Signal processing device and signal processing method

Abstract

The present invention achieves fast convergence of optimization and enables high-speed sound source separation. A signal processing device according to the present invention outputs a separated signal for one target sound source from a collected mixed signal. The signal processing device has a separated signal calculation unit that uses a separation filter to calculate a separated signal from a mixed signal, a filter updating unit that uses a first covariance matrix, a second covariance matrix, auxiliary information for a target sound source, and a prescribed parameter to update the separation filter, and a separated signal outputting unit that uses the updated separation filter to generate and output a separated signal for the target sound source designated by the auxiliary information from the mixed signal.

Description

Signal processing device and signal processing method

[0001] The present disclosure relates to a sound source separation technique for separating a pre-mixed sound source signal from a mixed signal observed using multiple microphones.

[0002] Active research has been conducted into sound source separation technology that separates pre-mixed sound source signals from collected mixed signals. A method called Geometrically-Regularized Independent Vector Analysis (GR-IVA) is known as a method for separating sound sources using spatial information, such as the direction of the sound source to be separated. To date, Vector-Wise Coordinate Descent (VCD) has been developed as an optimization method for GR-IVA (see, for example, Non-Patent Document 1).

[0003] Y. Mitsui, N. Takamune, D. Kitamura, H. Saruwatari, Y. Takahashi and K. Kondo, "Vectorwise Coordinate Descent Algorithm for Spatially Regularized Independent Low-Rank Matrix Analysis," 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Calgary, AB, Canada, 2018, pp. 746-750.

[0004] However, when the number of microphones is d, VCD uses d separation filters w １ ,w ２ ,…,w ｄ By updating sequentially, the separation matrix W = [w １ ,w ２ ,…,w ｄ ], there was a problem that the optimization convergence was slow and the time required for sound source separation was long.

[0005] Therefore, the present disclosure has been made to solve the above-mentioned problems, and aims to provide a signal processing device that enables fast optimization convergence and high-speed sound source separation when the target sound to be separated and extracted is a single sound source.

[0006] In order to solve the above problem, a signal processing device according to one aspect of the present disclosure is a signal processing device that outputs a separation signal of one target sound source from a collected mixed signal, and includes: a separation signal calculation unit that calculates the separation signal from the mixed signal using a separation filter; a filter update unit that updates the separation filter using a first covariance matrix, a second covariance matrix, auxiliary information of the target sound source, and predetermined parameters; and a separation signal output unit that generates and outputs a separation signal of the target sound source specified by the auxiliary information from the mixed signal using the updated separation filter.

[0007] According to the signal processing device of the present disclosure, when the target sound to be separated and extracted is one sound source, the entire separation matrix can be optimized in one go, which enables fast convergence of optimization and high-speed sound source separation.

[0008] FIG. 1 is a diagram showing an example of the functional configuration of a signal processing device according to the first embodiment. FIG. 2 is a diagram showing an example of the processing flow of a signal processing method according to the first embodiment. FIG. 3 is a diagram showing an example of the processing flow of step S50-1. FIG. 4 is a diagram showing an example of the functional configuration of a signal processing device according to the second embodiment. FIG. 5 is a diagram showing an example of the processing flow of a signal processing method according to the second embodiment. FIG. 6 is a diagram showing an example of the processing flow of step S50A-1. FIG. 7 is a diagram showing an example of the functional configuration of a computer.

[0009] <Character notation> Symbols used in the text: ～ " (superscript tilde) should be written directly above the character immediately following it, but due to limitations in text notation, it is written immediately before the character in question. In mathematical formulas, these symbols are written in their proper position, that is, directly above the character. For example, " ～ "x" is expressed in the formula as follows: The symbol "^" (superscript hat) used in text should normally be written directly above the character immediately following it, but due to limitations in text notation, it is written immediately before the character in question. In mathematical formulas, these symbols are written in their proper position, i.e., directly above the character. For example, "^y" is expressed in a mathematical formula as follows: The symbol "￣" (overline) used in text should normally be written directly above the character immediately following it, but due to limitations in text notation, it is written immediately before the character in question. In mathematical formulas, these symbols are written in their proper position, i.e., directly above the character. For example, "￣b" is expressed in a mathematical formula as follows: Characters used in sentences with subscripts, e.g. " Ｈ " (superscript H) should be written directly above the subscript that follows it, but due to limitations in text notation, it is written immediately after the subscript. In mathematical formulas, these characters are written in their proper position, i.e., directly above the subscript. For example, "w ｉ Ｈ " is expressed in the following formula: Hereinafter, embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. Components having the same functions will be assigned the same reference numerals, and duplicated descriptions will be omitted.

