JP2007178590A - Object signal extracting device and method therefor, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an object signal extracting device and method therefor, and program which can extract the object signals from mixed signals at high speed, as well as with accuracy. <P>SOLUTION: The matrix-diagonalization calculation section 11 of the object signal extracting device 10 obtains an isolated matrix by independent component analysis, based on the maximum criteria of the posterior probability, and the object signal extracting section 12 extracts the object signals from the mixed signals by using the matrix diagonalization. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a target signal extraction apparatus, a target signal extraction method, and a program for extracting a target signal from a mixed signal.

One technique for separating and extracting a target signal from a mixed signal is a technique based on independent component analysis (ICA) (hereinafter also referred to as “ICA”). This is a technique for estimating a plurality of linearly mixed signals without using any knowledge about the original signal and mixing process, and is called blind source separation (BSS). First, blind sound source separation will be described.
Mixed Signal (Observed Signal) Model in Real Environment FIG. 8 is a diagram showing a blind source separation model by the ICA method. s _i is the signal from the signal source 11 _i , h _ji is the impulse response (frequency response) from the signal source 11 _i to the sensor 12 _j , P is the order of the impulse response, and N is the number of the signal sources 11 _i (N > 1) ), the sensor 12 _j number of M (M ≧ n) pieces of, when the discrete time n, the signal y _j observed by the sensor 12 _j is

と表現される。ここでＮ個の信号ｓ_iは統計的に互いに独立であると仮定する。観測信号ｙ_j（ｎ）は一定周期で標本化され、ディジタル信号系列とされている。
・分離信号のモデル
ブラインド音源分離では、式（１）の形で得られる観測信号と、長さがＱタップ、インパルス応答がｗ_ijのＮ×Ｍ個の分離フィルタ群１３_ijから成る分離系を用いて分離する。この分離フィルタ群１３_ijを用いて、分離して得られる信号ｙ_i（ｎ）は

It is expressed. Here, it is assumed that the N signals s _i are statistically independent of each other. The observation signal y _j (n) is sampled at a constant period to form a digital signal sequence.
Separation signal model In blind sound source separation, a separation system consisting of an observation signal obtained in the form of equation (1), N × M separation filter groups 13 _ij having a length of Q taps and an impulse response of w _ij Use to separate. A signal y _i (n) obtained by separation using the separation filter group 13 _ij is

と表される。図８には、Ｎ＝Ｍ＝２の場合について、信号源１１₁，１１₂とセンサ１２₁，１２₂間の混合過程と、センサ１２₁，１２₂の出力信号ｙ₁，ｙ₂から２×２個のフィルタ群１３_ijを用いるＩＣＡ法により分離信号ｙ₁（ｎ），ｙ₂（ｎ）を出力端子１４₁，１４₂に得る分離過程を示している。
分離フィルタ係数（周波数応答）ｗ_ijの推定には、独立成分分析（ＩＣＡ）と呼ばれる技術が広く用いられる。時間領域信号を周波数領域へ変換し、各周波数において分離行列を求める手法（周波数領域ＢＳＳ）が広く用いられている（例えば、非特許文献１参照）。

It is expressed. In FIG. 8, in the case of N = M = 2, the mixing process between the signal sources 11 ₁ and 11 ₂ and the sensors 12 ₁ and 12 ₂ and the output signals y ₁ and y _{2 of} the sensors 12 ₁ and 12 ₂ are 2 The separation process of obtaining separation signals y ₁ (n) and y ₂ (n) at output terminals 14 ₁ and 14 ₂ by the ICA method using two filter groups 13 _ij is shown.
A technique called independent component analysis (ICA) is widely used for estimating the separation filter coefficient (frequency response) w _ij . A technique (frequency domain BSS) for converting a time domain signal into a frequency domain and obtaining a separation matrix at each frequency is widely used (for example, see Non-Patent Document 1).

FIG. 9 shows a functional configuration of the frequency domain BSS method by ICA based on the conventional maximum likelihood method. The observed signals y ₁ (n) and y ₂ (n) are subjected to a frequency domain transform unit, for example, a short-time discrete Fourier transform (DFT) (multiplied by a window function, for example, discrete Fourier transform frame by frame while shifting every 1/2 frame) ) To convert it to a frequency domain signal Y (ω).
In the separation matrix estimation unit based on the likelihood maximization criterion, the separation matrix W (f) is estimated as in the following equation so that the frequency domain signals Y _i (f, m) of each output signal are independent of each other. .

