KR100647826B1 - The blind dereverberation models considering measured noises and the deriving method thereof - Google Patents

Abstract

A blind echo cancel model based on a measured noise and a method for deriving the same are provided to improve a quality of an audio signal acquired by a convergent speed and a process result by deriving a learning rule of each model through an analysis of an independent element. A method for deriving a blind echo cancel model based on a measured noise includes the steps of: defining a blind minimum square cost function for an estimated independent signal source in a frequency region in a blind signal process structure having a signal observed by an input terminal and an independent signal source estimated by an output terminal; defining a blind minimum square cost function in an output terminal by adding the blind minimum square cost function defined for an independent signal source in a frequency region; and differentiating the blind minimum square cost function for each adaptive filter; and adapting a natural gradient for the differentiated blind minimum square cost function to a frequency region.

Description

The blind dereverberation models considering measured noises and the deriving method approximation

1 is a structural diagram of a conventional echo model and a blind echo cancellation model.

2 is a structural diagram of a conventional speech synthesis channel and speech interpretation channel.

3 is a diagram illustrating a reflection channel model including a conventional speech synthesis channel.

4 is a diagram illustrating a learning model of a pure blind echo cancellation filter to which a conventional whitening filter is applied.

FIG. 5 is a view illustrating an observation signal model in which white Gaussian measurement noise is added to a signal including echo in a conventional echo signal recording process.

FIG. 6 is a diagram illustrating a blind echo cancellation model for learning two independent output signals when using one microphone of the present invention. FIG.

FIG. 7 is a diagram illustrating a blind echo cancellation model for learning two independent output signals when using two microphones of the present invention. FIG.

8 is a diagram illustrating a blind echo cancellation model for learning three independent output signals when using two microphones of the present invention.

The present invention relates to a blind echo cancellation model and its derivation method considering the measured noise. In particular, an independent component analysis that can consider the second order or higher statistical characteristics is applied to an echo cancellation filter adaptation algorithm to perform echo cancellation performance under measured noise. To improve the model and its derivation method.

(10)가 미지의 반향필터

(11)를 통해서 관찰된 신호

(12)만을 알고 있을 때 원래의 음원신호

(10)를 복원하는 기술이다.As shown in Fig. 1, the blind echo separation is an unknown sound source signal independent of each other between samples.

(10) Unknown echo filter

Signal observed through (11)

Original sound source signal when only (12) is known

(10) is a technique to restore.

(11)로 모델링할 수 있으며, 센서로 측정한 관찰신호

(12)는 다음의 수학식 1과 같다.Linear filter with various time delays for echo channel effects in real environment

Can be modeled as (11), observation signal measured by sensor

(12) is shown in Equation 1 below.

(13)를 통과한 출력신호

(14)는 다음의 수학식 2와 같다.Where * is a convolution operator. Meanwhile, echo cancellation filter

Output signal passed through (13)

(14) is shown in Equation 2 below.

(10)는 샘플들 사이에 독립이고 비정규분포(non-Gaussian)인 확률밀도함수를 갖는다고 가정하고, 복원된 신호

(14)가 이러한 가정을 만족하는 가에 대한 기준으로 삼을 수 있는 비용함수(cost function)를 만듦으로써 반향제거 필터

(13)를 학습할 수 있는 적응 알고리즘(15)을 얻을 수 있다. 비용함수를 최소화시키는 형태로 학습이 진행되고 나면, 반향제거 필터

(13)는 반향필터

(11)의 역필터 형태로 얻어진다.Unknown source

(10) assumes that the sample has an independent and non-Gaussian probability density function

Echo cancellation filters by creating a cost function that can be used as a basis for whether (14) meets these assumptions.

An adaptive algorithm 15 that can learn (13) can be obtained. Once the learning is done in a form that minimizes the cost function, the echo cancellation filter

13, echo filter

Obtained in the form of an inverse filter of (11).

(13)의 학습법칙은 다양한 비용함수에 따라 서로 다른 알고리즘들이 존재하는데, 근사화와 좀 더 효율적인 학습을 위한 변형을 통해서 이러한 학습법칙들은 다음의 수학식 3과 같은 형태로 얻어진다.Echo cancellation filter

The learning law of (13) has different algorithms according to various cost functions. Through the approximation and transformation for more efficient learning, these learning laws are obtained in the form of Equation 3 below.

