JP2015055843A - Number-of-signal-source estimation device, number-of-signal-source estimation method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accurately estimate the number of signal sources even when the influence of reflection or reverberation is large.SOLUTION: A frequency region conversion unit 10 converts an observation signal into a time frequency region, and divides each frequency bin into a plurality of separation signals. An activity calculation unit 12 determines activities from the separation signals. A clustering unit 14 clusters the activities into a plurality of activity clusters. A number-of-signal-sources determination unit 16 determines an estimated value of the number of signal sources on the basis of the number of activities included in each activity cluster.

Description

  The present invention relates to a technique for estimating the number of signal sources using observation signals.

It is assumed that signals from N signal sources are mixed and observed by M sensors in a reverberant environment (N is an integer of 1 or more). At this time, the time-frequency conversion x _j (τ, f) of the signal observed by the j (1 ≦ j ≦ M) -th sensor is modeled by Expression (1).

Here, τ represents a time frame number, f represents a frequency bin number, and s _k (τ, f) represents a time-frequency conversion of the k-th signal source. H _jk (f) represents the transfer function from the k-th signal source to the j-th sensor, and models the effects of signal delay and attenuation due to the distance between the signal source and the sensor, and reflection and reverberation on the wall. Yes.

The problem of estimating the number of signal sources handled in the present invention is a problem of estimating the number N of signal sources using observation signals x _j (τ, f) (j = 1,..., M) from M sensors. . In particular, in the framework of estimating the number of signal sources based on clustering, a feature quantity f _ζ (ζ = 1,..., L) that takes a unique value for each signal source is used, and a histogram of the feature quantity corresponds to each signal source. Based on having a peak, one probability model is applied to each peak by clustering and model selection, and the number of signal sources is estimated by counting the number of probability models.

  As a conventional signal source number estimation technique, a sound source number estimation technique based on clustering of sound arrival directions is known (for example, Patent Document 1). FIG. 1 shows an example of an apparatus for estimating the number of sound sources by clustering described in Patent Document 1. This corresponds to a case in which the arrival direction is used as a feature amount and a Dirichlet distribution is given to the prior distribution of the mixture weight as a model selection mechanism in the general framework of signal source number estimation based on the clustering described above.

The sound source number estimation apparatus 400 shown in FIG. 1 includes J (J ≧ 2) sound collection means 20 _j (for example, microphones) (j = 1,..., J), a frequency domain conversion unit 30, a power estimation unit 32, The arrival direction estimation unit 34, the direction information distribution estimation device 200, and the sound source number estimation unit 36 are included. Observation signals observed by the J sound collecting means 20 _j are converted into the frequency domain by the frequency domain conversion unit 30. The arrival direction estimation unit 34 calculates the arrival direction using the observation signal converted into the frequency domain. The power estimation unit 32 calculates a power that is a weight given to the likelihood function during clustering. This weight is given to reduce the influence of data with low power and low reliability on clustering. The direction information distribution estimation apparatus 200 clusters the arrival direction into clusters corresponding to each sound source using the power and the arrival direction. The sound source number estimation unit 36 calculates the estimated value of the number of sound sources by counting the number of clusters.

JP 2010-145836 A

  There is a problem that the estimated value of the sound arrival direction calculated using the observation signal is easily affected by reflection or reverberation on a wall or the like. If the influence of reflection and reverberation on the wall is small like an anechoic room and the sound propagation can be approximated to plane wave propagation, the direction of sound arrival can be estimated accurately, but the reflection and reverberation on the wall are not. If it is large, it is difficult to estimate accurately. Therefore, it is difficult for the sound source number estimation device of Patent Document 1 to accurately estimate the number of sound sources in such a case.

  An object of the present invention is to accurately estimate the number of signal sources even when the influence of reflection or reverberation on a wall or the like is large.

  In order to solve the above-described problem, the signal source number estimation device of the present invention includes a clustering unit and a signal source number determination unit. The clustering unit clusters the activity of the separated signal acquired from each of the separated signals into a plurality of activity clusters when the observed signal is separated into a plurality of separated signals for each frequency bin. The signal source number determination unit obtains the number of activity clusters determined to correspond to the signal source among the activity clusters as an estimated value of the number of signal sources.

  The signal source number estimation technique according to the present invention estimates the number of signal sources using an activity as a feature amount in a framework of signal source number estimation based on clustering. An activity is a feature amount indicating a unique pattern for each signal regardless of the presence or absence of reflection or reverberation on a wall or the like. Therefore, according to the signal source number estimation technique of the present invention, the number of signal sources can be accurately estimated even when the influence of reflection or reverberation on a wall or the like is large.

