KR20080095422A - Echo canceller and method thereof - Google Patents

Abstract

The present invention focuses on an adaptive filter technique for echo cancellation, in particular speech echo cancellation, and proposes a moving-average least mean square (MVLMS) detection algorithm that modifies the LMS detection algorithm. An acoustic echo canceller with an applied MVLMS filter and a method thereof are provided.

In addition, the MVLMS filter according to the present invention generalizes the moving average of the mean square, to increase the step size of the MVLMS filter for robust results of the coefficient detection algorithm, and to compensate for the drawback that the convergence speed depends on the step size. By using the moving average filter, the convergence speed and stability are improved by applying the MVLMS algorithm.

Description

Acoustic echo canceller and method thereof

1 is a block diagram showing the configuration of a conventional acoustic echo canceller.

2 is a block diagram showing a configuration of an acoustic echo canceller according to an embodiment of the present invention.

The present invention relates to acoustic echo cancellation, and more particularly, to an acoustic echo canceller and a method to which a detection algorithm based on a moving average predictor is applied among adaptive filter techniques for acoustic echo cancellation.

The development of wireless communication and Internet communication is rapidly progressing according to the demand of various services, and its application fields such as long distance conference system, hands-free phone for automobile, speakerphone system, and video conferencing are gradually expanding.

The effect of acoustic echo can be greatly influenced by the spatial environment between the speaker and the phone apparatus, which is a common device basically used in the field of voice communication, and such acoustic echo is an important factor that greatly degrades the call quality. .

In this case, when making a call using a communication terminal, the signal of the far-end speaker is introduced into the microphone through the speaker of the near-end speaker terminal and returned to the transmitting end of the near-end speaker.

In other words, when talking in the speakerphone mode in the conference room or hands-free mode in the car, the acoustic echo deteriorates the sound quality of the call, which causes the call to be disturbed.

Therefore, an echo canceller is used between the speaker and the microphone of the communication terminal to estimate the path of the echo signal using the adaptive filter and to remove the influence of the noise and the echo signal from the expected signal.

1 is a block diagram showing the configuration of a conventional acoustic echo canceller.

Referring to FIG. 1, the adaptive filter 14 continuously estimates the impulse response of the echo generation path using the far-end speaker signal x (k) and the error signal e (k) input as reference signals. Here k denotes a time index and the unit is determined according to the number of filtering taps.

In addition, the echo generation path means a path from the speaker 10 of the near-end talker terminal to the microphone 12 of the near-end talker terminal. At this time, as the detection algorithm used for estimation, a least mean square (LMS), a normalized LMS (NLMS), an orthographic projection (AP), and the like are used.

The adaptive filter 14 generates an estimated echo signal y '(k) using the estimated impulse response and passes it to the subtractor 16. The subtractor 16 then subtracts the estimated echo signal y '(k) from the expected signal d (k) containing echo to generate the error signal e (k) from which echo is removed.

The adaptive algorithms used in the acoustic echo canceller according to the related art configured as described above do not accurately estimate the echo due to the filtering coefficients diverging when the magnitude of the near-end talker signal suddenly changes, that is, when a large noise occurs. It was. Especially in the case of the simultaneous call where the near-end talker and the far-end talker talk simultaneously, the near-end talker signal is present together with the echo signal of the far-end talker signal, so the adaptive filter accurately estimates the echo signal under the influence of the near-end talker signal. The problem arises that the coefficients of the filter diverge.

In addition, the adaptive filter to which the least-squares mean algorithm is most widely used, because it has the advantage of simplicity of implementation.

However, in a system having a long impulse response, the performance of the conventional echo canceller has a high coefficient coefficient of the filter when a long time delay is included, such as a communication system, and has a high correlation with an input signal for an unknown path. When composed of a negative pattern, the performance is reduced as the side effects of the convergence speed due to a large coefficient is increased.

The present invention focuses on an adaptive filter technique for echo cancellation, in particular speech echo cancellation, and proposes a moving-average least mean square (MVLMS) detection algorithm that modifies the LMS detection algorithm. An object of the present invention is to provide an acoustic echo canceller equipped with an applied MVLMS filter and a method thereof.

