Spatial correlation matrix estimation device, spatial correlation matrix estimation method, and spatial correlation matrix estimation program

1. A non-transitory spatial correlation matrix estimation device comprising:

a memory; and

a processor coupled to the memory and programmed to execute a process comprising:

estimating, in a situation in which N first acoustic signals associated with N target sound sources (where, N is an integer equal to or greater than 1) and a second acoustic signal associated with background noise are present in a mixed manner, based on observation feature value vectors calculated based on M observation signals (where, M is an integer equal to or greater than 2) each of which is recorded at a different position, a first mask that is the proportion of the first acoustic signal included in a feature value of the observation signal for each time-frequency point and a second mask that is the proportion of the second acoustic signal included in a feature value of the observation signal for each time-frequency point and that estimates a spatial correlation matrix of the target sound sources based on the first mask and the second mask,

wherein the estimating estimates the spatial correlation matrix of the target sound sources based on a first spatial correlation matrix obtained by weighting, by a first coefficient, a first feature value matrix calculated based on the observation signals and the first masks and based on a second spatial correlation matrix obtained by weighting, by a second coefficient, a second feature value matrix calculated based on the observation signals and the second masks.

2. The spatial correlation matrix estimation device according to claim 1 , wherein the estimating calculates the first coefficient and the second coefficient such that, under the condition that a spatial correlation matrix of background noise is not temporally changed, a component derived from the background noise included in an estimation value of the spatial correlation matrix of the target sound sources becomes zero.

3. The spatial correlation matrix estimation device according to claim 1 , wherein the estimating calculates the first coefficient and the second coefficient such that the ratio of the first coefficient to the second coefficient is equal to the ratio of the reciprocal of a time average value of the first masks to the reciprocal of a time average value of the second masks.

4. The spatial correlation matrix estimation device according to claim 1 , wherein, when N=1, the first spatial correlation matrix is a time average, for each frequency, of an observation feature value matrix calculated based on the observation feature value vectors.

5. The spatial correlation matrix estimation device according to claim 1 , further comprising:

applying a short-time signal analysis to the observation signals, extracting a signal feature value for each time-frequency point, and calculating, for each time-frequency point, the observation feature value vector that is an M-dimensional column vector having the signal feature value as a component;

calculating, based on the observation feature value vector, for each time-frequency point, an observation feature value matrix by multiplying the observation feature value vector by Hermitian transpose of the observation feature value vector;

calculating, regarding each of the target sound sources, the time average, for each frequency, of a matrix obtained by multiplying, for each time-frequency point, the observation feature value matrix by the first mask as the first feature value matrix and that estimates the first spatial correlation matrix by multiplying the first coefficient by the first feature value matrix; and

calculating, regarding the background noise, the time average, for each frequency, of a matrix obtained by multiplying, for each time-frequency point, the observation feature value matrix by the second mask as the second feature value matrix and estimating the second spatial correlation matrix by multiplying the second coefficient by the second feature value matrix, wherein

the spatial correlation matrix of the target sound sources being estimated by subtracting the second spatial correlation matrix from the first spatial correlation matrix, and

the ratio of the first coefficient to the second coefficient is equal to the ratio of the reciprocal of the time average value of the first mask to the reciprocal of the time average value of the second mask.

6. The spatial correlation matrix estimation device according to claim 1 , further comprising modeling, for each frequency, a probability distribution of the observation feature value vectors by a mixture distribution composed of N+1 component distributions each of which is a zero mean M-dimensional complex Gaussian distribution with a covariance matrix represented by the product of a scalar parameter that has a time varying value and a positive definite Hermitian matrix that has time invariant parameters as its elements and setting, to the first mask and the second mask, each of posterior probabilities of the component distributions obtained by estimating the parameters of the mixture distributions such that the mixture distributions approach the distribution of the observation feature value vectors.

7. The spatial correlation matrix estimation device according to claim 6 , wherein, from among the component distributions, estimating sets, to the second mask, the posterior probability of an component distribution that has the most flat shape of the distribution of eigenvalues of the positive definite Hermitian matrix that has the time invariant parameters as the elements.

