JP5387442B2 - Signal processing device - Google Patents

Description

  The present invention relates to a technique for emphasizing (separating or extracting) sound from a specific sound source among a plurality of mixed sound generated by different sound sources.

  Conventionally, there is a technology (sound source separation technology) that separates sound from each sound source by performing sound source separation on multiple observation signals obtained by collecting mixed sounds of multiple sounds (speech and noise) with multiple sound collection devices. Proposed. The separation matrix (inverse mixing matrix) applied to sound source separation is calculated for each frequency by a learning process (iterative update) using, for example, frequency-domain independent component analysis (FDICA). The

  In Non-Patent Document 1, a separation matrix is generated by learning processing for each frequency selected from a plurality of frequencies for each predetermined number, and a separation matrix for each non-selected frequency is supplemented using the separation matrix after learning processing. Technology is disclosed. A blind spot control type beam forming (NBF (Null Beam Former)) is used to generate a separation matrix of non-selected frequencies. In other words, a non-selected frequency separation matrix is set so that a dead angle of sound collection is formed in the sound source direction estimated from the separation matrix after learning processing. According to the technique of Non-Patent Document 1, it is possible to reduce the amount of calculation compared to a case where learning processing by independent component analysis is executed for all frequencies from the beginning.

Osako et al., “Fast method of blind source separation using frequency band interpolation by blind spot control type beamformer”, Proceedings of the Acoustical Society of Japan, Acoustical Society of Japan, March 2007, p.549-p.550

  However, in the configuration in which the separation matrix generated by the blind spot control type beam forming as in Non-Patent Document 1 is used as the separation matrix of the non-selected frequency, the accuracy of sound source separation cannot be sufficiently ensured for the non-selected frequency. There is sex. In view of the above circumstances, an object of the present invention is to achieve both reduction in the amount of calculation required for generating a separation matrix and high accuracy in sound source separation.

  In order to solve the above problems, the signal processing apparatus according to the present invention has a plurality of frequencies for a plurality of observation signals obtained by collecting a plurality of sound mixed sounds generated by different sound sources by a plurality of sound collecting devices. A signal separation unit that generates a plurality of separation signals for each sound source by applying each separation matrix, a frequency selection unit that selects a plurality of frequencies into a first frequency and a second frequency, and a first in the plurality of observation signals First learning means for generating a separation matrix of the first frequency in a primary learning process using learning data corresponding to a component of one frequency, and estimating the direction of each sound source from the separation matrix generated by the first learning means Corresponding to direction estimation means, initial matrix setting means for generating an initial separation matrix so that a dead angle or beam of sound collection is formed in the direction estimated by the direction estimation means, and components of the second frequency in a plurality of observation signals Learning The second frequency separation process is performed by using the initial separation matrix generated by the initial matrix setting means as the initial value with a smaller number of iterations than the first learning process, thereby generating the second frequency separation matrix. Second learning means.

  In the above configuration, a separation matrix is generated in the primary learning process for each frequency selected as the first frequency, and a separation matrix generated in the primary learning process for each frequency selected as the second frequency. The separation matrix is generated by executing the secondary learning process with a smaller number of iterations than the primary learning process with the initial separation matrix corresponding to the initial value as the initial value. Therefore, there is an advantage that the calculation amount is reduced as compared with the case where the primary learning process is executed for all frequencies. Also, a configuration (secondary learning process) in which a separation matrix set so that a dead angle or a beam of sound collection is formed in the direction of a sound source estimated from the separation matrix generated by the primary learning process is applied to the second frequency. Compared with a configuration that does not execute the above, there is an advantage that a separation matrix capable of high-accuracy sound source separation can be generated.

  By the way, in an environment where the sound collection conditions are poor, there is a case where a high-precision separation matrix can be generated without performing the secondary learning process for the second frequency. Considering the above tendency, the signal processing apparatus according to a preferred aspect of the present invention includes a condition determination unit that determines whether the sound collection condition is good or not for each frequency, and the second learning unit has a good sound collection condition. For the frequencies determined by the condition determination means, a separation matrix is generated by the second learning process using the initial separation matrix as an initial value, and for the frequencies determined by the condition determination means that the sound collection condition is bad, the initial separation matrix is Adopt as a separation matrix. In the above aspect, since the secondary learning process is not executed for frequencies with poor sound collection conditions, it is compared with the configuration in which the secondary learning process is executed for all frequencies selected as the second frequency regardless of the sound collection conditions. Then, it becomes possible to generate a highly accurate separation matrix. In addition, the specific example of the above aspect is later mentioned as 2nd Embodiment, for example.

  In addition, when only the sound of one sound source is included in the frequency component selected as the second frequency in the observation signal, the separation matrix is set so that the frequency component does not change excessively before and after sound source separation. A configuration to be set is preferable. Therefore, condition determination means for determining whether the sound collection condition is good (single / plurality of sound sources) for each frequency, and the frequency determined by the condition determination means that the sound collection condition is bad among the frequencies selected as the second frequency. For example, a configuration including matrix setting means (for example, the matrix setting unit 76 in FIG. 14) that sets a separation matrix so that incoming sounds from sound source directions estimated from a plurality of observation signals are emphasized may be employed. In the above aspect, from the viewpoint of reducing the amount of computation required for generating the separation matrix, initial matrix setting is performed for the frequencies determined by the condition determination means that the sound collection condition is bad among the frequencies selected as the second frequency. A configuration in which the generation of the initial separation matrix by the means and the secondary learning process by the second learning means is stopped is particularly suitable. In addition, the specific example of the above aspect is later mentioned as 3rd Embodiment, for example.

  The signal processing apparatus according to a preferred aspect of the present invention includes a significant index calculation means for calculating a significant index value indicating the significance of the learning process to which the learning data of each frequency is applied for each frequency from a plurality of observation signals, The frequency sorting means sorts a plurality of frequencies into a first frequency and a second frequency according to a significant index value of each frequency. In the above aspect, since the plurality of frequencies are sorted into the first frequency and the second frequency according to the significant index value indicating the significance of the learning process, the selection of the first frequency and the second frequency is performed in the learning process. Compared to a configuration that is executed regardless of significance (for example, a configuration in which a frequency selected for each predetermined number from a plurality of frequencies is selected as a first frequency and a remaining frequency is selected as a second frequency) It is possible to generate an accurate separation matrix. Specifically, the condition determination means determines that the sound collection condition is bad when the differences in the intensity of a plurality of sounds generated by different sound sources are large, and the sound collection condition is determined when the difference in the intensity of each sound is small. Judge as good.

  In the aspect including the significant index calculation means, a configuration in which the condition determination means determines the quality of the sound pickup condition for each frequency according to the significant index value of each frequency calculated by the significant index calculation means is particularly suitable. . In the above aspect, there is an advantage that the amount of calculation required for generating the separation matrix can be reduced as compared with the configuration in which the index of the sound collection condition is calculated separately from the significant index value. Specifically, the determinant of the covariance matrix of the observation vector whose element is the intensity at each frequency in each of the plurality of observation signals is used as an index indicating the significance of the learning process, and different sound sources are used. It changes according to the difference in the intensity of the plurality of generated sounds (good or bad sound collection conditions). Therefore, a configuration in which the determinant of the covariance matrix of the observation vector is used for selecting the frequency and determining the quality of the sound collection condition is preferable.

  Note that the learning process by independent component analysis is equivalent to the process of specifying the independent bases by the number of sound sources, so the total number of observation vector bases whose elements are the intensities at each frequency in each of the plurality of observation signals is It is preferably used as an index of the significance of learning using learning data. Therefore, the significant index calculation means in a preferred aspect of the present invention calculates the index value of the total number of bases in the distribution of observation vectors whose elements are the intensities at each frequency in each of the plurality of observation signals, and the frequency selection means The frequency having the large total number of bases indicated by the index value is selected as the first frequency.

  Examples of the index value of the total number of bases include determinants of the covariance matrix of observation vectors and condition numbers. Therefore, the significance index calculation means in a preferred aspect of the present invention is a covariance matrix calculation means for calculating an observation vector covariance matrix having each frequency component in a plurality of observation signals as an element. And index calculation means (for example, determinant calculation unit 424 in FIG. 6) for calculating a significant index value from the covariance matrix of each frequency. The index calculation means calculates a significant index value according to, for example, the determinant of the covariance matrix and the condition number. Considering the tendency that the distribution region (basis) of the observation vector is clearly identified for each sound source as the trace (power) of the observation vector covariance matrix is larger, A configuration in which the significant index calculation means calculates the significant index value is also suitable.

  In addition, the definition of a significant index value and the calculation method are arbitrary. For example, considering the tendency that the lower the kurtosis in the intensity distribution of the observed signal, the more the observed signal contains sound from the sound source, the significant index value corresponding to the kurtosis in the intensity distribution of the observed signal is significant. A configuration in which the index calculating unit calculates and the frequency selecting unit selects a frequency having a low kurtosis as the first frequency is preferable. In addition, considering the tendency that the higher the independence between multiple observation signals (the lower the correlation), the higher the significance of learning using learning data, it depends on the independence between multiple observation signals. A configuration in which the significant index value is calculated by the significant index calculation means, and the frequency selecting means selects the first independent frequency indicated by the significant index value is preferable. Examples of the index value of independence among a plurality of observation signals include cross-correlation and mutual information.