[0010] First Embodiment A signal processing device 1 according to a first embodiment of the present disclosure uses, for example, GR-IVE (Geometrically Regularized Independent Vector Extraction) shown in equations (1) to (7) as a method for calculating a separation filter W(ω) (see, for example, Reference 1).

[0011] (Reference 1) A. Brendel, T. Haubner and W. Kellermann, "A Unified Probabilistic View on Spatially Informed Source Separation and Extraction Based on Independent Vector Analysis," in IEEE Transactions on Signal Processing, vol. 68, pp. 3545-3558, 2020, doi: 10.1109 / TSP.2020.3000199.

[0012] In equation (1), the frequency bin ω should be used to express W(ω), V(ω), and b(ω), but ω is omitted. W is a separation filter. i is the separation filter for extracting the i-th sound source. ｊ (ω) is the estimated steering vector of the jth sound source. x(ω,τ) is the mixture signal as the observed signal. ^y ｉ (τ) is the separated signal estimated (calculated) as the sound source i. Ｈ denotes the Hermitian transpose. Tr(·) denotes the trace of a matrix. In equation (6), ||·|| ２ denotes the L2 norm. ||·|| ２ ２ denotes the square of the L2 norm. ｉ (ω) and V ｘ (ω) and R ｉ (ω) and R ｘ (ω) is the d×d covariance matrix. ｉ (ω) is a d-dimensional complex vector. G′(·) is the derivative of G(·), and for G(·), for example, equations (8) and (9) can be used.

[0013] In the GR-IVE separation filter calculation method, when the target sound to be separated and extracted is one sound source (k=1), the above equations (1) to (7) are written as equations (10) to (12). Here, the separation filter W is a regular matrix with real coefficients. １ is a separation filter for extracting the target sound source. ２ is a separation filter that extracts sounds other than the target sound source. ｎ W ２ , V ｘ V ２ , b １ =η^a １ is represented as b.

[0014] The signal processing device 1 of the present disclosure efficiently optimizes the separation filter w(ω) by globally solving the optimization problem expressed by the above equations (10) to (12).

[0015] As shown in FIG. 1 , the signal processing device 1 includes, for example, a filter initialization unit 10, a first covariance matrix generation unit 20, a separated signal calculation unit 30, a second covariance matrix generation unit 40, a filter update unit 50, and a separated signal output unit 60. The signal processing device 1 is a device that receives as input a mixed signal x(ω, τ) that is an observation signal collected by d microphones, an estimated value ^a(ω) of a steering vector of one target sound source that is the target of extraction, and a regularization parameter η, and extracts and outputs one separated signal y(ω, τ) that is the target sound source. Note that the estimated value ^a(ω) of the steering vector of the target sound source is auxiliary information, and this auxiliary information may be an estimated value of the sound source direction of the target sound source. In the present disclosure, both the first embodiment and the second embodiment will be described by unifying the auxiliary information into the estimated value ^a(ω) of the steering vector of the target sound source.

[0016] The signal processing device 1 performs the signal processing method of this embodiment by carrying out the processing flow shown in FIG.

[0017] In this disclosure, the mixed signal x(ω, τ) is a d-dimensional complex vector, where ω (ω=1,...,N ω ) is the frequency bin. τ (τ=1,...,N τ) is a time frame index. The estimated value ^a(ω) of the steering vector of the target sound source is a d-dimensional complex vector. The regularization parameter η is a positive real number. In the following description, it is assumed that the mixed signal x(ω, τ), the estimated value ^a(ω) of the steering vector of the target sound source, and the regularization parameter η are input to the signal processing device 1. In the signal processing device 1, the mixed signal x(ω, τ) is used, for example, by the first covariance matrix generation unit 20, the separated signal calculation unit 30, the second covariance matrix generation unit 40, and the separated signal output unit 60. The estimated value ^a(ω) of the steering vector of the target sound source is used, for example, by the first covariance matrix generation unit 20, the second covariance matrix generation unit 40, and the filter update unit 50. The regularization parameter η is used, for example, by the first covariance matrix generation unit 20, the second covariance matrix generation unit 40, and the filter update unit 50.

