JP6724905B2 - Signal processing device, signal processing method, and program - Google Patents

Description

The present disclosure relates to a signal processing device, a signal processing method, and a program, and particularly, when learning a spatial property of noise in a beam forming process, while using an observation signal of a time section including a target sound, The present invention relates to a signal processing device, a signal processing method, and a program that can provide a method for suppressing the influence.

There is a technique called beamforming that emphasizes a target sound coming from a certain direction by suppressing noise by using a multi-channel input audio signal of a microphone array. There are methods designed and developed for beamforming with various purposes and policies, but a typical method that learns spatial characteristics of noise in the usage environment and performs adaptive processing is the minimum dispersion beamformer ( For example, see Non-Patent Document 1).

The minimum dispersion beamformer learns spatial characteristics of noise to form a directional characteristic so as to suppress noise (sound coming from other directions) while maintaining a gain of 1 with respect to the arrival direction of the target sound. .. However, when the observation signal of the time section including the target sound is used when learning the noise, the learning of the noise is hindered by the target sound itself, and the performance of the beamformer deteriorates. As a countermeasure for that, there is also a method of learning the spatial property of noise using an input signal in a time section in which the target sound is not included.

Harry L. Van Trees, “Optimum Array Processing”, Wiley-Interscience, 2002.

However, in the method that learns the spatial properties of noise using the input signal in the time interval that does not include the target sound, the interfering sound that did not exist in the time interval used for learning is generated while the target sound is playing. When it begins to ring, it cannot suppress the disturbing sound.

The present disclosure has been made in view of such a situation, and when learning the spatial property of noise in the beam forming process, while using the observation signal of the time section including the target sound, the influence of the target sound It is to be able to provide a method of suppressing.

The signal processing device according to the first aspect of the present disclosure is a covariance matrix calculation unit that calculates a covariance matrix used in a minimum-dispersion beamformer from an observation signal in a time section including a target sound, and a steering vector of the target sound. And a steering vector generator for calculating both the noise steering vector and a steering vector of the target sound, a component orthogonal to the subspace spanned by the noise steering vector is obtained, and projected to the subspace orthogonal to the component. a projection matrix, using the covariance matrix and the projection matrix generation unit for calculating a projection matrix for processing a processing-obtained processed covariance matrix the covariance matrix by the projection matrix, the minimum variance beam And a filter generation unit that calculates a filter of the former.
A signal processing device according to the second aspect of the present disclosure is a covariance matrix calculation unit that calculates a covariance matrix used in a minimum variance beamformer from an observation signal in a time section including a target sound, and a steering vector of the target sound. And a steering vector generation unit that calculates both the steering vector of noise and the subspace spanned by the steering vector of the noise, and the projection matrix that projects onto the subspace is calculated as a projection matrix that processes the covariance matrix. And a filter generation unit that calculates a filter of the minimum dispersion beamformer using a processed covariance matrix obtained by processing the covariance matrix with the projection matrix.

The signal processing method according to the first aspect of the present disclosure, the signal processing device, from the observation signal of the time section including the target sound, calculates the covariance matrix used in the minimum variance beamformer, the steering vector of the target sound and calculating both the noise of steering vectors among the steering vector of the target sound, obtains a component orthogonal to the subspace steering vector of the noise stretched, a projection matrix for projecting the subspace orthogonal to the component, wherein calculated as projection matrix for processing the covariance matrix, the projection matrix using the processing covariance matrix obtained by processing the covariance matrix by, comprising the step of calculating a filter of the minimum variance beamformer.

The program according to the first aspect of the present disclosure includes a covariance matrix calculation unit that calculates a covariance matrix used in a minimum-dispersion beamformer from an observation signal in a time period in which a target sound is included, and steering of the target sound. A steering vector generating unit that calculates both a vector and a steering vector of noise, and a steering vector of the target sound, obtains a component orthogonal to a subspace spanned by the noise steering vector, and a subspace orthogonal to the component is obtained. The projection matrix to be projected is calculated as a projection matrix for processing the covariance matrix, and a processing covariance matrix obtained by processing the covariance matrix with the projection matrix is used to obtain the minimum variance. This is for functioning as a filter generation unit that calculates a filter of the beam former.

