JP2011170190A - Device, method and program for signal separation - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a signal separation device that separates a signal with high accuracy, from composite signals containing a target signal and an interference signal. <P>SOLUTION: The signal separation device calculates an output signal parameter of a conventional Wiener filter of a time frequency domain in each short time frame, based on a time frequency component sequence calculated from the composite signals containing the target and interference signals, and on a time frequency amplitude component sequence for every source signal estimated. A weighted distance between the time frequency component sequence of a time series signal and the output signal parameter, which are obtained when restoring a separation signal parameter of time frequency domain for estimating the signals of interest by an inverse short-time Fourier transform, is used as a modified Wiener criterion. The signal separation device estimates the separation signal parameter value so that the modified Wiener criterion becomes smaller than a value when the separation signal parameter in the modified Wiener criterion is replaced with the output signal parameter from the input signal. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a method for separating and extracting individual signals in a situation where an estimated value of a time-frequency amplitude component array of each signal and a mixed signal thereof are given, and relates to, for example, speech enhancement and noise suppression.

Conventionally, regarding signal separation, various separation methods based on a time-frequency representation of a signal or a representation obtained by dividing a signal into short-time frames are widely known.
The separated signal parameters obtained as a result of the separation are in the time-frequency domain. However, as a result of calculation in the time-frequency domain, the combination of the separated signal parameters is a combination that cannot actually exist as a signal in the time domain. There is a problem that a solution corresponding to a signal in the time domain by some kind of post-processing does not necessarily give the most likely separation result.
For example, Wiener filtering (for example, Non-Patent Document 1) is known as a separation method for the purpose of separating individual signals when an estimated value of a time-frequency amplitude component array of each signal and a mixed signal thereof are given. .

EJ Diethorn, "Subband noise reduction methods for speech enhancement," in Audio Signal Processing for Next-Generation Multimedia Communication Systems, Y. Huang and J. Benesty, Eds., Pp. 91--115, Boston, MA: Kluwer, 2004 .

  However, in the technique of Non-Patent Document 1, there is a possibility that the time-frequency component array estimated in the time-frequency domain may be a combination that cannot actually exist as a signal in the time domain, and an accurate signal separation result cannot be obtained. There was a problem that there was.

  The present invention has been made in view of the above problems, and a signal separation device, a signal separation method, and a signal separation method capable of performing signal separation processing with high accuracy from a mixed signal of a target signal and an interference signal, and It is to provide a signal separation program.

  The present invention has been made to solve the above-described problem, and the invention according to claim 1 is a time-frequency analysis unit that calculates a time-frequency component from an input signal that is a mixed signal of a target signal and an interference signal; A time frequency amplitude component array estimation unit that estimates a time frequency amplitude component array for each source signal of the target signal and interference signal included in the input signal from the input signal, and calculated from the input signal A Wiener filter that calculates output signal parameters of a conventional Wiener filter that is a Wiener filter in the time-frequency domain in each short-time frame based on the time-frequency component and the estimated value of the time-frequency amplitude component array for each source signal. An output signal parameter calculation unit; and a signal parameter in a time-frequency domain for estimating the target signal as a separation signal parameter; The weighted distance between the time-frequency component array of the time-series signal when the parameter is restored by inverse short-time Fourier transform and the output signal parameter of the conventional Wiener filter is a modified Wiener criterion, and the separated signal parameter in the modified Wiener criterion is A modified Wiener criterion reducing unit that estimates the separated signal parameter value from the input signal such that the modified Wiener criterion is smaller than a value when the output signal parameter of the conventional Wiener filter is replaced, and the modified Wiener criterion And a separation signal parameter output unit that outputs the separation signal parameter value estimated by the reduction unit.

  The invention according to claim 2 is characterized in that the modified Wiener criterion reducing unit includes a block dividing unit that divides the conventional Wiener filter output signal parameter into blocks having a length of a plurality of frames, and the conventional Wiener filter divided into the blocks. For a plurality of frames of output signal parameters, an intra-block separated signal parameter calculation unit for calculating a separated signal parameter value in the block, so that the modified Wiener criterion is smaller than in the case of the conventional Wiener filter, The signal separation apparatus according to claim 1, further comprising: an entire separated signal parameter calculation unit that calculates a separated signal parameter for the entire block from the separated signal parameter value.

  According to a third aspect of the present invention, the modified Wiener criterion reducing unit includes an information storage unit that stores the conventional Wiener filter output signal parameter and the time frequency amplitude component array estimated value, and an initial value for the separated signal parameter. A separation signal parameter initial value generation unit for generating a separation signal parameter convergence determination unit for determining whether or not the separation signal parameter value has converged, a separation signal parameter update unit for updating the separation signal parameter value, The signal separation device according to claim 1, further comprising: a separation signal parameter output unit that outputs the separation signal parameter value estimated by the separation signal parameter update unit.

  According to a fourth aspect of the present invention, the separated signal parameter update unit calculates a consistent time frequency component array calculation unit that calculates a consistent time frequency component array with a consistent time frequency component array of the signal for the separated signal parameter. A time-frequency component array weighted sum calculation unit that calculates a time-frequency component array weighted sum having the consistent time-frequency component array, the separated signal parameter, and the conventional Wiener filter output signal parameter as elements, The signal separation device according to claim 3, further comprising:

  The invention according to claim 5 is a time-frequency analysis process for calculating a time-frequency component array from an input signal which is a mixed signal of a target signal and an interference signal, and the target included in the input signal from the input signal. A time frequency amplitude component array estimating process for estimating a time frequency amplitude component array for each source signal of a signal and an interference signal, a time frequency component array calculated from the input signal, and a time frequency amplitude component for each source signal In order to estimate the target signal, a Wiener filter output signal parameter calculation process for calculating an output signal parameter of a conventional Wiener filter that is a Wiener filter in the time-frequency domain in each short-time frame, based on the estimated value of the array The signal parameter in the time-frequency domain is set as a separation signal parameter, and the separation signal parameter is obtained by inverse short-time Fourier transform. The weighted distance between the time-frequency component array of the time-series signal and the output signal parameter of the conventional Wiener filter is a corrected Wiener criterion, and the separation signal parameter in the corrected Wiener criterion is the output signal of the conventional Wiener filter. A modified Wiener criterion decreasing process for estimating the separated signal parameter value from the input signal so that the modified Wiener criterion is smaller than a value when the parameter is replaced with a parameter, and the separation estimated in the modified Wiener criterion decreasing process And a separation signal parameter output process for outputting a signal parameter value.

  In the invention according to claim 6, the modified Wiener criterion decreasing process further includes a block dividing process of dividing the conventional Wiener filter output signal parameter into blocks each having a length of a plurality of frames, and the conventional technique divided into the blocks. An intra-block separated signal parameter calculation process for calculating a separated signal parameter value in the block so that the modified Wiener criterion is reduced as compared with the conventional Wiener filter for a plurality of frames of Wiener filter output signal parameters; 6. The signal separation method according to claim 5, further comprising: a total separation signal parameter calculation step of calculating a separation signal parameter for the entire block from a separation signal parameter value in the block.

