JP4933975B2 - Signal extraction apparatus, method thereof, and program thereof - Google Patents

Description

  The present invention relates to a signal extraction apparatus, a method thereof, and a program thereof for satisfactorily extracting a target signal from mixed signals emitted from a plurality of signal sources.

  In recent years, with the advancement of multimedia technology, communication conferences such as a video conference using a loudspeaker call form using a sensor (for example, a microphone) and a speaker have become widespread. In that case, it is possible to make a natural call without conscious of microphones without installing microphones for the number of speakers on the desk, and to suppress only noise and reverberation that degrade voice quality and to observe only the target voice signal There is a need for an extraction device.

As such a conventional technique, there is a signal extraction device that suppresses reverberation and noise using a plurality of microphones. The details of this signal extraction device are described in Patent Document 1. FIG. 1 shows a functional configuration example of a conventional signal extraction device 500. The number of signal sources is A, of which the number of target signal sound sources (target sound sources) is one and the number of noise signal sound sources (noise sources) is A-1. Let B be the number of microphones. However, A and B are integers of 2 or more. B microphones 1 _b (b = 1,..., B) are connected to the signal extraction device 500. The signal extraction apparatus 500 includes a whitening filter coefficient calculation unit 110, an observed correlation function calculation unit 4, an inverse filter coefficient calculation unit 5, B whitening filter units 100 _b (b = 1,..., B), B The filter unit 2 _b (b = 1,..., B) and the adder unit 3 are included. Also, using the discrete time p, the target signal is M (p), the noise signal is N (p), and the observed signal R _b (p) observed by the microphone 1 _b . The observation signal R _b (p) is a signal obtained by mixing the target signal M _b (p) and the noise signal N _b (p).

First, the observed signal _R b (p) of a microphone _{1 b} is input to the whitening filter coefficient calculating section 110 and filter section _{2 b.} The whitening filter coefficient calculation unit 110 estimates an average spectrum of the target signal M (p) from the observation signal R _b (p). And the coefficient of the whitening filter which makes the average spectrum flat is calculated. The calculation of the whitening filter coefficient is performed as follows. First, a signal R ₁ (p) of a channel (for example, channel 1) specified in advance from the observation signal R _b (p) is divided into F frames. The frame length of all the frames is G (for example, 256). Then, the autocorrelation function U _1f (p) of the f (f = 1,..., F) -th frame is calculated by the following equation.
_{U 1f (p) = Σ q} R 1 (p) R 1 (p + q) (1)
However, q = G × (f−1), G × (f−1) +1,. . . , G × f-2, G × f-1
p = -G,. . . , 0,. . . , G. Taking an average of the frames which the average autocorrelation function U by the following equation ^- calculating a _{1 (p).
^{U - 1 (p) = Σ}} f = 1 F U 1f (p) (2)

The average autocorrelation function U ^- _{1 (p)} to convert the frequency domain, i.e., the following equation (3), the average spectrum V ₁ of the observation signal _(k). For example, Fourier transform or the like may be used as the method of transforming to the frequency domain.
_{^{V 1 (k) = FFT (}} U - 1 (p)) (3)
However, FFT (U ^- ₁ (p)) is performing Fourier-transform with respect to U ^- ₁ (p), and k shows a frequency.

As another average spectrum calculation method, there is a method in which the spectrum of the signal R ₁ (p) is calculated for each frame and averaged for the frame, as shown in the following equation (4).
V ₁ (k) = Σ _{f = 1} ^F | FFT (u ₁ (p)) | (4)

Next, the spectrum W ₁ (k) of the whitening filter is obtained by calculating the reciprocal of the calculated average spectrum as in the following formula (5).
W ₁ (k) = 1 / V ₁ (k) (5)
The filter coefficient w ₁ (p) of the whitening filter is calculated by performing an inverse Fourier transform IFFT on the spectrum W ₁ (k) and opening a window, as shown in the following equation (6). There are Hanning windows, Hamming windows, rectangular windows, triangular windows, Kaiser windows, and so on.
w ₁ (p) = window (IFFT (W ₁ (k))) (6)
These processes are performed for each channel b (b = 1,..., B) to obtain a whitening filter coefficient w _b (p).