In this way, separation at each frequency component is achieved.
In the time domain transform unit, the signal Y _i (f, m) output in the frequency domain is transformed into a time domain signal by, for example, inverse Fourier transform. The target signal is separated and extracted by selecting the target signal from the separated signals obtained in this way by using some method.

As one of the problems of performing ICA in the frequency domain, the independence of the outputs Z ₁ (f, m) and Z ₂ (f, m) is maintained even if the separation matrix rows are switched. That is, if the first and second rows of the separation matrix are switched, Z ₂ (f, m) is obtained as the _first output and Z ₁ (f, m) is obtained as the second output. Again, the two output signals are independent. That is, ICA makes output signals independent of each other, but does not constrain the output order.

Thus, when any two frequencies f ₁ and f ₂ are considered, for example, the output signals Z ₁ (f ₁ , m) and Z ₁ (f ₂ , m) are estimated signals for the same signal s _i. Not necessarily. Therefore, in the frequency domain ICA, the rows (each channel) of the separation matrix are correctly arranged so that Z _i (f ₁ , m) and Z _i (f ₂ , m) are estimates of the signal s _i of the same signal source. It is necessary to change. This is called a problem of permutation.
In addition, since the conventional ICA calculates the separation matrix based on the maximum likelihood criterion, it is necessary to repeatedly perform the operation until the value of the separation matrix converges to the optimum value, which requires time. Further, the calculated optimum value is a local optimum value and is not guaranteed to be a global optimum value.

Patent Document 1 describes a method for solving the problem of ICA Permutation using prior knowledge. In Patent Document 1, one component of the separation matrix is calculated by ICA based on the maximum likelihood criterion using the initial value of the separation vector (one component of the separation matrix) calculated from prior knowledge (see FIG. 10). . Prior frequency response information H ₁ (f) as prior knowledge is obtained by knowing the direction (target signal arrival direction) θ of the target signal source as H _j1 (f) = exp (j2πfτ _j1 ), τ _j1 = (d _j / c ) Sin θ ₁ calculated or measured in advance. Let the known frequency response be H ′ _j1 (f) = exp (j2πfτ _j1 ). For example, the constraint condition is given by
Σ _{j = 1} ^M H ′ _j1 W _1j (f) = W ₁ (f) H ₁ (f) = 1

A coefficient W ′ _1j (f) that minimizes the power of the error signal E (f, m) while satisfying the condition of this equation is obtained. Here, H ₁ (f) = [H ′ ₁₁ (f), H ′ ₂₁ (f)] ^T , W ₁ (f) = [W ₁₁ (f), W ₁₂ (f)]. The above equation is a constraint that the frequency response from the target signal to the output is 1 at all frequencies. This is a condition for outputting the target signal without distortion.

Further, Patent Document 2 solves the problem of permutation that occurs remarkably when the number of signal sources is smaller than the number of sensors. In Patent Document 2, independent component analysis is sequentially performed in the frequency domain (FD) and time domain (TD), and in particular, the signal identification process in FDICA is divided into a plurality of sub-blocks. The number of active sensors and the number of signal sources are substantially matched using the result of estimating the number of sources.
Patent Document 3 discloses a technique for estimating an average vector of a phoneme model (HMM) using a maximum a posteriori (MAP) estimation method.

JP 2004-302122 A JP 2005-91560 A JP-A-8-22296 A. Hyvaerinen and J. Karhunen and E. Oja, "Independent Component Analysis," John Wiley & Sons, 2001, ISBN 0-471-40540

However, in Patent Documents 1 and 2, since the separation matrix is obtained by ICA based on the maximum likelihood criterion, calculation time is required, and the accuracy of the obtained separation matrix and the extracted target signal is low. The problem is not solved.
Also, as in Patent Document 3, when applying the maximum posterior probability estimation method to HMM, the voice of an unspecified speaker can be used as known prior information, but ICA has no known prior information. For this reason, prior information needs to be estimated in order to apply the maximum posterior probability estimation method to ICA. In addition, in order to reduce the amount of calculation and perform the calculation at high speed, it is necessary to estimate the prior information with high accuracy.
The present invention has been made in view of the above problems, and provides a target signal extraction device, a target signal extraction method, and a program capable of extracting a target signal from a mixed signal at high speed and with high accuracy. Is an issue.