  In the invention according to claim 7, the modified Wiener criterion reduction process further includes an information storage process for storing the conventional Wiener filter output signal parameter and the time frequency amplitude component array estimated value, and the separated signal parameter. A separation signal parameter initial value generation process for generating an initial value, a separation signal parameter convergence determination process for determining whether or not the separation signal parameter value has converged, and a separation signal parameter update process for updating the separation signal parameter value 6. The signal separation method according to claim 5, further comprising: a separated signal parameter output step of outputting the separated signal parameter value estimated in the separated signal parameter update step.

  According to an eighth aspect of the present invention, in the separated signal parameter update process, a consistent time frequency component array for calculating a consistent time frequency component array in which the time frequency component array of the signal is consistent for the separated signal parameter is further calculated. Time-frequency component array weighted sum calculation process for calculating a time-frequency component array weighted sum including the calculation process, the consistent time-frequency component array, the separated signal parameter, and the conventional Wiener filter output signal parameter The signal separation method according to claim 7, further comprising:

  According to the ninth aspect of the present invention, there is provided a computer as a signal separation device, a time frequency analysis process for calculating a time frequency component array from an input signal that is a mixed signal of a target signal and an interference signal, and the input signal from the input signal. A time frequency amplitude component array estimation process for estimating a time frequency amplitude component array for each source signal of the target signal and interference signal included in the signal, a time frequency component array calculated from the input signal, and the respective Wiener filter output signal parameter calculation process for calculating output signal parameters of a conventional Wiener filter that is a Wiener filter in the time-frequency domain in each short-time frame based on the estimated value of the time-frequency amplitude component array for each source signal; A signal parameter in a time-frequency domain for estimating the target signal is a separated signal parameter, and the separated signal parameter is Using the weighted distance between the time-frequency component array of the time-series signal and the output signal parameter of the conventional Wiener filter when the meter is restored by inverse short-time Fourier transform as the modified Wiener criterion, the separated signal parameter in the modified Wiener criterion is A modified Wiener criterion reduction process for estimating the separated signal parameter value from the input signal so that the modified Wiener criterion is smaller than a value when the output signal parameter of the conventional Wiener filter is replaced, and the modified Wiener criterion And a separation signal parameter output process for outputting the separation signal parameter value estimated in the reduction process.

  According to the present invention, in parameter estimation, more accurate estimation is performed using the restriction that the time-frequency component arrangement of the signal is consistent. Thereby, this invention can perform a signal separation process with high precision from the mixed signal of an object signal and an interference signal. In addition, the present invention can perform signal separation at the same speed as in the prior art.

It is explanatory drawing explaining the relationship between a time domain and a time frequency domain. It is a block diagram which shows an example of a structure of the signal separation apparatus 1 by one Embodiment of this invention. FIG. 3 is a block diagram illustrating a first example of a configuration of a modified Wiener criterion reducing unit 40 in FIG. 2. It is a block diagram which shows the 2nd example of a structure of the correction Wiener reference | standard reduction part 40 of FIG. FIG. 5 is a block diagram illustrating an example of a configuration of a separated signal parameter update unit 431 in FIG. 4. It is a figure which shows an example of the spectrogram of the speech signal in noise. It is a figure which shows an example of the spectrogram of the conventional Wiener filter output signal. It is a figure which shows an example of the spectrogram of the signal obtained by the method based on an auxiliary function method.

<Principle>
As a known method, a signal separation method called “Wiener filtering” is known. Here, the formulation of Wiener filtering is performed as an optimization problem of the likelihood function in the time frequency domain.

It is assumed that the observation signal (input signal) x is a mixed signal composed of two signals of the target signal s ₁ and the interference signal s ₂ . Also, consider calculating the temporal frequency component array of the observation signal by short-time Fourier transform (hereinafter also referred to as STFT) of frame shift R. Furthermore, it at each time frame t and frequency bin omega, signal s ₁ short-time Fourier transform coefficients S ₁ of the s ₂ and S ₂ are each mean 0 and variance sigma ₁ ² and sigma ₂ ² independent Gaussian variables Assuming For convenience, ν ⁽ⁱ⁾ = 1 / σ _i ² is written below. Here, it is assumed that the observation signal is a mixed signal composed of two signals, but the principle to be described is generally valid even in the case of a mixed signal composed of I signals.

If there are multiple interfering signals s ₂ , ..., s _I , it is assumed that the sound sources are uncorrelated with each other, so it can be restored when there is only one interfering signal without losing generality. it can. That is, s ₂ ,..., S _I can be handled as one global interference signal ~ s ₂ as the following equation (1).

Note that the description “˜s ₂ ” indicates that the symbol “˜” is added on the variable “s ₂ ” as shown in the following equation (1). The same applies to the description such as “^” described later.

Similarly to the above, it can be said that the short-time Fourier transform coefficient of ~ s ₂ is an average of 0 and an independent Gaussian variable with variance (~ σ ₂ ) ² according to the following equation (2).

The short-time Fourier transform time-frequency component array of the observation signal is called X. Conventional Wiener filtering maximizes the log-likelihood function of STFT coefficients S ₁ and S ₂ . Under the constraint of X = S ₁ + S ₂ , the likelihood function can be written as a function of S = S ₁ as in the following equation (3).

The estimated value of conventional Wiener filtering for S ₁ is defined as the following equation (4).

  Also, if the sign of the log likelihood function (3) is reversed and constants that do not depend on S are omitted, the likelihood function maximization problem is defined as the objective function minimization problem defined by the following equation (5): Can be handled.

However, α _{ω, t} is the following equation (6).

Without any constraints on S, it is clear that the minimum solution of the objective function (5) is S = ^ S. However, it should be noted here that the short-time Fourier transform is a redundant expression with a special structure. The number of frequency bins is N, and the number of frames is T, a short time is obtained from the time series signal Fourier transform time-frequency component sequence is (are C set of all complex numbers) C ^NT is an element of all the C ^NT Elements are not obtained in that form (eg, Non-Patent Document 2: DW Griffin and JS Lim, “Signal estimation from modified short-time Fourier transform,” IEEE Trans. Acoustics, Speech, and Signal Processing, vol. 32 , no.2, pp.236-243, Apr. 1984 and Non-Patent Document 3: J. Le Roux, N. Ono and S. Sagayama, “Explicit consistency constraints for STFT spectrograms and their application to phase reconstruction,” in Proc See SAPA, Sep. 2008).

A short-time Fourier transform time-frequency component array obtained from a time-series signal is called a consistent time-frequency component array. Here, the necessary and sufficient conditions for a consistent time frequency component array with respect to the array of complex numbers in the time frequency domain will be described. Assuming that the inverse STFT is performed so that all time series signals can be completely reconstructed from the STFT time-frequency component array by the inverse STFT, an array W is a consistent time-frequency component array A necessary and sufficient condition is that the STFT of the inverse STFT of W is W itself. That is, a set of consistent time frequency component sequence is zero space defined by C R (the set of all R is a real number) in the ^NT linear mapping F as: (7).

  However, the mapping G in the time frequency domain is defined as the following equation (8).

  Here, STFT (W) represents W STFT, and iSTFT (W) represents W inverse STFT. The relationship between the time domain and the time frequency domain is shown in FIG.

