KR100545832B1 - Sound echo canceller robust to interference signals - Google Patents

Abstract

The present invention relates to an acoustic echo canceller. In particular, the present invention relates to an acoustic echo canceller for removing acoustic echo generated in a voice communication device such as a hands-free and video conferencing system without interference of the near end speaker's speech and background noise.

According to the present invention, an acoustic echo canceller having a speaker stage which is received and output from a far-end speaker to a near end speaker, a microphone stage which is transmitted from the near end speaker to a far end speaker, and an acoustic echo path, A FIR filter for generating an acoustic echo signal using updated adaptive filter coefficients from the microphone signal; A background noise estimator for calculating a background noise level by inputting an estimated error signal obtained by subtracting the estimated echo signal from a microphone signal of a far-end speaker; A step size adjuster for determining a step size according to the background noise level and whether a simultaneous call is present; A simultaneous call detector for determining a simultaneous call state from a cross correlation between the speaker signal of the far-end speaker and the estimated error signal calculated in advance from the adaptive filter coefficient updater; And an adaptive filter coefficient updater for updating the adaptive reflection coefficient from the speaker signal, the step size, and the estimated error signal of the far-end speaker.

Acoustic, Echo, Canceller, Background Noise Estimator, Simultaneous Call Detector

Description

Interference robust acoustic canceller echo removal apparatus

1 is a block diagram of an acoustic echo canceller according to a conventional embodiment.

2 is a block diagram of an acoustic echo canceling device according to another exemplary embodiment.

3 is a block diagram of the acoustic echo canceller according to the present invention.

4 is a detailed block diagram of the background noise estimator illustrated in FIG. 3.

FIG. 5 is a detailed structural diagram of the adaptive filter coefficient updater shown in FIG. 3.

<Description of Symbols for Major Parts of Drawings>

300,560: FIR filter 310: adaptive filter coefficient updater

320: background noise estimator 330,550: step size adjuster

340,540: simultaneous call detector 400: background noise estimation unit

500: cross-correlation calculator 510: adaptive filter coefficient storage unit

520,530: time delay

The present invention relates to an acoustic echo canceller that is robust to interference signals. More particularly, the present invention relates to an acoustic echo canceller for removing acoustic echo generated by a voice communication device such as a hands-free and video conferencing system without interference between the near end speaker's voice and background noise.

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. .

The echo canceller is generally used for the adaptive filter to remove the acoustic echo. The interference between the caller (near end speaker) and the background noise input from the microphone degrades the adaptive filter and ultimately degrades the echo cancellation performance. It becomes a factor. Therefore, there is a need for an acoustic echo canceller having a structure robust to interference signals.

1 is a block diagram of an acoustic echo canceller according to a conventional embodiment. 2 is a block diagram of an acoustic echo canceling device according to another exemplary embodiment.

The acoustic echo canceller shown in FIG. 1 tracks the linear impulse response coefficient of the echo path and estimates the same echo signal at the output of the adaptive filter. Thereafter, the echo signal is canceled by canceling the estimated echo signal 140 from the microphone input signal.

Here, the most important part is a process of estimating the linear impulse response coefficient of the echo path, and the estimation accuracy and convergence speed of the actual echo signal vary according to this process. Conventionally used methods include the Normalized Least Mean Square (NLMS) algorithm of the Least Mean Square (LMS) series, and are preferred because they can be easily implemented in a relatively simple structure. The NLMS obtains the linear impulse response coefficient w (k + 1) of the echo path for the k + 1 th sample in the fixed step adaptive filter coefficient updater 120. This is obtained by multiplying the linear impulse response coefficient w (k) by the far end speaker's audio signal x (k), the echo estimation error signal e (k), and the step term μ (k), as shown in Equations 1 and 2. Is calculated. In particular, the step term μ (k) is a value obtained by dividing the step coefficient β by the average power P _x (k) of the far-end speaker signal x (k), which is an important factor in determining the estimated accuracy and the convergence speed.