In order to solve the above-mentioned problem, the invention according to claim 1 observes a mixed signal including a target signal, which comes from a plurality of directions, by a plurality of sensors, and determines each frequency based on the observation signals from the plurality of sensors. A separation matrix calculation means for obtaining an independent component analysis of the time series and obtaining a separation matrix for each frequency, and a target signal extraction means for extracting a target signal from the mixed signal using the separation matrix for each frequency. An objective signal extraction device provided, wherein the separation matrix calculation means obtains a separation matrix by independent component analysis based on a criterion of maximum posterior probability.
According to the present invention, since the target signal extraction apparatus obtains a separation matrix by independent component analysis based on the maximum posterior probability criterion, the target signal extraction device performs separation at higher speed and higher accuracy than the conventional independent component analysis based on the maximum likelihood criterion. A matrix can be obtained, and a target signal can be extracted from the mixed signal at high speed and with high accuracy using the obtained separation matrix.

According to a second aspect of the present invention, in the target signal extraction device according to the first aspect, the separation matrix calculation means uses the separation matrix estimated based on the frequency response from the target signal source to the sensor as prior information. The separation matrix is obtained by independent component analysis based on the criterion of maximum posterior probability.
According to the present invention, the target signal extraction apparatus obtains the separation matrix by independent component analysis based on the criterion of the maximum posterior probability, using the separation matrix estimated based on the frequency response from the target signal source to the sensor as prior information. Therefore, prior information can be calculated with high accuracy, and the problem of permutation of the separation matrix channel does not occur. In addition, the separation matrix can be obtained at high speed and with high accuracy, and the target signal can be extracted from the mixed signal at high speed and with high accuracy using the obtained separation matrix.

According to a third aspect of the present invention, in the target signal extraction apparatus according to the first or second aspect, the separation matrix calculation means uses a signal in a section where only an interference signal exists to estimate a separation matrix estimated in advance. And a separation matrix is obtained by independent component analysis based on a criterion of maximum posterior probability.
According to the present invention, the target signal extraction apparatus uses, as a priori information, a separation matrix estimated by using a signal in a section where only an interference signal exists, and obtains a separation matrix by independent component analysis based on a criterion of maximum posterior probability. Since it is characterized by obtaining, the prior information can be estimated with high accuracy, and the separation matrix can be obtained with high speed and high accuracy. Therefore, the target signal can be extracted from the mixed signal with high speed and high accuracy using the obtained separation matrix.

According to a fourth aspect of the present invention, in the target signal extraction device according to any one of the first to third aspects, the separation matrix calculation means uses a signal in a section where only an interference signal exists, and minimizes the energy. A separation matrix is obtained by independent component analysis based on a criterion of maximum posterior probability, using a separation matrix estimated based on the constraint condition of (2) as prior information.
According to the present invention, the target signal extraction device uses a signal in a section where only an interference signal exists, uses a separation matrix estimated based on a constraint condition with the minimum energy as a priori information, and uses it as a criterion for maximizing the posterior probability. Since the separation matrix is obtained by independent component analysis based on it, prior information can be estimated with high accuracy, and the separation matrix can be obtained at high speed and with high accuracy. Therefore, the target signal can be extracted from the mixed signal with high speed and high accuracy using the obtained separation matrix.

According to a fifth aspect of the present invention, in the target signal extraction device according to any one of the first to fourth aspects, the separation matrix calculation means is a separation matrix estimated using the correlation of the observation signals of each channel. Is used as a priori information, and a separation matrix is obtained by independent component analysis based on a criterion of maximum posterior probability.
According to the present invention, the target signal extraction apparatus obtains a separation matrix by independent component analysis based on the criterion of the maximum posterior probability using the separation matrix estimated using the correlation of the observation signals of each channel as prior information. The prior information can be estimated with high accuracy, and the separation matrix can be obtained with high speed and high accuracy. Therefore, the target signal can be extracted from the mixed signal with high speed and high accuracy using the obtained separation matrix.