観測相関関数計算部４は、白色化観測信号Ｒ’_ｎ（ｐ）（ｎ＝１，．．．，Ｎ）の間の相関関数ｒ’_１１（ｐ），ｒ’_１２（ｐ），．．．，ｒ’_１Ｂ（ｐ），ｒ’_２１（ｐ），．．．，ｒ’_２Ｂ（ｐ），．．．，ｒ’_Ｂ１（ｐ），．．．，ｒ’_ＢＢ（ｐ）を計算する。この相関関数を以下では、ｒ’_ｉｊ（ｐ）（ｉ＝１，．．．，Ｂ，ｊ＝１，．．．，Ｂ）と表す。
ここで、ｒ’_ｉｊ（ｐ）＝Σ_ｑＲ’_ｉ(ｑ) Ｒ’_ｊ(ｑ＋ｐ) （８）である。 Next, the whitening filter unit 100 _b (b = 1,..., B) applies the whitening filter coefficient w _b (p) to the observation signal R _b (p) as shown in the following equation (7). The observation signal R _b (p) is whitened by convolution. Hereinafter, the whitened observation signal is referred to as a whitened observation signal.

The observed correlation function calculator 4 correlates the correlation functions r ′ ₁₁ (p), r ′ ₁₂ (p),... Between the whitened observation signals R ′ _n (p) (n = 1,..., N). . . , R ′ _1B (p), r ′ ₂₁ (p),. . . , R ′ _2B (p),. . . , R ′ _B1 (p),. . . , R ′ _BB (p). Hereinafter, this correlation function is expressed as r ′ _ij (p) (i = 1,..., B, j = 1,..., B).
Here, a _{r 'ij (p) = Σ} q R' i (q) R 'j (q + p) (8).

The inverse filter coefficient calculation unit 5 calculates an inverse filter by solving the following simultaneous linear equations. The inverse filter coefficient is a filter that can suppress the noise signal and extract the target signal by convolution with an observation signal in which the target signal and the noise signal are mixed.

Here, R is a matrix of correlation functions between channels, R _ij is a matrix of correlation functions of i-th microphone 1 _i and j-th microphone 1 _j , c is a vector of inverse filter coefficients to be obtained, and c _b is an n-th vector of inverse filter coefficients, d is the blind object impulse response coefficient vector, d _b is the b th blind purposes impulse response vector, c b _(L) is the b-th inverse filter coefficient, B is the number of microphones, L is This is the number of taps of the inverse filter. [delta] _b is "1" when the target sound source ₆₁ is closest to the b-th microphone 1 _b in the microphone, in other cases a "0".

The simultaneous linear equations of Equation (9) are solved and the inverse filter coefficient vector c is calculated to obtain the inverse filter coefficient c _b (p). In order to solve the simultaneous linear equations of the equation (4), the same conditions as the MINT theory (the following equations (10) and (11) must be satisfied).
B> A + 1 (10)
L = A (K-1) / (NA) (11)
However, A is the number of sound sources, and K is the number of impulse response taps. For MINT theory, see “M. Miyoshi and Y. Kaneda,” “Invese Filtering of acoustics,” IEEE Trans Acoust. Speech Signal Process., Vol. ASSP-36, no2, pp. 145-152, Feb. 1998. "It is described in.

Each filter unit 2 _b obtains a signal S _b (p) by convolving an inverse filter coefficient c _b (p) with the observation signal R _b (p). All the outputs S _b (p) of the filter units 2 _b are added by the adder 3, and the addition result is output as the target signal S (p). This output signal S (p) suppresses noise and reverberation sound and extracts only the target sound.
JP 2006-66989 A

  When the conventional signal extraction device 500 is implemented, it must be performed in an environment where the microphone closest to the target sound source and the microphone closest to the noise source exist individually. Without this environment, in the simultaneous linear equations of Equation (9), the number of equations was insufficient, and appropriate inverse filter coefficients could not be obtained. Therefore, there is a problem that the reverberation of the voice that is the extraction purpose is suppressed, but the noise is not removed.

  An object of the present invention is to provide a signal extraction apparatus, a method thereof, and a program thereof that effectively remove reverberation and noise of a target signal by obtaining an inverse filter coefficient in consideration of a noise signal.

  The present invention originates from a specific signal source from B channel observation signals obtained by observing a mixed signal of signals emitted from a plurality of signal sources with B (B ≧ 2) sensors, respectively. It is a signal extraction device for extracting a target signal. The signal extraction device includes a noise correlation function calculation unit, an observation correlation function calculation unit, a weighted inverse filter coefficient calculation unit, a filter unit, and an addition unit. The noise correlation function calculation unit calculates a noise correlation function that is an inter-channel correlation function of a noise signal included in each B channel observation signal. The observation correlation function calculation unit calculates an observation correlation function that is an inter-channel correlation function of the B channel observation signal. The weighted inverse filter coefficient calculation unit obtains an inverse filter coefficient using the observed correlation function and the noise correlation function. The filter unit convolves the inverse filter coefficient with the observation signal. The adding unit generates a target signal by adding the outputs of the filter unit.