  The method for calculating the inverse STFT will be described in more detail. The inverse STFT of W generally multiplies the synthesized window function s in the inverse Fourier transform of each frame of W, overlaps and adds the obtained short-time signals, and calculates the time series signals of the entire interval. Since the inverse STFT of the STFT time-frequency component array X of a certain time-series signal x is equal to x itself (complete reconstruction is established), normalization must be performed. When the frame shift of STFT is called R, the normalization term is expressed by the following equation (9), and it can be seen that it is a periodic function of period R.

  Where l is a time-series index. Therefore, normalization may be performed after overlap addition, but a composite window function with normalization term ~ s is defined, and the composite window function with normalization term directly in the windowing before overlap addition is performed. Use ~ s. In this way, complete reconstruction can be achieved only by performing overlap add after that, and normalization is not necessary.

  Note that there can generally be a plurality of composite window functions that can be completely reconstructed in a certain analysis window function. However, as explained in Non-Patent Document 2, a time-series signal obtained from a complex array W 1 has a time-frequency component array closest to that array W in the sense of least square error. In order to ensure that, it can be assumed that the composite window function s before entering the normalization term is regarded as s = w. In other words, assuming s = w, G (W) is a consistent time-frequency component array closest to the complex array W in the sense of least square error. Here, s = w is assumed.

  As described above, if no constraint is applied to S, the minimum solution of the objective function (5) is S = ^ S. However, the minimum of ψ so that the solution is essentially a consistent time-frequency component array. Should be performed under the constraint of F (S) = 0.

  Next, the reason why it is necessary to “search for a solution from consistent time-frequency component arrays” will be described. Generally, conventional Wiener filter output signal parameters are not always consistent. The STFT time-frequency component array of a signal obtained by inverse STFT from a non-consistent time-frequency component array is different from the original time-frequency component array. Therefore, when the conventional Wiener filter output signal parameter that should be the minimum solution is not consistent, the time-frequency component array of the time-domain signal obtained therefrom is not actually the minimum solution of ψ. That is, the actually heard signal obtained from the conventional Wiener filter output signal parameters does not generally maximize the Wiener likelihood function. What we actually want to estimate is a time-series signal whose STFT time-frequency component array minimizes the Wiener criterion ψ, and if it is formulated in the time-frequency domain, it is defined as It is the minimum solution of the “standard”.

  If S is constrained to be consistent, the modified Wiener criterion ~ ψ (S) and the conventional Wiener criterion ψ (S) are equivalent, so the minimization problem of ~ ψ is subject to the constraint of F (S) = 0. It can be considered as a problem of minimizing ψ in (Policy 1). It can also be considered as a problem of minimizing ψ (STFT (s)) directly in the time domain using the time series signal s as a parameter (policy 2). It is also possible to relax the consistency constraint and introduce an objective function as a penalty function (policy 3). If the weight of the consistent penalty function is sufficiently large or gradually increased during the optimization process, the finally estimated time-frequency component array is consistent, and ψ is set within the consistent time-frequency component array. Minimize.

  The three optimization policies described above will be specifically described. First, the above-described policy 2 will be described.

<Policy 2: Optimization using time-series signals as parameters>
First, a method for directly minimizing ψ (STFT (s)) in the time domain using the time series signal s as a parameter will be described. Return ψ (STFT (s)) to the time domain by Parseval's theorem. Here, the weighting factor (α _{ω, t} ) _, where N × N diagonal matrix with diagonal coefficients as _{ω = 0, ..., N-1} is A _t , N × N Fourier is F, and the length of signal x is L Let W _{t be} the N × L matrix that calculates the t-th frame multiplied by the window from the signal, and let ^ s _t be the inverse Fourier transform of the t-th STFT frame of ^ S. Then, the objective function ψ (S) in the equation (5) can be rewritten as the following equation (11).

  A signal s such that the derivative of ψ (STFT (s)) with respect to s becomes 0 satisfies a linear equation such as the following equation (12).

The right side of equation (12) shows the inverse Fourier after applying the weighting factor (α _{ω, t} ) _{ω = 0, ..., N-1} for each frequency in the Fourier transform domain to the output signal ^ s of the normal Wiener filter Represents an operation to transform and further window.

Alternatively, the right side of Equation (12) is obtained by convolving an inverse Fourier transform of weighting factors (α _{ω, t} ) _{ω = 0} ,. Represents a windowing operation.

  Then, the signal s that minimizes ψ (STFT (s)) can be analytically calculated by the following equation (13).

Without the above equation in matrix A _t, as described in Non-Patent Document 2, it is readily solutions calculations, with the solution is conventional in the sense of least square error Wiener filter output signal parameters ^ S one It becomes a time-series signal having the closest time-frequency component arrangement. Relative In this problem which, can obtain an optimum signal s by the influence of the weighting coefficient matrix A _t, the problem of finding a solution of the linear equations following formula (14), by huge matrix of size L × L.

  However, since the matrix of the following formula (15) is actually a Hermitian band matrix, it can be calculated with a calculation time and a memory amount within a possible range.

  With regard to the amount of memory in particular, in order to calculate the whole signal at once, if the memory is insufficient, the signal is divided into overlapping blocks containing several frames and optimized within each block. If the optimal solution in the block is combined last, the calculation can be performed with a small amount of memory. In this case, there is a possibility that the obtained solution is different from the optimum solution of the whole signal in order to perform block division and superimposition of the optimum signal in the block, but it is experimentally confirmed that the same result can be obtained by adjusting the overlap. Recognize.

<Policy 1: Optimization by Douglas-Rachford splitting>
Policy 1 considers the minimization problem of ~ ψ as a minimization problem of φ under the constraint of F (S) = 0. As described above, inverse STFT is defined so that a time series signal obtained from an array of complex numbers is a time series signal having a time frequency component array closest to the array in the sense of least square error. Assume. here,

Write. However, i _{Ker (F)} is an indicator function of Ker (F) and is defined by the following equation (17).

It can be easily seen that finding a consistent time-frequency component array S that minimizes ψ is equivalent to finding a minimum solution of f ₁ + f ₂ . Both f ₁ and f ₂ are proper lower semi-continuous convex functions. The minimization problem of f ₁ + f _{2 in} that case has been studied by convex optimization theory, and it is known that it can be solved efficiently by an algorithm called Douglas-Rachford splitting method for monotone operators ( For details, see Non-Patent Document 4: PL Combettes and J.-C. Pesquet, “A Douglas-Rachford splitting approach to nonsmooth convex variational signal recovery,” IEEE Journal of Selected Topics in Signal Processing, vol.1, no.4. , pp.564-574, Dec. 2007). A function of variable Z for each X∈S

Has a unique minimum solution, which is called prox _fi (X). Thus defined monovalent operator to (uniquely-valued operator) is referred to as a proximity operator of f _i. In this problem setting, the probability operators for βf ₁ and βf ₂ can be calculated specifically as in the following equations (19) and (20), respectively. However, β> 0 is a positive constant used in the future.