According to a sixth aspect of the present invention, in the target signal extracting device according to any one of the first to fifth aspects, the separation matrix calculating means is based on a constraint condition that the correlation of the observation signals of each channel is zero. Using the separation matrix estimated in this way as a priori information, the separation matrix is obtained by independent component analysis based on the criterion of maximum posterior probability.
According to the present invention, the target signal extraction device uses the separation matrix estimated based on the constraint condition that the observation signal of each channel is 0 as the prior information, and performs independent component analysis based on the criterion of the maximum posterior probability. In order to obtain the separation matrix, it is possible to extract features independent from the mixed signal with high accuracy and estimate the prior information with high accuracy, and to obtain the separation matrix with high speed and high accuracy. Therefore, the target signal can be extracted from the mixed signal with high speed and high accuracy using the obtained separation matrix.

The invention according to claim 7 is a signal observation step of observing a mixed signal including a target signal, which comes from a plurality of directions, by a plurality of sensors, and a time series for each frequency based on the observation signals from the plurality of sensors. A frequency domain transformation step to be obtained; a separation matrix calculation step for obtaining a separation matrix for each frequency by performing independent component analysis based on a criterion of maximum posterior probability for the time series obtained in the frequency domain transformation step; and the separation matrix There is provided a target signal extraction method comprising a target signal extraction step of extracting a target signal from a mixed signal using the separation matrix obtained in the calculation step.
According to the present invention, by obtaining a separation matrix by independent component analysis based on a criterion with the maximum posterior probability, it is possible to obtain a separation matrix at higher speed and with higher accuracy than independent component analysis based on a criterion with the maximum likelihood. . Therefore, the target signal can be extracted from the mixed signal with high speed and high accuracy using the obtained separation matrix.

According to an eighth aspect of the present invention, a prior information estimation step for estimating, as prior information, an initial value of a separation matrix for extracting a target signal from a mixed signal, and independent component analysis based on a maximum posterior probability are performed on a computer. Using the separation matrix calculation step for obtaining a separation matrix by sequentially updating the prior information estimated in the prior information estimation step, and using the separation matrix obtained in the separation matrix calculation step, A program for executing a target signal extraction step of extracting a target signal is provided.
According to the present invention, by performing independent component analysis based on the criterion of maximum posterior probability using the estimated prior information, the separation matrix can be obtained at higher speed and higher accuracy than the independent component analysis based on the criterion of maximum likelihood. Can be sought. Therefore, the target signal can be extracted from the mixed signal with high speed and high accuracy using the obtained separation matrix.

  According to the present invention, since the target signal extraction apparatus obtains a separation matrix by independent component analysis based on the maximum posterior probability criterion, the target signal extraction device performs separation at higher speed and higher accuracy than the independent component analysis based on the conventional maximum likelihood criterion. The matrix can be obtained, and the target signal can be extracted from the mixed signal at high speed and with high accuracy using the obtained separation matrix.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
(First embodiment)
FIG. 1 is a block diagram showing a functional configuration of a target signal extraction apparatus 10 according to the first embodiment of the present invention. As shown in the figure, the target signal extraction device 10 includes a separation matrix calculation unit 11 and a target signal extraction unit 12.

The separation matrix calculation unit 11 obtains a time series for each frequency based on observation signals from a plurality of sensors, and obtains a separation matrix for each frequency by performing independent component analysis on the time series based on a criterion of maximum posterior probability. .
The separation matrix calculation unit 11 includes a prior information estimation unit 111. The prior information estimation unit 111 estimates a separation matrix (its initial value) as prior information for independent component analysis based on the maximum posterior probability criterion.

The prior knowledge holding unit 112 holds prior knowledge. In this embodiment, the prior measurement value of the frequency response from the target signal source to the sensor is held as prior knowledge. In the present embodiment, the initial value of the separation matrix is estimated based on the premeasured value of the frequency response.
These functional configurations are realized by hardware such as a CPU and a memory included in the target signal extraction apparatus 10 and software such as a program and data stored in the memory.

Next, a processing procedure performed by the target signal extraction apparatus 10 will be described with reference to FIG. Here, there will be described a case where there is one noise source, the target sound source is the user's voice, and these sounds are input to the mobile phone as the target signal extraction device 10. The separation matrix is assumed to be a 2 × 2 (2 rows × 2 columns) matrix. Furthermore, the observation signals are Y ₁ (n), Y ₂ (n), the separated signals are Z ₁ (n), Z ₂ (n), and Z ₁ (n) is extracted as the separated target signal. And In addition, the frequency response from the user's mouth (target signal source) to the microphone (sensor) of the mobile phone is measured in advance, and is stored in advance in the prior knowledge holding unit 112 as a prior measurement value (prior knowledge). .
First, in Step 1, the prior measurement value stored in the prior knowledge holding unit 112 is acquired.
In Step 2, this value is set as one column of the separation matrix (that is, a ₁ and a _{2 in} the second column of the separation matrix of the following expression (1)).