FIG. 1 is a diagram illustrating a functional configuration of a conventional sound source number estimation apparatus. FIG. 2 is a diagram for explaining the principle of the invention. FIG. 3 is a diagram illustrating a functional configuration of the signal source number estimation apparatus. FIG. 4 is a diagram illustrating a processing flow of the signal source number estimation method. FIG. 5 is a diagram illustrating a functional configuration of the clustering unit. FIG. 6 is a diagram illustrating a functional configuration of the signal source number determination unit. FIG. 7 is a diagram for explaining the experimental environment.

  Hereinafter, embodiments of the present invention will be described in detail. In addition, the same number is attached | subjected to the component which has the same function in drawing, and duplication description is abbreviate | omitted.

[Principle of the invention]
Prior to the description of the embodiments, the principle of the present invention will be described.

  In the present invention, in the conventional framework for estimating the number of signal sources in clustering, the number of signal sources is estimated by clustering the activities using the activity as the feature amount. The signal activity is time series information corresponding to the time change of the amplitude of the separated signal. For example, the time series of the amplitude value of the separated signal for each time interval, the time series of the contribution degree of the separated signal for each time interval, etc. It is. Signal activity has the property of showing a unique pattern for each signal, regardless of the presence or absence of reflection or reverberation on a wall or the like. By clustering the signal activity based on this property, the number of signal sources can be accurately estimated even when the influence of reflection or reverberation on a wall or the like is large.

  In order to obtain signal activity, a conventionally known sound source separation method is applied. However, although the conventional sound source separation method is premised on the fact that the number of sound sources is known, in the present invention, the number of signal sources is unknown. Therefore, a separated signal is obtained by a conventional sound source separation method, assuming that the number is larger than the assumed number of signal sources.

  The separated signal thus obtained includes a signal such as noise that does not correspond to the signal source. Therefore, it is necessary to prevent the number of signal sources from being overestimated by treating a separated signal that does not correspond to the signal source as corresponding to the signal source. In order to realize this, measures such as removing a separated signal that does not correspond to the signal source or reducing the influence of the separated signal that does not correspond to the signal source on the estimation of the number of signal sources can be considered.

  Therefore, the present invention takes advantage of the fact that the activity of the separated signal that does not correspond to the signal source often takes a smaller value compared to the activity of the separated signal that corresponds to the signal source, and therefore the separated signal that does not correspond to the signal source. Eliminate signal effects. That is, when clustering the activities of the separated signals, small activities are not used for clustering, or clusters consisting of small activities are excluded.

  The separated signals originating from the same signal source are expected to be classified into the same cluster because their activities are similar to each other. Moreover, since the activity of the separated signal not corresponding to the signal source has a small value, it is not used for clustering or excluded from the clustering result. Therefore, it is expected that the number of clusters finally obtained by clustering activities will coincide with the number of signal sources with high accuracy. In the embodiment described later, a cluster obtained by clustering activities is called an activity cluster.

FIG. 2 conceptually shows the principle of the invention described above. "Speech activity sequences" on the left side of the figure is the activity obtained from the separated signal obtained by separating the observation signal for each frequency bin. "Speech activity clusters" on the right side of the figure is an activity cluster obtained by clustering activities. As will be described in detail later, the time series of activities is clustered, and an activity cluster (represented as “Active” in the figure) having a large activity representative value ρ _i is determined to correspond to a signal source. On the other hand, it is determined that an activity cluster (represented as “Garbage” in the figure) having a small activity representative value ρ _i does not correspond to a signal source. Then, the number of the activity clusters determined to correspond to the signal source is determined as the number of signal sources.

[Points of invention]
The first point of the present invention is that activities are clustered using activities as feature quantities in the conventional framework for estimating the number of signal sources by clustering.

  The second point is to obtain activity by performing signal source separation assuming a larger number of signal sources than actual in order to obtain activity from the observed signal in a situation where the number of signal sources is not known in advance. It is.

  The third point is that a means for removing an activity not corresponding to the signal or reducing the influence thereof is introduced. This is because an activity that does not correspond to the signal occurs due to the assumption that the number of signal sources is larger than the actual number, and an activity that does not correspond to the signal is mistakenly recognized as an activity that corresponds to the signal, thereby overestimating the number of signal sources. This is to prevent this.