는 추정하고자하는 음원신호

(10)의 확률밀도함수(probability density function)이고,

은 반향제거필터

(13)의 시간 지연 탭 수이다.here,

Is the sound source signal to be estimated

Probability density function of (10)

Silver echo cancellation filter

The number of time delay taps in (13).

(10)가 샘플들 사이에 독립이고, 비정규분포를 이루는 가정을 만족하는 경우에 반향필터

(11)의 영향을 제거하는 반향제거 필터

(13)의 학습에 관한 내용이었다. 그러나, 음원신호

(10)가 음성신호(speech signal)인 경우에는 샘플들 사이가 독립이라는 가정에 맞지 않는 문제점이 있기 때문에 수학식 3과 같은 반향제거 필터

(13)에 관한 학습방법을 바로 적용하면 결과적으로 얻어진 적응필터

는 반향채널의 역채널특성을 갖을 뿐만 아니라 음성 샘플들 자체의 의존성(dependency)을 제거함으로써 음성을 백색화(whitening)하려는 특성도 포함하는 형태로 학습하게 된다.Until now, the unknown sound source signal as shown in FIG.

Echo filter when (10) is independent between samples and satisfies the assumption of non-normal distribution

(11) Echo cancellation filter to remove the effect

It was contents about learning of (13). However, the sound source signal

If (10) is a speech signal, there is a problem that does not fit the assumption that the samples are independent, so the echo cancellation filter as shown in Equation (3)

Adaptive filter obtained as a result of directly applying the learning method of (13)

Not only has the inverse channel characteristics of the echo channel, but also learns to include the characteristic of whitening the speech by removing the dependency of the speech samples themselves.

(11) 및 반향제거 필터

(13)에 적용하게 되면 도 3과 도 4와 같이 음성합성 채널(30)을 적용한 반향필터

(31)과 백색화 필터(40)를 적용한 순수한 반향제거 필터

(41)를 얻을 수 있다.Therefore, in the case where the voice signal is a signal source, it is assumed that the non-noise voice signal as shown in FIG. 2 is a signal passed through the voice synthesis channel 20 from mutually independent synchronization signals, and from the observed voice signal. By learning to obtain the independent source signal as much as possible, a speech analysis channel or a whitening filter 21, which is an inverse channel form of the speech synthesis channel 20, is obtained. The echo synthesis channel 20 and the speech interpretation channel 21 thus obtained are echo filters of FIG.

11 and echo cancellation filter

When applied to (13), the echo filter to which the speech synthesis channel 30 is applied as shown in FIGS. 3 and 4.

Pure echo cancellation filter with 31 and whitening filter 40

(41) can be obtained.

(14)의 독립성을 이용하여 반향제거 필터

(13)를 학습하는 대부분의 방법에서는 도 4와 같은 반향제거 필터

(41)을 이용한다. 실제로, 도 3과 같은 반향필터

(31)의 영향만이 존재하는 이상적인 경우로 수학식 3을 이용하여 음성 신호원인 경우의 암묵 반향제거 실험을 진행하면 반향을 충분히 제거할 수 있는 반향제거 필터

(41)를 성공적으로 학습해 낼 수 있으며, 결과적으로 얻어진 추정된 음성신호

(14)도 반향필터

(31)의 영향이 상당히 없어졌음을 확인할 수 있다.Finally, the estimated sound source signal

Echo cancellation filter using independence of 14

In most methods of learning 13, the echo cancellation filter as shown in FIG.

(41) is used. In fact, the echo filter as shown in FIG.

As an ideal case where only the influence of (31) exists, an echo cancellation filter capable of sufficiently removing echoes by performing an echo cancellation experiment in the case of an audio signal source using Equation 3

(41) can be successfully learned and the resulting estimated speech signal

14 degree echo filter

It can be seen that the effect of (31) is considerably eliminated.