[0018] (Filter Initialization Unit 10) The filter initialization unit 10 initializes the separation filter w(ω) used to calculate the separated signals and generate the target sound source (described later) (step S10). The initialization is performed by using e １ is set as the initial value of the separation filter w(ω). Here, e １ is any d-dimensional complex vector, such as the first column vector of a d×d unit matrix. The initialized separation filter w(ω) is output to the separated signal calculation unit 30.

[0019] (First Covariance Matrix Generator 20) The first covariance matrix generator 20 generates a covariance matrix V using the mixed signal x(ω, τ), the estimated value ^a(ω) of the steering vector of the target sound source, and the regularization parameter η. ｘ (ω) is generated (step S20). ｘ (ω) is generated by the weighted sum of two covariance matrices, for example, as shown in Equation (14), which is a modification of Equation (3). ｘ (ω) is the d×d covariance matrix. Here, Rx(ω) is the covariance matrix of the mixed signal x(ω, τ) calculated using equation (7). Ｈis a d×d covariance matrix calculated by the estimated steering vector ^a(ω) of the target sound source. ｘ (ω) is output to the filter update unit 50. Note that the order of the processing of steps S10 and S20 may be reversed, or these steps may be processed in parallel.

[0020] (Separated Signal Calculation Unit 30) The separated signal calculation unit 30 calculates the separated signal ^y(ω,τ) from the mixed signal x(ω,τ) using the separation filter w(ω) (step S30). The calculation of the separated signal ^y(ω,τ) is performed, for example, by the calculation of equation (15). The separated signal ^y(ω, τ) is an estimated signal of the target sound source at this stage and is a one-dimensional complex number. The calculated separated signal ^y(ω, τ) is output to the second covariance matrix generation unit 40.

[0021] (Second covariance matrix generation unit 40) The second covariance matrix generation unit 40 generates a covariance matrix V of a noise signal that becomes noise for the target sound source from the separated signal ^y(ω, τ), the mixed signal x(ω, τ), the estimated value ^a(ω) of the steering vector of the target sound source, and the regularization parameter η. １ (ω) is generated (step S40). １ The generation of (ω) is calculated using, for example, equations (5), (6), and (2), which are modifications of equations (16) to (18). Here, the covariance matrix V １ (ω) is a d×d matrix. R1(ω) is the covariance matrix expressed by equation (16). G'(·) is the derivative of G(·), and G(·) can be expressed by equations (8) and (9), for example. The generated covariance matrix V １ (ω) is output to the filter update unit 50 .

[0022] (Filter Update Unit 50) The filter update unit 50 updates the covariance matrix V ｘ (ω) and covariance matrix V １ (ω), the regularization parameter η, and the estimated value ^a(ω) of the steering vector of the target sound source are used to update the separation filter w(ω) (step S50-1).

[0023] 3 is a diagram showing an example of the processing flow of step S50-1. In step S50-1, the optimum separation filter w(ω) at that stage is calculated for each frequency bin ω by performing the processing of the following steps S50-1-1 to S50-1-7. When calculating equations (10) to (12), the covariance matrix V １ (ω) is V in equation (12) １ and the covariance matrix V ｘ (ω) is V in equation (12) ２ Let η ≡ a(ω) be b in equation (12). Here, V １ , V ２ is a d×d matrix, and b is a d-dimensional complex vector.

[0024] To calculate the optimum separation filter, first, a joint diagonalization matrix P and a diagonal matrix D are calculated (step S50-1-1). P and D are calculated using, for example, equations (20) and (21). where diag(·) is a diagonal matrix and α １ ≦・・・≦α ｄ are sorted in ascending order, and I is the identity matrix.

[0025] Next, as shown in equation (22), P Ｈ Calculate b, － b (step S50-1-2).

[0026] Next, as shown in equation (23), | － b ｉ | ２ Calculate and β ｉ (Step S50-1-3).

[0027] Next, as shown in equations (24) and (25), the undetermined multiplier λ ０ is calculated and set as the initial value of λ (step S50-1-4). where α ｉ , β ｉ ≧0 (i=1,...,d), and α１ is the data of one element of the diagonal matrix D (Equation (20)), and β １ is β in equation (23) i (i=1).

[0028] Next, as shown in equation (26), λ-g(λ) / (g'(λ)) is set as a new λ (step S50-1-5). Here, g(λ) is defined by equation (27).