[First embodiment]
An example of a functional configuration of the signal source number estimation apparatus 100 according to the first embodiment will be described with reference to FIG. The signal source number estimation apparatus 100 includes a frequency domain conversion unit 10, an activity calculation unit 12, a clustering unit 14, and a signal source number determination unit 16. The signal source number estimation device 100 is configured by, for example, reading a special program into a known or dedicated computer having a central processing unit (CPU), a main storage device (Random Access Memory, RAM), and the like. Special equipment. For example, the signal source number estimation device 100 executes each process under the control of the central processing unit. The data input to the signal source number estimation device 100 and the data obtained in each process are stored in, for example, a main storage device, and the data stored in the main storage device is read out as necessary to perform other processing. Used for

  With reference to FIG. 4, an example of the processing flow of the signal source number estimation method executed by the signal source number estimation apparatus 100 will be described in the order of procedures actually performed.

In step S1, M number of microphones 20 ₁ to the frequency domain conversion unit 10, ..., 20 mixed signals observed time domain by _M ~ x (t) is input. The mixed signal ~ x (t) in the time domain is defined by equation (2).

Here, ~ x _m (t) represents a time-domain mixed signal observed by the m-th microphone 20 _m , t represents a time index, and superscript T represents transposition of a vector.

  The frequency domain transform unit 10 performs time frequency transform such as short-time Fourier transform on the input time domain mixed signal to x (t) to generate a time frequency domain mixed signal x (τ, f). And output. The output mixed signal x (τ, f) in the time frequency domain is input to the activity calculator 12. The mixed signal x (τ, f) in the time-frequency domain is defined by Equation (3).

Here, x _m (τ, f) represents the time-frequency conversion of the time domain mixed signal to x _m (t) observed by the m-th microphone 20 _m , τ represents the time frame number, and f represents the frequency. Represents the bin number.

  In step S2, the activity calculation unit 12 calculates and outputs an activity y (ζ) for the input mixed signal x (τ, f) in the time frequency domain. The output activity y (ζ) is input to the clustering unit 14.

In the first embodiment, as the activity of the separated signal, the time series of the contribution (posterior probability) of the separated signal for each time interval, that is, the mixed signal x (τ, f) in the time frequency domain is k (1 ≦ k ≦ K) Use the posterior probability η _kτf to which the th cluster contributes. As a method for obtaining the posterior probability η _kτf , for example, a method disclosed in “JP 2009-053349 A (Reference Document 1)” can be used.

The total number K of clusters is set to be equal to or greater than the upper limit of the number of signal sources assumed. A time series [η _k1f ... η _kTf ] ^{T of} posterior probabilities can be defined for each cluster k in each frequency bin f. Here, T appearing in the subscript represents the total number of frames, and T of the superscript represents the transposition of the vector. A serial number defined by ζ = k + (f−1) K is assigned to these, and activity y (ζ) is defined by equation (4).

  The total number of activities is L: = K × F. Here, F is the total number of frequency bins.

  The posterior probability may be obtained for each frame as in the above example, or may be obtained for every several frames.

  In steps S3 to S9, the clustering unit 14 applies the input activity y (ζ) to a probability model representing the distribution of activities, and calculates model parameters of the probability model using a predetermined evaluation function for evaluating the probability model. Thus, clustering is performed.

  Since the rise and fall of the amplitude of the signal shows a unique pattern for each signal source, the activities of the separated signals corresponding to the same signal source are similar to each other. In the present invention, this property is used to cluster activities. By collecting the activities of separated signals originating from the same signal source into one cluster, the number of signal sources can be estimated based on the number of activity clusters (hereinafter also referred to as activity clusters).

  The probability model representing the activity distribution is modeled as follows, for example. For example, the activity distribution relating to the i-th cluster is modeled by a Watson distribution as shown in Expression (5).

  Here, Γ is a gamma function. Also, z (ζ) is normalized as shown in the equation (6) so that the norm is 1 as preprocessing.

Here, || · || represents the Euclidean norm. w _i represents the center of the distribution of activity z (ζ) after normalization with respect to cluster i, which is called a mean orientation, and satisfies the constraint condition || w _i || = 1. к _i represents the degree of concentration of the distribution of activity z (ζ) after normalization with respect to cluster i, and is called a density parameter. M (a, b, x) is a Kummer function. For details on the Kummer function, see, for example, “S. Sra and D. Karp,“ The multivariate Watson distribution: Maximum-likelihood estimation and other aspects ”, Journal of Multivariate Analysis, vol. 114, pp. 256-269, Feb. See 2013. (Reference 2).

  As described above, the likelihood function of the activity z (ζ) after normalization is given by the mixed model represented by the equation (7).

Here, α _i is a mixing weight, and α _i satisfies the constraint condition that Σ _{i = 1} ^K α _i = 1. Θ is a parameter set shown in the equation (8).

