can be assumed.

となる。（式１０）は、（式７）よりも、平均的に上限値を狭めることができることを意味する。但し、（式９）の近似において、再生信号は、平均的に白色で、Ｅ（Ｘ_＊ ^Ｔ（ｋ）Ｘ_＊（ｋ））の非対角要素は零、対角要素は全てＬ・｜ｘ（ｋ）⁻｜^２と想定し、また、誤差信号ベクトルｅ_＊（ｋ）の各要素の絶対値の平均は全て｜ｅ（ｋ）⁻｜、従って、Ｅ（‖ｅ_＊（ｋ）‖）＝√ｐ・｜ｅ（ｋ）⁻｜であると想定している。
このとき、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、反響路の平均利得‖ｈ_＊‖を正確に知り得るとは限らないため、反響路の平均利得‖ｈ_＊‖の値の信頼度に応じ、βは、更新の抑圧量を調整するための係数である。反響路の平均利得‖ｈ_＊‖の値が十分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。

It becomes. (Equation 10) means that the upper limit value can be narrowed on average than (Equation 7). However, in the approximation of (Equation 9), the reproduction signal is white on average, the non-diagonal elements of E (X _* ^T (k) X _* (k)) are zero, and all the diagonal elements are L · | x (k) ⁻ | ² and the average of the absolute values of the elements of the error signal vector e _* (k) are all | e (k) ⁻ |, and therefore E (‖e _* (k) ‖ ) = √p · | e (k) ⁻ |.
At this time, as an example of an appropriate limiter function ψ (a),

Is mentioned. Here, since not be known exactly the average gain ‖H _* ‖ the echo path, according to the reliability of the average gain ‖H _* ‖ value of echo path, beta adjusts the amount of suppression update Is a coefficient for If the average gain of the echo path ‖h _* ‖ is sufficiently reliable, β = 1 can be said to be the most reasonable value.

  As examples of the present invention, outlines of characteristic characteristics of functions other than the example of the limiter function ψ (a) described above (FIG. 2A) are shown in FIGS. 2B to 2F. When the influence of the disturbance is small, the norm division error does not exceed the norm division error upper limit value. In a sufficiently reliable implementation environment, the norm division error of a predetermined magnitude or more as in Alternative 5 shown in FIG. On the other hand, a characteristic having a decreasing tendency may be given.

または、

を更新式としてもよい。ここで、δは、正の実数であり、

が零となった場合の、零除算による数値不安定性を防止するものである。 Instead of updating the simulated characteristic h ′ _* (k) by (Equation 4),

Or

May be an update formula. Where δ is a positive real number,

This prevents numerical instability due to division by zero when becomes zero.

で置き換え、正規化手段２０２においては、それら置き換えられたノルムにより正規化し、正規化模擬誤差射影ベクトルとして、

を得ても、本発明の効果は維持される。 At this time, the norm calculation unit 201 calculates the norm of the simulated error projection vector as follows:

In the normalization means 202, normalization is performed using the replaced norm, and a normalized simulated error projection vector is obtained.

Even if obtained, the effect of the present invention is maintained.

と与えてもよい。但し、δは非負の実数とする。 In the error signal vectorization means 105 shown in FIG. 1, the error vector e _* (k) is

You may give. Where δ is a non-negative real number.

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

The block diagram for demonstrating one Example of the echo cancellation apparatus which has an echo path estimation means by this invention. The graph which shows the alternative example of the limiter function used for the Example shown in FIG. The block diagram for demonstrating the structure of the echo cancellation apparatus which has a conventional echo path estimation means. The block diagram for demonstrating the basic composition of an echo cancellation apparatus.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Sound reproduction means 2 Sound pickup means 100 Echo canceling apparatus 101 Reproduction signal vectorization means 102 Simulated echo path 103 Output error calculation means 104 Reproduction signal vector delay accumulation means 105 Error signal vectorization means 106 Echo path error projection estimation means 201 Norm Calculation means 202 Normalization means 203 Non-linear processing means 204 Simulated echo path updating means

Claims (6)

