JP2005181813A - Signal separating method, signal separating system and signal separating program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method and a system to reduce the number of repetitions required to learn separating coefficients in a signal separating system that conducts a separation process in a frequency region and to prevent occurrence of signal replacement after the separation. <P>SOLUTION: In a separation coefficient initial value setting section 8, phase information is extracted from the separating coefficients of frequency that are already learned and the information is converted into optimum information as preliminary information of the separating coefficients of other frequencies. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a signal separation method, a signal separation system, and a computer program for causing a computer to perform signal separation. In particular, the present invention relates to a signal separation method, a signal separation system, and a signal separation signal capable of separating a plurality of mixed sound and audio signals. It relates to computer programs.

  Conventionally, this type of signal separation system is used to separate two or more input time-series signals in which two or more plural sound / audio signals are spatially mixed into signals before mixing. It has been.

  As an example of a conventional signal separation system, for example, the descriptions of Patent Document 1 and Non-Patent Document 1 described later are referred to. In Patent Document 1 described later, it is possible to blindly separate a source signal from a signal (audio, image, radio wave, etc.) in a real environment in which background noise exists and is mixed by convolution. There has been proposed a blind signal separation processing device capable of performing a calculation process for solving the above-mentioned problem by a real-time process.

  FIG. 35 is a diagram illustrating a configuration of a blind signal separation processing device disclosed in Patent Document 1 described later. The configuration shown in FIG. 35 will be outlined below based on the description in Patent Document 1 below. A plurality of sensors 91 for detecting separated signals are provided, a separated signal storage unit 92 for storing the separated signals detected by each sensor 91, and a separated signal stored in the separated signal storage unit 92 are input. And a signal processing unit 93 that performs processing of taking out the separated signal and separating the separated signal into signals of respective signal sources. The signal processing unit 93 performs a discrete Fourier transform on the signal to be separated and converts it into a time-frequency domain component, and a mixing matrix based on the signal after the discrete Fourier transform process. Based on the mixing matrix estimation processing unit 95 that estimates each frequency component and the mixing matrix estimated for each frequency component, a separation matrix is calculated for each frequency component, and the separation matrix and the signal after the discrete Fourier transform processing are calculated. A signal separation processing unit 96 for calculating the product for each frequency component, a discrete inverse Fourier transform based on the product of the separation matrix for all frequency components obtained by the signal separation processing unit 96 and the signal after the discrete Fourier transform processing And an inverse discrete Fourier transform (I-DFFT) processing unit 97 for separating and reproducing the separated signal.

  The mixing matrix estimation processing unit 95 uses the estimated value of the mixing matrix of the frequency component (frequency number n) adjacent to the target frequency component (for example, frequency number n + 1) as a setting initial value when estimating the mixing matrix in each frequency component. Used. The separated signal separated by the signal processing unit 93 is output to the separated signal storage unit 98, and signals from each signal source are output to the subsequent circuit.

  In this way, in the signal separation processing device shown in FIG. 35, the value of the mixing matrix estimated at the previous frequency (frequency number n) is used as the initial value for the subsequent frequency component (frequency number n + 1). “Relay processing of separation matrix at adjacent frequency” is applied.

JP2003-78423A (5th page, FIG. 2) Ding Mathematics, Masafumi Otsuka, Masaki Serizawa, Teruo Niizuma, Kazuyoshi Sugai, “Blind Separation of Acoustic Signals in Real Environment Using ICA in Time-Frequency Domain”, IEICE Technical Report, SP2001-1, pp. 1-8, April 2001 (s. Ding, M. Otsuka, M. Ashizawa, T. Niitsuma, and K. Sugai, “BIND SOURCE SEPARATION OF REAL-WORD ACCOUNTIC SIGNAL A TED Q Report of IEICE, EA2001-1, SP2001-1, pp.1-8 (2001))

  However, the conventional signal separation system described with reference to FIG. 35 and the like is intended to avoid replacement of channel numbers of separated signals, but has the following problems. .

  The first problem is that the channel numbers of the separated signals are switched.

For example, it is assumed that voices simultaneously produced by two speakers are collected by two microphones, and a mixed voice signal of the collected two speakers is separated into voices of the respective speakers. In the following, the voices of the two speakers are called “signal sources” and the mixed voice signal picked up by the first microphone is picked up as “input time series signal of channel number 1” and picked up by the other microphone. The mixed signal is referred to as “input time series signal of channel number 2”. Further, the frequency spectrum of each speaker's voice after separation is referred to as “frequency-specific output time-series signal”. The mixed signal of the two signal sources δ1 and δ2 is λ1 (input time series signal of channel number 1) and λ2 (input time series signal of channel number 2), and the output time series by frequency separated based on λ1 and λ2. The signals are Y ₁ (f, t) (frequency-specific output time-series signal of channel number 1) and Y ₂ (f, t) (frequency-specific output time-series signal of channel number 2). When the channel number of the separated signal is switched, Y ₁ (n, t) at a certain frequency component (also referred to as frequency number or frequency bin) f = n on the frequency domain (frequency spectrum) is the signal source δ1. When the component Y ₂ (n, t) is a component of the signal source δ2, Y ₁ (m, t) at the other frequency component f = m is a component of the signal source δ2, Y ₂ (m, t) Is a component of the signal source δ1.

  Thus, the reason why the channel numbers of the separated signals are switched is that, in the configurations described in Patent Document 1 and Non-Patent Document 1, the separation coefficient learned (optimized) at the adjacent frequency is used as the initial learning stage. This is because it is used as a value.

  When the separation coefficient learned at a different frequency is used at another frequency, a phase shift occurs. Then, when a phase shift occurs in the separation coefficient and a plurality of frequencies that cannot be corrected by learning are successively selected, the phase shift is accumulated and the channel number of the separated signal is switched. In addition, a large phase shift occurs.

That is, with respect to the cost function of the adjacent frequency (see Non-Patent Document 1 above), there are two local minimum solutions, one of which is the signal source δ1 → Y ₁ (m, t); the signal source δ2 → Y _2. The other corresponds to the signal source δ1 → Y ₂ (m, t); the signal source δ2 → Y ₁ (m, t). When learning is started from an initial value in which the phase of the separation coefficient is greatly shifted, there is a case where convergence to another optimum solution (signal source δ2 → Y ₁ (m, t); Y ₂ (m, t) → signal source δ1). is there.

  The second problem is that the number of repetitions necessary for learning the separation coefficient is large.

  The reason is the same as the reason for the first problem. The greater the difference between the initial value having phase shift and the separation coefficient after learning, the greater the number of iterations required for learning (the number of iterations until optimization is acquired).

  The third problem is that it is necessary to continuously process the frequency. That is, basically, a restriction is imposed that the frequency to be processed next is limited only to the frequency (frequency number = n ± 1) adjacent to the processed frequency (frequency number = n).

  The reason is the same as the reason for the first problem described above. The phase shift increases as the frequency becomes discontinuous. That is, for example, when a separation coefficient learned at a frequency (frequency number = n) is used as an initial value of a discontinuous frequency (frequency number n ± r, where r is an integer of 2 or more), the phase of the separation coefficient is The deviation becomes larger.

  The fourth problem is that the number of repetitions required for learning the separation coefficient is large at the frequency to be processed first.

  The reason is that no consideration is given to the frequency to be processed first. For the frequency to be processed first, learning starts from a predetermined initial value, but at a frequency where the difference between the initial value and the separation factor after learning is large, it is necessary to learn the separation factor compared to a smaller frequency. There are many repetitions. In addition, since the amount of information useful for learning included in the signal before separation varies depending on the frequency, the number of repetitions necessary for learning the separation coefficient also varies.

  Accordingly, an object of the present invention is to provide a signal separation method, system, and computer program that can prevent the channel number of a separated signal from being changed.

  Another object of the present invention is to provide a signal separation method, system, and computer program that can reduce the number of repetitions necessary for learning of a separation coefficient.

  Still another object of the present invention is to provide a signal separation method, system, and computer program that can be processed discontinuously with respect to frequency.

  The invention disclosed in the present application will be summarized as follows.

  A signal separation system according to one aspect of the present invention includes a separation coefficient initial value setting unit that corrects a phase shift when using a learned frequency separation coefficient at another frequency, and a processing frequency. And a processing frequency control unit for controlling.

  A signal separation system according to another aspect of the present invention includes a Fourier transform unit that obtains an input signal of each frequency of each channel by converting an input time-series signal of a plurality of channels into a frequency domain, and a selection of a frequency to be processed. A separation frequency initial value setting unit that calculates a separation frequency initial value of a frequency to be processed, which is selected by the processing frequency control unit, based on signal source and sensor arrangement information; Based on the input signal and separation factor of each channel at the frequency to be processed, a separation output signal calculation unit for obtaining a separation output signal, and separation factor learning to learn the separation factor by using the separation output signal at the frequency to be processed Part.