  Equation (19) can be easily derived by minimizing the quadratic function. For the above equation (20),

Minimum solution can be derived by always the ‖ZX‖ ² is an element of the minimum to Ker (F), that is, consistent time frequency component arranged closest to X in the sense of minimum squared error.
When the Douglas-Rachford splitting method is applied to this problem, the following optimization algorithm is obtained. The initial time-frequency component array is an arbitrary S ⁽⁰⁾ ∈ S, an arbitrary positive constant β> 0, and (0,2) such that Σ _p λ _p (2-λ _p ) = + ∞. If an arbitrary sequence (λ _p ) _p∈N is set in advance and a recurrence rule is defined as in the following equation (22), the sequence S ^(p) converges, and the convergence value ^V S is corrected Wiener It can be proved that this is the minimum standard solution and that G ( ^V S) is a consistent time-frequency component array that minimizes ψ.

Here, if λ _p = 1 and rewrite as γ = 1 / (2β) as the special case, the update equation (22) can be rewritten as follows.

  As will be described below, the above update formula is very similar to the update formula obtained by the formulation in which the penalty function described later is introduced into the objective function.

  In the above iterative calculation, G (S) for a certain S may be strictly calculated by inverse STFT and STFT. For speeding up, approximate calculation may be performed in the frequency domain instead of G (S). An approximate calculation will be described. The linear map G can be calculated directly in the time-frequency domain (see, for example, Non-Patent Document 3). The STFT analysis window function is called w, and the composite window function with the inverse STFT normalization term is called ~ s. Here, since it is assumed that the composite window function s before entering the normalization term is regarded as s = w, ~ s is defined by the following equation (24).

  When the calculation performed by the inverse STFT and STFT is specifically written, the expression of the mapping G in the time-frequency domain is obtained by the following equation (25).

However, Q = N / R is the overlap rate of STFT frames. The weighting coefficient η _{q, p} is defined by the following equation (26). Further, the index of the frequency bin of S is assumed to be mod N.

As explained in Non-Patent Document 3, the main contribution of the weighting coefficient η is concentrated on the coefficient with a small index ω ′ (for example, −2 ≦ ω ′ ≦ 2). Even if the range of the sum of is narrowed to | ω ′ | ≦ l << N / 2, the result of the calculation hardly changes. Therefore, the calculation of G (S ^(p) ) in the update equation (23) at step p + 1 may be approximately replaced with the following equation (27).

  Further, in the calculation of Expression (27), the number of multiplications can be further reduced by using various symmetries of the weighting coefficient η. Further, all the time frequency bins may be updated. Alternatively, the update may be calculated for each bin according to some criterion or may be selectively updated to keep the value of the previous step. For example, a certain threshold value may be determined, and only the bin of the time frequency amplitude component array larger than the threshold value may be updated. Further, the threshold value may be fixed or may be changed for each repetitive step.

<Policy 3: Optimization by auxiliary function method>
The consistency of an array S of complex numbers can be evaluated numerically by using an arbitrary vector norm of F (S), not the strict constraint of F (S) = 0. Here, the L ² norm of F (S) is used. Also, as described above, inverse STFT is defined so that the time series signal obtained from an array of a complex number is a time series signal having a time frequency component array closest to that array in the sense of least square error. Assuming that The consistency penalty function defined as such is newly introduced into the objective function as shown in the following equation (28).

However, γ represents the weighting coefficient of the consistency term of the objective function. When γ is fixed to 0 from beginning to end, a method and apparatus equivalent to the known Wiener filtering is obtained. In addition, γ may be a fixed coefficient that is not 0 from the beginning, but the value of γ may be arbitrarily changed for each process of iterative calculation. If γ is gradually increased in the iterative calculation process, it can be expected that a minimum solution of the Wiener criterion that gradually takes into account consistency is obtained. Here, in order to show a specific calculation method of the signal separation parameter, an auxiliary function for the newly introduced consistency term is introduced. For details on the auxiliary function method, see Non-Patent Document 5 (Non-Patent Document 5: DD Lee and HS Seung, “Learning of the parts of objects by non-negative matrix factorization,” Nature, vol. 401, pp. 788-791, 1999). Here, the concept will be briefly explained. If you want to minimize an objective function f (θ) by θ, f ⁺ (θ, ￣θ) that satisfies the condition of the following equation (29) is called an auxiliary function of f (θ), and ￣θ is Called auxiliary variable.

  If the minimization of f + with respect to θ and ￣θ can be performed analytically, respectively, θ and そ れ ぞ れ θ can be updated alternately as shown in the following equations (30) and (31), respectively. It can be proved that (θ) converges without increasing.

The concept is applied to ψ _γ (S). From the assumption about the inverse STFT described above, it can be seen that G (S) is a consistent time-frequency component array closest to S in the sense of least square error with respect to an array S of arbitrary complex numbers. That is, the following equation (32) is satisfied.

From the results of equation (32), the following equation is defined by the following equation (33) is a function of (34), it can be seen that an auxiliary function of [psi _gamma.

However, C is a constant that depends only on ν ₁ and ν ₂ .

From the above discussion, iterative optimization algorithms can be designed. If we start with the array of complex numbers S ^(p) , which is step p, we first update ￣S to G (S ^(p) ) and then update the updated value of S ^{(p + 1)} to the diagonal of the coefficients. It can be calculated as a minimum solution of the form In addition, S ^(p) can be calculated iteratively by an update equation such as the following equation (35).

The update equation of Equation (35) is different from the update equation of Equation (23) derived by the Douglas-Rachford splitting method in that F (S ^(p) ) is eliminated, but is very similar. It is worth noting. As described in the description of Douglas-Rachford splitting, in the above iterative calculation, G (S) may be strictly calculated for a certain S by inverse STFT and STFT. For speeding up, approximate calculation may be performed in the frequency domain instead of G (S). Specifically, the calculation of G (S ^(p) ) in the update formula (35) at step p + 1 may be approximately replaced with formula (27). Further, as described in the description of Douglas-Rachford splitting, reduction of the number of multiplications due to symmetry of weighting factors, selective update of time frequency bins, and the like may be performed.

<Example 1>
A signal separation device according to an embodiment of the present invention will be described. FIG. 2 is a functional block diagram of the signal separation device 1. The first embodiment is configured to perform signal separation in accordance with the “policy 2”.

  As shown in FIG. 2, the signal separation device 1 includes a time frequency analysis unit 10, a time frequency amplitude component array estimation unit 20, a conventional Wiener filter output signal parameter calculation unit 30, a modified Wiener criterion reduction unit 40, and a separation signal parameter output. Part 50 is provided.

  Further, as shown in FIG. 3, the modified Wiener criterion reducing unit 40 includes a block dividing unit 400, an intra-block separated signal parameter calculating unit 410, and an entire separated signal parameter calculating unit 420.

The time frequency analysis unit 10 _obtains time frequency components X _{ω, t} (ω = 0,..., N−1, t = 0,. The index corresponding to the frequency and time is calculated and output to the conventional Wiener filter output signal parameter calculation unit 30. More specifically, the time-frequency analysis unit 10 receives the time-series signal x and performs time-frequency analysis using a short-time Fourier transform (STFT), thereby performing the time-frequency component _{Xω, t.} And a matrix X = (X _{ω, t} ) _{N × T} storing the time frequency components X _{ω, t} is output.