  The signal processing apparatus according to each aspect described above is realized by hardware (electronic circuit) such as a DSP (Digital Signal Processor) dedicated to voice processing, and general-purpose arithmetic processing such as a CPU (Central Processing Unit). This is also realized by cooperation between the apparatus and the program. The program according to the present invention applies a separation matrix of each of a plurality of frequencies to a plurality of observation signals obtained by collecting a plurality of sound mixed sounds generated by different sound sources by a plurality of sound collecting devices. A signal separation process for generating a plurality of separated signals for each, a frequency sorting process for sorting a plurality of frequencies into a first frequency and a second frequency, and learning data corresponding to a component of the first frequency in the plurality of observation signals. A first process for generating a separation matrix of the first frequency in the applied primary learning process, a direction estimation process for estimating the direction of each sound source from the separation matrix generated in the first process, and a direction estimated by the direction estimation process Initial matrix setting processing for generating an initial separation matrix so that a dead angle or beam of sound collection is formed in the second stage, and secondary learning processing using learning data corresponding to components of the second frequency in a plurality of observation signals, The initial separation matrix column setting means is generated as an initial value by executing a small number of iterations than the primary learning process is executed and a second process of generating the second frequency of the separating matrix in the computer. According to the above program, the same operation and effect as the signal processing apparatus according to the present invention are exhibited. The program of the present invention is provided to a user in a form stored in a computer-readable recording medium and installed in the computer, or provided from a server device in a form of distribution via a communication network and installed in the computer. Is done.

1 is a block diagram of a signal processing device according to a first embodiment. It is explanatory drawing of an observation vector and learning data. It is a block diagram of a signal separation unit. It is a block diagram of a separation matrix production | generation part. It is explanatory drawing of operation | movement of a separation matrix production | generation part. It is a block diagram of a significant index calculation part. It is a conceptual diagram which shows the relationship between the determinant of the covariance matrix of an observation vector, and a basis number. It is a graph which shows the relationship between the number of 1st frequencies, and the repetition frequency of a learning process. It is a graph which shows the relationship between the number of 1st frequencies, and a noise suppression rate. It is a graph which shows the relationship between the number of 1st frequencies, and cepstrum distortion. It is a graph which shows the relationship between the number of 1st frequencies, and a noise suppression rate. It is a graph which shows the relationship between the number of 1st frequencies, and cepstrum distortion. It is a block diagram of the separation matrix production | generation part in 2nd Embodiment. It is a block diagram of the separation matrix production | generation part in 3rd Embodiment. It is a conceptual diagram which shows the relationship between the trace of a covariance matrix, and the distribution range of an observation vector. It is a graph which shows the relationship between kurtosis before correction | amendment, and a weight value.

<A: First Embodiment>
FIG. 1 is a block diagram of a signal processing apparatus 100 according to the first embodiment. The sound collecting device M1 and the sound collecting device M2 arranged in the plane PL with a space therebetween are connected to the signal processing apparatus 100. The sound source S1 and the sound source S2 exist at different positions around the sound collection device M1 and the sound collection device M2. The sound source S1 is located in the direction of the angle θ1 with respect to the normal line Ln of the plane PL, and the sound source S2 is located in the direction of the angle θ2 (θ2 ≠ θ1) with respect to the normal line Ln. Angle θ1 and angle θ2 are unknown. Note that the number of sound collecting devices M (M1, M2) and the number of sound sources S (S1, S2) can be arbitrarily changed.

  The mixed sound of the sound SV1 generated by the sound source S1 and the sound SV2 generated by the sound source S2 reaches the sound collection device M1 and the sound collection device M2. The sound collecting device M1 generates an observation signal V1 (t), and the sound collecting device M2 generates an observation signal V2 (t). Each of the observation signal V1 (t) and the observation signal V2 (t) is an acoustic signal representing a time waveform of a mixed sound of the sound SV1 and the sound SV2.

  The signal processing apparatus 100 generates a separation signal Y1 (t) and a separation signal Y2 (t) by sound source separation (filter processing) for the observation signal V1 (t) and the observation signal V2 (t). The separated signal Y1 (t) is an acoustic signal that emphasizes the sound SV1 from the sound source S1 (suppresses the sound SV2 from the sound source S2), and the separated signal Y2 (t) emphasizes the sound SV2 (suppresses the sound SV1). Sound signal. That is, the sound SV1 and the sound SV2 are separated (sound source separation).

  The separated signal Y1 (t) and the separated signal Y2 (t) are reproduced as sound by being supplied to a sound emitting device (not shown) such as a speaker or headphones. A configuration in which only one of the separated signal Y1 (t) and the separated signal Y2 (t) is reproduced (for example, a configuration in which the separated signal Y2 (t) is discarded as noise) is also employed. An A / D converter that converts the observation signal V1 (t) and the observation signal V2 (t) from analog to digital, and a D that converts the separation signal Y1 (t) and the separation signal Y2 (t) from digital to analog. The illustration of the / A converter is omitted for convenience.

  As shown in FIG. 1, the signal processing device 100 is realized by a computer system including an arithmetic processing device 12 and a storage device 14. The storage device 14 stores a program and various data for generating the separation signal Y1 (t) and the separation signal Y2 (t) from the observation signal V1 (t) and the observation signal V2 (t). A known recording medium such as a semiconductor recording medium or a magnetic recording medium or a combination of a plurality of types of recording media is arbitrarily employed as the storage device 14.

  The arithmetic processing unit 12 functions as a plurality of elements (frequency analysis unit 22, signal separation unit 24, signal synthesis unit 26, and separation matrix generation unit 28) by executing a program stored in the storage device 14. A configuration in which an electronic circuit (DSP) dedicated to sound source separation realizes each element in FIG. 1 or a configuration in which each element in FIG. 1 is distributed over a plurality of integrated circuits may be employed.

  The frequency analyzer 22 generates a frequency spectrum (complex spectrum) Q1 of the observation signal V1 (t) and a frequency spectrum (complex spectrum) Q2 of the observation signal V2 (t) for each of a plurality of frames on the time axis. . As shown in FIG. 2, the frequency spectrum Q1 has component values x1 (m, f1) to x1 (m, fK) at each of K frequencies (actually frequency bands) f1 to fK set on the frequency axis. ) Series. Similarly, the frequency spectrum Q2 is a series of component values x2 (m, f1) to x2 (m, fK) at each of the K frequencies f1 to fK. The symbol m means the frame number (each time point set discretely on the time axis). A known technique (for example, short-time Fourier transform) is arbitrarily employed for calculating the frequency spectrum Q1 and the frequency spectrum Q2.

  The frequency spectrum Q1 and the frequency spectrum Q2 generated by the frequency analysis unit 22 are supplied to the signal separation unit 24 of FIG. The signal separation unit 24 has a component (component value x1 (m, fk)) of the frequency fk (k = 1 to K) in the observation signal V1 (t) and a component (component value x2) of the frequency fk in the observation signal V2 (t). (m, fk)) and the frequency spectrum R1 of the separated signal Y1 (t) and the frequency spectrum R2 of the separated signal Y2 (t) by separately performing the sound source separation for each of the K frequencies f1 to fK. Generate. The frequency spectrum R1 is a sequence of component values y1 (m, f1) to y1 (m, fK), and the frequency spectrum R2 is a sequence of component values y2 (m, f1) to y2 (m, fK).

  FIG. 3 is a block diagram of the signal separation unit 24. As shown in FIG. 3, the signal separation unit 24 includes K processing units P1 to PK corresponding to different frequencies fk (f1 to fK). The processing unit Pk having the frequency fk includes a filter 32 that generates a component value y1 (m, fk) of the separated signal Y1 (t) from the component value x1 (m, fk) and the component value x2 (m, fk), and a component value and a filter 34 for generating a component value y2 (m, fk) of the separated signal Y2 (t) from x1 (m, fk) and the component value x2 (m, fk).

The filter 32 and the filter 34 execute delay-added type (DS (delay-sum) type) beam forming. That is, the filter 32 of the processing unit Pk includes a delay element 321 that adds a delay corresponding to the coefficient w11 (fk) to the component value x1 (m, fk), as defined by the following formula (1A), Delay element 323 that adds a delay corresponding to w12 (fk) to component value x2 (m, fk), and addition that generates component value y1 (m, fk) by adding the outputs of delay element 321 and delay element 323 Part 325. Similarly, the filter 34 includes a delay element 341 that adds a delay corresponding to the coefficient w21 (fk) to the component value x1 (m, fk) and a coefficient w22 (fk) as defined by the following formula (1B). ) And a delay element 343 that adds a delay corresponding to the component value x2 (m, fk), and an adder 345 that generates a component value y2 (m, fk) by adding the outputs of the delay element 341 and the delay element 343; including. Note that blind spot control type (null) beam forming can also be applied to the processing unit Pk.
y1 (m, fk) = w11 (fk) x1 (m, fk) + w12 (fk) x2 (m, fk) (1A)
y2 (m, fk) = w21 (fk) x1 (m, fk) + w22 (fk) x2 (m, fk) (1B)

  1 generates a time domain acoustic signal by inverse Fourier transform of the frequency spectrum R1 (y1 (m, f1) to y1 (m, fK)) generated by the signal separation unit 24 for each frame. At the same time, the separated signal Y1 (t) is generated by connecting the preceding and following frames to each other. Similarly, the signal synthesizer 26 generates a time-domain separation signal Y2 (t) from the frequency spectrum R2 (y2 (m, f1) to y2 (m, fK)) generated by the signal separation unit 24.

As shown in FIG. 1, the frequency spectrum Q1 and the frequency spectrum Q2 generated by the frequency analysis unit 22 are supplied to the signal separation unit 24 and stored as observation vectors X (m, f1) to X (m, fK). 14. As shown in FIG. 2, the observation vector X (m, fk) is a vector (X (m, fk) = (x1) having the component value x1 (m, fk) and the component value x2 (m, fk) as elements. (m, fk), x2 (m, fk)) ^T ). The symbol T means transposition of the matrix. The observation vectors X (m, f1) to X (m, fK) stored in the storage device 14 are learned for each unit interval TU composed of a predetermined number (for example, 50) of frames as shown in FIG. Data is divided into D (f1) to D (fK). That is, the learning data D (fk) is a time series of the observation vector X (m, fk) of the frequency fk calculated for each frame in the unit interval TU.