According to a method of another aspect of the present invention, in a signal separation method for processing an input time-series signal of a plurality of channels in a frequency domain to separate signals,
Estimating signal source and sensor placement information (physical placement) from at least one processed frequency separation factor;
Calculating an initial value of a frequency separation coefficient to be processed based on the arrangement information.

According to the method of another aspect of the present invention, the step of obtaining the input signal of each frequency of each channel by converting the input time-series signals of a plurality of channels into the frequency domain,
Controlling the frequency to be processed;
Calculating a separation factor initial value of a frequency to be processed based on the arrangement information;
Obtaining a separated output signal based on the input signal and separation factor of each channel at the frequency to be processed;
Learning a separation coefficient using the separated output signal at the frequency to be processed.

  In the present invention, the step of controlling the frequency to be processed includes information on a predetermined processing order, statistical information obtained from an input signal of a plurality of frequencies of at least one channel among the input signals of each frequency of each channel, The frequency to be processed may be controlled based on at least one of the statistical information obtained from the separated output signal and the information on the already learned frequency separation factor.

  According to still another aspect of the present invention, there is provided a program comprising a computer constituting an apparatus for separating a plurality of input channel time-series signals in a frequency domain, a signal source and a sensor based on at least one frequency separation factor that has been processed. And a program for executing an initial value of a frequency separation coefficient to be processed based on the arrangement information.

  According to the present invention, in the processing frequency selected and controlled by the processing frequency control unit, the phase shift is corrected and the separation factor of the learned frequency is used, thereby avoiding the switching of the channel number of the separated signal. can do. This is because, in the present invention, after the phase shift of the separation coefficient is corrected, the initial value for learning of the separation coefficient is used, and accumulation of the phase shift can be prevented.

  According to the present invention, it is possible to reduce the number of repetitions necessary for learning of the separation coefficient by correcting the phase shift at the processing frequency controlled by the processing frequency control unit and using the already learned frequency separation coefficient. . This is because, in the present invention, since the phase shift of the separation coefficient is corrected, a separation coefficient closer to the learned separation coefficient can be used as an initial value.

  According to the present invention, the processing frequency controlled by the processing frequency control unit can be processed discontinuously with respect to the frequency by correcting the phase shift and using the already-learned frequency separation coefficient. This is because in the present invention, since the phase shift of the separation coefficient is corrected, it is not necessary to continuously process the frequency.

  For a detailed description of the present invention, reference will now be made to the accompanying drawings.

[First Embodiment]
FIG. 1 shows a configuration of a signal separation system according to a first embodiment of the present invention. Referring to FIG. 1, the signal separation system of the present embodiment includes an input terminal 15, an input terminal 16, a Fourier transform unit 1, a Fourier transform unit 2, an input signal multiplexing unit 3, and a processing frequency control unit. 41, the separation output signal calculation unit 5, the separation coefficient learning unit 61, the separation coefficient storage unit 71, the separation output signal calculation unit 9, the separation coefficient learning unit 10, the output terminal 17, and the output terminal 18. Have. The operation of the signal separation system of this embodiment will be described below.

  The input terminal 15 and the input terminal 16 are supplied with channel-specific input time-series signals. The input time series signal supplied to the input terminal 15 is subjected to an N-point discrete Fourier transform in the Fourier transform unit 1 and the input time series signal supplied to the input terminal 16 is subjected to an N-point discrete Fourier transform. Channel-specific frequency-specific input time-series signal Xk (f, t) converted into regions, channel number k = 1, 2, frequency number f = 0, 1,..., N / 2 -1, frame number t = 0, 1,...) Is output. In the frequency spectrum, the frequency of the nth component is fs × n / N (phase is 2πfs × n / N), the frequency number (frequency bin) n = 0 is a DC component, and n = N / 2 is a Nyquist frequency ( Nyquist frequency) (= fs / 2). In the following, frequency numbers (frequency bins) f = n, f = p on the frequency spectrum are described as frequencies n, p, etc. for simplicity.

  The channel-specific frequency-specific input time-series signal Xk (f, t) is supplied to the input signal multiplexing unit 3. The input signal multiplexing unit 3 multiplexes the channel-specific frequency-specific input time-series signals Xk (f, t) in different channels and outputs the frequency-specific input time-series signals X (n, t).

  The processing frequency control unit 41 supplies the frequency-dependent input time series signal X (n, t) having an arbitrary frequency f = n to the separated output signal calculation unit 5.

  The separated output signal calculation unit 5 reads the frequency-specific separation coefficient W (n), which is predetermined as an initial value, from the separation coefficient learning unit 61, and inputs the frequency-specific separation coefficient W (n), which is predetermined as an initial value, and the frequency Based on the time-series signal X (n, t), the frequency-specific output time-series signal Y (n, t) is calculated according to the following equation (1).

In the above equation (1), Y _j (n, t) represents the frequency-specific output time-series signal Y (n, t), which is a multiplexed signal, for channel numbers j = 1 and 2, and W _jk (N) shows frequency-specific separation coefficients W (n), which are multiplexing coefficients, for channel numbers j = 1, 2, k = 1, 2.

  The separation coefficient learning unit 61 learns the frequency-specific separation coefficient W (n) based on the frequency-specific output time series signal Y (n, t) obtained by the separation output signal calculation unit 5, and learns the frequency-specific separation coefficient. W (n) is supplied to the separated output signal calculator 5.

  The separated output signal calculation unit 5 and the separation coefficient learning unit 61 repeat the above-described operation until the learning is completed.

  When the learning is completed, the separated output signal calculation unit 5 supplies the frequency-specific output time series signal Y (n, t) to the output terminal 17, and the separation coefficient learning unit 61 separates the frequency-specific separation coefficient W (n). It supplies to the coefficient memory | storage part 71. FIG.

  When the processing at the first frequency f = n is completed, the processing frequency control unit 41, for example, the frequency f = n + 1 (or f = n−1) adjacent to the first processed frequency (the processed frequency) f = n. ) To switch the frequency to be processed. In this embodiment, the frequency f = n + 1 (or f = n−1) adjacent to f = n will be described for the sake of simplicity, but f = n ± r ( r may be an integer of 2 or more.

  Since the processing for frequencies other than the frequency f = n processed first is the same, the operation at f = n + 1 will be described using the separated output signal calculation unit 9 and the separation coefficient learning unit 10 of FIG.

  At the frequency f = n + 1, the separation coefficient learning unit 10 reads the frequency-specific separation coefficient W (n) already learned by the separation coefficient learning unit 61 from the separation coefficient storage unit 71, and separates the frequency-specific separation coefficient W for f = n. A value obtained by correcting the phase shift of (n) is supplied to the separated output signal calculation unit 9 as an initial value.

  As described above, in Patent Document 1 and Non-Patent Document 1 described above, in order to avoid channel replacement, the optimal solution W (n) for the frequency f = n is changed to the separation coefficient W for the adjacent frequency f = n + 1. Although the initial value of (n + 1) learning (optimization) is used, the optimum solution of the separation coefficient for the frequency f = n + 1 may converge to a solution different from the original due to the phase shift. It cannot be avoided.

  The operation of the separated output signal calculation unit 9 is the same as that of the separated output signal calculation unit 5, and the frequency-dependent input time-series signal X (n + 1, t) and the frequency-specific separation coefficient W (n + 1) supplied as the initial value are used. Based on this, the output time-series signal Y (n + 1, t) for each frequency is calculated.

  The separation coefficient learning unit 10 learns the frequency-specific separation coefficient W (n + 1) based on the frequency-specific output time-series signal Y (n + 1, t). The subsequent operation is the same as the operation at the frequency f = n at the beginning of the processing.

  As described above, in the frequency processed after the first frequency, the frequency-specific separation coefficient already learned in the adjacent frequency is used as the initial value of the separation coefficient. The learning is started from the value).

  By repeating the above-described processing up to the upper limit of the frequency to be processed, and similarly repeating from f = n−1 to the lower limit of the frequency to be processed, the output time-series signal Y (f, t) for each frequency at each frequency is obtained. obtain.

Each channel Y ₁ (f, t), Y ₂ (f, t) of the obtained output time-series signal Y (f, t) for each frequency at each frequency is a separated signal.

  FIG. 2 is a diagram illustrating a configuration of the processing frequency control unit 41 of FIG. Referring to FIG. 2, in this embodiment, the processing frequency control unit 41 has the same number of switches as the total number of frequencies. Of all the switches, only one is closed at a time. Assuming that the frequency at which the switch is closed first is f = n, the frequency at which the switch is closed next is f = n + 1 (or f = n−1), and the next is f = n + 2 (or f = n−2). ) And continuously change with respect to the frequency. Note that, as will be described later, it may be configured to select non-adjacent frequencies.

  FIG. 3 is a diagram showing a configuration of the separation coefficient learning unit 61 of FIG. Referring to FIG. 3, the separation coefficient learning unit 61 includes a switch 610, a separation coefficient update unit 611, a separation coefficient initial value storage unit 612, a terminal 6100, and a terminal 6101.