  The time-frequency amplitude component array estimation unit 20 calculates an estimated value of the time-frequency amplitude component array of each source signal (target signal and interference signal) from the input signal that is a mixed signal of the target signal and interference signal. Output to the filter output signal parameter calculation unit 30.

  This process can be realized by a known technique. For example, the time-frequency amplitude component array estimation unit 20 performs non-negative matrix factorization on the input (see, for example, Non-Patent Document 5), and converts the time-frequency amplitude component array of the input signal into the time frequency of several elements. Decompose into amplitude component array. Further, it may be considered that the oracle information of the estimated value of the time frequency amplitude component array of the target signal and the interference signal is given in advance.

  Alternatively, a case may be considered in which only one of the amplitude information is given instead of the time frequency amplitude component array estimated values of both the target signal and the interference signal. For example, if an estimated value of the time-frequency amplitude component array of an interference signal is given, spectral subtraction (for example, Non-Patent Document 6: SF Boll, “Suppression of acoustic noise in speech using spectral subtraction,” IEEE Trans. Acoustics , Speech, and Signal Processing, vol.27, pp.113-120, Apr. 1979) and estimating the time-frequency amplitude component array of the target signal and outputting both to the conventional Wiener filter output signal parameter calculation unit 30 You may think.

  When the input signal includes a speech signal, all-pole modeling is performed on the input signal (for example, Non-Patent Document 7: JS Lim and AV Oppenheim, “All-pole modeling of degraded speech,” IEEE Trans. Acoustics, Speech, and Signal Processing, vol. 26, pp. 197-210, Jun. 1978), and the time frequency amplitude component of the interference signal by spectral subtraction using the estimated audio time frequency amplitude component array It may be considered that the arrangement is estimated and both are output to the conventional Wiener filter output signal parameter calculation unit 30.

The conventional Wiener filter output signal parameter calculation unit 30 includes the time frequency component array of the input signal calculated by the time frequency analysis unit 10 and the time frequency of each signal given by the time frequency amplitude component array estimation unit 20. The output signal parameter of the conventional Wiener filter is calculated using the amplitude component array estimated value. More specifically, the conventional Wiener filter output signal parameter calculation unit 30 includes the time frequency component array ( _{Xω, t} ) of the input signal obtained by the time frequency analysis unit 10 and the time frequency amplitude component array estimation unit 20. The output signal parameter (^ S _{ω, t} ) of the conventional Wiener filter is calculated by the following equation (36) using the time frequency amplitude component array estimated value σ _i ^{2 of} each signal given by

However,

It is.

The modified Wiener criterion reduction unit 40 is provided by the output signal parameter ^ S _{ω, t of the} conventional Wiener filter calculated by the conventional Wiener filter output signal parameter calculation unit 30 and the time frequency amplitude component array estimation unit 20. Using a time frequency amplitude component array estimated value of each signal, a signal parameter that further lowers the modified Wiener criterion than the output signal parameter of the conventional Wiener filter is estimated as a separated signal parameter value that is a time domain signal parameter. .

In other words, the modified Wiener criterion reduction unit 40 uses the input signal to obtain a separation signal parameter value that makes the modified Wiener criterion smaller than the value when the separation signal parameter in the modified Wiener criterion is replaced with the output signal parameter of the conventional Wiener filter. presume.
For example, the modified Wiener criterion reduction unit 40 inserts the output signal parameter (^ S) of the conventional Wiener filter into the part where the separation parameter of the modified Wiener criterion is entered (S of G (S) ω, t in Equation 10). Rather, the separation signal parameter value such that the value of the modified Wiener criterion is smaller is estimated from the input signal.

Here, the signal parameter in the time-frequency domain for estimating the target signal is a separated signal parameter. The weighted distance between the time-frequency component array of the time-series signal and the output signal parameter of the conventional Wiener filter when the separated signal parameter is restored by inverse short-time Fourier transform is the modified Wiener criterion.
In addition, “the time-frequency component array of the time-series signal when the separated signal parameter is restored by inverse short-time Fourier transform” here means “the time-frequency component array of the time-series signal when the separated signal parameter is restored”. Or “a time-frequency component array of a time-series signal when the separated signal parameters are restored by inverse transformation of time-frequency analysis”.

  The separated signal parameter output unit 50 outputs the separated signal parameter value estimated by the modified Wiener criterion reducing unit 40. The separated signal parameter value is a signal parameter in the time frequency domain for separating the target signal to be estimated from the input signal. By performing inverse STFT on the separated signal parameter value, an estimated value of the target signal included in the input signal can be obtained.

Next, an example of the configuration of the modified Wiener criterion reduction unit 40 will be described in detail with reference to FIG.
The block dividing unit 400 divides the conventional Wiener filter output signal parameter into blocks of a plurality of frames. A small number of frames may be combined into one block so that the blocks overlap. Further, the number of frames included in each block may be fixed, or the number of frames may be different for each block. The same applies to the number of overlapping frames. Here, the index of the block is called b, the number of blocks is called B, the number of frames contained in one block is assumed to be constant, the number is called n _B, and the number of overlapping frames of adjacent blocks _Is called nO. The separated signal in the block is called s ^(b) . The first index of the frame corresponding to the b-th block is called n _b, and the frame index corresponding to the block is t = n _b , ..., n _{b + 1} + n _O −1. Also, using the estimated value of the time-frequency amplitude component array of each signal, the weighting coefficient is calculated as in the following equation (38).

The intra-block separation signal parameter calculation unit 410 calculates a separation signal s ^(b) that is a time-series signal that minimizes the modified Wiener criterion in the block b divided by the block division unit 400, and separates each intra-block separation Output as a signal parameter. More specifically, the intra-block separated signal parameter calculation unit 410 calculates the separated signal s ^(b) , which is a time series signal in the block b, by the following equation (39).

  Specifically, the following formula (40) is a numerator of formula (39)

And the following equation (41), which is the matrix of the denominator of equation (39)

It is possible to efficiently calculate the solution of the linear system by using the property that the denominator matrix is Hermitian and banded matrix without calculating the inverse matrix of.

Details, for example, Non-Patent Document 8: WH Press, WT Vetterling, SA Teukolsky and BP Flannery, Numerical Recipes in C ++: see ^{The Art of Scientific Computing, 2 nd} ed Cambridge University Press, 2002..

The total separation signal parameter calculation unit 420 calculates the total separation signal parameter using each intra-block separation signal parameter calculated by the intra-block separation signal parameter calculation unit 410. Here, it is assumed that n _{b + 2} ≧ n _{b + 1} + n _O and that the number of blocks that can be simultaneously overlapped is two or less. In the b-th block, when n _b + n _O <n _{b + 1} −1 holds, there is a section that does not overlap with any adjacent block, and in that section, the total separation signal parameter is expressed by the following equation (42): Thus, the intra-block separation signal parameter itself is used.

In a section where the b-th block and the b + 1-th block overlap, for example, the entire separation signal parameter is synthesized as follows. Since the part close to the boundary of the block may be affected by the boundary, for example, n _C <n _O / 2 is defined so as to exclude the part, and the length n _C at the beginning of the b + 1-th block In this part, the total separation signal parameter is set to the b-th intra-block separation signal parameter itself as shown in the following equation (43), and in the part of the length n _C at the end of the b-th block, the total separation signal parameter is As shown in the following equation (44), the b + 1-th intra-block separation signal parameter itself is used.