(12)를 마이크와 녹음장비를 사용해서 측정해야 한다. 이 경우에 불가피하게 녹음된 신호

(12)에는 녹음장비에 따라 다르긴 하지만, 대체적으로 10~30dB 정도의 신호대 잡음비(SNR, signal-to-noise ratio)를 갖게 하는 백색 가우시안 잡음(white Gaussian noise)들이 측정 잡음으로 포함된다. 따라서, 이상적인 경우의 모델인 도 3은 실제의 적용 환경에서는 도 5와 같이 반향신호에 백색 가우시안 잡음이 더해진 형태로 변경되어야 한다.However, in order to confirm and apply the blind echo cancellation performance in a real environment, the echo signal affected by the echo in the space where the echo exists.

(12) should be measured using a microphone and recording equipment. In this case, the signal inevitably recorded

(12) includes white Gaussian noise as measurement noise, which, depending on the recording equipment, typically has a signal-to-noise ratio (SNR) of around 10 to 30 dB. Therefore, FIG. 3, which is an ideal model, should be changed to a form in which white Gaussian noise is added to an echo signal as shown in FIG. 5 in an actual application environment.

(41)를 학습하면, 가산잡음의 영향으로 학습 수렴속도가 매우 느려지며, 그 결과로 추정된 음원신호의 품질도 매우 떨어진다.However, the echo cancellation filter shown in Figs. 5 to 4 shows a modified echo model in the presence of such noise.

Learning (41) slows down the learning convergence speed due to the addition noise, and the quality of the estimated sound source signal is also very low.

Accordingly, an object of the present invention is to provide a model and a method of deriving a method of improving the blind echo cancellation performance in the presence of measured noise and deriving a learning law using independent component analysis. It is done.

The present invention provides a blind echo cancellation model considering measured noise, using a microphone, and estimating output signals independent of each other by observing an input signal in order to improve echo cancellation performance when measurement noise is present. present.

The present invention also provides a first method for defining a blind least square cost function in the frequency domain for an estimated independent signal source in a blind signal processing structure having a signal observed at an input stage and an independent signal source estimated at an output stage using a microphone. step; A second step of defining an implicit least-squares cost function at all output stages by adding the implicit least-squared cost functions respectively defined in the frequency domain for the independent signal source; Deriving the blind least squares cost function for each adaptive filter; And a fourth step of applying a natural gradient method in the frequency domain with respect to the differentiated blind squared cost function.

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

(14)는 시간영역에서는 관찰신호

(12)와 반향제거 필터

(13)의 컨벌루션 (convolution)으로 상기 수학식 1과 같이 표현되지만, 주파수 영역에서는 다음의 수학식 4와 같이 표현된다.First, the blind least squares cost function for blind echo separation, which is simple to derive the learning law and is applicable to various structures, is presented in the frequency domain. Using the implicit least squares cost function, a learning rule is derived from the method of using one conventional microphone as shown in FIG. 1. Estimated Sound Source Signal

(14) is the observation signal in the time domain

12 and echo cancellation filter

The convolution of (13) is expressed by Equation 1 above, but is expressed by Equation 4 below in the frequency domain.

는 각각

의 주파수 영역을 표현한 신호이고,

는 성분별 곱셈(component-by-component multiplication) 연산이다. 주파수 영역에서의 암묵 최소제곱 비용함수는 다음의 수학식 5와 같다.here,

Are each

Is a signal representing the frequency range of,

Is a component-by-component multiplication operation. The implicit least square cost function in the frequency domain is given by Equation 5 below.

는 이산 푸리에 변환(discrete Fourier Transform)을 구하는 함수를 나타낸다. 추정된 음원신호

(14)의 샘플들 사이에 독립성을 최대화시키는 반향제거 필터

(13)의 학습법칙은 반향제거 필터

에 대한 경사 강 하법(gradient decent method)을 사용함으로써

를 최소화하는 형태로의 반향제거 필터

의 학습법칙을 얻을 수 있다.here,

Denotes a function for obtaining a discrete Fourier transform. Estimated Sound Source Signal

Echo cancellation filter to maximize independence between samples of 14

Learning rule of (13) is echo cancellation filter

By using the gradient decent method

Echo cancellation filter in a form that minimizes

The learning law of

의 학습식은 다음의 수학식 6과 같다.Echo cancellation filter obtained by differentiating Equation 5

The learning equation of Equation 6 is as follows.