  The switch 610 is first connected to the terminal 6100, and outputs the frequency-specific separation coefficient initial value stored in the separation coefficient initial value storage unit 612. Thereafter, the switch 610 is connected to the terminal 6101.

  The separation coefficient updating unit 611 updates the frequency-specific separation coefficient using the frequency-specific output time-series signal and the frequency-specific separation coefficient used to obtain the signal. At this time, since the switch 610 is connected to the terminal 6101, the updated frequency-specific separation coefficient is output.

  FIG. 4 is a diagram illustrating a configuration of the separation coefficient learning unit 10 of FIG. Referring to FIG. 4, the separation coefficient learning unit 10 includes a switch 100, a separation coefficient update unit 101, a terminal 1000, and a terminal 1001.

  The switch 100 is first connected to the terminal 1000 and outputs the input frequency-specific separation factor as it is. Thereafter, the switch 100 is connected to the terminal 1001.

  The separation coefficient update unit 101 updates the frequency-specific separation coefficient using the frequency-specific output time-series signal and the frequency-specific separation coefficient used when obtaining the frequency-specific output time-series signal. At this time, since the switch 100 is connected to the terminal 1001, the updated separation factor for each frequency is output.

  Hereinafter, various preferred embodiments of the present invention including means for correcting a phase shift of a separation coefficient will be described in detail.

[Second Embodiment]
FIG. 5 is a diagram for explaining a preferred embodiment of the present invention. As shown in FIG. 5, the signal separation system according to the second exemplary embodiment of the present invention includes an input terminal 15, an input terminal 16, a Fourier transform unit 1, a Fourier transform unit 2, and an input signal multiplexing unit 3. , Processing frequency control unit 4, separation output signal calculation unit 5, separation coefficient learning unit 6, separation coefficient storage unit 7, separation coefficient initial value setting unit 8, separation output signal calculation unit 9, separation coefficient The learning unit 10, the output terminal 17, and the output terminal 18 are included. The present embodiment shown in FIG. 5 is different from the signal separation system of the first embodiment shown in FIG. 1 in the following points.

(1) A processing frequency control unit 4 is provided instead of the processing frequency control unit 41 of FIG.
(2) Instead of the separation coefficient learning unit 61, a separation coefficient learning unit 6 is provided.
(3) Instead of the separation coefficient storage unit 71, a separation coefficient storage unit 7 is provided.
(4) A separation coefficient initial value setting unit 8 is provided.

  FIG. 31 is a flowchart showing the operation of the second exemplary embodiment of the present invention. The operation of this embodiment will be described with reference to FIGS.

  The processing frequency control unit 4 to which the frequency-specific input time-series signal (step S101 in FIG. 31) is supplied controls selection of the frequency to be processed in accordance with a predetermined processing order (step S102 in FIG. 31).

  The separation coefficient learning unit 6 of the frequency to be processed first (step S103 in FIG. 31) supplies the frequency-specific separation coefficient determined as an initial value to the separation output signal calculation unit 5 (step S105 in FIG. 31). The calculation unit 5 calculates the frequency-specific output time-series signal (step S106 in FIG. 31), and learns the frequency-specific separation coefficient based on the frequency-specific output time-series signal (step S107 in FIG. 31).

  When the learning is completed, the separation coefficient learning unit 6 supplies the learned frequency-specific separation coefficient to the separation coefficient storage unit 7.

  The separation coefficient learning unit 6 re-learns the frequency-specific separation coefficient when the frequency-specific separation coefficient is supplied from the separation coefficient initial value setting unit 8.

  The separation coefficient storage unit 7 stores the frequency-specific separation coefficients of the frequencies for which learning has been completed, and supplies the separation coefficient obtained by multiplexing these to the separation coefficient initial value setting unit 8.

  Based on the separation coefficient supplied from the separation coefficient storage unit 7, the separation coefficient initial value setting unit 8 calculates a frequency-specific separation coefficient that is optimum for the next frequency (steps S102 and S103 in FIG. 31) (FIG. 31). In step S104), this is output to the separation coefficient learning unit 10 as an initial value of learning.

  The separation factor learning unit 10 learns the separation factor by frequency based on the supplied separation factor by frequency.

  The above operation is repeated until the frequency-specific separation coefficients for all the frequencies to be processed are learned (step S108 in FIG. 31).

  In the following description, the same or equivalent elements as those shown in FIGS. 1 to 4 are denoted by the same reference numerals. Hereinafter, the signal separation system according to the second embodiment of the present invention will be described with a focus on differences from the first embodiment described above.

  FIG. 6 is a diagram illustrating an example of the configuration of the processing frequency control unit 4 of FIG. Referring to FIG. 6, the processing frequency control unit 4 includes the same number of switches as the total number of frequencies, and a switch control unit 40.

  The switch control unit 40 stores a predetermined processing order, and first closes (turns on) the switch of the frequency f = m according to the order.

  When the processing at the frequency f = m ends (when the corresponding separation coefficient learning unit stores the learned separation coefficient in the separation coefficient storage unit and receives a completion notification), the switch control unit 40 determines in advance. According to the processing order, the switch of the other frequency f = 1 is closed. At this time, the other frequency f = 1 does not need to be one and may be plural. That is, except for the first frequency f = m, the switches of a plurality of frequencies may be closed simultaneously.

  FIG. 7 is a diagram illustrating an example of the configuration of the separation coefficient learning unit 6 of FIG. Referring to FIG. 7, the separation coefficient learning unit 6 includes a switch 60, a separation coefficient update unit 62, a switch 63, a separation coefficient initial value storage unit 64, a terminal 600, a terminal 601, a terminal 630, and a terminal. 631.

  First, when the switch 60 is connected to the terminal 600 and the switch 63 is connected to the terminal 631, the separation coefficient learning unit 6 outputs the frequency-specific separation coefficient initial values stored in the separation coefficient initial value storage unit 64.

  Thereafter, the switch 60 is connected to the terminal 601. The separation coefficient updating unit 62 updates the frequency-specific separation coefficient using the frequency-specific output time-series signal and the frequency-specific separation coefficient used when obtaining the frequency-dependent output time-series signal. At this time, since the switch 60 is connected to the terminal 601, the updated frequency-specific separation coefficient is output.

  When the frequency-specific separation coefficient is supplied to the separation coefficient learning unit 6, the switch 63 is connected to the terminal 630, the switch 60 is connected to the terminal 600, and the separation coefficient learning unit 6 displays the supplied frequency-specific separation coefficient. Output.

  Thereafter, similarly to the above, the switch 60 is connected to the terminal 601, and the separation coefficient updating unit 62 updates the separation coefficient.

  FIG. 8 is a diagram illustrating an example of the configuration of the separation coefficient initial value setting unit 8 in FIG. Referring to FIG. 8, the separation coefficient initial value setting unit 8 includes a separation coefficient phase calculation unit 80 and a separation coefficient initial value calculation unit 81.

  The separation factor phase calculation unit 80 calculates a separation factor phase based on the supplied separation factor, and the separation factor initial value calculation unit 81 calculates each frequency based on the separation factor phase and the supplied separation factor. Calculate the optimal frequency separation factor. Thereafter, the separation coefficient initial value setting unit 8 outputs the calculated frequency separation coefficient.

The separation coefficient phase calculation unit 80 divides the supplied separation coefficient into frequency-specific separation coefficients W (p). Here, p is a frequency that has already been learned. The divided frequency separation factor W (p) is further divided into channel-specific frequency separation factors W _jk (p). However, channel numbers j = 1, 2 and k = 1, 2.

Here, as shown in FIG. 9, it is assumed that two signal sources and two sensors (sensors 1 and 2) are spatially arranged, and the frequency-specific separation factor W _jk (p) is expressed as follows: Model as follows.

Here, p represents the p-th frequency component (fs × p / N), π is the pi, fs is the sampling frequency, N is the number of samples, dk is the position of the sensor in FIG. 9, and θj is the signal The direction of the source, c, is the propagation speed of the signal. Further, j ² = −1.

The signals (s ₁ (t), s ₂ (t)) from the two signal sources are transferred functions H (t) (between signal source 1 and sensor 1: H ₁₁ , between signal source 1 and sensor 2: H ₂₁ , between the signal source 2 and the sensor 1: H ₁₂ , between the signal source 2 and the sensor 2: H ₂₂ ), and reaches the sensors 1 and 2, and the input signal of the sensor 1 x ₁ (t) = H ₁₁ * s ₁ (t) + H ₁₂ * s ₂ (t) and input signal x ₂ (t) of sensor 2 = H ₂₁ * s ₁ (t) + H ₂₂ * s ₂ (T) (where * represents a convolution operation). The separation factor is given by the inverse matrix of the transfer function, and the separation factor W (ω) (where ω = 2πfs × n / N) on the frequency domain is given by the inverse matrix of the transfer function H (ω). When the above expression (1) of the convolution operation is Fourier transformed, Y (ω) = W (ω) · X (ω) is obtained on the frequency axis, and W _jk (p) is a 2 × 2 matrix W ( corresponds to the frequency component p of ω). det (p) is a determinant of the transfer function H (p). Further, d · sin θ ₁ / c and d · sin θ ₂ / c on the shoulders of the exponential functions (exp) in the expressions (2) to (5) are delay times for the signals of the signal sources 1 and 2 to reach the sensors 1 and 2. (Phase difference).