Here, for example, a notation such as {α _i } _i represents a set {α _i | ∀i}. Note that i represents an index of an activity cluster and should be distinguished from k, which is an index of a cluster of time frequency components.

  In the first embodiment, for the sake of simplicity, it is assumed that the number of activity clusters is the same number K as the number of clusters of time frequency components, but it is not always necessary to have the same number. The number of activity clusters can be freely set as long as the number of activity clusters is equal to or greater than the upper limit of possible signal sources.

  As already mentioned, since the activity shows a unique pattern for each signal source, the activity distribution has a plurality of peaks corresponding to each signal source. In order to estimate the number of signal sources based on the number of activity clusters, it is important to fit one probability distribution model to each peak of the activity distribution. Therefore, the Dirichlet prior distribution given by the equation (9) is given to the mixture weight as a prior distribution.

  Here, Γ is a gamma function, and φ is a hyperparameter. By defining φ in the range of 0 <φ <1, it is possible to make it easy for only a small number of mixing weights to take a large value. A uniform prior distribution is assumed for parameters other than the mixture weight. Therefore, the prior distribution of the parameter set Θ is expressed by Expression (10).

  Here, the symbol ∝ represents proportionality.

  The clustering unit 14 applies the activity z (ζ) after normalization to the probability model modeled as described above, and uses a predetermined evaluation function for evaluating the probability model, and uses the posterior probability and the optimum parameter set Θ. And clustering activities based on the result.

  Hereinafter, the processing of the clustering unit 14 will be described more specifically.

  The clustering unit 14 includes, for example, a normalization unit 140, a posterior probability calculation unit 142, a parameter update unit 144, a parameter holding unit 146, and a cluster determination unit 148, as shown in FIG. The parameter updating unit 144 further includes a mixed weight updating unit 1440, a correlation matrix updating unit 1442, an average direction updating unit 1444, a density parameter updating unit 1446, and a weight updating unit 1448.

In step S <b> 3 shown in FIG. 4, the clustering unit 14 stores the initial value of the parameter set Θ composed of the mixture weight α _i , the average direction w _i , and the density parameter к _i in the parameter holding unit 146. For example, the initial value may be set as shown in Expression (11).

In addition, w _i may be selected at random from {z (ζ)} _ζ .

  In step S4, the normalization unit 140 normalizes the input activity y (ζ) so that the Euclidean norm becomes 1 as shown in Expression (12) as pre-processing prior to clustering, and the calculation result is obtained. Output as activity z (ζ) after normalization.

In step S <b> 5, the posterior probability calculation unit 142 outputs the activity z (ζ) after normalization output from the normalization unit 140 and the parameter set Θ ′: = {{α ′ _i } _i stored in the parameter holding unit 146. , {w ′ _i } _i , {к ′ _i } _i }, the posterior probability γ _i (ζ) related to the activity z (ζ) after normalization is calculated by the equation (13) and output. Here, the posterior probability γ _i (ζ): = p (i | z (ζ), Θ ′) is the posterior probability that the activity z (ζ) after normalization belongs to the i-th cluster.

In step S6, the mixture weight updating unit 1440 of the parameter updating unit 144 outputs the posterior probability γ _i (ζ) output from the posterior probability calculating unit 142 and the weight stored in the parameter holding unit 146, as shown in Expression (14). The mixing weight α _i is updated using λ ′.

In step S 6, the correlation matrix updating unit 1442 of the parameter updating unit 144 uses the posterior probability γ _i (ζ) output from the posterior probability calculating unit 142 and the activity z (ζ) output from the normalizing unit 140 to use the cluster. Compute the correlation matrix R _i for _i . The correlation matrix R _i is defined by equation (15).

In step S6, the average direction updating unit 1444 of the parameter updating unit 144 updates the average direction w _i assuming that the eigenvector corresponds to the maximum eigenvalue of the correlation matrix R _i and the Euclidean norm is 1.

In step S6, the density parameter updating unit 1446 of the parameter updating unit 144 updates the density parameter к _i using the average direction w _i and the correlation matrix R _i as shown in Expression (16).

Here, r _i : = w _i ^T R _i w _i .

In step S6, the weight updating unit 1448 of the parameter updating unit 144 updates the weight λ using the mixture weight α _i , the average direction w _i , and the density parameter к _i as shown in Expression (17).

  Hereinafter, the basis for deriving the update formula for each parameter in the parameter update unit 144 will be described. The parameter update is performed based on the EM algorithm. The index i of the Watson distribution is treated as a hidden variable in the EM algorithm. First, an evaluation function G (Θ) for MAP (maximum a posteriori) estimation is given by equations (18) and (19).