부분은 수렴 성능의 향상을 위해서 피셔 점수 함수(Fisher score function),

로 대체하는 것이 좋으며, 학습법칙은 다음의 수학식 7과 같이 얻어진다.Where * is a complex conjugate operater. Corresponding to the error in Equation 6

The part uses Fisher score function,

It is better to replace with, and the learning law is obtained as shown in Equation 7 below.

Here, an echo cancellation filter learning method using an implicit least square cost function having improved convergence performance by adding a natural gradient method is derived as in Equation (8).

By applying the learning law of the echo cancellation filter derived as in Equation 8, the convergence performance can be improved more than in the conventional Equation 3 method. Several implicit echo cancellation models will be proposed in the future, and we will derive the learning law for each model using the implicit least square cost function as shown in Equation (5).

(12)가 존재할 때 1개의 음원신호

(14)만을 추정하게 되는데, 도 6에서는 2개의 반향제거 필터

(65)과

(66)를 통해서 2개의 서로 독립적인 신호

(67)와

(68)가 추정된다. 이는 백색 가우시안 잡음이 더해진 반향 영향을 받은 음성신호

(63)로부터 서로 독립인 2개의 신호

(67)과

(68)를 추정하는 문제이다.First, an echo cancellation model using one microphone is shown in FIG. 6. In the conventional model, one observation signal as shown in FIG.

1 sound source signal when (12) is present

Only 14 is estimated. In FIG. 6, two echo cancellation filters are shown.

65 and

2 mutually independent signals through 66

With 67

(68) is estimated. This is an echo affected voice signal with white Gaussian noise.

2 signals independent of each other from 63

(67) and

It is a matter of estimating (68).

(67)과

(68)로부터 상기 수학식 5와 같은 암묵 최소제곱 비용함수를 각각 구해보면, 다음의 수학식 9와 같다.Two output signals in the case of the 1 × 1 × 2CH echo cancellation model shown in FIG.

(67) and

The implicit least square cost functions of Eq. (5) can be obtained from Eq.

과

의 합으로 정의할 수 있고, 최종적으로 도 6에서 제시된 1 ×1 ×2CH 모델의 반향제거 필터

(65)와

(66)의 학습법칙은 다음의 수학식 10과 수학식 11과 같이 유도된다.Overall, the implicit least squares cost function

and

It can be defined as the sum of, and finally the echo cancellation filter of the 1 × 1 × 2CH model shown in Figure 6

65 and

The learning law of (66) is derived as in Equations 10 and 11 below.

Using the Equation 10 and Equation 11, two independent signals estimated in the presence of additive noise as an echo cancellation filter learning result are 0 to 4 kHz for the sound source signal having a sampling rate of 16 kHz and 16 kHz, respectively. An echo canceled speech signal having a signal bandwidth of 4 to 8 kHz can be obtained. In the case of the output speech signal having a bandwidth of 0 to 4 kHz mainly used in speech signal processing, the proposed method shows good performance even when comparing the same band signal of the echo canceled speech signal obtained by the conventional method. In this case, an echo canceled speech signal having a bandwidth of 0 to 4 kHz can be obtained for a sound source signal having a sampling rate of 16 kHz, but if a speech signal having a bandwidth of 4 kHz or more is desired, This can be solved by using a method that increases the sampling rate.

Shadow echo cancellation models using two microphones are presented in FIGS. 7 and 8. In the case of the conventional blind echo cancellation model, only one microphone is used to estimate one sound source signal. In the present invention, a model for estimating an echo canceled sound source signal using two microphones is also included. Suggesting.