  Finally, the entire separation signal parameter is calculated by cross-fading as shown in the following equation (45), for example, in the middle part of the overlapping section.

  In the modified Wiener criterion reducing unit 40 as a whole, the block dividing unit 400 does not divide the block (assuming that the number of blocks B = 1), and the entire separation signal parameter can be calculated at once by the equation (39). In that case, the process of the total separated signal parameter calculation unit 420 is omitted.

<Example 2>
The second embodiment is configured to perform signal separation in accordance with the “policy 1”.
The signal separation device 1 according to the embodiment of the present invention described here is the signal separation device described in Example 1 shown in FIG. 2, and the modified Wiener criterion reduction unit 40 is described in Example 1 (FIG. 3). Instead of this, the configuration shown in FIG. 4 is used.

  The modified Wiener criterion reduction unit 40 of the second embodiment includes an information storage unit 401, a separated signal parameter initial value generation unit 411, a separated signal parameter convergence determination unit 421, a separated signal parameter update unit 431, and a separated signal parameter output unit 441. ing.

  The information storage unit 401 is a storage area for storing the separation signal parameter, the conventional Wiener filter output signal parameter, and the time frequency amplitude component array estimated value of the target signal and the interference signal. The information storage unit 401 also stores a weighting coefficient that can be calculated as in the following equation (46) from the time frequency amplitude component array estimated values of the target signal and the interference signal.

The separated signal parameter initial value generation unit 411 sets an initial value of the separated signal parameter S≡ (S _{ω, t} ) _{N × T} , and outputs the set value to the information storage unit 401 as S ′. For example, the initial value S ′ of S may be set to ^ S = (^ S _{ω, t} ) _{N × T} defined by Equation (4).

  The separated signal parameter updating unit 431 calculates a new separated signal parameter from the separated signal parameter S ′ stored in the information storage unit 401, the weighting factor α, and the conventional Wiener filter output signal parameter ^ S, and information The separation signal parameter S ′ stored in the storage unit 401 is replaced with a new separation signal parameter. Further, the separated signal parameter update unit 431 generates a new signal when the separated signal parameter convergence determination unit 421 determines that the separated signal parameter S ′ stored in the information storage unit 401 does not satisfy a predetermined criterion. The separation signal parameter is calculated again. More specifically, the separated signal parameter updating unit 431 uses the separated signal parameter S ′ stored in the information storage unit 401 to calculate a consistent time frequency component array G (S ′) according to the following equation (47). Then, a new separated signal parameter S is calculated by the following equation (48).

  Where γ is a positive constant. Also, the calculated new value S is reflected in the information storage unit 401 as S ′.

  The separated signal parameter convergence determination unit 421 determines whether or not the separated signal parameter S ′ stored in the information storage unit 401 satisfies a predetermined criterion. More specifically, the separated signal parameter convergence determination unit 421 counts the number of times (the number of iterations) that the separated signal parameter update unit 431 has been executed, and whether or not the number of iterations has reached a predetermined number, The rate of change of the separation signal parameter S ′ stored in the information storage unit 401 and the separation signal parameter S newly calculated in the separation signal parameter update unit 431 is calculated, and whether or not the rate of change is below a predetermined threshold value Or whether the rate of change of the objective function value of the following equation (49) is less than or equal to a predetermined threshold, or whether the objective function value of the following equation (49) is less than or equal to a predetermined threshold. To determine whether the separated signal parameter S ′ satisfies a predetermined criterion.

  The separation signal parameter output unit 441 outputs the analysis parameter when the separation signal parameter convergence determination unit 421 determines that the analysis parameter satisfies a predetermined criterion.

<Example 3>
The third embodiment is configured to perform signal separation in accordance with the above-described “policy 3”. In the third embodiment, the processes of the separated signal parameter update unit 431 and the separated signal parameter convergence determination unit 421 in the second embodiment are replaced with the following processes.

  Here, as shown in FIG. 5, the separated signal parameter update unit 431 includes a consistent time frequency component array calculation unit 4310 and a time frequency component array weighted sum calculation unit 4311.

  The separated signal parameter updating unit 431 calculates a new separated signal parameter from the separated signal parameter S ′ stored in the information storage unit 401, the weighting factor α, and the conventional Wiener filter output signal parameter ^ S. The separated signal parameter updating unit 431 calculates a new separated signal parameter again when the separated signal parameter convergence determining unit 421 determines that the separated signal parameter does not satisfy a predetermined criterion.

  The consistent time frequency component array calculation unit 4310 calculates a new consistent time frequency component array from the separated signal parameters stored in the information storage unit 401. More specifically, the consistent time frequency component array calculation unit 4310 calculates the consistent time frequency component array G (S ′) by the following equation (50) using S ′ stored in the information storage unit 401. To do.

  The time-frequency component array weighted sum calculation unit 4311 is made consistent with the weight coefficient α and the conventional Wiener filter output signal parameter ^ S stored in the information storage unit 401 and the non-consistent time-frequency component array calculation unit 4310. A new separation signal parameter is calculated from the time-frequency component array. More specifically, the time-frequency component array weighted sum calculation unit 4311 uses the ~ S obtained by the consistent time-frequency component array calculation unit 4310 and the conventional Wiener filter output signal parameter ^ S to obtain the following equation (51): The separation signal parameter S is calculated by

  Where γ is a positive constant. The calculated new value S is reflected in the information storage unit 401 as S ′.

The separated signal parameter convergence determination unit 421 determines whether or not the separated signal parameter stored in the information storage unit 401 satisfies a predetermined criterion. More specifically, the separated signal parameter convergence determination unit 421 counts the number of times (the number of iterations) that the separated signal parameter update unit 431 has been executed, and whether or not the number of iterations has reached a predetermined number, Whether the rate of change of the separation signal parameter S ′ stored in the information storage unit 401 and the separation signal parameter S newly calculated in the separation signal parameter updating unit 431 is equal to or less than a predetermined threshold value, or The separation signal parameter S depends on whether the rate of change of the objective function value in (52) is less than or equal to a predetermined threshold value or whether the objective function value in the following equation (52) is less than or equal to a predetermined threshold value. It is determined whether or not ′ satisfies a predetermined criterion.

<Example 4>
In the third embodiment, the consistent time frequency component array calculation unit 4310 may be performed as follows as an alternative process.

  The consistent time frequency component array calculation unit 4310 calculates a new consistent time frequency component array from the separated signal parameter S ′ stored in the information storage unit 401. More specifically, the consistent time frequency component array calculation unit 4310 uses S ′ stored in the information storage unit 401 to approximately calculate the consistent time frequency component array ~ S according to the following equation (53). To do.

However, the weighting coefficients η _{t ′ and ω ′} are defined by the following equation (54). Further, the index of the frequency bin of S is assumed to be mod N.

  Where w is an analysis window function of STFT, and ~ s is a composite window function including a normalization term of inverse STFT. Here, since it is assumed that the composite window function s before entering the normalization term is regarded as s = w, ~ s is defined by the following equation (55).