  The separation matrix generation unit 28 in FIGS. 1 and 3 generates separation matrices W (f1) to W (fK) that the signal separation unit 24 applies to sound source separation. As shown in FIG. 3, the separation matrix W (fk) of the frequency fk is obtained by applying a coefficient w11 (fk) and a coefficient w12 (fk) applied to the filter 32 of the processing unit Pk and a coefficient w21 (fk) applied to the filter 34. ) And coefficient w22 (fk) as a matrix. The separation matrix W (fk) is sequentially generated for each unit interval TU by a learning process (repetitive update) using the learning data D (fk) stored in the storage device 14. FIG. 4 is a block diagram of the separation matrix generation unit 28, and FIG. 5 is an explanatory diagram of the operation of the separation matrix generation unit 28. As illustrated in FIG. 4, the separation matrix generation unit 28 includes a significant index calculation unit 42, a frequency selection unit 44, a first processing unit 50, and a second processing unit 60.

  The significant index calculation unit 42 calculates K significant index values Z (fk) (Z (f1) to Z (fK)) as a measure of the significance of the learning process using the learning data D (fk) of the frequency fk. Calculation is made for each of the frequencies f1 to fK. The significant index value Z (fk) corresponds to a numerical value indicating the degree to which the accuracy of sound source separation by the separation matrix W (fk) is improved as a result of the learning process using the learning data D (fk). As shown in FIG. 5, the frequency sorting unit 44 sorts (classifies) each of the K frequencies f1 to fK into the first frequency fA and the second frequency fB according to the significant index value Z (fk). The first frequency fA is a frequency at which the significance of the learning process using the learning data D (fk) is higher than that of the second frequency fB.

  The significant index calculation unit 42 of the first embodiment uses the determinant z1 (fk) of the covariance matrix Rxx (fk) of the learning data D (fk) (observation vector X (m, fk)) as the significant index value Z (fk). ) And is calculated for each frequency fk, and includes a covariance matrix calculation unit 422 and a determinant calculation unit 424 as shown in FIG.

The covariance matrix calculation unit 422 calculates the covariance matrix Rxx (fk) (Rxx (f1) to Rxx (fK)) of the learning data D (fk) for each of the K frequencies f1 to fK. The covariance matrix Rxx (fk) of the frequency fk is a matrix having the covariance of a plurality of observation vectors X (m, fk) in the learning data D (fk) as elements. That is, the covariance matrix Rxx (fk) is calculated by the following formula (2), for example.
Rxx (fk) = E [X (m, fk) X (m, fk) ^H ] (2)
The symbol H means transposition (conjugate transposition) of the matrix, and the symbol E [] means an average value or an addition value over a plurality of frames (the entire learning data D (fk)) in the unit interval TU. That is, the covariance matrix Rxx (fk) is a 2-by-2 square matrix generated for each unit interval TU (each learning data D (fk)).

  The determinant calculating unit 424 in FIG. 6 determines the determinant z1 (fk) (z1 (f1) to z1) for each of the K covariance matrices Rxx (f1) to Rxx (fK) calculated by the covariance matrix calculating unit 422. (fK)) is calculated. A known method is arbitrarily employed for calculating the determinant z1 (fk). For example, the following method using singular value decomposition of the covariance matrix Rxx (fk) is preferable. Hereinafter, for convenience, the covariance matrix Rxx (fk) is generalized to J rows and J columns (J = 2 in the present embodiment).

The covariance matrix Rxx (fk) is subjected to singular value decomposition as shown in Equation (3) below. The matrix F in Equation (3) is an orthogonal matrix of 2 rows and 2 columns, and the matrix D is a singular value matrix of J rows and J columns in which elements other than the diagonal components (d1,..., DJ) are zero. Diagonal matrix).
Rxx (fk) = FDF ^H (3)

Considering the singular value decomposition of Equation (3), the determinant z1 (fk) of the covariance matrix Rxx (fk) is expressed by Equation (4) below. In order to derive Equation (4), the product of the conjugate transpose matrix F ^H of the matrix F and the matrix F is a J-th order unit matrix (F ^H F = I), or the determinant det (AB of the matrix AB ) Is equal to the determinant det (BA) of the matrix BA.
z1 (fk) = det (Rxx (fk))
= Det (FDF ^H )
= Det (D)
= D1 · d2 · · · dJ (4)

  As can be understood from the equation (4), the determinant z1 (fk) of the covariance matrix Rxx (fk) is represented by J pieces of the singular value matrix D specified by the singular value decomposition of the covariance matrix Rxx (fk). This corresponds to the multiplication value of the diagonal components (d1,..., DJ). The determinant calculating unit 424 in FIG. 6 calculates the determinants z1 (f1) to z1 (fK) by executing the calculation of the equation (4) for each of the K frequencies f1 to fK.

  FIG. 7 is a scatter diagram of each observation vector X (m, fk) in the unit interval TU. The horizontal axis represents the component value x1 (m, fk), and the vertical axis represents the component value x2 (m, fk). Part (A) in FIG. 7 is a scatter diagram when the determinant z1 (fk) is large, and part (B) in FIG. 7 is a scatter diagram when the determinant z1 (fk) is small. When the determinant z1 (fk) is large as in part (A) of FIG. 7, the axis (base) of the region in which the observation vector X (m, fk) is distributed is clearly distinguished for each sound source S. . Specifically, an observation vector X (m, fk) in which the observation vector X (m, fk) from which the sound SV1 from the sound source S1 is dominant is distributed along the axis α1 and the sound SV2 from the sound source S2 is dominant. ) Is clearly distinguished from the region A2 in which it is distributed along the axis α2. On the other hand, when the determinant z1 (fk) is small, the number of regions (number of axes) of the distribution of the observed vectors X (m, fk) that can be clearly distinguished in the scatter diagram is less than the total number of the actual sound sources S. For example, there is no clear area A2 (axis α2) corresponding to the sound SV2 from the sound source S2 as shown in part (B) of FIG.

  As can be understood from the above tendency, the determinant z1 (fk) of the covariance matrix Rxx (fk) is the basis (observation) in the distribution of each observation vector X (m, fk) constituting the learning data D (fk). It functions as an index of the total number of axes in the region in which the vector X (m, fk) is distributed. That is, there is a tendency that the frequency fk with the larger determinant z1 (fk) has more bases. The frequency fk at which the determinant z1 (fk) is zero includes only one independent basis. The independent component analysis applied to the learning process of the separation matrix W (fk) is equivalent to the process of specifying the independent bases by the number of the sound sources S. Therefore, the covariance matrix Rxx ( It can be said that the learning data D (fk) of the frequency fk having a small determinant z1 (fk) of fk) has low learning significance (the degree to which the accuracy of sound source separation improves before and after the learning process).

  Using the above property of the determinant z1 (fk), the frequency selecting unit 44 in FIG. 4 selects one or more frequencies fk having a large determinant z1 (fk) among the K frequencies f1 to fK as the first frequency. The remaining frequency fk having a small determinant z1 (fk) is selected as the second frequency fB. Specifically, the frequency selection unit 44 includes a predetermined number of frequencies fk positioned in descending order of the determinants z1 (f1) to z1 (fK) among the K frequencies f1 to fK, and K frequencies f1. One or more frequencies fk in which the determinant z1 (fk) exceeds a predetermined threshold value are selected as the first frequency fA, and frequencies fk other than the first frequency fA are selected as the second frequency fB. The selection of the first frequency fA / the second frequency fB by the frequency selection unit 44 is executed sequentially for each unit interval TU, for example.

  The first processing unit 50 and the second processing unit 60 in FIG. 4 generate a separation matrix W (fk) (W (f1) to W (fK)) used in the signal separation unit 24 for each frequency fk. The first processing unit 50 generates a separation matrix W (fk) (hereinafter sometimes referred to as “separation matrix WA (fk)”) for each frequency fk selected by the frequency selection unit 44 as the first frequency fA. The second processing unit 60 then separates each frequency fk selected by the frequency selecting unit 44 into the second frequency fB (hereinafter, sometimes referred to as “separation matrix WB (fk)”). Is generated.

As shown in FIG. 4, the first processing unit 50 includes an initial matrix setting unit 52, a first learning unit 54, and a correction processing unit 56. The initial matrix setting unit 52 sets an initial value (hereinafter referred to as “initial separation matrix”) WA ^[0] (fk) of learning processing for generating the separation matrix WA (fk). Although the method of setting the initial separation matrix WA ^[0] (fk) is arbitrary, for example, a unit matrix can be set as the initial separation matrix WA ^[0] (fk). As described above, according to the configuration in which the initial separation matrix WA ^[0] (fk) is set regardless of the observation vector X (m, fk), there is an advantage that prior information on the sound source S1 and the sound source S2 is unnecessary. .