Where k is the time axis sample number. P _x (k) is defined as the average power (1/2 power of the mean of the squares) of the surrounding samples of the audio signal x (k).
The update of the adaptive filter may cause interference when the near end speaker's voice signal comes from the microphone input, resulting in an incorrect adaptive filter coefficient. Therefore, the simultaneous call detector 130 which detects a state in which the voice signal of the near-end caller is also used is used. Thus, the adaptive filter is not updated in the simultaneous call state.

The acoustic echo canceller shown in FIG. 2 uses the Sum-LMS method to add the power P _x (k) of the estimated error signal to decrease the step term μ (k) as shown in Equation 3 so that the interference signal increases. Used. Also, as shown in Equation 4, the step size is reduced to the interference signal by using c (k), which is low-pass filtered on the speaker output signal x (k) from the far-end speaker and the estimated error signal e (k). A method of making it affected is also presented.

delete

Summarizing the conventional techniques shown in FIGS. 1 and 2, a method using a simultaneous call detector, a Sum-LMS method, and a method using cross-correlation values are used.

However, if only the simultaneous call detector is used, performance degradation against background noise is concerned, and the simultaneous call detection for most acoustic reflections is correlated between the speaker output of the far end speaker and the microphone input of the far end speaker or the speaker output of the far end speaker. Because it is based on the correlation between the estimated error signal and the estimated error signal, a large amount of computation is required.

In addition, in the case of the Sum-LMS method, the size of the adaptation constant decreases due to the influence of the estimated error signal, resulting in a slow convergence speed. In addition, the use of the cross-correlation value also has the disadvantage of requiring a large amount of computation in an environment with long echo paths, and the influence of the power of the estimated error signal is not completely eliminated, thus affecting the interference signal while satisfying a lower computation amount. There is a need for a new way to ensure convergence performance.

Accordingly, an object of the present invention is to solve the above-mentioned problems, and an object of the present invention is to estimate an interference signal such as an acoustic signal and a background noise of a near-end speaker when an acoustic signal received from a far-end speaker is input into a microphone through an echo path. It is to provide a low-acoustic acoustic echo canceller that prevents the degradation.

As a technical idea for achieving the above object of the present invention

In the acoustic echo cancellation device having a speaker stage that is received and output from the speaker of the far end to the speaker of the near end, a microphone stage transmitted from the speaker of the far end to the speaker of the far end, and the acoustic echo path,

A FIR filter for generating an acoustic echo signal using updated adaptive filter coefficients from the far-end speaker's microphone signal;

A background noise estimator for calculating a background noise level by inputting an estimated error signal obtained by subtracting the estimated echo signal from a microphone signal of a far-end speaker;

A step size adjuster for determining a step size according to the background noise level and whether a concurrent call is present;                         

A simultaneous call detector for determining a simultaneous call state from a cross-correlation view between a speaker signal and an estimated error signal of a far-end speaker previously calculated from an adaptive filter coefficient updater; And

And an adaptive filter coefficient updater for updating the adaptive reflection coefficient from the speaker signal, the step size, and the estimated error signal of the far-end speaker.

Hereinafter, with reference to the accompanying drawings, the configuration and operation of the embodiment of the present invention will be described in detail.

3 is a block diagram of the acoustic echo canceller according to the present invention. 4 is a detailed block diagram of the background noise estimator shown in FIG. 3. FIG. 5 is a detailed structural diagram of the adaptive filter coefficient updater shown in FIG. 3.

In the present invention, n denotes an nth frame (current frame), and n-1 denotes an n-1th frame (past frame). An expression without n and n-1 basically means an expression for the nth frame (the current frame). For example, x (k) is an acoustic signal of the kth sample received from the far-end speaker for the nth frame (the current frame).
The acoustic echo cancellation apparatus of the present invention is based on a block adaptive filter, and the speaker stage is received from the far-end speaker as the near-end speaker and is output from the far-end speaker to the far-end speaker as shown in FIG. A finite stimulus response (FIR) filter for generating an acoustic echo signal by filtering the acoustic signal x (k) received from the far-end speaker with the updated adaptive filter coefficient w (k) in the state where a predetermined voice input / output device is provided. Group 300; A background noise estimator (320) for obtaining a background noise level by inputting an estimated error signal obtained by subtracting the estimated echo signal y (k) from the far-end speaker's microphone signal d (k); A step size adjuster 330 for determining a step size μ _n according to the background noise coefficient R _n and whether g _n is a simultaneous call; A simultaneous call detector 340 for determining a simultaneous call state from a cross-correlation diagram (composed of the speaker signal x (k) and the estimated error signal e (k) of the far-end speaker) previously calculated from the adaptive filter coefficient updater; And an adaptive filter coefficient updater 310 for updating the adaptive echo coefficient w (k) from the speaker signal x (k), the step size mu (k) and the estimated error signal e (k).