  Further, in the calculation of Expression (53), the number of multiplications can be further reduced by using various symmetries of the weighting coefficient η. Further, all the time frequency bins may be updated. Alternatively, the update may be calculated for each bin according to some criterion or may be selectively updated to keep the value of the previous step. For example, a certain threshold value may be determined, and only the bin of the time frequency amplitude component array larger than the threshold value may be updated. Further, the threshold value may be fixed or may be changed for each repetitive step. See Non-Patent Document 3 for details.

<Experimental example to verify the effect>
In order to show the effects and operations of the present invention, experimental examples using the embodiments of the present invention will be described below. In each method, an evaluation experiment is performed on the value of the corrected Wiener criterion ~ ψ when the calculation time and the iterative calculation are finished. A consistent time-frequency component array having the amplitude component closest to the amplitude component of ^ S and the output signal parameter ^ S of the conventional Wiener filter in the sense of least square error (see Non-Patent Document 2, for example) ) The calculated value is the comparison target of the present invention in this experiment.

  The sampling frequency was 16 kHz, and the time frequency component array was created by shifting the sine window function with a frame length of N = 1024 (64 ms) every R = 512 (32 ms) unless otherwise specified. The sine window function was also used for the inverse STFT composite window.

Each method was implemented as follows. As described in the first embodiment, the time domain method is processed for each block in which n _B = 64 frames are combined. N _O = n _B / 2 was set so that adjacent blocks overlapped half. That is, n _b = bn _B / 2. In addition, the number of frames in the excluded section near the boundary is set to n _C = n _B / 8. That is, the length of the section in which crossfading is performed is n _O −2n _C = n _B / 4. Parameters n _B , n _O , and n _C related to block partitioning are preliminary experiments, and the calculation time and required memory amount are kept as much as possible while maintaining almost the value of the modified Wiener criterion when solving the analysis with the entire interval as one block. Was set to reduce.

For the method based on the Douglas-Rachford splitting method and the method based on the auxiliary function method, a better solution can be obtained intuitively, but the convergence speed becomes slower. In the method based on the auxiliary function method, when the update formula (35) using γ ₁ larger than γ ₀ is further performed from the point obtained by the update formula (35) using a certain γ ₀ , the modified Wiener criterion ~ ψ Experimentally noticed that the monotonic decrease. Therefore, starting from a very small value (for example, γ = 10 ⁻⁵ ), while keeping γ fixed, the separation signal parameter S is updated by the update equation (35) until the decrease of ~ ψ becomes smaller than 1%, When the decrease in ~ ψ was smaller than 1%, we changed the γ to 2γ and restarted the S update. The algorithm was aborted when the value of γ was raised twice consecutively without a decrease of ~ 1% in ~ ψ. Alternatively, the algorithm was aborted when the number of iterations reached 200. Also, with the method based on the Douglas-Rachford splitting method, a phenomenon such as monotonic decrease was not as obvious as the method based on the auxiliary function method, so γ was experimentally fixed at 10 ⁴ and the number of iterations of the algorithm was 4000 times. I broke it off. Finally, the phase estimation algorithm by Griffin-Lim was terminated after 4000 iterations.

  The mixed signal of two female speakers in 5.5s was used as the input signal. The case where the time frequency amplitude component arrangement | sequence of each utterance is given previously (Oracle condition) was considered. The experimental results are shown in Table 1. Although the performance of the conventional Wiener filter is already very high, it can be seen that all the proposed methods have achieved significant improvements both in improving the modified Wiener criterion and in improving the signal-to-noise ratio (SNR). The reason for improving SNR may not be obvious, but it can be understood from the fact that the time-frequency component array of the re-synthesized time-series signal is closer to the solution of maximum likelihood estimation. Although the calculation of the optimal solution in the time domain is very time-consuming, it can be seen that the estimated value obtained in a shorter time by the method based on the auxiliary function method is very close to the optimal solution.

  Experiments on the effect of frame shift on STFT time-frequency component array were also conducted. Table 2 shows the experimental results. It can be seen that when the frame shift is reduced, that is, the amount of overlap between frames is increased, the SNR is improved particularly in the case of an optimal solution by analytical calculation. This result can be interpreted by the consistency constraint becoming stronger as the overlap increases. However, the calculation amount also increases in proportion to the overlap. In all methods, the amount of calculation increases in a linear proportion with the number of frames of the temporal frequency component array.

  Finally, noise suppression evaluation experiments were also conducted. Here, a method based on the auxiliary function method was used. Three conditions were considered by using the voice with Gaussian white noise added as the input signal: 1) The actual values of the time frequency amplitude component arrays of the speech target signal and the noise interference signal are known ("oracle" condition) 2) the actual value of the time frequency amplitude component array of the speech target signal is known, and only the variance is known for the noise ("variance" condition); 3) only the variance of the noise interference signal is given, The time frequency amplitude component array of the speech target signal is estimated by spectral subtraction ("subtraction" condition). Here, the method based on the auxiliary function method is compared with the conventional Wiener filter. Three SNRs were used in each condition: -10dB, 0dB and 10dB. Table 3 summarizes the results of averaging for 10 independently generated noise signals for each setting. Especially in spectral subtraction, a significant improvement can be seen by the method based on the auxiliary function method. Perceptually, although the musical noise is still strong, the residual noise signal remaining in the Wiener filter has become very weak.

  The spectrogram of the speech signal in noise is shown in FIG. Moreover, the spectrogram of the conventional Wiener filter output signal is shown in FIG. Moreover, the spectrogram of the time domain signal obtained by the technique based on the auxiliary function method is shown in FIG.

  As described above, according to the present embodiment, in parameter estimation, more accurate estimation is performed using the restriction that the time-frequency component arrangement of signals is consistent. As a result, there is an advantage that high-precision and high-speed signal separation processing can be performed. That is, according to the present invention, signal separation can be performed at the same speed as in the case of the prior art, and more accurate separation results can be obtained.

  Each configuration provided in the signal separation device 1 described above may be realized by dedicated hardware, and each configuration provided in the signal separation device 1 includes a memory and a CPU (central processing unit). The program may be realized by loading a program for realizing the function of each component included in the signal separation device 1 into a memory and executing the program.

  Further, by recording a program for realizing the function of each component included in the signal separation device 1 on a computer-readable recording medium, causing the computer system to read and execute the program recorded on the recording medium, You may perform the process of each structure with which the signal separation apparatus 1 is provided. Here, the “computer system” includes an OS and hardware such as peripheral devices.

Further, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used.
The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Furthermore, the “computer-readable recording medium” dynamically holds a program for a short time like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, a volatile memory in a computer system serving as a server or a client in that case, and a program that holds a program for a certain period of time are also included. The program may be a program for realizing a part of the functions described above, and may be a program capable of realizing the functions described above in combination with a program already recorded in a computer system.

  The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and includes design and the like within a scope not departing from the gist of the present invention.