  Furthermore, you may have a whitening filter coefficient calculation part and a whitening filter part. The whitening filter coefficient calculation unit calculates a whitening filter coefficient using the average spectrum of the observation signal. The whitening filter unit whitens the input signal by convolving whitening filter coefficients. In this case, the observation correlation function calculation unit calculates an inter-channel correlation function of the B channel observation signal whitened by the whitening filter unit, and outputs the correlation function.

  Furthermore, the noise correlation function calculation unit may calculate an inter-channel correlation function of the noise signal that has been whitened by the whitening filter unit and output it as a noise correlation function.

  Furthermore, you may provide the noise area detection part which detects the noise signal contained in an observation signal. In this case, the noise correlation function calculation unit calculates a noise correlation function between the detected noise signals. Further, in this case, the whitening filter coefficient calculation unit may calculate the whitening filter coefficient using the average spectrum of the noise signal detected by the noise section detection unit and the average spectrum of the B channel observation signal.

  In addition, the signal extraction apparatus of the present invention may include a noise section detection unit, a noise correlation function calculation unit, an observation correlation function calculation unit, a weighted inverse filter coefficient calculation unit, a filter unit, and an addition unit. The noise section detection unit detects noise section information including only a noise signal from the B channel observation signal. The noise correlation function calculation unit calculates a noise correlation function that is an interchannel correlation function of a noise signal included in each of the B channel observation signals from the B channel observation signals and the noise interval information. The observation correlation function calculation unit, the weighted inverse filter coefficient calculation unit, the filter unit, and the addition unit are the same as described above.

  Further, in the case of the configuration of the signal extraction device, a whitening filter coefficient calculation unit and a whitening filter unit may be provided. The whitening filter coefficient calculation unit calculates the whitening filter coefficient using the average spectrum of the B channel observation signal. The whitening filter unit convolves whitening filter coefficients with the input signal. The observation correlation function calculation unit calculates a correlation function between channels of the B channel observation signal whitened by the whitening filter unit, and outputs the correlation function as an observation correlation function. The noise correlation function calculation unit calculates a noise correlation function from the whitened B channel observation signal and the noise interval information.

  In this case, the noise section detection unit may also detect a noise signal. Then, the whitening filter coefficient calculation unit calculates the whitening filter coefficient using the detected average spectrum of the noise signal and the average spectrum of the B channel observation signal.

  With the above configuration, an observation correlation function that is a correlation function between channels of an observation signal and a noise correlation function that is a correlation function between channels of a noise signal are obtained. Then, an inverse filter coefficient is obtained by constraining the observed correlation function with a noise correlation function. That is, an equation reflecting the characteristics of the noise signal increases. As a result, even if the microphone closest to the noise source and the microphone closest to the target sound source are the same, not only the reverberation of the target signal but also noise can be effectively suppressed.

  The best mode for carrying out the invention will be described below. In addition, the same number is attached | subjected to the process which performs the same part and the process which has the same function, and duplication description is abbreviate | omitted.

In the description of the signal extraction device of this embodiment, the signal to be extracted is described as an audio signal, but is not limited thereto. According to the signal extraction apparatus of this embodiment, for example, a target electromagnetic wave can be clearly extracted from an electromagnetic wave buried in noise. The sensor for observing the signal is, for example, a microphone. FIG. 2 shows a functional configuration example of the signal extraction apparatus 600-1 of the first embodiment, and FIG. 3 shows a main processing flow of the signal extraction apparatus 600-1. In the same manner as described in [Background Art], the number of signal sources is A, of which the number of target signal sound sources is one and the number of noise signal sound sources is A-1. Let B be the number of microphones. However, A and B are integers of 2 or more. B microphones 1 _b (b = 1,..., B) are connected to the signal extraction device 600-1. The signal extraction device 600-1 includes a noise section detection unit 120, a whitening filter coefficient calculation unit 110, a whitening filter unit 100 _b (b = 1,..., B), and a filter unit 2 _b (b = 1,. , B), an observation correlation function calculation unit 4, a noise correlation function calculation unit 130, and a weighted inverse filter coefficient calculation unit 140.