In Step 3, the voice data is measured and converted to the frequency domain.
In Step 4, the fluctuation of W is absorbed by ICA (independent component analysis based on the maximum posterior probability criterion) of MAP estimation. Specifically, prior knowledge about W is introduced instead of the conventional W = arg maxP (W / data), and W is estimated by Equation (2) by MAP estimation.
W = arg maxP (W / data) P (W) (2)
Assuming a probability distribution of W with a ₁ and a ₂ obtained in the above steps as an average value of W, estimation is performed by sequentially updating W according to Equation (3) based on the maximum posterior probability criterion.

  For example, when assuming a normal distribution, the following equation is obtained.

Here, μ is a prior estimated value, and σ ² is a variance value representing a change in the prior estimated value of the separation matrix.
In Step 5, the target signal is extracted by performing separation using the separation matrix estimated in the above Step.
In Step 6, the target signal is restored to a time domain signal.
(Second Embodiment)
Next, a second embodiment will be described. FIG. 3 is a block diagram showing a functional configuration of the target signal extraction apparatus 10 according to the second embodiment of the present invention. As shown in the figure, the separation matrix calculation unit 11 according to the second embodiment includes the prior information estimation unit 111 and the prior knowledge holding unit 112 included in the separation matrix calculation unit 11 according to the first embodiment. The silent section detecting unit 113 is provided.

The prior information estimation unit 111 according to the second embodiment uses a signal in a section (silent section) in which only an interference signal exists, in addition to the functions of the prior information estimation unit 111 according to the first embodiment. Thus, a function of estimating one component of the separation matrix is provided.
Next, a processing procedure performed by the target signal extraction apparatus 10 according to the present embodiment will be described with reference to FIG. The preconditions here are the same as the preconditions described with reference to FIG. 2 in the first embodiment.
First, in Step 1, a prior measurement value is acquired from the prior knowledge holding unit 112.
Next, in Step 2, the value acquired from the prior knowledge holding unit 112 is set to one column of the separation matrix (that is, a ₁ and a _{2 in} the second column of the separation matrix shown in the following formula (1)).

In Step 3, the voice data is measured and converted to the frequency domain.
In Step 4, a section where the user is not speaking is detected.
In Step 5, w ₁ is obtained based on the energy minimization criterion of the output Z ₁ using observation data (Y ₁ and Y ₂ ) of the section not uttered by the user and the following equation (2).

As a result, w ₁ is estimated by equation (4):
w ₁ = a ₁ * (R ₂₂ -R ₂₁ ) / (R ₁₁ -R ₁₂ ) (4)
Here, R = E [y _i y _j ] (5)
In Step 6, the fluctuation of W is absorbed by ICA of MAP estimation. That is, prior knowledge about W is introduced instead of the conventional W = arg maxP (W / data), and W is estimated by Equation (6) by MAP estimation.
W = arg maxP (W / data) P (W) (6)
By assuming the probability distribution of W with w ₁ and a ₁ , a ₂ obtained in the above step as the average value of W, and sequentially updating W by equation (7) based on the maximum posterior probability criterion presume.

  For example, when assuming a normal distribution, the following equation (8) is obtained.

Here, μ is a prior estimated value, and σ ² is a variance value representing a change in the prior estimated value of the separation matrix.
In Step 7, separation is performed using the separation matrix estimated in the above Step, and a target signal is extracted.
In Step 8, the target signal is restored to a time domain signal.
(Third embodiment)
Next, a third embodiment will be described. FIG. 5 is a block diagram showing a functional configuration of the target signal extraction apparatus 10 according to the third embodiment of the present invention.