As shown in FIG. 5, the first learning unit 54 in FIG. 4 performs sequential updating (hereinafter referred to as “primary learning”) using the initial separation matrix WA ^[0] (fk) set by the initial matrix setting unit 52 as an initial value. In the processing, a separation matrix WA (fk) of each frequency fk selected as the first frequency fA is generated. For the primary learning process by the first learning unit 54, a known technique can be arbitrarily adopted. For example, the updated (n + 1) -th updated separation matrix W ^{[n + 1]} (fk) The calculation of Equation (5) calculated from the separation matrix W ^[n] (fk) (when calculating the separation matrix W ^[1] (fk), the initial separation matrix WA ^[0] (fk)) is preferable.
W ^{[n + 1]} (fk) = W ^[n] (fk) −η {off-diag (E [φ (m, fk) Y ^[n] (m, fk) ^H ] W ^[n] (fk) ……(Five)

Symbol η in equation (5) is a predetermined constant (step size), and symbol off-diag () is an operator that replaces the diagonal component with zero. The symbol φ () means a nonlinear function. For example, a hyperbolic tangent function (tanh: hyperbolic tangent) can be applied as the nonlinear function φ (). The symbol Y ^[n] (m, fk) in the formula (5) is calculated by the calculation of the formula (1A) and the formula (1B) to which the immediately preceding separation matrix W ^[n-1] (m, fk) is applied. A vector (Y ^[n] (m, fk) = (y1 (m, fk), y2 (m, fk)) ^{T having the} component value y1 (m, fk) and the component value y2 (m, fk) as elements ). The first learning unit 54 determines the separation matrix W ^[NA] (fk) as the separation matrix WA (fk) when the calculation of Expression (5) is repeated NA times. However, if it is determined that the separation matrix W ^{[n + 1]} (fk) calculated by Equation (5) has converged, the first learning unit 54 performs primary learning before the iteration reaches NA times. The process is terminated, and the separation matrix W ^{[n + 1]} (fk) at that time is determined as the separation matrix WA (fk).

  By the way, the separation matrix WA (fk) calculated by independent component analysis (primary learning processing) has a problem that the amplitude of each signal after execution of sound source separation is indefinite (scaling problem), and There is a problem (permutation problem) that the combination of each signal and each sound source can change for each frequency fk. 4 corrects each separation matrix WA (fk) generated by the first learning unit 54 for each frequency fk selected as the first frequency fA so that the scaling problem and the permutation problem are solved. To do.

  Known techniques are arbitrarily employed to solve the above scaling problem and permutation problem. For example, the scaling problem is solved by multiplying the separation matrix WA (fk) by a diagonal matrix composed of the diagonal components of the inverse matrix of the separation matrix WA (fk), and is estimated from the separation matrix WA (fk). The permutation problem is solved by replacing each row of the separation matrix WA (fk) with each other so that the directions of the sound sources are matched. Each separation matrix WA (fk) corrected by the correction processing unit 56 is applied to the processing unit Pk of each frequency fk selected as the first frequency fA among the K processing units P1 to PK of the signal separation unit 24. Is done. Saruwatari et al., “Blind Source Separation Combininb Independent Component Analysis and Beamforming”, EURASIP Journal on Applied Signal Processing Vol.2003, No.11, p.1135-1146, 2003 (below) (Referred to as “Non-Patent Document 2”).

  The second processing unit 60 in FIG. 4 uses the separation matrix WB (fk) of each frequency fk selected by the frequency selection unit 44 as the second frequency fB and the separation matrix WA (fk) generated by the first processing unit 50. And generate. As shown in FIG. 4, the second processing unit 60 includes a direction estimation unit 62, an initial matrix setting unit 64, and a second learning unit 66.

  The direction estimation unit 62 estimates the direction θ1 of the sound source S1 and the direction θ2 of the sound source S2 from each separation matrix WA (fk) generated by the first processing unit 50. For estimating the direction θ1 and the direction θ2, a known technique (for example, a method disclosed in Non-Patent Document 2) is arbitrarily adopted. For example, the following method is preferable. First, as shown in FIG. 5, the direction estimation unit 62 estimates the direction θ1 (fk) and the direction θ2 (fk) from the separation matrix WA (fk) for each frequency fk selected as the first frequency fA. To do. For example, the direction θ1 (fk) is specified from the coefficient w11 (fk) and the coefficient w12 (fk) of the separation matrix WA (fk), and the direction θ2 (fk) is determined from the coefficient w21 (fk) and the coefficient w22 (fk). Identified. Second, as shown in FIG. 5, the direction estimation unit 62 determines the direction θ1 of the sound source S1 and the direction θ2 of the sound source S2 from the direction θ1 (fk) and the direction θ2 (fk) of each frequency fk (first frequency fA). And calculate. For example, the representative value (average value or median value) of each direction θ1 (fk) is specified as the direction θ1, and the representative value of each direction θ2 (fk) is specified as the direction θ2.

As shown in FIG. 5, the initial matrix setting unit 64 in FIG. 4 determines the initial separation matrix WB ^[0] (fk) of the separation matrix WB (fk) according to the direction θ1 and the direction θ2 estimated by the direction estimation unit 62. Set. For generating the initial separation matrix WB ^[0] (fk), for example, blind spot control type beam forming disclosed in Non-Patent Document 2 is applied. Specifically, the initial matrix setting unit 64 calculates the coefficient w11 (fk) and the coefficient calculated so that the dead angle of sound collection (region where sound collection sensitivity is low) is formed in the direction θ2 estimated by the direction estimation unit 62. Initial separation matrix WB ^[ w12 (fk) and coefficient w21 (fk) and coefficient w22 (fk) calculated so that a dead angle of sound collection is formed in direction θ1 estimated by direction estimation unit 62 ^0] (fk) is generated. The initial separation matrix WB ^[0] (fk) is individually generated for each frequency fk selected by the frequency selection unit 44 as the second frequency fB.

In the above example, the initial separation matrix WB ^[0] (fk) is generated by blind-angle-controlled beam forming. However, the initial separation matrix WB ^[0] (fk) is generated in a region where the sound collection sensitivity is high ( Beam forming (for example, delay-added beam forming) may be employed. That is, the coefficient w11 (fk) and the coefficient w12 (fk) of the initial separation matrix WB ^[0] (fk) are set so that the collected beam is directed in the direction θ1, and the collected beam is directed in the direction θ2. Thus, the coefficient w21 (fk) and coefficient w22 (fk) of the initial separation matrix WB ^[0] (fk) are set.

As shown in FIG. 5, the second learning unit 66 in FIG. 4 sequentially updates the initial separation matrix WB ^[0] (fk) set by the initial matrix setting unit 64 (hereinafter referred to as “secondary learning”). In the process, a separation matrix WB (fk) of each frequency fk selected as the second frequency fB is generated. A known technique can be arbitrarily employed for the secondary learning process, but the calculation of Expression (5) is preferably employed as in the primary learning process. That is, the second learning unit 66 uses the initial separation matrix WB ^[0] (fk) set by the initial matrix setting unit 64 as an initial value, and learns data D (fk) of each frequency fk selected as the second frequency fB. Then, the calculation of Formula (5) is repeated using the vector Y ^[n] (m, fk) calculated by Formula (1A) and Formula (1B).

The number of iterations NB of Equation (5) by the second learning unit 66 is set to be less than the number of iterations NA of the first learning unit 54 (NB <NA). The second learning unit 66 calculates the separation matrix W ^[NB] (fk) at the time when the calculation of the formula (5) is repeated NB times as the separation matrix WB (fk) of the second frequency fB. Similar to the first learning unit 54, when the separation matrix W ^{[n + 1]} (fk) converges, the second learning unit 66 ends the secondary learning process before the iteration reaches NB times, The separation matrix W ^{[n + 1]} (fk) at that time is determined as the separation matrix WB (fk). The separation matrix WB (fk) generated by the above-described secondary learning processing is processed by the processing unit Pk of each frequency fk selected as the second frequency fB among the K processing units P1 to PK of the signal separation unit 24. Applied. In addition, the structure which made the content of calculation differ by the primary learning process and the secondary learning process may be employ | adopted.

The separation matrix WB (fk) calculated from the initial separation matrix WB ^[0] (fk) corresponding to the direction θ1 and the direction θ2 is compared with the separation matrix WA (fk) generated without applying prior information. Then, the above-mentioned scaling problem and permutation problem are unlikely to occur. Therefore, correction for solving the scaling problem and the permutation problem is not performed on the separation matrix WB (fk) generated by the second learning unit 66. However, a configuration in which the correction processing unit 56 corrects the separation matrix WB (fk) generated by the second learning unit 66 may be employed.

As described above, in the present embodiment, for each frequency fk selected as the first frequency fA, the separation matrix WA (fk) is generated by the primary learning process with the number of iterations NA, and is selected as the second frequency fB. Each frequency fk is separated by the secondary learning process of the number of iterations NB (NB <NA) using the initial separation matrix WB ^[0] (fk) generated according to the separation matrix WA (fk) as an initial value. A matrix WB (fk) is generated. Accordingly, the calculation amount of the arithmetic processing unit 12 (separation matrix generation unit 28) is reduced as compared with the configuration in which the calculation of the equation (5) is repeated NA times for all K frequencies f1 to fK. There is. In addition, since the separation matrix WB (fk) of the second frequency fB is generated by the secondary learning process using the initial separation matrix WB ^[0] (fk) generated from the separation matrix WA (fk) as an initial value, Compared with the configuration of Patent Document 1 in which the separation matrix WB ^[0] (fk) is used as the separation matrix WB (fk) for sound source separation (that is, the configuration in which the secondary learning process is omitted), high-accuracy sound source separation is achieved. There is an advantage that a possible separation matrix WB (fk) can be generated.

  FIG. 8 shows the relationship between the number of frequencies fk (horizontal axis) selected as the first frequency fA out of K (K = 513) and the total number of calculations of Equation (5) (hereinafter referred to as “the total number of learning”). It is a graph which shows. In the first embodiment, the total value of the number of primary learning processes and the number of secondary learning processes corresponds to the total number of learnings. In FIG. 8, the iteration number NA of the primary learning process is set to 500 and the iteration number NB of the secondary learning process is set to 100, and learning is performed when convergence of the separation matrix W (fk) is detected. It is assumed that processing will be stopped.