[0029] Next, if |g(λ)| is greater than ε', which is a predetermined small value (Yes in step S50-1-6), the process returns to step S50-1-5 and repeats the process. On the other hand, if |g(λ)| is equal to or less than ε' (No in step S50-1-6), the separation filter w(ω) is calculated and returned (step S50-1-7). The separation filter w(ω) is calculated, for example, using equation (28). That is, steps S50-1-5 and S50-1-6 find the minimum λ that satisfies the Secular Equation g(λ) = 0. In other words, the calculation of the above-described equations (10) to (12) is reduced to a method of determining λ that is the minimum solution for g(λ) = 0. As a result, the optimum separation filter w(ω) at that stage is calculated for the frequency bin ω being calculated.

[0030] On the other hand, the filter update unit 50 determines whether or not a predetermined repetition condition is met (step S50-2).

[0031] As a result, if the filter update unit 50 determines that the predetermined iteration condition is not satisfied (No in step S50-2), the process returns to step S30 and repeats the process. Examples of the predetermined iteration condition include whether or not the loss coefficient calculated from the calculated separation filter w(ω) and the immediately preceding separation filter w(ω) reaches a predetermined threshold value, whether or not a predetermined number of calculations has been reached, etc.

[0032] If the filter update unit 50 determines that the predetermined iteration condition is satisfied (Yes in step S50-2), the separation filter w(ω) for extracting the target sound source is determined, and the determined separation filter w(ω) is output to the separation signal output unit 60.

[0033] (Separated Signal Output Unit 60) The separated signal output unit 60 generates and outputs a separated signal y(ω,τ) of the target sound source specified by the estimated value ^a(ω) of the steering vector of the target sound source from the mixed signal x(ω,τ) using the determined (updated) separation filter w(ω) (step S60). The separated signal y(ω,τ) is generated by calculating a one-dimensional complex number shown in equation (29), for example.

[0034] The signal processing device 1 of the present disclosure has been described above. The VCD of the conventional method uses d separation filters w １ ,w ２ ,…,w ｄ were updated sequentially. In the signal processing device 1, equations (10) to (12) are calculated using two covariance matrices as in steps S50-1-1 to S50-1-7. That is, according to the above-described signal processing device 1, when there is one target sound source to be separated and extracted, the entire separation matrix can be optimized all at once. Therefore, optimization converges quickly, enabling high-speed sound source separation. For example, one example of a usage scenario is a scene in which one speaker speaks to an audio terminal (e.g., a conversation robot) in a noisy living environment (e.g., a coffee shop). In this case, the signal processing device 1 can be used to remove unnecessary sounds and emphasize the target speaker's voice so that the audio terminal can clearly hear the speech.

[0035] Second Embodiment In a signal processing device 1A according to a second embodiment of the present disclosure, a case will be considered in which the separation filter W(ω) is constrained by the condition of equation (30). Here, U is a unitary matrix. In the GR-IVE separation filter calculation method, if the target sound to be separated and extracted is one sound source (k=1), equations (31) and (32) are solved repeatedly.

[0036] Here, the whitening matrix R ｘ －１／２ Using ～ R １ If we use equation (33) to express the separation filter U(ω), it is sufficient to solve equations (31) and (32) by solving equations (34) and (35). That is, equations (31) and (32) can be reduced to equations (34) and (35). However, R １ is defined by equation (16). Here, u １ is a separation filter for extracting the target sound source. In equations (34) and (35), for simplicity, u １ u, ～ R １ V, ～ When b is expressed as b, the equations (36) and (37) are obtained.

[0037] The signal processing device 1 of the present disclosure efficiently performs the calculations of the above-mentioned equations (36) and (37).

[0038] As shown in FIG. 4 , the signal processing device 1A includes, for example, a whitening unit 11A, an auxiliary information conversion unit 12A, a filter initialization unit 10A, a separated signal calculation unit 30A, a covariance matrix generation unit 40A, a filter update unit 50A, and a separated signal output unit 60A. Like the signal processing device 1, the signal processing device 1A is a device that receives as input a mixed signal x(ω, τ) that is an observation signal, an estimated value ^a(ω) of a steering vector of one target sound source, and a regularization parameter η, and extracts and outputs one separated separated signal y(ω, τ) that is the target sound source. The signal processing device 1A performs the signal processing method of this embodiment by implementing the processing flow shown in FIG. 5 . In the signal processing device 1A, the mixed signal x(ω, τ) is used, for example, in the whitening unit 11A and the auxiliary information conversion unit 12A. The estimated value ^a(ω) of the steering vector of the target sound source is used, for example, in the auxiliary information conversion unit 12A. The regularization parameter η is used, for example, in the auxiliary information conversion unit 12A.