DESCRIPTION OF SYMBOLS 1 ... Signal separation apparatus, 10 ... Time frequency analysis part, 20 ... Time frequency amplitude component arrangement | sequence estimation part, 30 ... Conventional Wiener filter output signal parameter calculation part, 40 ... Modified Wiener criterion reduction part, 50 ... Separation signal parameter output part, 400: Block division unit, 401: Information storage unit, 410: Intra-block separation signal parameter calculation unit, 411: Separation signal parameter initial value generation unit, 420 ... Whole separation signal parameter calculation unit, 421 ... Separation signal parameter convergence determination unit, 431 ... Separation signal parameter update unit, 441 ... Separation signal parameter output unit, 4310 ... Contradiction time frequency component array calculation unit, 4311 ... Petrogram weighted sum calculation unit

Claims (9)

A time-frequency analysis unit that calculates a time-frequency component array from an input signal that is a mixed signal of a target signal and an interference signal;
A time frequency amplitude component array estimation unit for estimating a time frequency amplitude component array for each source signal of the target signal and the interference signal included in the input signal from the input signal;
Based on the time frequency component array calculated from the input signal and the estimated value of the time frequency amplitude component array for each source signal, the Wiener filter of the conventional Wiener filter in the time frequency domain in each short time frame Wiener filter output signal parameter calculation unit for calculating output signal parameters;
The time frequency domain signal parameter for estimating the target signal is set as a separation signal parameter, and the time frequency component array of the time series signal and the output of the conventional Wiener filter when the separation signal parameter is restored by inverse short-time Fourier transform Using the weighted distance with the signal parameter as a modified Wiener criterion, the modified Wiener criterion is smaller than the value when the separated signal parameter in the modified Wiener criterion is replaced with the output signal parameter of the conventional Wiener filter. A modified Wiener criterion reduction unit for estimating the separated signal parameter value from the input signal;
A separation signal parameter output unit that outputs the separation signal parameter value estimated in the modified Wiener criterion reduction unit;
A signal separation device comprising: The modified Wiener criterion reduction part is
A block dividing unit for dividing the conventional Wiener filter output signal parameter into blocks each having a length of a plurality of frames;
Intra-block separation for calculating a separation signal parameter value in the block so that the modified Wiener criterion is reduced compared to the case of the conventional Wiener filter for a plurality of frames of the conventional Wiener filter output signal parameter divided into the blocks A signal parameter calculator,
An overall separated signal parameter calculation unit for calculating a separated signal parameter for the entire block from the separated signal parameter value in the block;
The signal separation device according to claim 1, further comprising: The modified Wiener criterion reduction part is
The conventional Wiener filter output signal parameter, and an information storage unit that stores the time frequency amplitude component array estimated value,
A separation signal parameter initial value generation unit for generating an initial value for the separation signal parameter;
A separation signal parameter convergence determination unit that determines whether or not the separation signal parameter value has converged;
A separation signal parameter update unit for updating the separation signal parameter value;
A separated signal parameter output unit that outputs the separated signal parameter value estimated in the separated signal parameter update unit;
The signal separation device according to claim 1, further comprising: The separated signal parameter update unit includes:
For the separated signal parameter, a consistent time frequency component array calculation unit that calculates a consistent time frequency component array that makes the time frequency component array of the signal consistent, and
A time-frequency component array weighted sum calculation unit that calculates a time-frequency component array weighted sum having the non-consistent time-frequency component array, the separated signal parameter, and the conventional Wiener filter output signal parameter as elements;
The signal separation device according to claim 3, further comprising: A time-frequency analysis process for calculating a time-frequency component array from an input signal that is a mixed signal of a target signal and an interference signal;
A time frequency amplitude component array estimation process for estimating a time frequency amplitude component array for each source signal of the target signal and the interference signal included in the input signal from the input signal;
Based on the time frequency component array calculated from the input signal and the estimated value of the time frequency amplitude component array for each source signal, the Wiener filter of the conventional Wiener filter in the time frequency domain in each short time frame Wiener filter output signal parameter calculation process for calculating output signal parameters;
The time frequency domain signal parameter for estimating the target signal is set as a separation signal parameter, and the time frequency component array of the time series signal and the output of the conventional Wiener filter when the separation signal parameter is restored by inverse short-time Fourier transform Using the weighted distance with the signal parameter as a modified Wiener criterion, the modified Wiener criterion is smaller than the value when the separated signal parameter in the modified Wiener criterion is replaced with the output signal parameter of the conventional Wiener filter. A modified Wiener criterion reduction process for estimating the separated signal parameter value from the input signal;
A separated signal parameter output process for outputting the separated signal parameter value estimated in the modified Wiener criterion decreasing process;
A signal separation method comprising: The modified Wiener criterion reduction process further includes:
A block division process of dividing the conventional Wiener filter output signal parameter into blocks of a length of a plurality of frames;
Intra-block separation for calculating a separation signal parameter value in the block so that the modified Wiener criterion is reduced compared to the case of the conventional Wiener filter for a plurality of frames of the conventional Wiener filter output signal parameter divided into the blocks Signal parameter calculation process;
A total separation signal parameter calculation process for calculating a separation signal parameter for the whole block from the separation signal parameter value in the block;
The signal separation method according to claim 5, further comprising: The modified Wiener criterion reduction process further includes:
Information storage process for storing the conventional Wiener filter output signal parameters and the time frequency amplitude component array estimated value,
A separation signal parameter initial value generation step of generating an initial value for the separation signal parameter;
A separation signal parameter convergence determination step of determining whether or not the separation signal parameter value has converged;
A separation signal parameter update process of updating the separation signal parameter value;
A separated signal parameter output process of outputting the separated signal parameter value estimated in the separated signal parameter update process;
The signal separation method according to claim 5, further comprising: The separated signal parameter update process further includes:
For the separated signal parameter, a consistent time frequency component array calculation process for calculating a consistent time frequency component array in which the time frequency component array of the signal is consistent, and
A time-frequency component array weighted sum calculation step of calculating a time-frequency component array weighted sum having the non-consistent time-frequency component array, the separated signal parameter, and the conventional Wiener filter output signal parameter as elements;
The signal separation method according to claim 7, further comprising: To a computer as a signal separation device,
A time-frequency analysis process for calculating a time-frequency component array from an input signal that is a mixed signal of a target signal and an interference signal;
A time frequency amplitude component array estimation process for estimating a time frequency amplitude component array for each source signal of the target signal and the interference signal included in the input signal from the input signal;
Based on the time frequency component array calculated from the input signal and the estimated value of the time frequency amplitude component array for each source signal, the Wiener filter of the conventional Wiener filter in the time frequency domain in each short time frame Wiener filter output signal parameter calculation process for calculating output signal parameters;
The time frequency domain signal parameter for estimating the target signal is set as a separation signal parameter, and the time frequency component array of the time series signal and the output of the conventional Wiener filter when the separation signal parameter is restored by inverse short-time Fourier transform Using the weighted distance with the signal parameter as a modified Wiener criterion, the modified Wiener criterion is smaller than the value when the separated signal parameter in the modified Wiener criterion is replaced with the output signal parameter of the conventional Wiener filter. A modified Wiener criterion reduction process for estimating the separated signal parameter value from the input signal;
A separated signal parameter output process for outputting the separated signal parameter value estimated in the modified Wiener criterion decreasing process;
Signal separation program for executing