(51)에 더해지는 형태로 나타나게 된다. 또한, 이때 더해지는 측정 잡음은 실험적으로 백색 가우시안 잡음의 특성을 갖고 있음을 확인할 수 있다. 이에, 도 7에서는 2개의 마이크로폰에 각각 더해진 백색 가우시안 잡음

(73)과

(74)를 추정하고자 하는 음성신호와 독립인 또 하나의 독립 신호원으로 간주하고, 출력단에서는 2개의 서로 독립적인 음성신호원과 측정잡음 신호원을 추정하는 문제로 간주하고 있다. 비록, 2개의 마이크로폰에 더해진 백색 가우시안 잡음이 동일하지 않지만, 통계적 특성은 같은 가우시안 분포를 가지기 때문에 1개의 출력에서만 측정 잡음에 해당하는 독립 신호원 성분을 추출할 수 있다.First, FIG. 7 is one of models using two microphones. In the real world, the audio signal recorded through the microphone is inevitably measured by the echo filter as shown in FIG.

It appears in the form added to (51). In addition, it can be confirmed that the measured noise added at this time has a characteristic of white Gaussian noise. Thus, in FIG. 7, the white Gaussian noise added to the two microphones, respectively.

73 and

(74) is regarded as another independent signal source independent of the audio signal to be estimated, and the output stage is regarded as a problem of estimating two independent audio signal sources and measurement noise signal sources. Although the white Gaussian noise added to the two microphones is not the same, since the statistical properties have the same Gaussian distribution, independent signal source components corresponding to the measurement noise can be extracted from only one output.

In the case of the 1 × 2 × 2CH echo cancellation model shown in FIG. 7, the observed echo signal affected by the noise is expressed in the following frequency domain.

{1, 2}는 마이크로폰 인덱스(index)이다.here,

{1, 2} is the microphone index.

If the estimated two echo canceled signals, which are output signals, are expressed in the frequency domain, the following equation (13) is given.

{1, 2}는 출력단의 인덱스이다.here,

{1, 2} is the index of the output terminal.

(82)와

(83)으로부터 수학식 5와 같은 암묵 최소제곱 비용함수를 구해보면 다음의 수학식 14와 같다.Two output signals in case of the 1 × 2 × 2CH echo cancellation model shown in FIG.

With 82

The implicit least square cost function of Eq. (5) is obtained from Eq.

들에 대한 합으로 정의하여 다음의 수학식 15와 같이 나타낼 수 있다.Overall, the implicit least squares cost function for all outputs is

It can be defined as the sum of the following equations as shown in the following equation (15).

Differentiate Equation 15 with respect to 1 × 2 × 2CH echo cancellation filters, respectively.

As a result of applying the soft gradient method, the learning law of the echo cancellation filters may be expressed as Equation 16 below.

Here, the natural gradient portion may be expressed as in Equation 17 below.

는 허미션 연산자(Hermitian operator)이다.here,

Is a Hermitian operator.

Applying the echo cancellation adaptive filter learning law of the 1 x 2 x 2CH model obtained in Equation 16, the speech signal transformed by the echo obtained through the two microphones and the echo signal from which the echo is removed from the speech signal having the measurement noise is 2 Can be recovered from one output stage. At this time, the white Gaussian noise is still added to the audio signal output from which the echo is removed, but the magnitude appears to be relatively reduced than that of the observed signal, and as a result, the quality of the speech is further improved. On the other hand, the white output Gaussian noise corresponding to the signal source of the white Gaussian measurement noise is learned.

(104),

(105) 및

(106)로 증가한 형태이다. 1 ×2 ×3CH 반향 제거 모델의 경우에 잡음의 영향을 받은 관찰된 반향신호는 주파수 영역에서 상기수학식 12와 같이 표시된다. 출력신호인 추정된 3개의 반향 제거된 신호를 주파수 영역에서 표현하면 상기 수학식 13과 같이 나타나고,

{1, 2, 3}는 출력단의 인덱스이다. 1 ×2 ×3CH 모델의 경우에 전제적인 암묵 최소제곱 비용함수를 구하면, 다음의 수학식 18과 같다.Another model for performing blind echo cancellation using two microphones is shown in FIG. 8 as a 1 × 2 × 3CH blind echo cancellation model. Compared with the 1 x 2 x 2CH blind echo model of FIG.

(104),

105 and

Increased to 106. In the case of the 1 × 2 × 3CH echo cancellation model, the observed echo signal affected by the noise is represented by Equation 12 in the frequency domain. When the estimated three echo canceled signals, which are output signals, are expressed in the frequency domain, they are represented by Equation (13).