[0039] (Whitening Unit 11A) The whitening unit 11A whitens the mixed signal x(ω, τ) which is the observed signal (step S11A). The whitening is calculated, for example, by using Equation (38), and the whitened mixed signal is ～ Let x(ω,τ). Whitening removes correlation between elements and performs variance normalization. Here, R ｘ (ω)^(-1 / 2) is the whitening matrix. ｘ (ω) is calculated using, for example, equation (7). ～ x(ω, τ) is output to the separated signal calculation unit 30A, the covariance matrix generation unit 40A, and the separated signal output unit 60A.

[0040] (Auxiliary Information Converter 12A) The auxiliary information converter 12A converts the estimated value ^a(ω) of the steering vector of the target sound source into a predetermined format using the mixed signal x(ω, τ) and the regularization parameter η. ～ In the conversion process, the estimated value ^a(ω) of the steering vector of the target sound source is converted into a d-dimensional complex vector, ^a(ω), as shown in equation (39), in accordance with the calculation process of the whitening unit 11A (step S12A). ～ b(ω). The auxiliary information conversion unit 12A obtains the whitening matrix R ｘ It is also possible to receive (ω)^(-1 / 2). Converted ～ b(ω) is output to the filter update unit 50A.

[0041] (Filter Initialization Unit 10A) The filter initialization unit 10A initializes the separation filter u(ω) used to calculate the separated signals and generate the target sound source (described later) (step S10A). The initialization is performed by using e １ is set as the initial value of the separation filter u(ω). Here, e １may be any d-dimensional complex vector, such as the first column vector of a d×d unit matrix. The initialized separation filter u(ω) is output to the separated signal calculation unit 30A. Note that the above-mentioned steps S11A, S12A, and S10 may be processed in reverse order or may be processed in parallel. The auxiliary information conversion unit 12A receives the whitening matrix R from the whitening unit 11A. ｘ Alternatively, (ω)^(−½) may be received. In this case, step S12A is performed after step S11A.

[0042] (Separated Signal Calculation Unit 30A) The separated signal calculation unit 30A calculates the whitened mixed signal. ～ A separation signal ^y(ω,τ) is calculated from x(ω,τ) using a separation filter u(ω) (step S30A). The calculation of the separation signal is performed by, for example, the calculation of equation (41). Here, the separated signal ^y(ω, τ) is an estimated signal of the target sound source at this stage and is a one-dimensional complex vector. The calculated separated signal ^y(ω, τ) is output to the covariance matrix generation unit 40A.

[0043] (Covariance Matrix Generation Unit 40A) The covariance matrix generation unit 40A generates a covariance matrix from the separated signal ^y(ω, τ) and the whitened mixed signal ～ x(ω, τ) and the covariance matrix, which is the covariance matrix of the noise signal that becomes noise with respect to the target sound source signal, ～ R １ (ω) to generate the covariance matrix ～ R １ The generation of (ω) is calculated using, for example, equation (42). where the covariance matrix ～ R １ (ω) is a d×d matrix. r(τ) is calculated as in equation (17). G'(·) is the derivative of G(·), and G(·) can be calculated using equations (8) and (9), for example. The calculated covariance matrix ～ R １ (ω) is output to the filter update unit 50 .

[0044] (Filter Update Unit 50A) The filter update unit 50A updates the covariance matrix～ R １ (ω) and the transformed auxiliary information, a d-dimensional complex vector ～ Using b(ω), the separation filter u(ω) is updated (step S50A-1).

[0045] 6 is a diagram showing an example of the processing flow of step S50A-1. In step S50A-1, the optimum separation filter u(ω) at that stage is calculated for each frequency bin ω by performing the processing of the following steps S50A-1-1 to S50A-1-7. When calculating equations (36) and (37), the covariance matrix ～ R １ (ω) is V, ～ Let b(ω) be b. Here, V is a d×d matrix, and b is a d-dimensional complex vector.

[0046] To calculate the optimum separation filter, first, a joint diagonalization matrix P and a diagonal matrix D are calculated (step S50A-1-1). P and D are calculated using, for example, equation (44). where diag(·) is a diagonal matrix and α １ ≦・・・≦α ｄ are sorted in ascending order.