First, when the observation signal R _b (p) is observed by the microphone 1 _b (step S2), the observation signal R _b (p) is input to the noise section detection unit 120. The noise section detection unit 120 detects the noise signal N _b (p) included in the observation signal R _b (p) (step S4). For example, the target signal M _b (p) is considered as speech, and a signal other than the speech interval is output as the noise signal N _b (p) using general VAD (Voice Activity Detection).

The whitening filter coefficient calculation unit 110 calculates a whitening filter coefficient using the average spectrum of the observation signal R _b (p) from the microphone 1 _b (step S6). Since the above formulas (1) to (6) may be used for obtaining the average spectrum and the whitening filter coefficient, they are omitted here. The whitening filter unit 100 _b convolves the whitening filter coefficient w _b (p) with the corresponding observation signal R _b (p) according to the above equation (7) to obtain the whitening observation signal R ′ _b (p). Generate (step S8). Then, the observed correlation number calculator 4 uses the observed correlation function r ′ _ij (i = 1,...), _Which is a correlation function between channels, for the observed signal R ′ _b (p) that has been whitened from the above equation (8). , B j = 1,..., B) are obtained (step S10).

On the other hand, the whitening filter unit 100 _b includes whitening filter coefficients _w b a (p) by convoluting corresponding to the noise signal _N b (p), to produce a whitened noise signal N _'b (p) (Step S12 ).
The noise correlation function calculator 130 is a noise correlation function n ′ _ij (p) (i = 1,..., B j = 1,...) _That is a correlation function between channels of the whitened noise signal N ′ _b (p). ., B) is obtained (step S14).

Here, an _{n 'ij (p) = Σ} q N' i (q) N 'j (q + p) (12). In this addition process, q is added for the reverberation time. In FIG. 2, the whitening filter unit that whitens the observation signal R _b (p) and the whitening filter unit that whitens the noise signal N _b (p) are individually described. Also good.
The weighted inverse filter coefficient calculation unit 140 calculates the weighted inverse filter coefficient c _b (p) using the noise correlation function N and the observed correlation function R. Specifically, the inverse filter c _b (p) is obtained by solving the following simultaneous equations (13).

ただし、λ_１は補正係数であり、Ｒはチャネル間相関関数の行列、Ｒ_ijはｉ番目のマイクロホン１_ｉとｊ番目のマイクロホン１_ｊのチャネル間相関関数の行列、ｃは逆フィルタ係数のベクトル、ｃ_ｎは第ｎ番目の逆フィルタ係数のベクトル、ｄはブラインド目的インパルス応答係数ベクトル、ｄ_ｎは第ｎ番目のブラインド目的インパルス応答ベクトル、ｃ_ｂ（Ｌ）は第ｂ番目の逆フィルタ係数、Ｂはマイクロホン数、Ｌは逆フィルタタップ数である。δ_ｂは、目的音源６_１がマイクロホンの中で第ｂ番目のマイクロホン１_ｂに最も近いときに１となり、それ以外では０とする。行列ｄの要素は、ｄ_１〜ｄ_ＢでＢＬ個であり、「０」の数もＢＬ個である。従って行列ｄの要素数は２ＢＬ個である。

Where λ ₁ is a correction coefficient, R is a matrix of inter-channel correlation functions, R _ij is a matrix of inter-channel correlation functions of i-th microphone 1 _i and j-th microphone 1 _j , and c is a vector of inverse filter coefficients , c _n is a vector of the n-th inverse filter coefficients, d is the blind object impulse response coefficient vector, d _n is the n-th blind purposes impulse response vector, c b _(L) is the b-th inverse filter coefficients, B is the number of microphones, and L is the number of inverse filter taps. [delta] _b is becomes 1 when the target sound source ₆₁ is closest to the b-th microphone 1 _b in the microphone, 0 to otherwise. Elements of the matrix d _is a BL number in d 1 ~d _B, the number of "0" is also a BL number. Therefore, the number of elements of the matrix d is 2BL.

  The above formula (13) is shown in detail as shown in FIG. In order to solve the equation (13), the MINT theory explained in [Background Art] must be established.

加算部３は、生成された信号Ｍ’_ｂ（ｐ）を全てのチャネル分加算することで、目的信号Ｍ（ｐ）が生成する。つまり、以下の式により求められる。
Ｍ（ｐ）＝Σ_ｂ＝１ ^ＢＭ’_ｂ（ｐ） （１５） The obtained weighted inverse filter coefficient c _b (p) is input to the corresponding filter unit 2 _b . The filter unit 2 _b convolves the observation signal R _b (p) with the weighted inverse filter coefficient c _b (p) to generate a signal M ′ (p). That is, the following formula (14) is performed.