  The separation factor phase calculation unit 80 calculates the separation factor phase based on the equations (2) to (5).

When

Is calculated as follows.

  Then, these separation coefficient phases φ 1 and φ 2 (physical or geometrical arrangement information) are supplied to the separation coefficient initial value calculation unit 81. Note that the separation factor phases φ1 and φ2 can be calculated for all frequencies that have already been learned. Therefore, when there are a plurality of already learned frequencies, an average value of a plurality of already learned frequencies, a mode value of a histogram, and the like can be calculated and supplied to the separation coefficient initial value calculation unit 81.

  The separation coefficient initial value calculation unit 81 divides the supplied separation coefficient into frequency-specific separation coefficients W (p). Here, p is a frequency that has already been learned. In the separation coefficient initial value calculation unit 81, based on these frequency-specific separation coefficients W (p) and the separation coefficient phases φ1 and φ2 supplied from the separation coefficient phase calculation unit 80, frequencies other than p q = p ± The frequency separation coefficient W (q) at r (where q is an arbitrary integer satisfying 0 ≦ q ≦ N / 2) is calculated as follows.

  From the above equations (2) to (6), when W (q) is divided for each channel, the following equations (9) to (13) are given.

However,

In addition, the separation factor phase

and

  Is substituted into equation (13) from equation (9), equation (18) is obtained from equation (14).

However,

By the above formulas (14) to (17), W (p) is multiplied by exp (j ± φ ₁ ) and exp (j ± φ ₂ ) to correct the phase shift of the separation coefficient. That is, the separation coefficient initial value calculation unit 81 multiplexes the channel-specific frequency separation coefficients represented by the above equations (14) to (17) into the frequency-specific separation coefficients W (q), and sets the separation coefficient initial values. The unit 8 outputs the frequency-specific separation coefficient W (q).

  Next, the function and effect of the second embodiment of the present invention will be described.

  In the present embodiment, when the separation coefficient in the already learned frequency is used for other frequencies, the separation coefficient initial value setting unit 8 is configured to calculate the optimum frequency-specific separation coefficient for each frequency. Therefore, it is possible to prevent the channel number of the separated signal from being changed, and to reduce the number of repetitions necessary for learning the separation coefficient.

  Further, in the present embodiment, since the processing frequency is controlled by the processing frequency control unit 4, it is not necessary to continuously process the frequency.

  Further, in the present embodiment, since the frequency to be processed first is determined by the prior information in the processing frequency control unit 4, it is possible to reduce the number of repetitions necessary for learning the separation coefficient at the frequency to be processed first.

[Third Embodiment]
Next, a third embodiment of the present invention will be described. FIG. 10 is a diagram illustrating the configuration of the signal separation system according to the third embodiment of this invention. When the third embodiment of the present invention is compared with the signal separation system of the second embodiment shown in FIG. 5, in the third embodiment of the present invention, the separation coefficient initial value setting unit of FIG. The difference is that a separation coefficient initial value setting unit 82 is provided instead of 8. Other configurations are the same as those of the signal separation system according to the second embodiment shown in FIG. Therefore, in the following, the third embodiment of the present invention will be described focusing on differences from the second embodiment shown in FIG.

  FIG. 11 is a diagram showing a configuration of the separation coefficient initial value setting unit 82 shown in FIG. Referring to FIG. 11, the separation coefficient initial value setting unit 82 includes a separation coefficient phase calculation unit 820 and a separation coefficient initial value calculation unit 821.

  The operation of the separation factor phase calculation unit 820 and the operation of the separation factor phase calculation unit 80 in FIG. 8 are the same, and the separation factor phases φ1 and φ2 are calculated and output. The separation factor phases φ1 and φ2 are calculated based on the above equations (7) and (8).

The separation coefficient initial value calculation unit 821 calculates the frequency-specific separation coefficient W (q) based on the separation coefficient phases φ1 and φ2 as follows. In the present embodiment, the initial value of W (q) = W (p ± r) obtained by correcting the phase of W (p) is not used, and W ₁₁ (p), W ₁₂ (p), W ₂₁ (P), W ₂₂ (p) is set to 1 and calculated as follows.

However,

  The separation coefficient initial value calculation unit 821 multiplexes the channel-specific frequency separation coefficients represented by the above equations (19) to (22) into the frequency-specific separation coefficient W (q), and the separation coefficient initial value setting unit 82. Outputs the frequency-specific separation factor W (q).

  Also in the third embodiment of the present invention, the same effect as that of the first embodiment can be obtained.

[Fourth Embodiment]
Next, a fourth embodiment of the present invention will be described. FIG. 12 is a diagram illustrating the configuration of the signal separation system according to the fourth embodiment of this invention. Referring to FIG. 12, the signal separation system according to the fourth embodiment of the present invention is different from the signal separation system according to the second embodiment shown in FIG. 5 in the following points.

(1) A processing frequency control unit 42 is provided instead of the processing frequency control unit 4 of FIG.
(2) A separation coefficient initial value setting unit 83 is provided instead of the separation coefficient initial value setting unit 8 in FIG. 5, and a relearning frequency flag is input from the separation coefficient initial value setting unit 83 to the processing frequency control unit 42. Has been.

  In the following, the fourth embodiment of the present invention will be described focusing on the differences from the second embodiment shown in FIG.

  FIG. 13 is a diagram showing a configuration of the separation coefficient initial value setting unit 83 in FIG. Referring to FIG. 13, the separation coefficient initial value setting unit 83 includes a separation coefficient phase calculation unit 830, a separation coefficient initial value calculation unit 831, and a relearning frequency determination unit 832.

  The separation coefficient phase calculation unit 830 calculates the separation coefficient phase and supplies the separation coefficient phase to the separation coefficient initial value calculation unit 831 in the same manner as the separation coefficient phase calculation unit 80 in FIG. The frequency-separated separation coefficient phase calculated in step S3 is supplied to the relearning frequency determination unit 832.

  The operation of the separation coefficient initial value calculation unit 831 is the same as the operation of the separation coefficient initial value calculation unit 81 in FIG.

  The relearning frequency determination unit 832 determines a frequency that deviates significantly from the average value of the frequency-dependent separation coefficient phases and the mode value of the histogram as the relearning frequency, and outputs a relearning frequency flag.

  FIG. 14 is a diagram illustrating a configuration of the processing frequency control unit 42 of FIG. Referring to FIG. 14, the processing frequency control unit 42 includes the same number of switches as the total number of frequencies, and a switch control unit 420. The switch control unit 420 stores a predetermined processing order, and first closes the switch of the frequency f = m according to the order.

  When the processing at the frequency f = m is completed, the switch of the other frequency f = 1 is closed according to a predetermined processing order. The other frequency f = 1 does not need to be one and may be plural. That is, except for the first frequency f = m, the switches of a plurality of frequencies may be closed simultaneously.

  Further, when the re-learning frequency flag is supplied from the re-learning frequency determination unit 832 of the separation coefficient initial value setting unit 83, the switch control unit 420 closes the switch of the frequency and inputs the frequency-specific input time series. Output a signal.

  In the fourth embodiment of the present invention, in addition to the effect of the second embodiment, the re-learning frequency determination unit 832 re-learns the separation coefficient at a frequency that cannot be separated well. There is an effect that it can be improved.

[Fifth Embodiment]
Next, a fifth embodiment of the present invention will be described. FIG. 15 is a diagram illustrating a configuration of a signal separation system according to the fifth embodiment of this invention. When the fifth embodiment of the present invention shown in FIG. 15 is compared with the signal separation system of the fourth embodiment shown in FIG. 12, instead of the separation coefficient initial value setting unit 83, the separation coefficient initial value setting is performed. The difference is that the portion 84 is provided. Hereinafter, a fifth embodiment of the present invention will be described focusing on this difference.

  FIG. 16 is a diagram showing a configuration of the separation coefficient initial value setting unit 84 of FIG. Referring to FIG. 16, the separation coefficient initial value setting unit 84 includes a separation coefficient phase calculation unit 840, a separation coefficient initial value calculation unit 841, and a relearning frequency determination unit 842.

  The difference between the separation coefficient initial value setting unit 84 of the present embodiment and the separation coefficient initial value setting unit 83 shown in FIG. 13 is the operation of the separation coefficient initial value calculation unit 831 and the separation coefficient initial value calculation unit 841. Is the difference.

  The separation coefficient initial value calculation unit 841 calculates the frequency-specific separation coefficient by the same method as the separation coefficient initial value calculation unit 821 of FIG. That is, the separation coefficient initial value calculation unit 841 (which does not input the separation coefficient) calculates the frequency-specific separation coefficient W (q) according to the above equations (19) to (22) based on the separation coefficient phases φ1 and φ2. To do.