Here, as already defined, p (z (ζ) | _i , w _i , к _i ) is given by equation (20), and const. Represents a constant that does not depend on the parameter set Θ.

  A Q function, which is an evaluation function used in the EM algorithm, is given by Expression (21).

The second term of the equation (21) is a term derived from the Dirichlet prior distribution, and has an effect that only a small number of mixture weights α _i take a large value. Furthermore, in order to adaptively determine a more suitable value for φ even if conditions such as the placement of microphones and signal sources change, “S. Medasani and R. Krishnapuram,“ Categorization of image databases for efficient retrieval using robust mixture Following “decomposition”, Computer Vision and Image Understanding, vol. 83, pp. 216-235, 2001. (reference 3), equation (21) is changed and hyperparameter φ is changed as shown in equation (22). Instead, an adaptive hyperparameter λ is introduced.

  The hyperparameter λ is updated by equation (23) for each iteration of the EM algorithm.

  The update equation for each parameter is derived by setting the partial derivative with respect to the parameters of the modified Q function to Q (Θ) as 0. However, if there is a constraint, it is derived by the Lagrange undetermined multiplier method.

First, the blend weight update formula is obtained by solving the simultaneous equations of Formula (25) with respect to Formula (24) while paying attention to the constraint condition Σ _{i = 1} ^K α _i = 1. Here, μ is Lagrange's undetermined multiplier.

The update equation in the average direction is obtained by solving the simultaneous equations of Equation (27) with respect to Equation (26) with the constraint condition || w _i || = 1.

In the density parameter update formula, first, the Q function ~ Q (Θ) is partially differentiated by the density parameter к _i to obtain formula (28).

Where M ′ (1/2, T / 2, к _i ) is the partial differentiation of M (1/2, T / 2, к _i ) by к _i , and R _i is It is defined by equation (29). Also, r _i : = w _i ^T R _i w _i .

“S. Sra and D. Karp,“ The multivariate Watson distribution: Maximum-likelihood estimation and other aspects ”, Journal of Multivariate Analysis, vol. 114, pp. 256-269, Feb. 2013. (reference 4) By the described approximation method, an approximate solution of κ _i is obtained as shown in equation (30).

  In step S <b> 7, the clustering unit 14 determines whether a preset parameter update count maximum value max_iter (for example, 200) has been reached or whether the convergence condition shown in Expression (31) is satisfied.

Here, thr_conv is a threshold value for convergence determination, and is set to 10 ⁻¹⁰ , for example.

  In step S <b> 8, the parameter holding unit 146 stores the parameter set Θ and the weight λ obtained by the update process in the parameter update unit 144. The parameter holding unit 146 provides the held parameter set Θ and weight λ as the parameter set Θ ′ and weight λ ′ in the next processing in the posterior probability calculation unit 142 and the parameter update unit 144.

The clustering unit 14 performs processing in the posterior probability calculation unit 142 and the parameter update unit 144 and updates to the parameter holding unit 146 until the maximum value of the parameter update count is reached in step S7 or the convergence condition shown in Expression (31) is satisfied. Repeatedly read and write data. The posterior probability γ _i (ζ) after the end of the iteration is input to the cluster determination unit 148.

In step S9, the cluster determination unit 148 uses the activity y (ζ) before normalization and the posterior probability γ _i (ζ) after the end of the iteration, to calculate the activity cluster C _i (1 ≦ i ≦ K). (32).

  In the example of the clustering unit 14 described above, as a clustering method having a model selection mechanism for fitting one probability distribution model to each peak of the activity distribution, the Dirichlet prior distribution having hyperparameters adaptive to the mixture weights. A mixed model was used. However, the clustering method having a model selection mechanism is not limited to this, and various methods may be used. For example, a Dirichlet prior distribution having a fixed hyperparameter may be used, or a nonparametric Bayes model ( For example, a Dirichlet process mixture model), information criterion (for example, Akaike information criterion or Bayesian information criterion), cross validation, hypothesis testing, and the like may be used.

  Further, in the example of the clustering unit 14 described above, an example is described in which clustering is performed using activities in all frequency bins, but clustering may be performed using only activities in some frequency bins. The frequency bins to be used are preferably 200 frequency bins included in the frequency band of 4 kHz to 5.56 kHz, but not limited to this, any frequency bin can be selected.