The adder 3 adds the generated signal M ′ _b (p) for all channels to generate the target signal M (p). That is, it is obtained by the following formula.
M (p) = Σ _{b = 1} ^B M ′ _b (p) (15)

Further, if the value of λ ₁ in the equation (13) is reduced, the equation (9) is approached, and the reverberation of the target signal can be suppressed, but the noise signal cannot be suppressed. by increasing the lambda ₁ value, it is possible to suppress the noise signal, the reverberation of the target signal is increased. Therefore, the value of λ ₁ may be changed as appropriate in consideration of the environment in which the target signal is extracted. Further, in the matrix R _E in the above formula (13), λ ₁ is multiplied by the noise correlation function matrix N, but λ ₁ may be multiplied by the observed correlation function matrix R. In this case, if the value of λ ₁ is increased, the reverberation of the target signal can be suppressed, but the noise signal cannot be suppressed. If the value of λ ₁ is decreased, the noise signal can be suppressed, but the reverberation of the target signal increases. Moreover, lambda ₁ was subjected to matrix N of the observation correlation function may be multiplied by a correction coefficient lambda _{1 'is} lambda ₁ and different values matrix R.

The observation signal R _b (p) includes a noise signal N _b (p) and a target signal M _b (p). Convolution of the whitening filter coefficient w _b (p) with the observation signal R _b (p) means that the target signal M _b (p) in the observation signal can be appropriately whitened. The noise signal N _b (p) cannot be appropriately whitened. Therefore, the noise signal N _b (p) from the noise section detection unit 120 is also whitened with w _b (p), and an appropriate weighted inverse filter is obtained by using the whitened observation signal and noise signal. be able to.

A matrix R used in obtaining the inverse matrix by the signal extracting unit 500 is compared with the matrix R _E used in signal extraction apparatus 600-1, the matrix R _E, and the matrix N of the noise correlation function is taken into account Yes. Therefore, the number of equations is increased compared to the prior art, and the characteristics of the noise signal N _b (p) are taken into consideration. Therefore, even in an environment where the microphone closest to the noise source and the microphone closest to the target sound source are the same, it is possible to obtain a more appropriate inverse matrix than before, and as a result, noise can be effectively suppressed.

[Modification 1]
Next, a signal extraction device 600-2 that is a first modification of the first embodiment will be described. An example of a functional configuration of the signal extraction device 600-2 is shown in FIG. The signal extraction device 600-2 is different from the signal extraction device 600-1 in that the noise signal N _b (p) is not whitened. That is, the noise correlation function calculation unit 130 obtains an interchannel correlation function of the noise signal N _b (p). For example, if the noise signal has a waveform close to a whitened signal, whitening is not necessary. Therefore, if the noise signal has a waveform close to that of a whitened signal, the noise signal whitening process can be omitted by using such a configuration, and the same effect as that of the signal extracting device 600-1 can be eliminated. Can be obtained.

[Modification 2]
Next, a signal extraction device 600-3 that is a second modification of the first embodiment will be described. A functional configuration example of the signal extraction device 600-3 is shown in FIG. The signal extraction device 600-3 is different from the signal extraction device 600-1 in that the noise section detection unit 120 is not provided and the noise signal N _b (p) is not detected. In an environment where the noise signal can be predicted to some extent, it may not be necessary to detect the noise signal. For example, in the case where it is desired to extract a voice signal target signal of a speaker in a conference room, the noise signal is often air conditioning in the conference room and the noise signal can be predicted in advance. Therefore, if a noise signal to be predicted is input in advance and the noise correlation function for the input noise signal is calculated by the noise correlation function calculation unit 130, the noise correlation function is obtained without detecting the noise signal. The same effect as that of the signal extraction device 600-1 can be obtained.

[Modification 3]
Next, a signal extraction device 600-4 of Modification 3 will be described. A functional configuration example of the signal extraction device 600-4 is shown in FIG. The signal extraction device 600-4 is different from the signal extraction device 600-1 in that the whitening filter unit 100 _b and the whitening filter coefficient calculation unit 110 are not provided. For example, in the case of a waveform close to a whitened signal for both the noise signal and the observation signal, it is not necessary to whiten the noise signal and the observation signal. Therefore, the calculation process of the whitening filter coefficient, the convolution process of the whitening coefficient with respect to the noise signal, and the convolution process of the whitening coefficient with respect to the observation signal can be omitted, and the same effect as that of the signal extraction device 600-1 can be obtained.