  This embodiment also has the same effect as the above-described fourth embodiment of the present invention.

[Sixth Embodiment]
Next, a sixth embodiment of the present invention will be described. FIG. 17 is a diagram illustrating a configuration of a signal separation system according to the sixth embodiment of this invention. Compared with the signal separation system of the second embodiment shown in FIG. 5, the sixth embodiment of the present invention shown in FIG. 17 has a processing frequency control unit 43 instead of the processing frequency control unit 4. The difference is provided. Hereinafter, a sixth embodiment of the present invention will be described focusing on the above differences.

  FIG. 18 is a diagram showing a configuration of the processing frequency control unit 43 in FIG. Referring to FIG. 18, the processing frequency control unit 43 includes the same number of switches as the total number of frequencies, and a switch control unit 430.

  The switch control unit 430 determines the frequency f = m to be processed first based on the information obtained by the statistical processing of the supplied input time-series signal by frequency, and closes the switch of the frequency f = m.

  When the process of the frequency f = m is completed, the switch control unit 430 closes another switch of the frequency f = 1 determined based on the statistical information obtained from the frequency-specific input time-series signal. The other frequency f = 1 does not need to be one and may be plural. That is, other than the first frequency f = m, the switches of a plurality of frequencies may be closed simultaneously.

Information obtained by statistical processing of input time-series signals by frequency
・ Power information of input time-series signals by frequency,
・ Information on correlation between channels of input time-series signals by frequency,
・ Information on eigenvalues and eigenvectors obtained when principal component analysis is performed on input time-series signals by frequency,
and so on.

  In the present embodiment, in addition to the effects of the second embodiment, the switch control unit 430 controls the processing frequency using the statistical information of the input time-series signals by frequency, so that the separation performance is improved in advance. There is no need to determine a good frequency, and there is an effect that the selection of the frequency to be processed can be optimally controlled with respect to the input time series signal inputted.

[Seventh Embodiment]
Next, a seventh embodiment of the present invention will be described. FIG. 19 is a diagram illustrating a configuration of a signal separation system according to a seventh embodiment of this invention. As shown in FIG. 19, the seventh embodiment of the present invention has a processing frequency control unit instead of the processing frequency control unit 4 as compared with the signal separation system of the third embodiment shown in FIG. The difference is that 43 is provided.

  The operation of the processing frequency control unit 43 according to the seventh embodiment of the present invention is the same as that described in the signal separation system according to the sixth embodiment. The effect of the seventh embodiment of the present invention is the same as the effect of the sixth embodiment described above.

[Eighth Embodiment]
Next, an eighth embodiment of the present invention will be described. FIG. 20 is a diagram illustrating the configuration of the signal separation system according to the eighth embodiment of this invention. Referring to FIG. 20, the eighth embodiment of the present invention has a processing frequency control unit 44 instead of the processing frequency control unit 42 as compared with the signal separation system of the fourth embodiment shown in FIG. 12. Is different. Hereinafter, an eighth embodiment of the present invention will be described focusing on the above differences.

  FIG. 21 is a diagram showing a configuration of the processing frequency control unit 44 of FIG. Referring to FIG. 21, the processing frequency control unit 44 includes the same number of switches as the total number of frequencies, and a switch control unit 440.

  The switch control unit 440 determines the frequency f = m to be processed first based on the statistical information obtained by the statistical processing of the supplied input time-series signal by frequency, and closes the switch of the frequency f = m.

  When the process of the frequency f = m is completed, the switch control unit 440 closes another switch of the frequency f = 1 determined based on the statistical information obtained from the frequency-specific input time-series signal. The other frequency f = 1 does not need to be one, and may be plural. That is, except for the first frequency f = m, the switches of a plurality of frequencies may be closed simultaneously.

  Further, when the relearning frequency flag is supplied, the switch control unit 440 closes the switch of the frequency and outputs the input time series signal for each frequency.

Information obtained by statistical processing of input time-series signals by frequency
・ Power information of input time-series signals by frequency,
・ Information on correlation between channels of input time-series signals by frequency,
・ Information on eigenvalues and eigenvectors obtained when principal component analysis is performed on input time-series signals by frequency,
and so on.

  In the eighth embodiment of the present invention, in addition to the effect of the sixth embodiment, the re-learning frequency determining unit 832 re-learns the separation coefficient at a frequency that cannot be separated well. There is an effect of improving.

[Ninth Embodiment]
Next, a ninth embodiment of the present invention will be described. FIG. 22 is a diagram illustrating the configuration of the signal separation system according to the ninth embodiment of this invention. Referring to FIG. 22, the ninth embodiment of the present invention has a processing frequency control unit 44 instead of the processing frequency control unit 42 as compared with the signal separation system of the fifth embodiment shown in FIG. 15. Is different.

  The processing frequency control unit 44 has the configuration shown in FIG. That is, the switch control unit 440 of the processing frequency control unit 44 determines the frequency f = m to be processed first based on the statistical information obtained by the statistical processing of the supplied frequency-specific input time-series signal, and the frequency f Close the = m switch. When the processing of the frequency f = m is completed, the switch control unit 440 closes another switch of the frequency f = 1 determined based on statistical information obtained from the frequency-specific input time-series signal. The other frequency f = 1 does not need to be one, and may be plural. That is, except for the first frequency f = m, the switches of a plurality of frequencies may be closed simultaneously. Further, when the relearning frequency flag is supplied, the switch control unit 440 closes the switch of the frequency and outputs the input time series signal for each frequency. Thus, the operation of the processing frequency control unit 44 is the same as the operation described in the signal separation system of the eighth embodiment.

  This embodiment has the same effect as the effect of the eighth embodiment described above.

[Tenth embodiment]
Next, a tenth embodiment of the present invention will be described. FIG. 23 is a diagram illustrating the configuration of the signal separation system according to the tenth embodiment of this invention. Referring to FIG. 23, the tenth embodiment of the present invention differs from the signal separation system of the second embodiment shown in FIG. 5 in the following points.

(1) A processing frequency control unit 45 is provided instead of the processing frequency control unit 4.
(2) Instead of the separation coefficient learning unit 10, a separation coefficient learning unit 11 is provided.

  First, the processing frequency control unit 45 outputs frequency-specific input time series signals of all frequencies.

  Each separated output signal calculation unit for all frequencies reads an initial value predetermined for each frequency from the separation coefficient learning unit, and outputs an output signal for each frequency.

  The processing frequency control unit 45 controls the processing frequency based on the frequency-specific output signal after learning zero or more times.

  The separation coefficient learning unit 11 has the same configuration as that of the separation coefficient learning unit 6.

  Hereinafter, a tenth embodiment of the present invention will be described focusing on the internal structure of the processing frequency control unit 45. FIG. 24 is a diagram showing a configuration of the processing frequency control unit 45 of FIG. Referring to FIG. 24, the processing frequency control unit 45 includes the same number of switches as the total number of frequencies and a switch control unit 450.

  The switch control unit 450 first closes all the switches, and outputs input time-series signals by frequency for all frequencies. Then open all switches.

  Then, the switch control unit 450 determines the frequency f = m to be processed first based on the information obtained by the statistical processing of the supplied output time-series signal by frequency, and closes the switch of the frequency f = m. When the process of the frequency f = m is completed, the switch control unit 450 closes the other switch of the frequency f = 1 determined based on the statistical information obtained from the frequency-specific output time series signal. The other frequency f = 1 does not need to be one and may be plural. That is, except for the first frequency f = m, the switches of a plurality of frequencies may be closed simultaneously.

Information obtained by statistical processing of output time series signals by frequency
・ Power information of output time series signal by frequency,
-Information on correlation between channels of output time-series signals by frequency,
-Eigenvalue and eigenvector information obtained when principal component analysis is performed on output time series signals by frequency,
and so on.

  In this embodiment, in addition to the operational effects of the second embodiment, the switch control unit 450 uses the statistical information of the output time-series signal for each frequency to control the processing frequency. As a result, the optimum frequency can be processed first.

[Eleventh embodiment]
Next, an eleventh embodiment of the present invention will be described. FIG. 25 is a diagram showing a configuration of a signal separation system according to the eleventh embodiment of the present invention. Referring to FIG. 25, the eleventh embodiment of the present invention is shown in FIG. 24 instead of the processing frequency control unit 4 as compared with the signal separation system of the third embodiment shown in FIG. The difference is that a processing frequency control unit 45 is provided.

  The operation of the processing frequency control unit 45 is the same as the operation described in the signal separation system of the tenth embodiment described with reference to FIG. This embodiment has the same effects as the tenth embodiment.

[Twelfth embodiment]
Next, a twelfth embodiment of the present invention will be described. FIG. 26 is a diagram showing a configuration of a signal separation system according to the twelfth embodiment of the present invention. Referring to FIG. 26, when compared with the signal separation system of the fourth embodiment shown in FIG. 12, the twelfth embodiment of the present invention replaces the processing frequency control unit 42 with a processing frequency control unit. The difference is that 46 is provided. Hereinafter, the twelfth embodiment will be described focusing on these differences.