In addition, the MVLMS filter according to the present invention generalizes the moving average of the mean square, to increase the step size of the MVLMS filter for robust results of the coefficient detection algorithm, and to compensate for the drawback that the convergence speed depends on the step size. By using the moving average filter, the convergence speed is improved by applying the MVLMS algorithm.

In order to achieve the above object, an acoustic echo canceller according to an embodiment of the present invention includes an echo canceller including a channel (Unknown Channel (h)) and an adaptive filter (Estimator) (hv), and an active detector. ), And a moving average filter,

The adaptive filter is a finite impulse response (FIR) adaptive filter to which the above-described moving-average least mean square (MVLMS) detection algorithm is applied.

In addition, the acoustic echo cancellation method according to an embodiment of the present invention,

Selecting a least square mean (LMS) forgetting factor in the range of? Of time k using

Updating a second step;

Determining an index set {b _ji } that satisfies the following; K time using

Updating a fourth step;

And repeating the second to fourth steps.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

The present invention focuses on an adaptive filter technique for echo cancellation, in particular speech echo cancellation, and proposes a moving-average least mean square (MVLMS) detection algorithm that modifies the LMS detection algorithm. The MVLMS filter to which the MVLMS detection algorithm is applied is a generalized moving average of the mean squares.The MVLMS filter increases the step size of the MVLMS filter for robust results of the coefficient detection algorithm, and moves to compensate for the drawback that the convergence speed depends on the step size. By using the average filter, the convergence speed is improved by applying the MVLMS algorithm.

Hereinafter, the LMS algorithm will be described before describing the MVLMS detection algorithm according to the present invention.

Referring to FIG. 1 mentioned above, the adaptive filter 14 uses the signal x (k) of the far-end speaker as an input and applies the filtering coefficient obtained by the speech echo cancellation algorithm to the x (k) to estimate the echo signal y. '(k) is subtracted from the expected signal d (k) to remove the far-end speaker signal. Here, the expected signal d (k) is an echo signal y (k) flowing from the speaker 10 of the near-end talker terminal to the microphone 12 of the near-end talker terminal, the voice signal s (k) of the near-end talker and the ambient noise signal. It consists of the sum of n (k). That is, Equation 1 below holds true.

Equation 1

The error signal e (k) is obtained according to Equation 2 below by subtracting the estimated echo signal y '(k) from the expected signal d (k).

Equation 2

Here, r (k) is smaller as the performance of the echo cancellation algorithm is better, and as r (k) is smaller, the echo signal transmitted to the far-end speaker is smaller.

The typical algorithm for minimizing r (k) is Least Mean Square (LMS). The update result of the LMS algorithm is shown in Equation 3 below.

Equation 3

Here, W (k) is a vector representing coefficients of the adaptive filter, and when the number of taps of the adaptive filter is L (normally 128 taps), the L coefficients w1, w2, ... wL are formed. Μ (k) is an adaptive constant determined experimentally, and e (k) is an error signal obtained by subtracting the estimated echo signal y '(k) from the adaptive filter from the expected signal input to the near-end speaker's microphone. ) Is the signal of the far-end speaker. X (k) is likewise made up of L samples, depending on the number of taps L of the adaptive filter.

The LMS algorithm is often referred to as the stochastic gradient algorithm, which is the most widely used adaptive filter algorithm because it works well in real environments in addition to simplicity.

In addition, the LMS algorithm requires a relatively small amount of calculation, and the aforementioned parameter μ plays a very important role in the LMS algorithm. The μ may vary over time but is determined to be a constant value after experiments for that application.

In the LMS algorithm represented by Equation 3, the signal X (k) of the far-end speaker has a great influence on the updating of coefficients. The normalized LMS (Normalized LMS: NLMS) developed to compensate for this drawback uses the adaptation constant μ (k) by normalizing the power of the far-end speaker signal as shown in Equation 4 below. Reduce the effects of changes in (k).

Equation 4

Here, the superscript T means transpose of the vector.