[Modification 4]
Next, a signal extraction device 600-5 of Modification 4 will be described. A signal extraction device 600-5 is shown in FIG. The signal extraction device 600-5 is different from the signal extraction device 600-4 in that the noise section detection unit 120 is not provided. As described in [Modification 2], if the noise signal is in an environment where the noise signal can be predicted in advance and both the noise signal and the target signal have a waveform close to the whitened signal, the signal extraction device 600-5 is used. With the configuration, the amount of calculation can be greatly reduced, and the same effect as the signal extraction device 600-1 can be obtained.

In the configuration of the signal extraction device 600-1 described in the first embodiment, when the power of the noise signal N _b (p) is larger than the power of the target signal M _b (p), the whitening filter unit 100 _b can generate a noise signal. As a result, the noise signal and the observation signal cannot be appropriately whitened, resulting in a problem that the reverberation suppression performance and noise suppression performance of the target signal deteriorate. Therefore, the signal extraction apparatus 600-6 of the second embodiment uses the observation signal and the noise signal instead of using only the observation signal when calculating the whitening filter coefficient (step S6 in FIG. 3). The whitening filter coefficient is obtained. The whitening filter coefficient obtained in this way is more accurate than the whitening filter coefficient obtained in the first embodiment in whitening the target signal.

FIG. 9 illustrates a functional configuration example of the signal extraction device 600-6 according to the second embodiment. The signal extraction device 600-6 is such that the observation signal R _b (p) and the noise signal N _b (p) from the noise section detection unit 120 are input to the whitening filter coefficient calculation unit. Different from 1. The reference number of this whitening filter coefficient calculation unit is 112.
The whitening filter coefficient calculation unit 112 uses the average spectrum VR (k) of the observation signal R _b (p) and the average spectrum VN (k) of the noise signal N _b (p) to obtain the target signal M _b (p). The average spectrum VM ₁ (k) of is estimated. Specifically, it is calculated | required by the following formula | equation (16).
VM ₁ (k) = VR (k) −λ ₂ VN (k) (16)
Here, λ ₂ is a correction coefficient, and 0 <λ ₂ <1. In Expression (16), VN (k) is multiplied by a correction coefficient, but VR (k) may be multiplied. Further, both VN (k) and VR (k) may be multiplied.
Then, by calculating V ₁ (k) in Expression (5) as VM ₁ (k) and using Expression (6), the whitening filter coefficient wM _b (p) of the target signal can be obtained. Using the noise signal N ′ _b (p) and the observation signal R ′ _b (p) whitened by wM _b (p), the noise correlation function calculation unit 130 and the observation correlation function calculation unit 4 perform noise correlation functions and observations. What is necessary is just to obtain | require a correlation function.

The observation signal R _b (p) includes a noise signal N _b (p) and a target signal M _b (p). Convolution of the observation signal R _b (p) with the whitening filter coefficient wM _b of the target signal means that the target signal M _b (p) in the observation signal can be appropriately whitened. The noise signal N _b (p) cannot be appropriately whitened. Therefore, the noise signal N _b (p) from the noise section detection unit 120 is also whitened with wM _b , and an appropriate weighted inverse filter can be obtained by using the whitened observation signal and noise signal. .

In the signal extraction device 600-1 of the first embodiment, since the whitening filter is calculated including the characteristics of the noise signal N _b (p), the target signal cannot be whitened. However, the configuration of the signal extraction device 600-6 subtracts the average spectrum VN (k) of the noise signal from the noise section detection unit 120 from the average spectrum VR (k) of the observation signal, so that the target signal with higher accuracy can be obtained. The average spectrum VM (k) of the target signal can be estimated, and the target signal can be whitened. As a result, the accuracy of the inverse filter coefficient is improved, and the performance of noise suppression and reverberation suppression of the target signal is improved. Therefore, even when the power of the noise signal N _b (p) is large, noise suppression and reverberation suppression of the target signal can be performed.

FIG. 10 shows a functional configuration example of the signal extraction device 600-7 of the third embodiment. The noise interval detector 120 in the signal extraction device 600-7 includes noise interval information T _b (b = 1,..., B) for each channel used to detect the noise signal N _b (p). A noise signal N _b (p) is also output. The noise segment information T _b, for example, a time interval t1~t2 of the noise signal contained in the observed signal. The noise interval information T _b is input to the noise correlation function calculation unit 130, and the noise signal N _b (p) is input to the whitening filter coefficient calculation unit 112. Then, the whitening filter coefficient calculation unit 112 and the whitening filter unit _100b are processed.