  FIG. 27 is a diagram showing a configuration of the processing frequency control unit 46 of FIG. Referring to FIG. 27, the processing frequency control unit 46 includes the same number of switches as the total number of frequencies, and a switch control unit 460.

  The switch control unit 460 first closes all the switches and outputs input time-series signals for all frequencies.

  Thereafter, the switch control unit 460 opens all the switches. Then, based on the information obtained by the statistical processing of the output time-series signal by frequency, the frequency f = m to be processed first is determined, and the switch of the frequency f = m is closed. When the processing of the frequency f = m is completed, the other frequency f = 1 switch determined based on the statistical information obtained from the frequency-specific output time series signal is closed. At this time, the other frequency f = 1 does not need to be one and may be plural. That is, except for the first frequency f = m, the switches of a plurality of frequencies may be closed simultaneously.

  Further, when the relearning frequency flag is supplied, the switch control unit 460 closes the switch of the corresponding frequency and outputs the input time-series signal for each frequency of the corresponding frequency.

Information obtained by statistical processing of output time series signals by frequency
・ Power information of output time series signal by frequency,
-Information on correlation between channels of output time-series signals by frequency,
-Eigenvalue and eigenvector information obtained when principal component analysis is performed on output time series signals by frequency,
and so on.

  In addition to the effect of the tenth embodiment, the effect of the present embodiment is that the re-learning frequency determination unit 832 re-learns the separation coefficient at a frequency that cannot be separated well, thereby improving the separation performance. , Has the effect.

[Thirteenth embodiment]
Next, a thirteenth embodiment of the present invention will be described. FIG. 28 is a diagram showing a configuration of a signal separation system according to the thirteenth embodiment of the present invention. Referring to FIG. 28, the thirteenth embodiment of the present invention has a processing frequency control unit 46 instead of the processing frequency control unit 42 as compared with the signal separation system of the fifth embodiment shown in FIG. Is different. The operation of the processing frequency control unit 46 is the same as that described in the signal separation system of the twelfth embodiment. This embodiment has the same effect as the twelfth embodiment.

As described above, in the second to thirteenth embodiments, in the processing frequency control units 4, 42, 43, 44, 45, 46,
Information on the processing order set in advance,
Statistical information of input time series signals by frequency, and
Statistical information of output time series signals by frequency,
The case where the processing frequency is controlled based on any one of the information has been described as an example, but may be controlled in combination.

  Further, the processing frequency control unit may control the processing frequency independently of these three pieces of information.

  Further, in each of the above embodiments, the signal separation system has been described for the case where the number of input time-series signals is two for simplification of explanation, but the same applies to the case where there are three or more input time-series signals. Of course you can.

[Fourteenth embodiment]
Next, a fourteenth embodiment of the present invention will be described. FIG. 29 is a diagram showing the configuration of the speech recognition system according to the fourteenth embodiment of the present invention. As shown in FIG. 29, the fourteenth embodiment of the present invention includes a signal separation system 12, an output signal divider 13, a voice recognizer 140, a voice recognizer 141, an output terminal 142, and an output terminal. 143. Such a speech recognition system operates as follows.

  The signal separation system 12 supplies the output time series signal Y (f, t) for each frequency to the output signal dividing unit 13. The output signal dividing unit 13 divides the frequency-specific output time-series signal Y (f, t) into channel-specific frequency-specific output time-series signals Yj (f, t).

  The divided frequency-specific output time-series signals Yj (f, t) for each channel are supplied to the speech recognition unit 140 and the speech recognition unit 141 for each channel. The speech recognition unit 140 and the speech recognition unit 141 perform speech recognition based on the supplied frequency-specific output time-series signal Yj (f, t). The speech recognition unit 140 outputs the recognition output for each channel to the output terminal 142, and the speech recognition unit 141 outputs the recognition output for each channel to the output terminal 143.

  The signal separation system 12 is configured by any of the signal separation systems described in the first to thirteenth embodiments of the present invention.

  The speech recognition system according to the fourteenth embodiment of the present invention is for the case where the number of channels of the output time-series signal for each frequency is two, but the same applies to the case where there are three or more.

  The fourteenth embodiment of the present invention has an effect that individual voices can be recognized even when a plurality of voices are temporally overlapped by separating the signals by the signal separation system 12. Play.

[Fifteenth embodiment]
Next, a fifteenth embodiment of the present invention is described. FIG. 30 is a diagram illustrating the configuration of the signal transmission system according to the fifteenth embodiment of the present invention. Referring to FIG. 30, the fifteenth embodiment of the present invention includes a signal separation system 12, an output signal dividing unit 13, an inverse Fourier transform unit 144, an inverse Fourier transform unit 145, and an output signal transmitting unit 146. , An output signal transmission unit 147, an output terminal 148, and an output terminal 149. The signal transmission system of this embodiment operates as follows.

  The signal separation system 12 supplies the output time series signal Y (f, t) for each frequency to the output signal dividing unit 13.

  The output signal dividing unit 13 divides the frequency-specific output time-series signal Y (f, t) into channel-specific frequency-specific output time-series signals Yj (f, t).

  The divided frequency-specific output time-series signal Yj (f, t) for each channel is supplied to the inverse Fourier transform unit 144 and the inverse Fourier transform unit 145 for each channel.

  The inverse Fourier transform unit 144 and the inverse Fourier transform unit 145 respectively perform inverse Fourier transform on the supplied frequency-specific output time series signals Yj (f, t).

  The inverse Fourier transform unit 144 supplies the output time series signal for each channel to the output signal transmission unit 146, and the inverse Fourier transform unit 145 supplies the output time series signal for each channel to the output signal transmission unit 147. The output signal transmission unit 146 outputs the channel-specific transmission signal to the output terminal 148, and the output signal transmission unit 147 outputs the channel-specific transmission signal to the output terminal 149.

  The signal separation system 12 is any one of the signal separation systems described in the first to thirteenth embodiments.

  The signal transmission system of the fifteenth embodiment of the present invention is for the case where the number of channels of the output time-series signal for each frequency is two, but the same is true for the case of three or more.

  According to the fifteenth embodiment of the present invention, the signals are separated by the signal separation system 12 so that, for example, the driver's seat and the passenger's seat in the car can transmit each voice separately in a hands-free manner. Play.

[Sixteenth embodiment]
Next, a sixteenth embodiment of the present invention will be described. FIG. 32 is a diagram showing the configuration of the sixteenth embodiment of the present invention. Referring to FIG. 32, in this embodiment, the signal separation that constitutes one of the input device P201, the output device P203, and the signal separation system of the first to thirteenth embodiments of the present invention described above. A system P202 is provided.

  The signal separation program P204 is read into the signal separation system P202, and controls the operation of the signal separation system P202.

  By the signal separation program P204, the signal separation system P202 executes the same processing as any one of the signal separation systems of the first to thirteenth embodiments of the present invention.

[Seventeenth embodiment]
Next, a seventeenth embodiment of the present invention will be described. FIG. 33 is a diagram showing the configuration of the seventeenth embodiment of the present invention. Referring to FIG. 33, the seventeenth embodiment of the present invention includes an input device P201, an output device P206, and the speech recognition system P205 of the fourteenth embodiment of the present invention described above.

  The speech recognition program P207 is read into the speech recognition system P205 and controls the operation of the speech recognition system P205. With the speech recognition program P207, the speech recognition system P205 executes the same processing as the speech recognition system according to the fourteenth embodiment of the present invention described above.

[Eighteenth embodiment]
Next, an eighteenth embodiment of the present invention will be described. FIG. 34 is a diagram showing the configuration of the eighteenth embodiment of the present invention. Referring to FIG. 34, an eighteenth embodiment of the present invention includes an input device P201, an output device P209, and the signal transmission system P208 of the fifteenth embodiment of the present invention described above.

  The signal transmission program P210 is read into the signal transmission system P208 and controls the operation of the signal transmission system P208. By the signal transmission program P210, the signal transmission system P208 executes the same processing as the signal transmission system of the fifteenth embodiment of the present invention described above.

  Although the present invention has been described with reference to the above embodiments, the present invention is not limited to the configurations of the above embodiments, and can be made by those skilled in the art within the scope of the principle of the present invention. Of course, various modifications and corrections will be included. The signal source is not limited to sound, but can be applied to signal separation of electromagnetic waves, light (infrared rays, etc.), or image signals (image analysis, etc.).

  According to the present invention, it can be applied to various uses such as separating and extracting a signal before mixing from a plurality of time-series signals in which a plurality of signals are mixed.