  The signal source number determination unit 16 determines and outputs the signal source number N using the activity cluster obtained by the clustering unit 14. Specifically, the number of remaining activity clusters excluding an activity cluster consisting of small activities (that is, an activity cluster assumed not to correspond to a signal source) is estimated as the signal source number N.

  Hereinafter, the processing of the signal source number determination unit 16 will be described more specifically.

  The signal source number determination unit 16 includes, for example, an activity total calculation unit 160, an activity total clustering unit 162, and an element number counting unit 164 as shown in FIG.

In step S10 shown in FIG. 4, the activity sum calculation unit 160 of the signal source number determination unit 16 uses the input activity cluster C _i (1 ≦ i ≦ K), and the activity sum ρ _{i according} to the equation (33). Calculate (1 ≦ i ≦ K) and output.

  Here, sum (•) represents the sum of the elements of the vector.

In step S11, the activity total clustering unit 162 of the signal source number determination unit 16 classifies the activity clusters C _i into two clusters based on the input activity totals ρ _i . Clustering is performed, for example, using k-means clustering described in “CM Bishop,“ Pattern Recognition and Machine Learning, ”Springer, 2006. (reference 5)” with two clusters. The one where the centroid value is not small is obtained as the cluster D _active corresponding to the signal source. The activity cluster corresponding to the signal source is considered to have a larger activity sum than the activity cluster not corresponding to the signal source. Therefore, the activity sum corresponding to the signal source and the activity sum not corresponding to the signal source are each considered to form a cluster and a cluster having a large centroid value corresponds to the signal source.

In step S <b> 12, the element number counting unit 164 of the signal source number determining unit 16 counts the number of input D _active elements and outputs the number as the signal source number N.

  In the example of the signal source number determination unit 16 described above, the process of clustering into two clusters based on the distance between the activity sums of the activity clusters has been described. However, the activity sum is an example of a feature amount corresponding to the activity size, and other feature amounts may be used instead of the activity sum. For example, the representative value of the activity included in the activity cluster, such as the size (norm) of any one of the activities included in each activity cluster and the average size (norm) of the activities in each cluster. It may be a feature amount calculated using.

  Instead of classifying the two clusters based on the distance between the activity clusters and then obtaining the cluster corresponding to the signal source, the representative value of each activity cluster (or the representative value is the size of the activity cluster). (The value obtained by multiplying the corresponding weight) is determined whether or not the size is equal to or larger than a predetermined threshold, and the number of activity clusters determined to be equal to or larger than the predetermined threshold is output as the number N of signal sources. Also good.

[Modification]
In the first embodiment, after clustering activities by the clustering unit 14, the signal source number determining unit 16 performs processing for eliminating activity clusters that do not correspond to signal sources (that is, activity clusters consisting of small activities). However, this order may be reversed. That is, first, the activity of the separated signal for each frequency bin is clustered into two clusters based on its magnitude (norm), and the cluster with the activity with the smaller norm is excluded as a cluster that does not correspond to the signal source. The cluster having the activity with the larger norm is held as the cluster corresponding to the signal source, and only the activities belonging to the cluster corresponding to the signal source are clustered into the activity cluster by the processing of the clustering unit 14. Good. In this case, since the activity not corresponding to the signal source has already been excluded, the signal source number determination unit 16 simply sets the number of activity clusters whose mixing weight α _i is equal to or greater than a predetermined threshold as the signal source number N. Output.

In the first embodiment, the activity that does not correspond to the signal source is excluded by using the property that the activity that does not correspond to the signal source is smaller than the activity that corresponds to the signal source. It goes without saying that other methods that take advantage of this property may exclude activities that do not correspond to the signal source. For example, first, the activity norm of the separated signal for each frequency bin is calculated as || y (ζ) ||, and the activity y (ζ) whose norm is less than a predetermined threshold is excluded as the activity not corresponding to the signal source. The activity y (ζ) having the norm equal to or greater than a predetermined threshold may be held as an activity corresponding to the signal source, and only the activity corresponding to the signal source may be clustered into an activity cluster by the processing of the clustering unit 14. . Also in this case, since the activity that does not correspond to the signal source has already been excluded, the signal source number determination unit 16 simply determines the number of activity clusters whose mixing weight α _i is equal to or greater than a predetermined threshold as the number N of signal sources. As output.

  Alternatively, instead of explicitly excluding activities that do not correspond to signal sources, activities that do not correspond to signals may be ignored in clustering in the clustering unit 14. That is, the evaluation function ~ Q (Θ) in the clustering unit 14 is weighted as shown in Expression (34) based on the size of the activity.

  Here, ν (ζ) is a weight based on the size of the activity, and is defined by Expression (35).