[0047] Next, as shown in equation (45), P Ｈ Calculate b, － b (step S50A-1-2).

[0048] Next, as shown in equation (46), | － b ｉ | ２ Calculate and β ｉ Let's say. Next, as shown in equation (47), α １ -(β １ )^(1 / 2) is calculated and set as the initial value of λ (step S50A-1-4). α １ is the data of one element of the diagonal matrix (Equation (44)), and β １ is the above-mentioned β i (i=1).

[0049] Next, as shown in equation (48), λ−h(λ) / (h′(λ)) is set as a new λ (step S50A-1-5). Here, h(λ) is defined by equation (49). Next, if |h(λ)| is greater than ε', which is a predetermined small value (Yes in step S50A-1-6), the process returns to step S50A-1-5 and repeats the process. On the other hand, if |h(λ)| is equal to or less than ε' (No in step S50A-1-6), the separation filter u(ω) is calculated and returned (step S50A-1-7). The separation filter u(ω) is calculated, for example, using equation (50). That is, steps S50A-1-5 and S50A-1-6 find the minimum λ that satisfies the Secular Equation h(λ) = 0. As a result, the optimum separation filter u(ω) at that stage is calculated for the frequency bin ω being calculated.

[0050] On the other hand, the filter update unit 50A determines whether or not a predetermined repetition condition is met (step S50A-2).

[0051] As a result, if the filter update unit 50A determines that the predetermined iteration condition is not satisfied (No in step S50A-2), the process returns to step S30A and the process is performed. Examples of the predetermined iteration condition are the same as those in the signal processing device 1, and therefore the description thereof will be omitted.

[0052] If the filter update unit 50A determines that the predetermined iteration condition is satisfied (Yes in step S50A-2), the separation filter u(ω) is determined, and the determined separation filter u(ω) is output to the separation signal output unit 60A.

[0053] (Separated Signal Output Unit 60A) The separated signal output unit 60A generates and outputs a separated signal y(ω,τ) of the target sound source specified by the estimated value ^a(ω) of the steering vector of the target sound source from the mixed signal x(ω,τ) using the determined (updated) separation filter u(ω) (step S60A). The separated signal y(ω,τ) is generated by calculating a one-dimensional complex vector shown in equation (51), for example. The above-described signal processing device 1A can also optimize the entire separation matrix in one go when the target sound to be separated and extracted is one sound source. Therefore, optimization converges quickly, enabling high-speed sound source separation.

[0054] The first and second embodiments of the present disclosure have been described above. Both signal processing devices are signal processing devices that output a separated signal y(ω,τ) of a target sound source from a collected mixed signal x(ω,τ), and include the following: A separated signal calculation unit (separated signal calculation unit 30 or separated signal calculation unit 30A) that calculates a separated signal ^y(ω,τ) from the mixed signal x(ω,τ) using a separation filter (w(ω) or u(ω)). A filter update unit (filter update unit 50 or filter update unit 50A) that updates the separation filter using two covariance matrices, auxiliary information (e.g., an estimated value ^a(ω) of the steering vector of the target sound source), and a predetermined regularization parameter η. A separated signal output unit (separated signal output unit 60 or separated signal output unit 60A) that generates and outputs a separated signal y(ω,τ) of the target sound source specified by the auxiliary information from the mixed signal x(ω,τ) using the updated separation filter (w(ω) or u(ω)).

[0055] Here, the two covariance matrices (here, referred to as the "first covariance matrix" and the "second covariance matrix") used by the filter updating unit 50 and the filter updating unit 50A are as follows. That is, in step S50-1, the filter updating unit 50 updates the separation filter w(ω) using the covariance matrix Vx(ω), the covariance matrix V1(ω), the regularization parameter η, and the estimated value ^a(ω) of the steering vector of the target sound source. Therefore, the two covariance matrices used by the filter updating unit 50 are the covariance matrix V ｘ (ω) and covariance matrix V １ (ω). In accordance with the processing of the signal processing device 1, the first covariance matrix is ​​the covariance matrix V ｘ (ω), and the second covariance matrix is ​​the covariance matrix V １ In step S50A-1, the filter update unit 50A calculates the covariance matrix ～ R １(ω) and the converted signal by the auxiliary information conversion unit 12A. ～ b(ω) is used to update the separation filter u (ω). ～ b(ω) is the covariance matrix R ｘ Therefore, the two covariance matrices used by the filter update unit 50A are the covariance matrix R ｘ (ω) and the covariance matrix R １ (ω). In accordance with the processing of the signal processing device 1A, the first covariance matrix is ​​the covariance matrix R ｘ (ω), and the second covariance matrix is ​​the covariance matrix R １ (ω).