FIG. 8 shows the case where the primary learning process of the number of iterations NA (500 times) is executed for all (513) frequencies fk (the learning process is terminated when the separation matrix W (fk) converges). REF1) and Comparative 2 (REF2). The proportional 1 is a case where a unit matrix is used as the initial separation matrix W ^[0] (fk) of the primary learning process, and the proportional 2 is a blind spot control type using the known direction θ1 and the direction θ2. This is a case where the separation matrix generated by beam forming is used as the initial separation matrix W ^[0] (fk) in the primary learning process. In contrast 1 and contrast 2, secondary learning processing is not executed. FIG. 8 shows each of a plurality of cases in which the amplitude ratio RA (RA = 1, 0.87, 0.71, 0.5, 0.32) between the sound SV1 generated by the sound source S1 and the sound SV2 generated by the sound source S2 is changed. The total number of learning is shown. The conditions in each case on the horizontal axis and the condition of the amplitude ratio RA are the same in FIGS. 9 to 12 described later.

  As can be understood from FIG. 8, in the first embodiment in which the primary learning process is selectively executed, compared to the comparative 1 and the comparative 2 in which the primary learning process is executed for all the frequencies fk, The total number of learnings required for generating the separation matrix W (fk) is greatly reduced. As the number of first frequencies fA to be subjected to the primary learning process decreases, the difference in the total number of learnings between the first embodiment and the comparison 1 or the comparison 2 increases. That is, according to the first embodiment, there is an advantage that the amount of calculation required for generating the separation matrices W (f1) to W (fK) is reduced as compared with the comparative 1 and the comparative 2. The above tendency can be similarly confirmed regardless of the amplitude ratio RA between the sound SV1 and the sound SV2.

  FIG. 9 is a graph of noise reduction rate (NRR: Noise Reduction Rate) in the present embodiment and each comparison, and FIG. 10 is a graph of cepstrum distortion in the present embodiment and each comparison. The noise suppression rate is a difference (SNR_OUT−SNR_IN) between the intensity ratio SNR_OUT of the sound SV1 with respect to the sound SV2 in the separated signal Y1 (t) and the intensity ratio SNR_IN of the sound SV1 with respect to the sound SV2 in the observation signal V1 (t). Therefore, the higher the noise suppression rate (upward in FIG. 9), the higher the accuracy of sound source separation. The cepstrum distortion is an index of the difference in cepstrum between the sound SV1 and the separated signal Y1 (t). It means that the smaller the cepstrum distortion (upper part of FIG. 10), the smaller the change in waveform (spectrum envelope) due to sound source separation (that is, the sound SV1 is faithfully separated).

  As understood from FIGS. 9 and 10, the first frequency fA to be subjected to the primary learning process is within a range where the amplitudes of the sound SV1 and the sound SV2 are not excessively different (RA = 1, 0.87, 0.71). Even when the amount of calculation is reduced by reducing, the noise suppression rate decreases and the cepstrum distortion hardly increases as compared with the proportional 1 and the proportional 2. When the number of first frequencies fA is set to 256 or 384, it can be confirmed that the noise suppression rate and the cepstrum distortion are improved as compared with the proportional 1 and the proportional 2. As described above, according to the first embodiment, it is possible to achieve high accuracy of sound source separation while reducing the amount of calculation required for generating the separation matrix W (fk).

11 and 12 are graphs of the noise suppression rate (FIG. 11) and the cepstrum distortion (FIG. 12) when the secondary learning process is omitted under the first embodiment (hereinafter referred to as “comparative 3”). is there. That is, in the comparative 3 (REF3), as in the technique of Non-Patent Document 1, the initial separation matrix WB ^[0] (fk) set by the initial matrix setting unit 64 is the separation matrix WB (fk) of the second frequency fB. As applied to sound source separation. Conditions other than the omission of the secondary learning process are the same as those in the first embodiment shown in FIGS.

  As shown in FIG. 11, under contrast 3, it appears that the noise suppression rate improves as the number of first frequencies fA decreases. However, referring to FIG. 12, it can be confirmed that the cepstrum distortion increases as the number of the first frequencies fA decreases. That is, in FIG. 11, when the number of first frequencies fA is small, the noise suppression rate is improved because the waveform of the separated signal Y1 (t) and the waveform of the original sound SV1 are different. Therefore, it can be understood that the accuracy of sound source separation is not maintained at a high level. On the other hand, as shown in FIGS. 9 and 10, under the first embodiment in which the secondary learning process is executed for the second frequency fB, the noise suppression rate is maintained at a high level while sufficiently suppressing the cepstrum distortion. It is possible. Therefore, the first embodiment is more advantageous than the comparative 3 from the viewpoint of achieving both the maintenance of the noise suppression rate and the reduction of the cepstrum distortion (that is, realizing high-accuracy sound source separation). Further, when the numerical values of the cepstrum distortion in FIG. 10 and FIG. 12 are compared focusing on the range (RA = 1, 0.87, 0.71) in which the amplitude ratio RA between the sound SV1 and the sound SV2 is high, in the first embodiment, It can be confirmed that the cepstrum distortion is suppressed as compared with the comparative 3. Therefore, the first embodiment is advantageous from the viewpoint of faithful separation of the sound SV1 and the sound SV2.

<B: Second Embodiment>
A second embodiment of the present invention will be described. In the following examples, elements having the same functions and functions as those of the first embodiment are referred to by the same reference numerals as above, and detailed descriptions thereof are appropriately omitted.

  In the first embodiment, the secondary learning process is executed for all the frequencies fk selected as the second frequency fB. However, the secondary learning process is executed depending on the sound collection conditions of the sound collection device M1 and the sound collection device M2. In some cases, it is possible to achieve high-precision sound source separation.

  For example, as can be seen from FIG. 9 and FIG. 10, in the first embodiment, the accuracy of sound source separation when the intensity (amplitude and power) is different between the sound SV1 and the sound SV2 (RA = 0.5, 0.32). Is less than proportional 1 or proportional 2. On the other hand, as can be understood from the comparison between FIG. 9 and FIG. 11 and the comparison between FIG. 10 and FIG. 12, in the configuration of the proportional 3 that does not execute the secondary learning process, even if the amplitude ratio RA is low, the proportional Sound source separation with an accuracy comparable to 1 or proportional 2 is realized. Therefore, when the difference in intensity between the sound SV1 and the sound SV2 (hereinafter referred to as “sound source intensity difference”) that is the target of sound source separation is large (that is, when the sound collection condition is bad), the secondary learning process is not executed. It can be understood that higher accuracy sound source separation can be realized. In consideration of the above tendency, in the second embodiment, execution / stop of the secondary learning process is variably controlled according to the quality of the sound collection condition.

The sound collection conditions for the sound SV1 generated by the sound source S1 and the sound SV2 generated by the sound source S2 will be discussed below. A vector S (m, fk) (S (m, fk) = (s1 (m, fk)) having the component value s1 (m, fk) of the sound SV1 and the component value s2 (m, fk) of the sound SV2 as elements. , s2 (m, fk)) ^T ), the observation vector X (m, fk) of the observation signal V1 (t) and the observation signal V2 (t) is expressed by the following equation (6). The matrix A (fk) in Expression (6) is a mixing matrix indicating acoustic characteristics that are given from the sound source S1 and the sound source S2 to the sound collection device M1 and the sound collection device M2.
X (m, fk) = A (fk) S (m, fk) (6)

Considering Equation (2) and Equation (6), the covariance matrix Rxx (fk) of the observation vector X (m, fk) is the covariance matrix Rss (m, fk) ( Rss (m, fk) = E [S (m, fk) S (m, fk) ^H ]) and a mixing matrix A (fk) are expressed by the following equation (7).
Rxx (fk) = E [X (m, fk) X (m, fk) ^H ]
= E [A (fk) S (m, fk) {A (fk) S (m, fk)} ^H ]
= E [A (fk) S (m, fk) S (m, fk) ^H A (fk) ^H ]
= A (fk) E [S (m, fk) S (m, fk) ^H ] A (fk) ^H
= A (fk) Rss (fk) A (fk) ^H (7)

On the other hand, the covariance matrix Rss (fk) is subjected to eigenvalue decomposition as shown in the following equation (8).
Rss (fk) = Q (fk) Λ (fk) Q (fk) ^H (8)
Since the matrix of Equation (8) Q (fk) is an orthonormal matrix, determinant of matrix Q (fk) Q (fk) H (det (Q (fk) Q (fk) H) is 1. Thus , The determinant det (Rss (fk)) of the covariance matrix Rss (fk) is equal to the determinant det (Λ (fk)) of the diagonal matrix Λ (fk) (det (Rss (fk)) = det ( Λ (fk))) In consideration of the above, the determinant det (Rxx (fk)) of the covariance matrix Rxx (fk) is expressed by the following equation (9) obtained by modifying equation (7). In the equation (9), the symbol 総 means the operator (Πλi (fk) = λ1 (fk) · λ2 (fk)).
det (Rxx (fk)) = det (A (fk) Rss (fk) A (fk) ^H )
= Det (A (fk)) det (Rss (fk)) det (A (fk) ^H )
= | Det (A (fk)) | ² det (Λ (fk))
= | Det (A (fk)) | ² Πλi (fk) …… (9)

  Since the determinant det (A (fk)) of Equation (9) corresponds to an element of a constant multiple in the linear map to which the mixing matrix A is applied, the sound SV1 and sound for each of the sound collection devices M1 and M2 The greater the degree of inhibition of SV2 propagation (hereinafter referred to as “propagation inhibition degree”) (the worse the sound collection condition), the smaller the determinant det (A (fk)) of Equation (9). On the other hand, the symbol λi (fk) in Equation (9) is a component of the diagonal matrix Λ (the eigenvalue of the covariance matrix Rss (fk)). That is, the eigenvalue λ1 (fk) corresponds to the intensity (power) of the frequency fk component of the sound SV1, and the eigenvalue λ2 (fk) corresponds to the intensity (power) of the frequency fk component of the sound SV2. Therefore, the greater the sound source intensity difference (the worse the sound collection condition), the smaller the total power λi (fk) in Equation (9).