That is, the NLMS algorithm has a better convergence rate than the LMS algorithm. As the convergence weighting factor μ increases, stability decreases. On the contrary, when the convergence weighting factor μ decreases, the stability increases.

An embodiment of the present invention proposes a moving-average least mean square (MVLMS) detection algorithm to improve both the stability and the convergence rate.

2 is a block diagram showing the configuration of an acoustic echo canceller according to an embodiment of the present invention.

Referring to FIG. 2, an acoustic echo canceller according to an exemplary embodiment of the present invention includes an echo canceller including an unknown channel (h) and an adaptive filter (hs), and a dynamic detector. and a moving average filter.

In this case, the adaptive filter is a finite impulse response (FIR) adaptive filter to which the aforementioned moving-average least mean square (MVLMS) detection algorithm is applied.

In addition, the moving average filter is used to smooth the fluctuating data before interpretation. A generalized concept of the moving average filter is a finite impulse response (FIR) filter.

The moving average filter is a simple linear time invariant system as defined in Equation 5 below.

Equation 5

The moving average filter defined in Equation 5 is also referred to as an L-point prediction error moving average because the output of time n is calculated as an average of e [n] and L-1 previous inputs.

Hereinafter, the operation of the MVLMS detection algorithm according to an embodiment of the present invention will be described with reference to FIG. 2.

m &lt;n; ₀ ≦ t ₁ ≦ t ₂ <.. When <t _m ≤ n-1, the unknown channel shown in FIG. 2 is weakly active. If the unknown channel is defined as an n-dimensional parameter vector, Equation 6 is obtained.

Equation 6

Where n> j _m > j _m-1 >... > j ₂ > j ₁ > 0, and θ _j is a zero matrix of size 1ⅹj. Also, b _jm is a non-zero matrix,

to be. here,

Is an estimate of m,

Wow

Is the deviation between μ (k) and noise (k).

In addition, the active parameter of the MVLMS algorithm has a larger magnitude than the noise value of the LMS adaptation algorithm, and each remaining parameter is defined as an inactive parameter.

In addition, the goal of the MVLMS detection algorithm is to locate m non-zero elements of h.

Here, the cost function based on the structurally consistent least squares (SCLS) of Equations 7, 8, and 9 will be used.

Equation 7

Equation 8

Equation 9

here,

ego,

silver

Is the deviation. m is the number of unknown activity parameters.

Also,

Is _equal to X _j (N)> T (N)

It is clear that it is minimized by the index.

Where X _j (N) is an activity measure, T (N) is an activity threshold, and T (N) is defined as in Equation 10 below.

Equation 10

Such a method can be summarized by the following algorithm.

first

Choose the least-squares forgetting factor from. (Step 1)

here,

, Where j = 0, 1,... , n-1.

Next, the time k

Update the. (Step 2)

next

Determine the index set {b _ji } that satisfies. (Step 3)

A vector g (k) of size n * 1 is generated in which 1 enters a position corresponding to the set {b _ji } of the index and 0 in the remaining positions.

Since k time using

Update the.

Where g _j (k) is the j th g (k) element.

Finally repeat from the second step. (Step 4)

According to the present invention, by applying the MVLMS detection algorithm, there is an advantage that the convergence performance and stability is improved compared to the conventional detection algorithm is applied.

The present invention described above is not limited to the above-described embodiments and the accompanying drawings, and various substitutions, modifications, and changes can be made without departing from the technical spirit of the present invention. It will be evident to those who have knowledge of.

Claims (2)

An echo cancellation unit including a channel (Unknown Channel (h)) and an adaptive filter (Estimator: hv); An active detector, Includes a moving average filter, And the adaptive filter is a finite impulse response (FIR) adaptive filter to which the moving-average least mean square (MVLMS) detection algorithm described above is applied.

Selecting a least square mean (LMS) forgetting factor in the range of? Of time k using

Updating a second step;

A third step of determining an index set {b _ji } that satisfies; K time using

Updating a fourth step;

And repeating the second step to the fourth step.

Images