Here, the index a is a predetermined positive number. The parameter update formula based on formula (34) can also be derived by the same method as in the first embodiment. In this way, if the weight based on the size of the activity is introduced, the clustering in the clustering unit 14 is greatly affected by the activity having a large ν (ζ) (that is, corresponding to the signal source), and a small ν (ζ). ) (I.e., not corresponding to the signal source). Therefore, even if activities that do not correspond to the signal source are not explicitly excluded, clustering can be performed while almost ignoring them. In this case, since the activity that does not correspond to the signal source is almost ignored in the clustering, the signal source number determination unit 16 simply determines the number of activity clusters whose mixing weight α _i is equal to or greater than a predetermined threshold as the signal source number N. As output.

[Second Embodiment]
In the signal source number estimation apparatus 100 according to the first embodiment, the activity calculation unit 12 uses a time series of posterior probabilities as activities. However, the activity used in the present invention is not limited to the time series of posterior probabilities, and a time frequency mask in each frequency bin may be used. Note that the time frequency mask in each frequency bin can be calculated using the power of the separated signal. In the following, a case where a Wiener mask is used as an example of a time frequency mask will be described, but any time frequency mask may be used. For example, “Y. Ephraim and D. Malah,“ Speech enhancement using a minimum-mean square error short-time spectral amplitude estimator ”, IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. ASSP-32, pp. 1109- 1121, Dec. 1984. (reference 6) ”may be used, or“ Y. Ephraim and D. Malah, “Speech enhancement using a minimum mean-square error log-spectral MMSE-logSTSA mask disclosed in "Amplitude estimator", IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. ASSP-33, no. 2, pp. 443-445, Apr. 1985. May be used.

  Hereinafter, the processing of the activity calculation unit when using a Wiener mask as an activity will be described.

The activity calculation unit 13 according to the second embodiment applies an arbitrary signal source separation technique to the input mixed signal x (τ, f) in the time-frequency domain to obtain K separated signals ξ _{kτf εC} . Ask. Here, k is the number of the separated signal, τ is the frame number, f is the frequency bin number, and C is the whole complex number. This signal source separation technique may be a signal source separation technique based on clustering, but any other signal source separation technique may be used. For example, a sound source separation technique based on independent component analysis can be applied. Here, the number K of separated signals is set to be equal to or more than the upper limit of the number of assumed signal sources.

Subsequently, the activity calculation unit 13 calculates the Wiener mask η _kτf using the separated signal ξ _{kτf according} to Expression (36).

Thereby, in each frequency bin f, the activity [η _k1f ... _ΗkTf ] ^T can be defined for the number k of each separated signal. Here, T appearing in the subscript represents the total number of frames, and T of the superscript represents the transposition of the vector. A serial number defined by ζ = k + (f−1) K is assigned to these, and activity y (ζ) is defined by equation (37).

[Experimental result]
An experiment was conducted to evaluate the estimation accuracy of the number of signal sources by the signal source number estimation technique of the present invention. The observation signal was generated by convolving the measured room impulse response with the sound source signal.

  The arrangement of sound sources and microphones in the experimental environment will be described with reference to FIG. Three microphones were placed near the center of the rectangular parallelepiped space so as to draw an equilateral triangle, and four sound sources were placed so as to draw a circle surrounding the microphones. The size of the rectangular parallelepiped space is 4.45 meters wide, 3.55 meters deep, and 2.5 meters high. The regular triangle drawn by the three microphones is 4 centimeters on a side. The radius of the circle drawn by the four sound sources is 1.2 meters. The four sound sources are arranged at positions of 70 °, 150 °, 245 °, and 315 ° counterclockwise with the downward direction in FIG. 7 as 0 °. Three microphones and four sound sources were installed horizontally, and the height from the floor was 1.2 meters.

  Table 1 shows other experimental conditions.

  Preliminary experiments showed that the frequency bin frequency band used for activity clustering was limited to 4 kHz to 5.56 kHz (including 200 frequency bins).

  The clustering of the time frequency component was performed by the method described in “JP 2009-053349 A (reference document 8)”.

  In the experiment, for each combination of the number of signal sources, the number of microphones, the reverberation time, and the number of clusters (K), the number of times that the number of signal sources was correctly estimated after eight trials with different combinations of sound source signals. Was calculated as the correct answer rate.

  Table 2 shows the experimental results when the number of signal sources is two and the number of microphones is three. The sound sources used are the sound sources 1 and 2 shown in FIG. 7, and the microphones used are the microphones 1, 2, and 3 shown in FIG.