  As understood from the above description, the worse the sound collection condition (the greater the propagation inhibition degree or the sound source intensity difference), the more the determinant z1 (fk) (z1 (fk) = det of the covariance matrix Rxx (fk) (Rxx (fk))) tends to be small. In consideration of the above tendency, in the second embodiment, the determinant z1 (fk) is applied to determine whether the sound collection condition is good or bad.

  FIG. 13 is a block diagram of the separation matrix generation unit 28A in the second embodiment. The separation matrix generation unit 28A of the second embodiment has a configuration in which a condition determination unit 72 is added to each element of the separation matrix generation unit 28 of the first embodiment. The condition determination unit 72 determines whether the sound collection condition is good or not for each frequency fk selected as the second frequency fB. In the determination by the condition determination unit 72, the determinant z1 (fk) calculated by the significant index calculation unit 42 for the selection of the frequency fk by the frequency selection unit 44 is used. That is, the condition determining unit 72 determines that the sound collection condition of the frequency fk is good (the propagation inhibition degree and the sound source intensity difference are small) when the determinant z1 (fk) exceeds a predetermined threshold, and the determinant z1 ( When fk) falls below the threshold, it is determined that the sound collection condition of the frequency fk is bad (a bad condition with a large degree of propagation inhibition and a difference in sound source intensity).

The second learning unit 66 in FIG. 13 determines the execution / stop of the secondary learning process for each frequency fk according to the determination result of the condition determination unit 72. That is, the second learning unit 66 is the same as in the first embodiment for the frequency fk determined that the determinant z1 (fk) is large (the sound collection condition is good) among the frequencies fk selected as the second frequency fB. In addition, the separation matrix WB (fk) is generated by the secondary learning process using the initial separation matrix WB ^[0] (fk) as an initial value. On the other hand, the second learning unit 66 stops the secondary learning process for the frequency fk determined that the determinant z1 (fk) is small (the sound collection condition is bad) among the frequencies fk selected as the second frequency fB. Then, the initial separation matrix WB ^[0] (fk) set by the initial matrix setting unit 64 is determined as the separation matrix WB (fk). Therefore, the learning data D (fk) having a small determinant z1 (bad condition) is not used to generate the separation matrix WB (fk).

  In the second embodiment, the same effect as in the first embodiment is realized. In the second embodiment, the execution / stop of the secondary learning process is controlled according to the quality of the sound collection condition. Therefore, the secondary frequency is selected for each frequency fk selected as the second frequency fB regardless of the sound collection condition. Compared to the case where the learning process is executed, a special effect is achieved in that the amount of computation for generating the separation matrix W (fk) is reduced while maintaining the accuracy of sound source separation.

<C: Third Embodiment>
In the second embodiment, the initial separation matrix WB generated by the initial matrix setting unit 64 by the blind spot control type beam forming for the frequency fk determined to have a poor sound collection condition among the frequencies fk selected as the second frequency fB. ^{[0] Although} (fk) is used as the separation matrix WB (fk), a method of setting the separation matrix WB (fk) for the frequency fk having a poor sound collection condition is exemplified as the third embodiment below. It is changed appropriately.

When the component of the frequency fk of the observation signal V1 (t) or the observation signal V2 (t) includes only the sound SV of one sound source S, the condition determination unit 72 determines that the sound collection condition of the frequency fk is bad. On the other hand, if only the sound SV of one sound source S exists at the frequency fk, it is less necessary to change the component of the frequency fk before and after sound source separation. Therefore, for the frequency fk determined by the condition determination unit 72 that the sound collection condition is bad (including only one sound source S), in addition to the secondary learning process by the second learning unit 66, the initial value by the initial matrix setting unit 64 The generation of the separation matrix WB ^[0] (fk) is also stopped, and a configuration in which the separation matrix WB (fk) is set so that the component of the frequency fk does not change excessively before and after the sound source separation can be adopted. A specific configuration will be described in detail below.

  FIG. 14 is a block diagram of the separation matrix generation unit 28B in the third embodiment. The separation matrix generation unit 28B of the third embodiment has a configuration in which a direction estimation unit 74 and a matrix setting unit 76 are added to the separation matrix generation unit 28A of the second embodiment.

  The direction estimation unit 74 uses the learning data D (fk) for each frequency fk determined by the condition determination unit 72 that the sound collection condition is bad among the frequencies selected as the second frequency fB, that is, the sound source direction (that is, The direction of one sound source that emits sound including a component of frequency fk) θe (fk) is estimated. A known technique can be arbitrarily employed for estimating the sound source direction θe (fk). The matrix setting unit 76 generates a separation matrix WB (fk) that separates the sound coming from the sound source direction θe (fk) estimated by the direction estimating unit 74 from the observation signal V1 (t) and the observation signal V2 (t). For example, the matrix setting unit 76 generates the separation matrix WB (fk) by executing the following processing for each frequency fk determined to have poor sound collection conditions.

First, the matrix setting unit 76 generates an extraction matrix C (fk) defined by the following mathematical formula. The symbol d means the interval between the sound collecting device M1 and the sound collecting device M2, and the symbol c means the speed of sound. Therefore, the symbol τ corresponds to the time difference at which the sound coming from the sound source direction θe (fk) reaches each of the sound collecting device M1 and the sound collecting device M2. The first row of the extraction matrix C (fk) emphasizes the incoming sound from the sound source direction θe (fk) when applied to delay-added beamforming.

  Second, the matrix setting unit 76 extracts the direction θ1 and the direction θ2 estimated from the separation matrix WA (fk) by the direction estimation unit 62 and the relationship between the sound source direction θe (fk) estimated by the direction estimation unit 74. A separation matrix WB (fk) is generated by replacing each row of the matrix C (fk) with each other. Specifically, the matrix setting unit 76 emphasizes the component of the frequency fk with the separated signal Y1 (t) when the sound source direction θe (fk) is close to the direction θ1, and the sound source direction θe (fk) is the direction θ2. If the frequency is close to, the position of each row of the extraction matrix C (fk) is adjusted so that the component of the frequency fk is emphasized by the separation signal Y2 (t).

  For example, the first row (w11 (fk), w12 (fk)) of the separation matrix WA (fk) acts to emphasize the acoustic SV1 in the direction θ1 (forms a blind spot in the direction θ2), and the separation matrix WA ( Assume that the second row (w21 (fk), w22 (fk)) of fk) acts to emphasize the sound SV2 in the direction θ2 (forms a blind spot in the direction θ1). When the sound source direction θe (fk) is close to the direction θ1, the matrix setting unit 76 determines the extraction matrix C (fk) described above as the separation matrix WB (fk). Therefore, the component value y1 (m, fk) of the frequency fk determined to have poor sound collection conditions in the separated signal Y1 (t) is set to a value that emphasizes the incoming sound from the sound source direction θe (fk), and is separated. The component value y2 (m, fk) of the frequency fk of the signal Y2 (t) is set to zero. On the other hand, when the sound source direction θe (fk) is close to the direction θ2, the matrix setting unit 76 determines a matrix obtained by replacing the first row and the second row of the extraction matrix C (fk) as the separation matrix WB (fk). . Therefore, the component value y1 (m, fk) of the frequency fk that is determined to be bad in the sound collection condition in the separated signal Y1 (fk) is set to zero, and the component value y2 of the frequency fk of the separated signal Y2 (t). (M, fk) is set to a numerical value that emphasizes the incoming sound from the sound source direction θe (fk).

In the third embodiment, the same effects as those of the first embodiment and the second embodiment are realized. In the third embodiment, the generation of the initial separation matrix WB ^[0] (fk) by the initial matrix setting unit 64 is also stopped in addition to the secondary learning process by the second learning unit 66 for the frequency fk having a poor sound collection condition. Therefore, there is also an advantage that the amount of calculation required for generating the separation matrix WB (fk) is reduced as compared with the second embodiment.

The contents of the extraction matrix C (fk) are not limited to the above examples. For example, if the extraction matrix C (fk) illustrated below is used, the component of the frequency fk included in the observation signal V1 (t) or the observation signal V2 (t) is directly used as the separation signal Y1 (t) or the separation signal Y2 ( The signal is output from the signal separation unit 24 as a component of the frequency fk of t).

<D: Fourth Embodiment (Exemplary Significant Index Value)>
In each of the above embodiments, the significant index value Z (fk) serving as a reference for selection by the frequency selection unit 44 is not limited to the determinant z1 (fk) of the covariance matrix Rxx (fk). Specifically, numerical values (statistics) exemplified in the following aspects can be adopted as the significant index value Z (fk).

<D-1: First mode (condition number z2 (fk))>
The condition number z2 (fk) of the covariance matrix Rxx (fk) of the plurality of observation vectors X (m, fk) constituting the learning data D (fk) is defined by the following equation (10). The operator ‖A‖ in Expression (10) means the norm (matrix distance) of the matrix A. The condition number z2 (fk) is small when the inverse matrix exists in the covariance matrix Rxx (fk) (when it is regular), and the condition number z2 (fk) when the inverse matrix does not exist in the covariance matrix Rxx (fk). ) Is a large number.
z2 (fk) = ‖Rxx (fk) ‖ ・ ‖Rxx (fk) ^-1 ………… (10)

The covariance matrix Rxx (fk) is subjected to eigenvalue decomposition as shown in the following equation (11A). The matrix U in Equation (11A) is an eigenmatrix (matrix having eigenvectors as elements), and the matrix Σ is a diagonal matrix having eigenvalues as elements. The inverse matrix of the covariance matrix Rxx (fk) is expressed by the following formula (11B) obtained by modifying the formula (11A).
Rxx (fk) = UΣU ^H (11A)
Rxx (fk) ^-1 = UΣ ^-1 U ^H (11B)

When the element of the matrix Σ includes zero, the matrix Σ ⁻¹ of the formula (11B) diverges infinitely, and therefore there is no inverse matrix of the covariance matrix Rxx (fk) (that is, the formula (10) Condition number z2 (fk) is a large numerical value). On the other hand, the fact that the elements of the matrix Σ (the eigenvalues of the covariance matrix Rxx (fk)) include values close to zero means that the total number of bases in the distribution of the observation vector X (m, fk) is small. Therefore, the condition number z2 (fk) of the covariance matrix Rxx (fk) is larger as the total number of bases of the observation vector X (m, fk) is smaller (the condition number z2 (fk) is smaller as the total number of bases is larger). The trend is grasped. That is, the condition number z2 (fk) of the covariance matrix Rxx (fk) functions as an index of the total number of bases of the observation vector X (m, fk), similarly to the determinant z1 (fk).