Here, the noise signal included in the whitened observation signal R ′ _b (p) from the whitening filter unit 100 _b is also whitened. In addition, the time interval of the noise signal included in the observation signal R _b (p) before whitening and the noise signal N ′ _b (p) included in the whitened observation signal R ′ _b (p) It is almost equal to the time interval. Therefore, if the noise section T detected by the noise section detection unit 120 is used, the whitened noise signal N ′ _b (p) can be detected from the whitened observation signal R ′ _b (p). This detection process is performed by the noise correlation function calculator 130 in the third embodiment, but is not limited to this. Therefore, it is not necessary to whiten the noise signal N _b (p) from the noise section detection unit 120, and the same effect as the signal extraction device 600-6 of the second embodiment can be obtained.

As a modification of the third embodiment, the observation signal R _b (p) is replaced with the whitening filter coefficient calculation unit 112 instead of the whitening filter coefficient calculation unit 112 as in the signal extraction device 600-8 illustrated in FIG. The whitening filter coefficient may be obtained using only Further, as signal extracting device 600-9 of FIG. 12 described, it may be from the signal extraction unit 600-8 by omitting the whitening filter coefficient calculating section 110 and the whitening filter unit 100 _b. In this case, the noise correlation function calculation unit 130 may detect an observation signal using the noise interval information T from the noise interval detection unit 120 and obtain a noise correlation function for the detected observation signal.

The figure which shows the function structural example of the conventional signal extraction apparatus 500. FIG. The figure which shows the function structural example of the signal extraction apparatus 600-1 of Example 1. FIG. The figure which shows the flow of the main processes of the signal extraction apparatus 600-1. The figure which showed the detail of Formula (13). FIG. 3 is a diagram illustrating a functional configuration example of a signal extraction device 600-2 according to the first embodiment. The figure which shows the function structural example of the signal extraction apparatus 600-3 of Example 1. FIG. The figure which shows the function structural example of the signal extraction apparatus 600-4 of Example 1. FIG. The figure which shows the function structural example of the signal extraction apparatus 600-5 of Example 1. FIG. The figure which shows the function structural example of the signal extraction apparatus 600-6 of Example 2. FIG. The figure which shows the function structural example of the signal extraction apparatus 600-7 of Example 3. FIG. The figure which shows the function structural example of the signal extraction apparatus 600-8 of Example 3. FIG. The figure which shows the function structural example of the signal extraction apparatus 600-9 of Example 3. FIG.

Claims (11)