  Table 3 shows the experimental results when the number of signal sources is two and the number of microphones is two. The sound sources used are the sound sources 1 and 2 shown in FIG. 7, and the microphones used are the microphones 1 and 2 shown in FIG.

  Table 4 shows the experimental results when the number of signal sources is three and the number of microphones is three. The sound sources used are the sound sources 1, 2, and 3 shown in FIG. 7, and the microphones used are the microphones 1, 2, and 3 shown in FIG.

  Table 5 shows the experimental results when the number of signal sources is three and the number of microphones is two. The sound sources used are the sound sources 1, 2, and 3 shown in FIG. 7, and the microphones used are the microphones 1 and 2 shown in FIG.

  Table 6 shows the experimental results when the number of signal sources is four and the number of microphones is three. The sound sources used are the sound sources 1, 2, 3, 4 shown in FIG. 7, and the microphones used are the microphones 1, 2, 3 shown in FIG.

  As shown in Tables 2 to 4, when the number of signal sources ≦ the number of microphones, an accuracy rate of 100% was obtained regardless of the number of clusters (K) and the reverberation time. Furthermore, as shown in Tables 5 and 6, even when the number of signal sources> the number of microphones, which is a more difficult condition, a high accuracy rate was obtained. In the case of 3 sound sources and 2 microphones in Table 5, when K = 12, a correct answer rate of 100% was obtained regardless of the reverberation time. Even in the most difficult case of 4 sound sources and 3 microphones in Table 6, when K = 8, 10, 12, a high accuracy rate of 75% or higher was achieved regardless of the reverberation time.

  From the above experimental results, it was confirmed that a sufficient effect can be obtained by the signal source number estimation technique of the present invention.

[Program, recording medium]
The present invention is not limited to the above-described embodiment, and it goes without saying that modifications can be made as appropriate without departing from the spirit of the present invention. The various processes described in the above-described embodiments are not only executed in time series according to the order described, but may be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes.

  When various processing functions in each device described in the above embodiment are realized by a computer, the processing contents of the functions that each device should have are described by a program. Then, by executing this program on a computer, various processing functions in each of the above devices are realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In this embodiment, the present apparatus is configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.

100 signal source number estimation device 10 frequency domain conversion unit 12, 13 activity calculation unit 14 clustering unit 140 normalization unit 142 posterior probability calculation unit 144 parameter update unit 146 parameter holding unit 148 cluster determination unit 16 signal source number determination unit 160 activity sum Calculation unit 162 Activity total clustering unit 164 Element number counting unit 400 Sound source number estimation device 20 Sound collection means 30 Frequency domain conversion unit 32 Power estimation unit 34 Arrival direction estimation unit 36 Sound source number estimation unit 200 Direction information distribution estimation device

Claims (8)

A clustering unit that clusters the activity of the separated signal acquired from each of the separated signals into a plurality of activity clusters when the observation signal is separated into a plurality of separated signals for each frequency bin;
A signal source number determination unit for obtaining the number of activity clusters determined to correspond to a signal source among the activity clusters as an estimated value of the number of signal sources;
An apparatus for estimating the number of signal sources. The signal source number estimation apparatus according to claim 1,
The signal source number determining unit is a feature corresponding to the activity size among two clusters obtained by clustering the activity cluster based on a feature amount corresponding to the activity size for each activity cluster. The total number of activity clusters included in the cluster having the larger amount is used as the estimated value of the number of signal sources. The signal source number estimation apparatus according to claim 1,
The signal source number determination unit uses the total number of activity clusters whose feature quantity corresponding to the activity size for each activity cluster is equal to or greater than a predetermined threshold as an estimate of the number of signal sources. Number estimation device. The signal source number estimation device according to any one of claims 1 to 3,
The activity is the contribution of the separated signal for each time interval. The signal source number estimation device according to any one of claims 1 to 3,
The activity is a time frequency mask obtained using the power of the separated signal. The signal source number estimation device according to any one of claims 1 to 5,
The clustering unit clusters the activity into a plurality of activity clusters by fitting the activity to a probability model representing an activity distribution and calculating model parameters of the probability model using an evaluation function for evaluating the probability model. The signal source number estimation device. When the clustering unit separates the observation signal into a plurality of separated signals for each frequency bin, a clustering step of clustering the activities of the separated signals acquired from each of the separated signals into a plurality of activity clusters;
A signal source number determination unit that determines the number of activity clusters determined to correspond to a signal source among the activity clusters as an estimate of the number of signal sources; and
A method for estimating the number of signal sources.   A program for causing a computer to function as the signal source number estimation apparatus according to claim 1.

Images