[0056] [Program, Recording Medium] The various processes described above can be implemented by loading a program that executes each step of the above method into the recording unit 2020 of the computer 2000 shown in FIG. 7, and operating the control unit 2010, input unit 2030, output unit 2040, display unit 2050, etc.

[0057] The program describing the processing contents can be recorded on a computer-readable recording medium, which may be, for example, a magnetic recording device, an optical disk, a magneto-optical recording medium, a semiconductor memory, or any other suitable recording medium.

[0058] The program may be distributed by, for example, selling, transferring, lending, etc. portable recording media such as DVDs and CD-ROMs on which the program is recorded. Furthermore, the program may be stored in a storage device of a server computer, and then transferred from the server computer to other computers via a network, thereby distributing the program.

[0059] A computer that executes such a program may first temporarily store the program recorded on a portable recording medium or transferred from a server computer in its own storage device. Then, when executing a process, the computer reads the program stored on its own recording medium and executes the process in accordance with the read program. Alternatively, the computer may read the program directly from a portable recording medium and execute the process in accordance with the program. Furthermore, the computer may execute the process in accordance with each program transferred from the server computer. Alternatively, the server computer may not transfer the program to the computer, but may instead execute the process through a so-called ASP (Application Service Provider) service, which realizes the processing function by issuing an execution instruction and obtaining the results. In this embodiment, the program includes information used for processing by a computer that is equivalent to a program (e.g., data that is not a direct instruction to the computer but has properties that define computer processing).

[0060] Furthermore, in this embodiment, the device is configured by executing a predetermined program on a computer, but at least a part of the processing contents may be realized by hardware.

[0061] 1, 1A Signal processing device 10, 10A Filter initialization unit 11A Whitening unit 12A Side information conversion unit 20 First covariance matrix generation unit 30, 30A Separation signal calculation unit 40 Second covariance matrix generation unit 40A Covariance matrix generation unit 50, 50A Filter update unit 60, 60A Separation signal output unit ^a Estimated value of steering vector of target sound source R ｘ (ω), V １ (ω), V ｘ (ω) Covariance matrix w (ω), u (ω) Separation filter x (ω, τ) Mixed signal y (ω, τ), ^y (ω, τ), Separation signal η Regularization parameter

Claims

1. A signal processing device that outputs a separation signal of one target sound source from a collected mixed signal, comprising: a separation signal calculation unit that calculates the separation signal from the mixed signal using a separation filter; a filter update unit that updates the separation filter using a first covariance matrix, a second covariance matrix, auxiliary information of the target sound source, and specified parameters; and a separation signal output unit that generates and outputs a separation signal of the target sound source specified by the auxiliary information from the mixed signal using the updated separation filter.

2. The signal processing device according to claim 1, further comprising: a first covariance matrix generation unit that generates the first covariance matrix using the mixed signal, the auxiliary information, and the parameters; and a second covariance matrix generation unit that generates the second covariance matrix, a covariance matrix of a noise signal that becomes noise for the target sound source, from the separated signals calculated by the separated signal calculation unit, the mixed signal, the auxiliary information, and the parameters.

3. The signal processing device according to claim 1, further comprising: a whitening unit that whitens the mixed signal using the first covariance matrix; an auxiliary information conversion unit that converts the auxiliary information into a predetermined format using the mixed signal and predetermined parameters; and a covariance matrix generation unit that generates the second covariance matrix, which is a covariance matrix of a noise signal that becomes noise for the target sound source, using the separation signals calculated by the separation signal calculation unit and the whitened mixed signal, wherein the separation signal calculation unit calculates the separation signals from the whitened mixed signal using a separation filter, and the filter update unit updates the separation filter using the generated covariance matrix and the auxiliary information converted by the auxiliary information conversion unit.

4. A signal processing method for outputting a separation signal of one target sound source from a collected mixed signal, comprising the steps of: calculating the separation signal from the mixed signal using a separation filter; updating the separation filter using a first covariance matrix, a second covariance matrix, auxiliary information of the target sound source, and specified parameters; and generating and outputting a separation signal of the target sound source specified by the auxiliary information from the mixed signal using the updated separation filter.

Images