The prior information estimation unit 111 according to the third embodiment, in addition to the function of the prior information estimation unit 111 according to the second embodiment, uses the correlation of the observation signals of each channel to obtain 1 of the separation matrix. It has a function to estimate components.
Next, with reference to FIG. 6, the procedure of the process performed by the target signal extraction apparatus 10 according to the present embodiment will be described. The preconditions here are the same as the preconditions described with reference to FIG. 2 in the first embodiment.
First, in Step 1, a prior measurement value is acquired from the prior knowledge holding unit 112.
Next, in Step 2, the pre-measurement value acquired in Step 1 is set to one column of the separation matrix (that is, a ₁ and a _{2 in} the second column of the separation matrix of the following formula (1)).

In Step 3, the voice data is measured and converted to the frequency domain.
In Step 4, a section where the user is not speaking (a section including only interference waves) is detected.
In Step 5, w ₁ is obtained based on the energy minimization criterion of the output Z ₁ using observation data (Y ₁ and Y ₂ ) of the section not uttered by the user and the following equation (2).

As a result, w ₁ is estimated by equation (4):
w ₁ = a ₁ * (R ₂₂ -R ₂₁ ) / (R ₁₁ -R ₁₂ ) (4)
Here, R = E [y _i y _j ]
In Step 6, w ₂ is obtained by using a correlation condition expression (5) for decorrelation using the voice data Y of all utterance sections (Expression (6)).
E [Z ₁ , Z ₂ ] = 0 (5)
w ₂ = − (a ₂ w ₁ E (y ₁ y ₂ ) + a ₁ a ₂ E (y ₂ y ₂ )) / (w ₁ E (y ₁ y ₁ ) + a ₁ E (y ₁ y ₂ )) ( 6)

In Step 7, the fluctuation of W is absorbed by ICA of MAP estimation. That is, prior knowledge about W is introduced instead of the conventional W = arg maxP (W / data), and W is estimated by Equation (7) by MAP estimation.
W = arg maxP (W / data) P (W) (7)
Assuming the probability distribution of W with w ₁ , w ₂ and a ₁ , a ₂ obtained in the above step as the mean value of W, W is sequentially updated by Equation (8) based on the maximum posterior probability criterion. To estimate.

  For example, when assuming a normal distribution, the following equation (9) is obtained.

Here, μ is a prior estimated value, and σ ² is a variance value representing a change in the prior estimated value of the separation matrix.
In Step 8, separation is performed using the separation matrix estimated in the above Step, and a target signal is extracted.
In Step 9, the target signal is restored to a time domain signal.
FIG. 7 shows a speech signal restored using the conventional ICA method based on the maximum likelihood criterion (FIG. 7A) and a speech signal restored using the ICA method based on the maximum posterior probability of the method of the present invention. An example of (FIG. 7B) is shown. The horizontal axis of the graph represents time, and the vertical axis represents the signal level. It can be seen that the signal level fluctuation is smaller in the graph shown in FIG. 7B than in the graph shown in FIG. 7A, and the noise signal is more removed from the audio signal. In other words, the SNR (Signal-to-Noise Ratio) was confirmed to be improved for the audio signal restored using the method of the present invention, compared to the audio signal restored using the conventional ICA.

  As explained above, since independent component analysis based on the criterion of maximum posterior probability is performed using prior information, a global optimum value can be obtained for the separation matrix, and the maximum likelihood criterion without prior information is used. The target signal can be extracted from the mixed signal at a higher speed and with higher accuracy than the independent component analysis based on the above. In addition, since the separation matrix estimated with high accuracy can be used as prior information, it is possible to calculate the separation matrix at high speed without causing channel replacement.

  In the above-described embodiment, the case where the separation matrix is a 2 × 2 matrix has been described. However, the present invention is not limited to this, and the present invention can also be applied to a matrix of 3 rows or more and 3 columns or more. Further, in the flowchart of the above-described embodiment, the target signal extracted in the frequency domain has been described as being converted into the time domain. However, as shown in the functional configuration diagram, the target signal is converted using the separation matrix converted into the time domain. May be extracted.

In the present invention, signals from a plurality of directions are mixed and received, and it is not possible to directly observe only a target signal to be observed, and in a situation where other noise (noise) is superimposed on the target signal and observed, This is effective when estimating the target signal. In particular, the method of the present invention is effective under conditions where the SNR is low.
For example, by applying the target signal extraction device 10 according to the present invention to a mobile phone, only the user's voice can be extracted even in a situation where ambient noise is recorded in the mobile phone in addition to the user's input voice in a noisy environment. Is possible. In addition to mobile phones, the present invention can be applied to all communication terminals such as PDAs (Personal Digital Assistance), PHS (Personal Handyphone System), fixed phones, and fixed phones. It can also be used to eliminate interference in an environment where disturbances such as wireless communication occur.