The present invention is based on a block adaptive filter and has a low complexity. This is because, in the block adaptive filter structure, the cross-correlation value of the far-end speaker's microphone signal and the estimated error signal used in the simultaneous call detection is calculated in advance by the adaptive filter coefficient updater 310.

In addition, the present invention has a feature of eliminating the interference of the background noise by separately estimating the background noise level and using a value extracted from the estimated error signal level.

Next, the steps of the present invention will be described in detail to explain the process of removing echo.

First, the speaker output signal x (k) received from the far-end speaker is transformed into an echo signal through the acoustic echo path and input into the microphone of the near-end speaker. In this case, the background noise from the outside and the voice signal of the near-end speaker may be input to the microphone together with the echo signal. In addition, by convolving the adaptive filter coefficients with the input of the speaker output signal x (k), the echo signal may be predicted in the FIR filter 300 as shown in Equation 5 below.

delete

Here, w (i) represents the first to Lth values of w (k) values representing kth adaptive filter coefficients for the current frame (nth frame) and is calculated by defining i to represent them. Accordingly, w (k) is also represented by w (k, n) in another expression, and w (k) is obtained by the equation (7). The L value is the number of adaptive filter coefficients, and the larger the value is, the more predicted the echo signal having the larger delay is. If the coefficient w (k) of the adaptive filter has an impulse response to the echo path, y (k) is close to the actual echo signal.
As shown in Equation 6, the signal e (k) obtained by subtracting the input signal from the microphone of the actual near end becomes an echo canceled error estimation signal which is canceled and is used again to update the adaptive filter coefficient of the echo canceller.

The estimated error signal e (k) thus obtained includes a large amount of unremoved echo signals at the beginning of echo cancellation, and then gradually decreases as the value of y (k) approaches the actual echo signal. The component of the included echo information is used in the adaptive filter coefficient updater 310 to obtain the correlation of the far-end speaker's speaker signal as shown in Equation 8 and used to update the adaptive filter coefficient as shown in Equation 7.

이외에도 잡음에 해당하는 상호 상관 값

항이 존재하게 되어 잡음성분이 클 경우 최적 계수 값과는 많이 다른 값을 얻을 수 있다.Here, the N value is an arbitrary value determined by the length of each block for each frame. w (k, n) denotes the kth adaptive filter coefficient for the nth frame (the current frame) and w (k, n-1) denotes the kth adaptive filter coefficient for the n-1th frame (previous frame). It means. Equations 7 and 8 are calculated only once per frame.
If the signal coming from the near-end microphone is a pure echo signal, the coefficient updater of the adaptive filter converges to the optimal adaptive filter coefficient value. However, when there is background noise, the cross-correlation value by the echo signal as shown in Equation 9

In addition, cross-correlation values corresponding to noise

If the term is present and the noise component is large, a value very different from the optimal coefficient value can be obtained.