8. A spatial correlation matrix estimation method for estimating, in a situation in which N first acoustic signals associated with N target sound sources (where, N is an integer equal to or greater than 1) and a second acoustic signal associated with background noise are present in a mixed manner, based on observation feature value vectors calculated based on M observation signals (where, M is an integer equal to or greater than 2) each of which is recorded at a different position, a first mask that is the proportion of the first acoustic signal included in a feature value of the observation signal for each time-frequency point and a second mask that is the proportion of the second acoustic signal included in a feature value of the observation signal for each time-frequency point and estimating a spatial correlation matrix of the target sound sources based on the first mask and the second mask, the spatial correlation matrix estimation method comprising:

a noise removal step of estimating the spatial correlation matrix of the target sound sources based on a first spatial correlation matrix obtained by weighting, by a first coefficient, a first feature value matrix calculated based on the observation signals and the first masks and based on a second spatial correlation matrix obtained by weighting, by a second coefficient, a second feature value matrix calculated based on the observation signals and the second masks.

9. The spatial correlation matrix estimation method according to claim 8 , wherein the noise removal step includes calculating the first coefficient and the second coefficient such that, under the condition that a spatial correlation matrix of background noise is not temporally changed, a component derived from the background noise included in an estimation value of the spatial correlation matrix of the target sound sources becomes zero.

10. The spatial correlation matrix estimation method according to claim 8 , wherein the noise removal step includes calculating the first coefficient and the second coefficient such that the ratio of the first coefficient to the second coefficient is equal to the ratio of the reciprocal of a time average value of the first masks to the reciprocal of a time average value of the second masks.

11. The spatial correlation matrix estimation method according to claim 8 , further comprising:

a time-frequency analyzing step of applying a short-time signal analysis to the observation signals, extracting a signal feature value for each time-frequency point, and calculating, for each time-frequency point, the observation feature value vector that is an M-dimensional column vector having the signal feature value as a component;

an observation feature value matrix calculating step of calculating, based on the observation feature value vector, for each time-frequency point, an observation feature value matrix by multiplying the observation feature value vector by Hermitian transpose of the observation feature value vector;

a noisy-environment target sound spatial correlation matrix estimating step of calculating, regarding each of the target sound sources, the time average, for each frequency, of a matrix obtained by multiplying, for each time-frequency point, the observation feature value matrix by the first mask as the first feature value matrix and estimating the first spatial correlation matrix by multiplying the first coefficient by the first feature value matrix; and

a noise spatial correlation matrix estimating step of calculating, regarding the background noise, the time average, for each frequency, of a matrix obtained by multiplying, for each time-frequency point, the observation feature value matrix by the second mask as the second feature value matrix and estimating the second spatial correlation matrix by multiplying the second coefficient by the second feature value matrix, wherein

the noise removal step includes estimating the spatial correlation matrix of the target sound sources by subtracting the second spatial correlation matrix from the first spatial correlation matrix, and

the ratio of the first coefficient to the second coefficient is equal to the ratio of the reciprocal of the time average value of the first mask to the reciprocal of the time average value of the second mask.

12. A non-transitory computer-readable recording medium having stored a spatial correlation matrix estimation program that causes a spatial correlation matrix estimation device to estimate, in a situation in which N first acoustic signals associated with N target sound sources (where, N is an integer equal to or greater than 1) and a second acoustic signal associated with background noise are present in a mixed manner, based on observation feature value vectors calculated based on M observation signals (where, M is an integer equal to or greater than 2) each of which is recorded at a different position, a first mask that is the proportion of the first acoustic signal included in a feature value of the observation signal for each time-frequency point and a second mask that is the proportion of the second acoustic signal included in a feature value of the observation signal for each time-frequency point and that estimates a spatial correlation matrix of the target sound sources based on the first mask and the second mask, and to estimate the spatial correlation matrix of the target sound sources based on a first spatial correlation matrix obtained by weighting, by a first coefficient, a first feature value matrix calculated based on the observation signals and the first masks and based on a second spatial correlation matrix obtained by weighting, by a second coefficient, a second feature value matrix calculated based on the observation signals and the second masks.