It is a block diagram which shows the function structure of the target signal extraction apparatus which concerns on the 1st Embodiment of this invention. It is a flowchart which shows the procedure of the process which the target signal extraction apparatus which concerns on the embodiment performs. It is a block diagram which shows the function structure of the target signal extraction apparatus which concerns on the 2nd Embodiment of this invention. It is a flowchart which shows the procedure of the process which the target signal extraction apparatus which concerns on the embodiment performs. It is a block diagram which shows the function structure of the target signal extraction apparatus which concerns on the 3rd Embodiment of this invention. It is a flowchart which shows the procedure of the process which the target signal extraction apparatus which concerns on the embodiment performs. (A) is a graph of a speech signal restored using an ICA method based on the conventional maximum likelihood criterion, and (b) is a speech signal restored using an ICA method based on the maximum posterior probability criterion of the method of the present invention. It is a graph of. It is a figure which shows the model of the blind sound source separation (BSS) by the conventional ICA method. It is a figure which shows the function structure of the frequency domain BSS method by ICA based on the conventional maximum likelihood method. It is a block diagram which shows the function structure of the target signal extraction apparatus in the conventional patent document 1. FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Target signal extraction apparatus 11 Separation matrix calculation part 12 Objective signal extraction part 111 Separation matrix initial value calculation part 112 Prior information holding part 113 Silent section detection part

Claims (8)

Observe mixed signals including target signals coming from multiple directions with multiple sensors, determine the time series for each frequency based on the observation signals from these multiple sensors, analyze the time series for independent components, A target signal extraction device comprising separation matrix calculation means for obtaining a separation matrix, and target signal extraction means for extracting a target signal from the mixed signal using the separation matrix for each frequency,
The separation matrix calculation means includes
An objective signal extraction apparatus characterized in that a separation matrix is obtained by independent component analysis based on a criterion of maximum posterior probability. The separation matrix calculation means includes
The separation matrix is obtained by independent component analysis based on a criterion of maximum a posteriori probability using a separation matrix estimated based on a frequency response from a target signal source to a sensor as prior information. Signal extraction device. The separation matrix calculation means includes
The separation matrix is obtained by independent component analysis based on a criterion of maximum a posteriori probability using a separation matrix estimated using a signal in a section in which only an interference signal exists as prior information. The target signal extraction device described. The separation matrix calculation means includes
Using the signal of the section where only the interference signal exists, using the separation matrix estimated based on the minimum energy constraint condition as a priori information, the separation matrix is obtained by independent component analysis based on the criterion of the maximum posterior probability The target signal extraction device according to any one of claims 1 to 3. The separation matrix calculation means includes
5. The separation matrix is obtained by independent component analysis based on a criterion of maximum posterior probability using a separation matrix estimated using correlation of observation signals of each channel as prior information. 2. The target signal extraction device according to item 1. The separation matrix calculation means includes
2. The separation matrix is obtained by independent component analysis based on a criterion of maximum posterior probability, using as a priori information a separation matrix estimated based on a constraint that the correlation of observation signals of each channel is zero. 6. The target signal extraction device according to any one of items 1 to 5. A signal observation step of observing a mixed signal including a target signal, coming from a plurality of directions, by a plurality of sensors;
A frequency domain transforming step for obtaining a time series for each frequency based on observation signals from the plurality of sensors;
A separation matrix calculation step for obtaining a separation matrix for each frequency by performing independent component analysis based on a criterion of maximum posterior probability for the time series obtained in the frequency domain transformation step;
A target signal extraction method comprising: a target signal extraction step of extracting a target signal from the mixed signal using the separation matrix obtained in the separation matrix calculation step. On the computer,
A prior information estimation step for estimating an initial value of a separation matrix for extracting a target signal from a mixed signal as prior information;
A separation matrix calculation step for obtaining a separation matrix by sequentially updating the prior information estimated in the prior information estimation step using independent component analysis based on a criterion of maximum posterior probability;
A program for executing a target signal extraction step of extracting a target signal from a mixed signal using the separation matrix obtained in the separation matrix calculation step.

Images