Therefore, the background noise estimator 320 estimates the energy of the background noise and uses the ratio of the estimated error signal energy as a factor for adjusting the step size. As the ratio of the background noise component to the estimated error signal increases, the interference is increased so that the adaptive step size is reduced. To this end, the background noise coefficient R _nx (k) is defined and used as shown in Equation (11).

delete

delete

delete

가 결정된다. 동시통화 검출의 결과인 동시통화 상태 g(k,n)는 수학식 13과 같이 계산되며 적응필터 계수 갱신기(310)에서 미리 계산된 상호 상관 값 Φ(k)를 사용함으로써 낮은 계산량을 가진다.
여기서, P_x(k)는 x(k)신호의 주변 값에 대한 파워 값으로 정의된다. 수학식 15에서와 같이 과거 M개의 상호상관도 값에 대한 절대 값이 사용되는데, 근단 화자의 음향성분이 같이 들어올 경우 이러한 절대 값들에 대한 합이 갑자기 커지 된다는 원리를 이용한 것이다. 수학식 12에서 β값은 수렴 상수를 뜻하고 g(k,n-1)은 n-1번째 프레임(이전프레임)에 대한 k번째 동시통화 상태 값을 표시한 것이다. Here, E _n (k) means the background noise level estimated by the background noise estimator 320 and is obtained from Equation 16 to be described later.
The step size adjuster 330 needs a simultaneous call detector because it needs to remove not only background noise but also interference of acoustic components of the near end speaker. Since the sound component of the near-end speaker has more severe interference than the background noise, the simultaneous call detector 340 is used separately to adjust the step size. Step size as shown in Equation 15

Is determined. The simultaneous call state g (k, n), which is the result of the simultaneous call detection, is calculated as in Equation 13 and has a low calculation amount by using the cross-correlation value Φ (k) previously calculated in the adaptive filter coefficient updater 310.
Here, P _x (k) is defined as the power value for the peripheral value of the x (k) signal. As shown in Equation 15, the absolute values of the past M cross-correlation values are used, and when the acoustic components of the near-end speaker come together, the sum of these absolute values suddenly increases. In Equation 12, β denotes a convergence constant and g (k, n-1) denotes a k-th simultaneous call state value for an n-1 th frame (previous frame).

delete

The background noise estimator 320 and the step size adjuster 330 may be applied together with a sub-band adaptive filter. When applied, the background noise estimator 320 and the step size adjuster 330 are divided into sub-bands to be normalized by power of each band, as shown in Equation (14). Have Such a structure can be expected to improve convergence performance by increasing background noise estimation accuracy.

Where m is the number for each frequency band. n-1 indicates the data of the previous frame. Therefore, R _nx (m, k) denotes the k-th background noise coefficient of the m-th band, and is obtained by Equation 11 from the estimated error energy level and the background noise energy value for the m-th frequency band. β (m) is the convergence constant for the m-th frequency band and g (m, k, n-1) is the simultaneous call of the k-th sample for the m-th frequency band and the n-1th frame (previous frame). The state value is expressed and obtained by applying Equation 13 to the m th frequency band.
4 is a detailed block diagram of the background noise estimator. The background noise level is estimated using the echo estimation error signal e (k) as shown in Equation (16). The speech level can be implemented as an average energy value, but each implementation uses a sum of absolute values to reduce the complexity and range of the background noise level.

The background noise energy level is estimated by simply selecting a smaller one of the term and the estimated error signal that increases with time as shown in Equation 20 using the characteristic that the energy change of the background noise is static.

Here, Equation 15 is the k-th estimation error signal with the average value of the absolute value from the k-th echo estimation error signal e (k) to the k- (N + 1) -th sample, e (k- (N + 1)). It is defined as the energy level E _e (k). Equation (20) is the estimated energy level E _e (k) is k-1-th calculated background noise energy level is less than E _n (k-1) k-th background noise energy levels E _n (k) of the k-th estimation error signal It is defined as the energy level E _e (k) of the error signal. On the contrary, when the energy level E _n (k-1) of the k-th estimated error signal is less than or equal to the energy level e (k) of the estimated error signal, the energy increase constant ρ is added to the background noise energy level E _n (k In small increments.
FIG. 5 is a block diagram illustrating simultaneous call detection and step size control of an adaptive filter updater. A correlation diagram of the estimated error signal e (k) and the speaker output signal x (k) of a far-end speaker are included in the adaptive filter updater. A correlation between 500 and the simultaneous call detector 540 and the step size controller 550 is shown.