It is a figure which shows the structure of the signal separation system of the 1st Embodiment of this invention. It is a figure which shows the structure of the processing frequency control part contained in the signal separation system of the 1st Embodiment of this invention. It is a figure which shows the structure of the separation factor learning part contained in the signal separation system of the 1st Embodiment of this invention. It is a figure which shows the structure of the other separation factor learning part contained in the signal separation system of the 1st Embodiment of this invention. It is a figure which shows the structure of the 2nd Embodiment of this invention. It is a figure which shows the structure of the processing frequency control part contained in the signal separation system of the 2nd Embodiment of this invention. It is a figure which shows the structure of the separation factor learning part contained in the signal separation system of the 2nd Embodiment of this invention. It is a figure which shows the structure of the separation factor initial value setting part contained in the signal separation system of the 2nd Embodiment of this invention. It is a figure which shows the example of spatial arrangement | positioning of the sensor which picks up the signal mixed with the signal source in this invention. It is a figure which shows the structure of the 3rd Embodiment of this invention. It is a figure which shows the structure of the separation factor initial value setting part contained in the signal separation system of the 3rd Embodiment of this invention. It is a figure which shows the structure of the 4th Embodiment of this invention. It is a figure which shows the structure of the separation factor initial value setting part contained in the signal separation system of the 4th Embodiment of this invention. It is a figure which shows the structure of the processing frequency control part contained in the signal separation system of the 4th Embodiment of this invention. It is a figure which shows the structure of the 5th Embodiment of this invention. It is a figure which shows the structure of the separation factor initial value setting part contained in the signal separation system of the 5th Embodiment of this invention. It is a figure which shows the structure of the 6th Embodiment of this invention. It is a figure which shows the structure of the processing frequency control part contained in the signal separation system of the 6th Embodiment of this invention. It is a figure which shows the structure of the 7th Embodiment of this invention. It is a figure which shows the structure of the 8th Embodiment of this invention. It is a figure which shows the structure of the processing frequency control part contained in the signal separation system of the 8th Embodiment of this invention. It is a figure which shows the structure of the 9th Embodiment of this invention. It is a figure which shows the structure of the 10th Embodiment of this invention. It is a figure which shows the structure of the processing frequency control part contained in the signal separation system of the 10th Embodiment of this invention. It is a figure which shows the structure of the 11th Embodiment of this invention. It is a figure which shows the structure of the 12th Embodiment of this invention. It is a figure which shows the structure of the processing frequency control part contained in the signal separation system of the 12th Embodiment of this invention. It is a figure which shows the structure of the 13th Embodiment of this invention. It is a figure which shows the structure of the 14th Embodiment of this invention. It is a figure which shows the structure of 15th Embodiment of this invention. It is a flowchart which shows the process sequence of the 2nd Embodiment of this invention. It is a figure which shows the structure of the 16th Embodiment of this invention. It is a figure which shows the structure of the 17th Embodiment of this invention. It is a figure which shows the structure of the 18th Embodiment of this invention. It is a figure which shows the structure of the conventional signal separation system.

Explanation of symbols

1, 2 Fourier transform unit 3 Input signal multiplexing unit 4, 41, 42, 43, 44, 45, 46 Processing frequency control unit 5, 9 Separation output signal calculation unit 6, 10, 11, 61 Separation coefficient learning unit 7, 71 Separation coefficient storage unit 8, 82, 83, 84 Separation coefficient initial value setting unit 12 Signal separation system 13 Output signal division unit 15, 16 Input terminal 17, 18 Output terminal 40, 420, 430, 440, 450, 460 Switch control Unit 60, 63, 100, 610 Switch 62, 101, 611 Separation coefficient update unit 64, 612 Separation coefficient initial value storage unit 80, 820, 830, 840 Separation coefficient phase calculation unit 81, 821, 831, 841 Separation coefficient initial value Calculation unit 91 Sensor 92 Separated signal storage unit 93 Signal processing unit 94 DFFT processing unit 95 Mixing matrix estimation processing unit 96 Signal separation processing Unit 97 I-DFFT processing unit 98 separated signal storage unit 140, 141 speech recognition unit 142, 143 output terminal 144, 145 inverse Fourier transform unit 146, 147 output signal transmission unit 148, 149 output terminal 600, 601, 630, 631, 1000, 1001, 6100, 6101 Terminals 832, 842 Relearning frequency determination unit P201 Input device P202 Signal separation system P203 Output device P204 Signal separation program P205 Speech recognition system P206 Output device P207 Speech recognition program P208 Signal transmission system P209 Output device P210 Signal transmission program

Claims (29)