Extraction of target signals from specific signal sources from B-channel observation signals obtained by observing mixed signals from multiple signal sources with B (B ≧ 2) sensors. In the signal extraction device
A noise correlation function calculation unit for calculating a noise correlation function that is an interchannel correlation function of a noise signal included in each of the B channel observation signals;
An observation correlation function calculator for calculating an observation correlation function which is an inter-channel correlation function of the B channel observation signal;
Using the observed correlation function and the noise correlation function, a weighted inverse filter coefficient calculation unit for obtaining an inverse filter coefficient;
A filter unit that convolves the inverse filter coefficient with the observation signal;
An adding unit that generates the target signal by adding the outputs of the filter unit;
A whitening filter coefficient calculation unit that calculates a whitening filter coefficient using an average spectrum of the observation signal;
A whitening filter unit that whitens the input signal by convolving the whitening filter coefficient;
A noise interval detection unit for detecting a noise signal included in the observation signal,
The noise correlation function calculator calculates a noise correlation function between the detected noise signals,
The whitening filter coefficient calculation unit uses an average spectrum of the observation signal average spectrum and the B channel of the detected noise signal in the noise interval detection section, the Der Rukoto adapted to calculate the whitening filter coefficients A characteristic signal extraction device. Certain mixing signals of the emitted signals from the plurality of signal sources in the channel correlation function of B-number (B ≧ 2) sensor noise signals included in each of the observation signal obtained B channels by observing each A noise correlation function calculator for calculating a noise correlation function;
An observation correlation function calculator for calculating an observation correlation function which is an inter-channel correlation function of the B channel observation signal;
When applied to the matrix consisting of the observed correlation function, an impulse is obtained for a specific channel, and a vector of 0 is obtained for the other channels, and when applied to a matrix weighted to the matrix consisting of the noise correlation coefficient the elements of 0 vector vector is obtained and calculated to Ruomomi with inverse filter coefficient calculating section as an inverse filter coefficients of each channel,
A filter unit that convolves the inverse filter coefficient with the observation signal;
An adding unit that generates the target signal by adding the outputs of the filter unit;
A signal extraction device comprising: The signal extraction device according to claim 2 ,
Furthermore, using the average spectrum of the observed signal, a whitening filter coefficient calculation unit that calculates a whitening filter coefficient;
A whitening filter unit that convolves the whitening filter coefficient with the input signal to whiten, and
The observation correlation function calculation unit calculates an inter-channel correlation function of the B channel observation signal whitened by the whitening filter unit, and outputs the correlation function as the observation correlation function Extraction device. The signal extraction device according to claim 3 ,
The signal extraction apparatus, wherein the noise correlation function calculation unit calculates an inter-channel correlation function of the noise signal whitened by the whitening filter unit and outputs it as the noise correlation function. The signal extraction device according to any one of claims 2 to 4 ,
Furthermore, a noise section detection unit for detecting a noise signal included in the observation signal is provided,
Before SL noise correlation function calculation unit, signal extraction apparatus, characterized in that to calculate the noise correlation function between said detected noise signal. The signal extraction device according to claim 5 ,
The whitening filter coefficient calculation unit calculates a whitening filter coefficient using an average spectrum of the noise signal detected by the noise section detection unit and an average spectrum of the observation signal of the B channel. A signal extraction device. Extraction of target signals from specific signal sources from B-channel observation signals obtained by observing mixed signals from multiple signal sources with B (B ≧ 2) sensors. In the signal extraction device
A noise interval detector that detects noise interval information including only noise signals from the B channel observation signal;
A noise correlation function calculator that calculates a noise correlation function that is an inter-channel correlation function of a noise signal included in each of the B channel observation signals from the B channel observation signal and the noise interval information;
An observation correlation function calculator for calculating an observation correlation function which is an inter-channel correlation function of the B channel observation signal;
When applied to the matrix composed of the observed correlation coefficients, an impulse is obtained for a specific channel, a vector of 0 is obtained for the other channels, and a matrix weighted to the matrix composed of the noise correlation function is obtained. the elements of 0 vector vector is obtained and calculated to Ruomomi with inverse filter coefficient calculating section as an inverse filter coefficients of each channel,
A filter unit that convolves the inverse filter coefficient with the observation signal;
An adder for generating the target signal by adding the outputs of the filter means;
A signal extraction device comprising: The signal extraction device according to claim 7 , wherein
A whitening filter coefficient calculation unit for calculating a whitening filter coefficient using an average spectrum of the observation signal of the B channel;
A whitening filter unit that convolves the whitening filter coefficient with the input signal;
The observed correlation function calculation unit calculates a correlation function between channels of the observation signal of the B channel that has been whitened by the whitening filter unit, and outputs the correlation function as the observed correlation function,
The signal extraction apparatus, wherein the noise correlation function calculation unit calculates a noise correlation function from the whitened B channel observation signal and the noise interval information. The signal extraction device according to claim 8 , wherein
The noise section detection unit also detects a noise signal,
The whitening filter coefficient calculation unit calculates a whitening filter coefficient using an average spectrum of the detected noise signal and an average spectrum of the observation signal of the B channel. . Extraction of target signals from specific signal sources from B-channel observation signals obtained by observing mixed signals from multiple signal sources with B (B ≧ 2) sensors. In the signal extraction method to
A process in which a noise section detection unit detects a noise signal included in each of the observation signals of the B channel;
A whitening filter coefficient calculation unit calculates a whitening filter coefficient using an average spectrum of the noise signal and an average spectrum of the observation signal;
A whitening filter unit that performs whitening by convolving the whitening filter coefficient with the observation signal of the B channel;
The whitening filter unit convolves the whitening filter coefficient with the noise signal to whiten;
Calculating an observation correlation function that is an inter-channel correlation function of the whitened B channel observation signal;
Calculating a noise correlation function that is an interchannel correlation function of the whitened noise signal;
When applied to the matrix comprising the observed correlation function , an impulse is obtained for a specific channel, and a vector of 0 is obtained for the other channels, and when applied to a matrix weighted to the matrix comprising the noise correlation function , 0 is obtained. The process of obtaining the vector elements from which the vectors are obtained as inverse filter coefficients for each channel ;
And a step of generating the target signal by convolving the inverse filter coefficient with the observed signal of the B channel and adding each of them. Signal extraction program for executing each processing of the signal extracting device according to the computer in any one of claims 1-9.

Images