The simultaneous call detector 540 reduces the amount of calculation using the output value of the cross-correlation calculator 500 previously calculated in the adaptive filter updater. In addition, the step size controller 550 adjusts the step size value μ (k) by inputting the output value of the simultaneous call detector 540 and the background noise coefficient value R _nx (k).

delete

delete

delete

delete

delete

As described above, according to the acoustic echo canceller according to the present invention, it is possible to cleanly remove acoustic echo generated by a hands-free phone or a video conference phone regardless of a sound signal or a background noise signal of a near-end speaker in a simultaneous call input from a microphone. Has an effect. In addition, since the device is implemented at a low complexity when the device is implemented, it is more advantageous to reduce development costs.

Claims (9)

In the acoustic echo cancellation device having a speaker stage that is received and output from the speaker of the far end to the speaker of the far end, a microphone stage transmitted from the speaker of the far end to the speaker of the far end, and an acoustic echo path, A FIR filter for generating an acoustic echo signal using updated adaptive filter coefficients from the far-end speaker's microphone signal; A background noise estimator for calculating a background noise level by inputting an estimated error signal obtained by subtracting the estimated echo signal from a microphone signal of a far-end speaker; A step size adjuster for determining a step size according to the background noise level and whether a simultaneous call is present; A simultaneous call detector for determining a simultaneous call state from a cross correlation between the speaker signal of the far-end speaker and the estimated error signal calculated in advance from the adaptive filter coefficient updater; And And an adaptive filter coefficient updater for updating the adaptive reflection coefficient from the speaker signal, the step size, and the estimated error signal of the far-end speaker. The method of claim 1, wherein the background noise level of the background noise estimator And selecting a minimum value between the background noise level of the previous step and the current block echo estimation error signal level according to Equation 18 below.

Where E _e (k) is the energy of the kth estimated error signal, E _n (k) is the k-th obtained background noise energy, and ρ is the energy increase constant. The method of claim 2, wherein the echo estimation error signal level is Acoustic echo canceller, characterized in that calculated using the absolute value according to equation (19) below.

(E (ki) is the ki-th echo estimation error signal, E _e (k) is the energy of the k-th estimation error signal, respectively.) The method of claim 1, wherein the simultaneous call detector is The fixed value for the simultaneous call detection parameter p (k) is expressed by Equation 20 below.

Acoustic echo canceller, characterized in that for detecting the simultaneous call state g (k, n).

(Where g (k, n) is the kth concurrent call state value for the nth frame that is the result of the simultaneous call detection, and p (k) means the kth simultaneous call detection parameter, respectively.) The method of claim 4, wherein the simultaneous call detection parameter p (k) is The cross-correlation value of the speaker output sound signal of the far-end speaker and the echo estimation error signal previously calculated from the adaptive filter coefficient updater by Equation 21 below.

And using the normalized value as a parameter of the speaker output sound signal of the far-end speaker.

Where Φ (ki) is the ki-th cross-correlation value, P _x (k) is the power for the peripheral value of the x (k) signal, and p (k) is the k-th simultaneous call detection parameter. ) The method of claim 1, wherein the step size adjuster The ratio of the energy of the background noise energy and the echo estimation error signal is expressed by Equation 22 below.

The step size using a value and the concurrent call state g (k, n-1)

Acoustic echo canceller, characterized in that to calculate the value.

(here,

Is the energy ratio of the k th echo estimation error signal, P _x (k) is the power of the surrounding value of the x (k) signal, β is the convergence constant, and g (k, n-1) is the n-1th frame. It means the k-th simultaneous call status value.) The method of claim 1, wherein the step size adjuster The energy ratio value of the background noise energy for each band and the echo estimation error signal for each band according to Equation 23 below when divided into subbands is used.

And the step size using the concurrent call state g (m, k, n-1)

Acoustic echo canceller, characterized in that to calculate the value.

Where m is the number for each frequency band, R _nx (m, k) is the k-th background noise coefficient of the m-th band, and β (m) is the convergence constant for the m-th frequency band, P _x (m, k) is the power of the peripheral value of the x (k) signal of the mth band, and g (m, k, n-1) is the simultaneous call state value of the kth sample of the n-1th frame of the mth frequency band. Each means.) delete delete

Images