In a signal separation method of separating signals by processing time-series input signals of a plurality of channels in the frequency domain,
Estimating signal and sensor placement information from the processed at least one frequency separation factor;
Calculating an initial value of a separation factor of a frequency to be processed based on the arrangement information;
A signal separation method comprising: Obtaining an input signal of each frequency of each channel by converting the input time-series signals of a plurality of channels into the frequency domain, and
Controlling the selection of frequencies to process;
Calculating an initial value of the separation factor of the frequency to be processed based on the arrangement information of the signal source and the sensor;
Obtaining a separated output signal based on the input signal and separation factor of each channel at the frequency to be processed;
Learning a separation factor using the separation output signal at the frequency to be processed;
A signal separation method comprising: Controlling the processing frequency comprises:
Information on the processing order set in advance,
Statistical information obtained from input signals of a plurality of frequencies of at least one channel among the input signals of each frequency of each channel;
Statistical information obtained from the separated output signal;
Information on the separation factor of the learned frequency,
The signal separation method according to claim 2, further comprising a step of controlling a frequency to be processed based on at least one of the information. A plurality of Fourier transform units that respectively input a plurality of channel-specific input time-series signals and a step of outputting channel-specific frequency-specific input time-series signals;
The step of multiplexing the input time series signal by frequency for each channel of the same frequency in different channels by the multiplexing unit, and outputting the input time series signal by frequency,
A step of performing a control for selecting a frequency to be processed by a processing frequency control unit that inputs an input time-series signal by frequency from the input signal multiplexing unit;
The separation coefficient learning unit corresponding to the frequency to be processed first derives the output time-series signal for each frequency based on the frequency-specific separation coefficient determined in advance as the initial value, and separates for each frequency based on the output time-series signal for each frequency. Learning a coefficient and storing the learned frequency-specific separation coefficient in the separation coefficient storage unit;
The separation coefficient storage unit stores a frequency-specific separation coefficient of a frequency for which learning has been completed, and supplies a separation coefficient obtained by multiplexing the frequency-specific separation coefficient for a frequency for which learning has been completed to the separation coefficient initial value setting unit; ,
The separation coefficient initial value setting unit derives a frequency-specific separation coefficient corresponding to a frequency to be processed next based on the separation coefficient supplied from the separation coefficient storage unit and corrected for phase shift, and performs the frequency separation. Outputting a coefficient as an initial value of learning to a separation coefficient learning unit corresponding to the frequency to be processed next;
The separation factor learning unit learning the frequency separation factor when the frequency separation factor is supplied from the separation factor initial value setting unit;
A signal separation method comprising: In a signal separation method for separating signals by processing input time series signals from a plurality of signal sources in the frequency domain,
Correcting the phase shift of the learned separation factor corresponding to the processed frequency;
And a step of setting the phase-corrected separation coefficient as an initial value of a separation coefficient of another frequency to be processed. In a signal separation method for separating signals by processing input time series signals from a plurality of signal sources in the frequency domain,
Obtaining a phase of the separation factor from the learned separation factor corresponding to the processed frequency, and calculating an initial value of the separation factor of the other frequency to be processed;
Learning a separation factor of the other frequency based on an initial value of the separation factor.   7. A speech recognition method comprising the step of recognizing the separated output signals of the respective channels obtained by the signal separation method according to claim 1. Obtaining the output signal of each channel by converting the separated output signal of each channel obtained by the signal separation method according to any one of claims 1 to 6 into a time domain, respectively.
Transmitting each of the output signals;
A signal transmission method comprising: In a signal separation system that separates signals by processing time-series input signals of multiple channels in the frequency domain,
It includes a separation coefficient initial value setting unit that estimates the arrangement information of the signal source and the sensor from the processed separation coefficient of at least one frequency and calculates an initial value of the separation coefficient of the frequency to be processed. Signal separation system. A Fourier transform unit that obtains an input signal of each frequency of each channel by converting input time-series signals of a plurality of channels into the frequency domain, and
A processing frequency control unit for controlling selection of a frequency to be processed;
A separation factor initial value setting unit that calculates a separation factor initial value of a frequency to be processed, selected by the processing frequency control unit, based on arrangement information of a signal source and a sensor;
A separated output signal calculation unit for obtaining a separated output signal based on an input signal and a separation factor of each channel at the frequency to be processed;
A separation coefficient learning unit that learns a separation coefficient using a separation output signal at the frequency to be processed;
A signal separation system comprising: A Fourier transform unit that obtains an input signal of each frequency of each channel by converting input time-series signals of a plurality of channels into the frequency domain, and
A processing frequency control unit for controlling selection of a frequency to be processed;
A separation factor initial value setting unit that calculates a value obtained by correcting the phase of the separation factor based on the already learned frequency separation factor selected by the processing frequency control unit as an initial value of the separation factor of the frequency to be processed; ,
A separated output signal calculation unit for obtaining a separated output signal based on an input signal and a separation factor of each channel at the frequency to be processed;
A separation coefficient learning unit that learns a separation coefficient using a separation output signal at the frequency to be processed;
A signal separation system comprising: A processing frequency control unit for controlling the frequency to be processed;
Information on the processing order set in advance,
Statistical information obtained from input signals of a plurality of frequencies of at least one channel among the input signals of each frequency of each channel;
Statistical information obtained from the separated output signal;
Information on the separation factor of the learned frequency,
The signal separation system according to claim 11, further comprising: a switch control unit that controls a frequency to be processed based on at least one of the information.   A speech recognition system comprising: a speech recognition unit that recognizes the separated output signals of each channel obtained by the signal separation system according to claim 9.   An inverse Fourier transform unit for obtaining an output signal of each channel by converting the separated output signal of each channel obtained by the signal separation system according to any one of claims 9 to 12 into a time domain, and the output An output signal transmission unit for transmitting each signal. In a computer constituting a device that separates input time-series signals of a plurality of channels in the frequency domain,
A program that estimates the arrangement information of a signal source and a sensor from at least one processed frequency separation coefficient, and executes a process of calculating an initial value of a frequency separation coefficient to be processed based on the arrangement information. A process for obtaining an input signal of each frequency of each channel by converting the input time-series signals of a plurality of channels into the frequency domain,
A process for controlling the selection of the frequency to be processed;
Processing for calculating the initial value of the separation coefficient of the frequency to be processed based on the signal source and sensor arrangement information;
A process for obtaining a separated output signal based on an input signal and a separation coefficient of each channel at the frequency to be processed;
A process of learning a separation coefficient using the separation output signal at the frequency to be processed;
A program that causes a computer to execute. The program according to claim 16, wherein
As a process for controlling the frequency to be processed,
Information on the processing order set in advance,
Statistical information obtained from input signals of a plurality of frequencies of at least one channel among the input signals of each frequency of each channel;
Statistical information obtained from the separated output signal;
A program that causes the computer to execute a process of controlling a frequency to be processed based on at least one piece of information on a frequency separation factor that has already been learned.   A program that causes a computer that performs signal separation using the program according to any one of claims 15 to 17 to perform processing for recognizing the separated output signal of each channel acquired by signal separation. A computer for performing signal separation by the program according to any one of claims 15 to 17,
Processing for obtaining an output signal of each channel by converting the separated output signal of each channel obtained by signal separation into a time domain, respectively;
A program for executing a process of transmitting each of the output signals. Means for converting time-series signals from a plurality of signal sources into signals in the frequency domain, and multiplexing each frequency component;
Means for controlling selection of a frequency to be processed from a plurality of frequency components;
Means for calculating a separation output signal based on the signal of the selected frequency component and the separation coefficient;
Means for setting the value obtained by correcting the phase of the separation coefficient based on the separation coefficient of the already selected and learned frequency component, as an initial value of learning of the separation coefficient of other unprocessed frequency components;
A signal separation device comprising: Means for converting time-series signals from a plurality of signal sources into signals in the frequency domain, and multiplexing each frequency component;
Means for controlling selection of a frequency to be processed from a plurality of frequency components;
Means for calculating a separation output signal based on the signal of the selected frequency component and the separation coefficient;
Based on the already selected and learned frequency component separation coefficients, the phase of the separation coefficients of other unprocessed frequency components is derived, and the initial value of learning of the separation coefficients of other frequency components is derived based on the phase of the separation coefficients. Means for deriving
A signal separation device comprising: A plurality of Fourier transform units that respectively input a plurality of input time series signals by channel and perform a Fourier transform and output an input time series signal by frequency by channel;
An input signal multiplexing unit that multiplexes input time-series signals by frequency for each channel of the same frequency in different channels and outputs input time-series signals by frequency,
A processing frequency control unit that performs control to input a frequency-specific input time-series signal from the input signal multiplexing unit and to select a frequency to be processed;
A plurality of separated output signal calculation units provided corresponding to a plurality of frequency-specific input time-series signals output from the processing frequency control unit;
A plurality of separation coefficient learning units provided corresponding to each of the plurality of separation output signal calculation units;
And means for correcting the phase of the separation coefficient to obtain an initial value of learning when the separation coefficient of the frequency learned by the separation coefficient learning unit is used at another frequency. apparatus. A plurality of Fourier transform units that respectively input a plurality of input time series signals by channel and perform a Fourier transform and output an input time series signal by frequency by channel;
An input signal multiplexing unit that multiplexes input time-series signals by frequency for each channel of the same frequency in different channels and outputs input time-series signals by frequency,
A processing frequency control unit that performs control to input a frequency-specific input time-series signal from the input signal multiplexing unit and to select a frequency to be processed;
A plurality of separated output signal calculation units provided corresponding to a plurality of frequency-specific input time-series signals output from the processing frequency control unit;
A plurality of separation coefficient learning units provided corresponding to each of the plurality of separation output signal calculation units;
At least one separation factor storage unit for storing separation factors;
At least one separation factor initialization setting unit for initializing the separation factor;
With
The one separation coefficient learning unit selected by the processing frequency control unit and corresponding to the first frequency to be processed has a frequency-specific separation coefficient determined as an initial value in advance and corresponds to the one separation coefficient learning unit. Supplying the separated output signal calculation unit, the one separated output signal calculation unit derives an output time-series signal by frequency,
The one separation factor learning unit learns a frequency separation factor based on the frequency-specific output time series signal calculated by the one separation output signal calculation unit, and the learned frequency separation factor is used as the separation factor. To the storage unit,
The separation coefficient storage unit stores the frequency-specific separation coefficient of the frequency for which learning has been completed, and supplies the separation coefficient obtained by multiplexing the frequency-specific separation coefficient for the frequency for which learning has been completed to the separation coefficient initial value setting unit. ,
The separation factor initial value setting unit derives a separation factor phase based on the separation factor supplied from the separation factor storage unit, and the frequency of each frequency based on the separation factor phase and the supplied separation factor. Deriving another separation factor,
The separation factor initial value setting unit is selected by the processing frequency control unit and corresponds to the frequency to be processed next, and the separation factor for each frequency in which the phase of the separation factor is corrected is used as the learning initial value. Output to other separation coefficient learning unit corresponding to the frequency to be processed,
The other separation factor learning unit learns the separation factor by frequency based on the separation factor by frequency supplied from the separation factor initial value setting unit.
A signal separation device.   The signal separation according to claim 23, wherein the one separation factor learning unit re-learns the separation factor by frequency when the separation factor by frequency is supplied from the separation factor initial value setting unit. apparatus. Each separation coefficient learning unit outputs a frequency-specific separation coefficient initial value stored in a separation coefficient initial value storage unit in each separation coefficient learning unit;
After that, the frequency-dependent output time-series signal and the frequency-specific separation coefficient used to obtain the frequency-specific output time-series signal are used to provide a separation coefficient update unit that updates the frequency-specific separation coefficient. The separation factor is output via the switch,
On the other hand, when a frequency-specific separation factor is supplied to the separation factor learning unit, the frequency separation factor supplied to the switch is output, and the separation factor update unit updates the separation factor. 24. The signal separation device according to claim 23. The separation factor initial value setting unit, based on the supplied separation factor, a separation factor phase calculation unit that calculates a separation factor phase;
A separation factor initial value calculation unit that calculates a separation factor for each frequency based on the separation factor phase and the supplied separation factor;
The separation factor phase calculation unit divides the supplied separation factor into frequency separation factors, further divides the divided frequency separation factor into channel frequency separation factors,
The signal separation device according to claim 23, wherein each channel-specific frequency separation coefficient is multiplexed with the frequency-specific separation coefficient and output. The separation coefficient initial value setting unit further includes a relearning frequency determination unit,
The separation factor phase calculation unit derives the separation factor phase, supplies the separation factor phase to the separation factor initial value calculation unit, and calculates the separation factor phase by frequency calculated at a plurality of already learned frequencies. Supplying the re-learning frequency determination unit;
The relearning frequency determination unit determines a relearning frequency based on a frequency-specific separation coefficient phase, and outputs a relearning frequency flag to the processing frequency control unit,
The processing frequency control unit performs control to output a frequency-specific input time-series signal when a relearning frequency flag is supplied from the relearning frequency determination unit of the separation coefficient initial value setting unit. 27. The signal separation device according to claim 26, further comprising a switch control unit.   The processing frequency control unit determines a frequency to be processed first based on information obtained by statistical processing of the input time-series signal by frequency supplied, and performs control to output the input time-series signal by frequency of the frequency 24. The signal separation device according to claim 23, further comprising a switch control unit for performing the operation. In the processing frequency control unit, the information obtained by the statistical processing of the input time-series signal by frequency,
Information on the power of input time-series signals by frequency,
Information on correlation between channels of input time-series signals by frequency, and
Eigenvalue and eigenvector information obtained when principal component analysis is performed on input time-series signals by frequency,
29. The signal separation device according to claim 28, wherein the signal separation device is at least one of the following.

Images