A reproduction signal reproduced from the sound reproducing means is transmitted by the sound collecting means via an echo path h _* which is an acoustic transmission path between the sound reproducing means and the sound collecting means existing in the same space. In the echo canceling method for canceling the echo signal obtained by collecting the sound from all the sound collecting signals collected by the acoustic sound collecting means,
A reproduction signal vectorization process for accumulating finite samples of the reproduction signal x (k), vectorizing the reproduction signal x (k), and outputting the reproduction signal vector x _* (k);
Simulated reverberation signal generation processing having simulation characteristics h ′ _* (k) of the reverberation path, inputting a reproduction signal vector x _* (k), and outputting a simulated reverberation signal d ′ (k) that simulates the reverberation signal When,
An output error calculation process for subtracting the simulated echo signal d ′ (k) from all the collected sound signals y (k) and outputting an output error signal e (k);
A reproduction signal vector delay accumulation process for accumulating the reproduction signal vector x _* (k) over a predetermined number of steps;
An error signal vectorization process for accumulating the output error signal e (k) over a predetermined number of steps and outputting an error signal vector e _* (k);
Based on the information of the error signal vector e _* (k), error characteristics of the echo path and the echo path to the vector subspace composed of the playback signal vector and the playback signal vector stored in the delay storage means h _* -h _'* projection component of _{(k) P p * (k} ) [h * -h' * (k)] a simulated error projection vector ^{_{X * (k) [X *}} T (k) X * (k )] ⁻¹ e _* (k) (where X _* (k) = [x _* (k), x _* (k−1),..., X _* (k−p + 1)] ^T ) An echo path error projection estimation process for outputting an error projection vector;
Norm of simulated error projection vector

A norm calculation process that outputs this norm,
The magnitude of the simulated error projection vector X _* (k) [X _* ^T (k) X _* (k)] ⁻¹ e _* (k) is normalized by the norm value, and the normalized simulated error projection vector

Normalization processing to obtain
Can size the Most than the value of a given magnitude before Symbol norm, giving a non-linear process having characteristics you suppress the value of the norm to the norm, the nonlinear norm

Norm nonlinear processing that outputs
The non-linear norm is multiplied by the normalized simulated error projection vector to obtain an updated vector of the simulated characteristic of the echo path, and the updated vector is added to the simulated characteristic of the echo path as it is, or is fixed or time-varying. Additionally multiplied updated adjustment factor μ a (k), the simulated echo path updating processing for updating the simulated characteristics h before Kihan sound path '* _(k),
The echo cancellation method characterized by including. A reproduction signal reproduced from the sound reproducing means is transmitted by the sound collecting means via an echo path h _* which is an acoustic transmission path between the sound reproducing means and the sound collecting means existing in the same space. In the echo canceling device for canceling the echo signal obtained by picking up the sound from all the sound pickup signals picked up by the sound pickup means,
Reproduction signal vectorization means for accumulating a finite number of samples of the reproduction signal x (k) and vectorizing the reproduction signal x _* (k);
A simulated reverberation path having simulated characteristics h ′ _* (k) of the reverberation path, receiving a reproduction signal vector x _* (k), and outputting a simulated reverberation signal d ′ (k) that simulates the reverberation signal;
Output error calculation means for subtracting the simulated echo signal d ′ (k) from all the collected sound signals y (k) and outputting an output error signal e (k);
Delay accumulation means for accumulating the reproduction signal vector x _* (k) over a predetermined number of steps;
Error signal vectorization means for accumulating the output error signal e (k) over a predetermined number of steps and outputting an error signal vector;
Based on the information of the error signal vector e _* (k), error characteristics of the echo path and the echo path to the vector subspace composed of the playback signal vector and the playback signal vector stored in the delay storage means h _* -h _'* projection component of _{(k) P p * (k} ) [h * -h' * (k)] a simulated error projection vector ^{_{X * (k) [X *}} T (k) X * (k )] ⁻¹ e _* (k) (where X _* (k) = [x _* (k), x _* (k−1),..., X _* (k−p + 1)] ^T ) Echo path error projection estimation means for outputting an error projection vector X _* (k) [X _* ^T (k) X _* (k)] ⁻¹ e _* (k);
Norm of simulated error projection vector

A norm calculation means for outputting the norm, and
The magnitude of the simulated error projection vector X _* (k) [X _* ^T (k) X _* (k)] ⁻¹ e _* (k) is normalized by the norm value, and the normalized simulated error projection vector

Normalization means to obtain
Can size the Most than the value of a given magnitude before Symbol norm, giving a non-linear process having characteristics you suppress the value of the norm to the norm, the nonlinear norm

Norm nonlinear processing means for outputting
The non-linear norm is multiplied by the normalized simulated error projection vector to obtain an update vector of the simulated characteristic of the echo path, and the update vector is added to the simulated characteristic of the echo path as it is, or is fixed or time-varying. Additionally multiplied updated adjustment factor μ a (k), the simulated echo path updating means for before updating the simulated characteristics h '* _(k) of Kihan Symphony path,
An echo canceling apparatus characterized by comprising:   The echo cancellation method according to claim 1,
  In the norm nonlinear processing, the nonlinear norm is

Is,
  An echo canceling method characterized by the above.   The echo canceling device according to claim 2,
  In the norm nonlinear means, the nonlinear norm is:

Is,
  An echo canceling device characterized by that.   5. An echo canceling program which is written in a computer-readable program language and causes the computer to function as the echo canceling device according to claim 2 or 4.   6. A recording medium comprising a computer-readable recording medium, wherein the echo canceling program according to claim 5 is recorded on the recording medium.

Images