{1, 2, 3} is the index of the output stage. In the case of the 1 × 2 × 3CH model, the implicit minimum square cost function is calculated by Equation 18 below.

는 상기 수학식 14와 같이 표현된다.here,

Is expressed as in Equation 14 above.

Differentiating Equation 18 with respect to the 1 × 2 × 3CH echo cancellation filters and applying the natural gradient method, the result is that the learning law of the echo cancellation filters is expressed as Equation 16 above. However, in the case of the 1 x 2 x 3CH model, unlike the case of the 1 x 2 x 2CH model shown in Equation 17, a natural slope portion is obtained as in Equation 19.

The advantage of the 1 × 2 × 3CH model is that, when compared to the 1 × 2 × 2CH model, three independent signal sources are learned from two observed signals, so that one of the three outputs is a white Gaussian source of measurement noise. In the case of the sound source signal having a learning rate of 16 kHz and the remaining two output stages, a reverberated speech signal source having a signal bandwidth of 0 to 4 kHz and 4 to 8 kHz, respectively, can be obtained. In the case of an output voice signal having a bandwidth of 0 to 4 kHz having important information in the voice signal, the method of using one conventional microphone and using one output signal, as well as the 1 × 1 × 2CH of FIG. Compared to the method using the model, a very high performance improvement can be expected in terms of voice quality. In addition, even in the case of the 1 × 2 × 2CH model of FIG. 7 presented in the present invention, improved results in convergence performance and voice quality can be obtained.

It will be apparent to those skilled in the art that various changes and modifications can be made without departing from the technical spirit of the present invention through the above description. Therefore, the technical scope of the present invention should not be limited only to the contents described in the embodiments, but should be defined by the claims.

As described above, the echo cancellation model and its induction method considering the measured noise according to the present invention improve the blind echo cancellation performance in the presence of the measured noise and derive the learning law of each model using independent component analysis. By doing so, it is possible to obtain a significantly improved performance in terms of convergence speed and quality of the speech signal obtained as a result of the conventional method.

Claims (10)

A blind echo cancellation model considering measured noise using a microphone and estimating output signals independent of each other by observing an input signal in order to improve echo cancellation performance when there is measurement noise. The method according to claim 1, Wherein the microphone is one, the input signal is one, the output signal source is two, and the echo cancellation model is 1 × 1 × 2CH. The method according to claim 1, The microphone has two microphones, the input signal is two, the output signal source is two, the echo cancellation model is 1 × 2 × 2CH characterized in that the blind echo cancellation model considering the noise. The method according to claim 1, The microphone has two microphones, the input signal is two, the output signal source is three, the echo cancellation model is 1 × 2 × 3CH, the blind echo cancellation model considering the measured noise. A first step of defining a blind least square cost function in the frequency domain with respect to the estimated independent signal source in the blind signal processing structure having the signal observed at the input and the independent signal source estimated at the output using a microphone; A second step of defining an implicit least-squares cost function at all output stages by adding the implicit least-squared cost functions respectively defined in the frequency domain for the independent signal source; Deriving the blind least squares cost function for each adaptive filter; And A fourth step of applying a natural gradient method in the frequency domain to the differentiated implicit least square cost function; Method of deriving the blind echo model considering the measured noise comprising a. The method according to claim 5, In the first step, there is one microphone, one observed signal, two independent signal sources, and the echo cancellation model is 1 × 1 × 2H. Way. The method according to claim 5, In the first stage, the microphone is 2, the observed signal is 2, the independent signal sources are 2, and the echo cancellation model is 1 x 2 x 2CH. . The method according to claim 5, In the first step, the microphone is 2, the observed signal is 2, the independent signal sources are 3, and the echo cancellation model is 1 x 2 x 3CH. . The method according to any one of claims 5 to 8, A method of deriving a blind echo model considering the measured noise, characterized in that the energy level of the adaptive filter exists by normalizing the estimated energy level of the independent signal source to a constant value. The method according to any one of claims 5 to 8, The method of deriving the blind echo cancellation model considering the measured noise, characterized in that to achieve a constant convergence performance by adapting the learning rate so that the ratio of the energy before the update of the adaptive filter and the amount of energy change after the update is constant.

Images