  Considering the above tendency, in the first mode, the condition number z2 (fk) of the covariance matrix Rxx (fk) is used as the significant index value Z (fk). In other words, the significant index calculation unit 42 performs the operation of Expression (10) on the covariance matrix Rxx (fk) of each of the K frequencies f1 to fK to thereby obtain the condition number z2 (fk) (z2 (f1) to z2 (fK)) is calculated. The frequency selection unit 44 selects one or more frequencies fk (for example, a predetermined number of frequencies fk positioned higher in ascending order or a frequency fk lower than the threshold) in the ascending order, with the condition number z2 (fk) calculated by the significant index calculation unit 42 being small. The first frequency fA is sorted and the remaining frequency fk is sorted to the second frequency fB.

<D-2: Second mode (cross-correlation z3 (fk), mutual information z4 (fk))>
The learning process of independent component analysis is a process of updating the separation matrix W (fk) so that each signal after the sound source separation is statistically independent. Therefore, the observation signal V1 (t) and the observation signal V2 (t) Therefore, it can be said that the learning frequency of the separation matrix W (fk) using the learning data D (fk) is higher as the frequency fk has a lower statistical correlation. Therefore, in the second mode, an index value (for example, cross-correlation z3 (fk)) corresponding to the independence between the observation signal V1 (t) and the observation signal V2 (t) is used as a significant index value Z (fk). Use as

The cross-correlation z3 (fk) between the frequency fk component of the observation signal V1 (t) and the frequency fk component of the observation signal V2 (t) is expressed by the following equation (12). The symbol σ1 in the equation (12) means the standard deviation of the intensity x1 (m, fk) in the unit interval TU, and the symbol σ2 means the standard deviation of the intensity x2 (m, fk) in the unit interval TU. .
z3 (fk) = E [{x1 (m, fk) -E (x1 (m, fk))} {x2 (m, fk) -E (x2 (m, fk))}] / σ1σ2 (12 )

  As understood from the equation (12), the cross-correlation z3 (fk) becomes a smaller numerical value as the frequency fk has higher independence (lower correlation) between the observation signal V1 (t) and the observation signal V2 (t). Considering the above tendency, in the second mode, the significant index calculation unit 42 performs the cross-correlation z3 (fk) (z3) by executing the calculation of the equation (12) for each of the K frequencies f1 to fK. (f1) to z3 (fK)), and the frequency selecting unit 44 selects one or more frequencies fk (for example, higher frequencies in ascending order) having a low cross-correlation z3 (fk) among the K frequencies f1 to fK. fk and the frequency fk below the threshold value are selected as the first frequency fA, and the remaining frequency fk is selected as the second frequency fB.

Further, the mutual information amount z4 (fk) defined by the following equation (13) can also be used as the significant index value Z (fk). Similarly to the cross-correlation z3 (fk), the mutual information z4 (fk) becomes a smaller numerical value as the frequency fk has a higher independence (lower correlation) between the observation signal V1 (t) and the observation signal V2 (t). Therefore, the frequency selection unit 44 selects one or more frequencies fk having a low mutual information amount z4 (fk) from the K frequencies f1 to fK as the first frequency fA.
z4 (fk) = (− 1/2) log (1-z3 (fk) ² ) (13)

<D-3: Third mode (trace z5 (fk))>
The trace (power) z5 (fk) of the covariance matrix Rxx (fk) is defined as the sum of the diagonal components of the covariance matrix Rxx (fk). The diagonal components of the covariance matrix Rxx (fk) are the variance σ1 ² of the intensity x1 (m, fk) of the observation signal V1 (t) in the unit interval TU and the intensity x2 () of the observation signal V2 (t) in the unit interval TU. m, fk) corresponding to the variance σ2 ² , the trace z5 (fk) of the covariance matrix Rxx (fk) is the variance σ1 ² of the strength x1 (m, fk) and the variance of the strength x2 (m, fk) is defined as .sigma. @ 2 ² and the addition value ^{(z5 (fk) = σ1 2} + σ2 2).

  FIG. 15 is a scatter diagram of each observation vector X (m, fk) in the unit interval TU. Part (A) in FIG. 15 is a scatter diagram when the trace z5 (fk) is large, and part (B) in FIG. 15 is a scatter diagram when the trace z5 (fk) is small. In the part (A) and the part (B) of FIG. 15, as in the part (A) of FIG. 7, an area A1 in which the observation vector X (m, fk) from which the sound SV1 from the sound source S1 is dominant is distributed, A region A2 in which the observation vector X (m, fk) in which the sound SV2 from the sound source S2 is dominant is distributed is schematically illustrated.

As can be understood from the definition of the added value of the variance σ1 ^{2 of the} intensity x1 (m, fk) and the variance σ2 ² of the intensity x2 (m, fk), the trace z5 (fk) of the covariance matrix Rxx (fk) The observation vector X (m, fk) is more widely distributed as the value of becomes larger. Therefore, when the trace z5 (fk) is large, the region (region A1 and region A2) where the observation vector X (m, fk) is distributed is clearly defined for each sound source S as shown in part (A) of FIG. If the trace z5 (fk) is small, the distinction between the region A1 and the region A2 tends to be ambiguous as shown in part (B) of FIG. That is, the trace z5 (fk) functions as an index value of the shape (expansion) of the region in which the observation vector X (m, fk) is distributed. Since the learning process (independent component analysis) of the separation matrix W (fk) is equivalent to the process of specifying the independent bases by the number of the sound sources S, the region in which the observation vector X (m, fk) is distributed (basis ) Is clearly distinguished for each sound source S (that is, the frequency fk at which the trace z5 (fk) is large), the learning of the separation matrix W (fk) using the learning data D (fk) is more significant. I can say that.

  Considering the above tendency, in the third mode, the trace z5 (fk) of the covariance matrix Rxx (fk) is used as the significant index value Z (fk). In other words, the significant index calculation unit 42 adds the diagonal components of the covariance matrix Rxx (fk) of each of the K frequencies f1 to fK to obtain the trace z5 (fk) (z5 (f1) to z5 (fK). ) Is calculated. The frequency sorting unit 44 sorts one or more frequencies fk (for example, a higher frequency fk or a frequency fk that exceeds a threshold value in descending order) having a large trace z5 (fk) calculated by the significant index calculation unit 42 into the first frequency fA. At the same time, the remaining frequency fk is selected as the second frequency fB.

<D-4: Fourth aspect (kurtosis z6 (fk))>
The kurtosis z6 (fk) in the frequency distribution of the intensity x1 (m, fk) of the observed signal V1 (t) (distribution function with the intensity x1 (m, fk) as a random variable) is expressed by the following formula (14 ).
z6 (fk) = μ4 (fk) / {μ2 (fk)} ² …… (14)

The symbol μ4 (fk) in the equation (14) means a fourth-order moment defined by the following equation (15A), and the symbol μ2 (fk) in the equation (14) is defined by the equation (15B). Means second moment. The symbol m (fk) in the equations (15A) and (15B) means the average value of the intensity x1 (m, fk) over a plurality of frames in the unit interval TU.
μ4 (fk) = E {x1 (m, fk) −m (fk)} ⁴ …… (15A)
μ2 (fk) = E {x1 (m, fk) −m (fk)} ² …… (15B)

  When only one of the component SV1 (fk) of the sound SV1 and the component SV2 (fk) of the sound SV2 is included (or dominant) in the observation signal V1 (t), the kurtosis z6 (fk) is a large numerical value. When the component SV1 (fk) and the component SV2 (fk) are included in the observation signal V1 (t) with substantially the same intensity, the kurtosis z6 (fk) is a small value (central limit theorem). Since the learning process (independent component analysis) of the separation matrix W (fk) is equivalent to the process of specifying the independent bases by the number of the sound sources S, the acoustic SV included in the observation signal V1 (t) with a significant volume is obtained. It can be said that the learning of the separation matrix W (fk) using the learning data D (fk) is more significant as the frequency fk having a larger number of sound sources S (that is, the frequency fk having a smaller kurtosis z6 (fk)).

  Considering the above tendency, in the fourth mode, the kurtosis z6 (fk) in the frequency distribution of the intensity x (m, fk) of the observation signal V1 (t) is used as the significant index value Z (fk). That is, the significant index calculation unit 42 calculates the kurtosis z6 (f1) to z6 (fK) by executing the calculation of the equation (14) for each of the K frequencies f1 to fK. The frequency selection unit 44 selects one or more frequencies fk (for example, a higher frequency fk in an ascending order or a frequency fk lower than the threshold value in ascending order) from the K frequencies f1 to fK as the first frequency fA. And the remaining frequency fk is selected as the second frequency fB.

  By the way, the kurtosis of human speech is a numerical value in the range of approximately 40 to 70. In addition, considering the reduction of kurtosis in the presence of noise (central limit theorem) and kurtosis measurement errors, the kurtosis of human speech is generally in the range of 20 to 80 (hereinafter referred to as “voice range”). ”). On the other hand, for the frequency fk where only stationary noise such as air-conditioning operation noise or crowded noise is present, the kurtosis of the observation signal V1 (t) is a sufficiently low value (for example, a value below 20). Therefore, there is a high possibility that the frequency selection unit 44 selects the first frequency fA. However, if the target sound (SV1, SV2) for sound source separation is a human voice, it can be said that the learning of the separation matrix W using the learning data D (fk) of the stationary noise frequency fk is low. .

  Therefore, a configuration in which the kurtosis of Equation (14) is corrected is preferably employed so that the stationary noise frequency fk is avoided from being selected as the first frequency fA. For example, the significant index calculation unit 42 calculates a multiplication value of a numerical value defined by Equation (14) (hereinafter referred to as “priority before correction”) and a weight value q as a corrected kurtosis z6 (fk). The weight value q is selected non-linearly with respect to the kurtosis before correction, for example, as illustrated in FIG. That is, for a range in which the kurtosis before correction is lower than the lower limit value (for example, 20) of the voice range, the kurtosis z6 (fk) after correction by multiplication of the weight value q exceeds the upper limit value (for example, 80) in the voice range. Thus, the weight value q is variably selected according to the kurtosis before correction, and the weight value q is set to a predetermined value (for example, 1) for the kurtosis in the speech range. In the range exceeding the upper limit of the voice range, the pre-correction kurtosis is sufficiently high (that is, the possibility that the frequency fk is selected as the first frequency fA is low), so the weight value q is equal to that in the voice range. Set to the number of. According to the above configuration, it is possible to generate a separation matrix W (fk) that can separate desired speech with high accuracy.

<E: Modification>
Various modifications are added to the above embodiments. Specific modifications are exemplified below. Two or more aspects arbitrarily selected from the following examples can be appropriately combined.

(1) Modification 1
The method of selecting the frequencies f1 to fK into the first frequency fA and the second frequency fB is appropriately changed. For example, a configuration in which the significant index value Z (fk) is calculated from the plurality of types of indexes exemplified above can be adopted. That is, the significant index calculation unit 42 calculates a weighted sum (for example, determinant z1 (fk) and trace z5 (fk) of a plurality of types of indices selected from the indices (z1 (fk) to z6 (fk)) exemplified above. Is calculated as a significant index value Z (fk).

  Note that the configuration (significant index calculating unit 42) that uses the significant index value Z (fk) for selection between the first frequency fA and the second frequency fB can be omitted. Specifically, a configuration in which the frequency fk is selected regardless of the observation vector X (m, fk) (learning data D (fk)) may be employed. For example, the frequency sorting unit 44 sorts each frequency fk selected for each predetermined number from the arrangement of the frequencies f1 to fK into the first frequency fA and sorts the remaining frequency fk into the second frequency fB. Further, the frequency fk having a high significance of the learning process is experimentally or statistically determined in advance, for example, due to circumstances such as the acoustic characteristics assumed for the observation signal V1 (t) and the observation signal V2 (t) and the contents of the learning process. If known, a configuration may be adopted in which the frequency fk is selected as the first frequency fA and the remaining frequency fk is selected as the second frequency fB. If the calculation of the significant index value Z (fk) is omitted as illustrated above, there is an advantage that the calculation amount of the arithmetic processing device 12 is reduced.

(2) Modification 2
In the second embodiment, the determinant z1 (fk) of the covariance matrix Rxx (fk) of the observation vector X (m, fk) is applied to determine the quality of the sound collection condition. The method is arbitrary. For example, the condition number z2 (fk) of the covariance matrix Rxx (fk) of the observation vector X (m, fk) functions as a measure of difficulty in numerical analysis. From the viewpoint that the sound collection condition is better as the numerical analysis of the learning data D (fk) is easier, the configuration for determining the quality of the sound collection condition according to the condition number z2 (fk) calculated by the significant index calculation unit 42 Can be employed. Since it can be evaluated that the sound collection condition is better as the condition number z2 (fk) is closer to 1, the condition determination unit 72 has a better sound collection condition of the frequency fk when the condition number z2 (fk) is lower than the threshold (good). If the condition number z2 (fk) exceeds the threshold, it is determined that the sound collection condition of the frequency fk is bad (bad condition). The secondary learning process is omitted for the frequency fk (second frequency fB) where the sound collection condition is bad.

  Referring to FIGS. 9 and 10, since the reduction of the noise suppression rate and the increase of the cepstrum distortion become apparent when the amplitude ratio RA is below 0.5, it is evaluated as an unfavorable condition when the amplitude ratio RA is below 0.5. Is reasonable. Since the condition number z2 (fk) corresponds to the relative ratio of the power of the sound SV1 and the sound SV2, when the amplitude ratio RA is 0.5 (the relative ratio of power is 0.25), the condition number z2 (fk ) Is expected to be 4. Therefore, when the condition number z2 (fk) is used to determine whether or not the sound collection condition is good, the threshold value for the sound collection condition is set to 4 (that is, when the condition number z2 (fk) is less than 4). A configuration in which a good condition is determined and a bad condition is determined when the condition number z2 (fk) exceeds 4 can be suitably employed.

  As described above, according to the configuration in which the significant index Z (fk) (z1 (fk), z2 (fk)) applied to the selection of the frequency fk is used for the determination of the sound collection condition, the frequency fk There is an advantage that the amount of calculation is reduced as compared with the configuration in which separate indicators are applied to the selection of sound and the determination of the sound pickup condition. However, a configuration in which separate indicators are applied to the selection of the frequency fk and the determination of the sound collection condition can also be adopted. For example, the determinant z1 (fk) is applied for the determination of the sound pickup condition, and the significant index Z (fk) (z2 (fk) to z6 (fk)) other than the determinant z1 (fk) is selected for selecting the frequency fk. The structure which applies is adopted.

(3) Modification 3
The method of generating the initial separation matrix WA ^[0] (fk) by the initial matrix setting unit 52 is arbitrary. For example, a configuration in which the initial matrix setting unit 52 generates the initial separation matrix WA ^[0] (fk) having random numbers as elements can be employed. In the above, the case where the angle θ1 of the sound source S1 and the angle θ2 of the sound source S2 are unknown (when the prior information is not used) is illustrated, but the initial separation matrix WA ^[0 is used using the prior information (angle θ1 and angle θ2). ^Arrangements for generating a (fk) is also suitable. The initial separation matrix WA ^[0] (fk) using prior information was generated by Tachibana et al., “Efficient Blind Source Separation Combining Closed-Form Second Order ICA and Nonclosed-Form Higher-Order ICA”, International Conference on Subspace methods such as principal component analysis and second order statistics ICA disclosed in Acoustics, Speech, and Signal Processing (ICASSP), Vol. 1, p. 45-48, Apr. 2007, or Japanese Patent No. 3990774 The adaptive beam forming disclosed in (1) can be suitably employed. In addition, the initial separation matrix WA ^[0] (fk) is generated from the direction of each sound source S estimated by the MUSIC (MUltiple SIgnal Classification) method or the minimum variance method using various beam forming (for example, adaptive beam forming). Alternatively, a method of generating an initial separation matrix WA ^[0] (fk) from a factor vector specified by factor analysis or a canonical vector specified by canonical correlation analysis may be employed.

DESCRIPTION OF SYMBOLS 100 ... Signal processing device, 12 ... Arithmetic processing device, 14 ... Memory | storage device, 22 ... Frequency analysis part, 24 ... Signal separation part, 26 ... Signal composition part, 28, 28A, 28B ... Separation matrix Generation unit 42... Significant index calculation unit 422... Covariance matrix calculation unit 424 .. Determinant calculation unit 44... Frequency selection unit 50... First processing unit 52. 54... First learning unit 56... Correction processing unit 60... Second processing unit 62 .. Direction estimation unit 64. Condition determination unit.

Claims (5)

A plurality of separated signals for each sound source by applying a separation matrix of each of a plurality of frequencies to a plurality of observation signals obtained by collecting a plurality of sound mixed sounds generated by different sound sources by a plurality of sound collecting devices Signal separating means for generating
Frequency sorting means for sorting the plurality of frequencies into a first frequency and a second frequency;
First learning means for generating the separation matrix of the first frequency by primary learning processing applying learning data corresponding to the component of the first frequency in the plurality of observation signals;
Direction estimating means for estimating the direction of each sound source from the separation matrix generated by the first learning means;
Initial matrix setting means for generating an initial separation matrix so that a dead angle or beam of sound collection is formed in the direction estimated by the direction estimation means;
The secondary learning process using the learning data corresponding to the component of the second frequency in the plurality of observation signals is less than the primary learning process with the initial separation matrix generated by the initial matrix setting means as an initial value. A signal processing apparatus comprising: a second learning unit that generates the separation matrix of the second frequency by executing the number of iterations. It comprises a condition determination means for determining the quality of sound collection conditions for each frequency,
The second learning means uses the initial separation matrix as an initial value for the frequencies determined by the condition determination means that the sound collection condition is good among the frequencies selected as the second frequency. The signal processing device according to claim 1, wherein a separation matrix is generated by processing, and the initial separation matrix is used as the separation matrix for the frequency determined by the condition determination unit when the sound pickup condition is bad. Significant index calculation means for calculating, for each frequency, a significant index value indicating the significance of learning processing using the learning data of each frequency from the plurality of observation signals,
The signal processing device according to claim 2, wherein the frequency sorting unit sorts the plurality of frequencies into the first frequency and the second frequency according to a significant index value of each frequency. The signal processing apparatus according to claim 3, wherein the condition determination unit determines whether the sound collection condition for each frequency is acceptable according to a significant index value of each frequency calculated by the significant index calculation unit. The signal processing device according to claim 3 or 4, wherein the significant index calculation means calculates, as the significant index, a determinant of a covariance matrix of an observation vector having an intensity at each frequency in each of a plurality of observation signals as an element. .

Images