In the first aspect of the present disclosure, from the observation signal of the time interval containing the target sound, the covariance matrix used in the minimum variance beamformer is calculated, the steering vector of the target sound and the steering vector of noise are calculated, Of the steering vector of the target sound, a component orthogonal to the subspace spanned by the noise steering vector is obtained, and a projection matrix projected onto the subspace orthogonal to the component is calculated as a projection matrix for processing the covariance matrix. Then, the filter of the minimum dispersion beamformer is calculated using the processed covariance matrix obtained by processing the covariance matrix by the projection matrix.
In the second aspect of the present disclosure, a covariance matrix used in the minimum variance beamformer is calculated from an observation signal in a time section including a target sound, and both the steering vector of the target sound and the steering vector of noise are calculated. Then, the subspace extended by the steering vector of the noise is obtained, the projection matrix projected onto the subspace is calculated as a projection matrix for processing the covariance matrix, and the covariance matrix is processed by the projection matrix. The filter of the minimum dispersion beamformer is calculated using the obtained processed covariance matrix.

The program can be provided by transmitting it via a transmission medium or by recording it on a recording medium.

The signal processing device may be an independent device, or may be an internal block forming one device.

According to an aspect of the present disclosure, it is possible to provide a method of suppressing the influence of a target sound while using the observation signal in the time section including the target sound when learning the spatial property of noise in the beamforming process. it can.

Note that the effects described here are not necessarily limited, and may be any effects described in the present disclosure.

1 is a block diagram showing a configuration example of an embodiment of a signal processing device according to the present disclosure. It is a flow chart explaining beam forming processing which a signal processing device performs. It is a figure explaining the characteristic of this beam forming processing. FIG. 19 is a block diagram showing a configuration example of an embodiment of a computer according to the present disclosure.

Hereinafter, modes for carrying out the present technology (hereinafter, referred to as embodiments) will be described. The description will be given in the following order.
1. 1. Description of minimum dispersion beamformer Block diagram of signal processing apparatus according to the present embodiment 3. Beamforming processing flowchart 4. Features of this beamforming process 5. Other simple projection method 1
6. Other simple projection method 2
7. Computer configuration example

<1. Description of minimum dispersion beamformer>
First, the minimum dispersion beamformer will be described.

The minimum-dispersion beamformer learns spatial properties of noise to keep the gain for the arrival direction of the target sound at 1, while suppressing noise (sound coming from directions other than the arrival direction of the target sound) as much as possible. Is a beamforming method for forming directional characteristics. The spatial property of the minimum dispersion beamformer is expressed by the covariance matrix R(ω) described below.

First, let x _m (n) be the acoustic signal observed by the m-th (0<m≦M) microphone of the microphone array configured by M microphones. n represents time. When a short time Fourier transform (STFT) is applied to this acoustic signal x _m (n), a narrow band observation signal X _m (ω,k) converted into the time frequency domain is obtained. Here, ω represents a frequency, and k represents a frame number that identifies a predetermined frame when the acoustic signal x _m (n) is composed of a plurality of frames.

A column vector X(ω,k) consisting of the observed signals X ₁ (ω,k),..., X _M (ω,k) of each of the M microphones is expressed by the following equation (1).

ここで、X^H(ω,ｋ)は、X(ω,ｋ)のエルミート転置を、E[]は、期待値を表す。In each microphone, an acoustic signal is observed with different gain and delay depending on the arrangement and the arrival direction of the sound wave. At this time, the covariance matrix R(ω) is defined by the following equation (2).

Here, X ^H (ω,k) represents the Hermitian transpose of X(ω,k), and E[] represents the expected value.

Now, beamforming can be considered as filtering processing of the column vector X(ω,k) of the observation signal. When the filter is represented by w(ω) and the output signal after the beam forming process is represented by Y(ω,k), between the observation signal X(ω,k) and the output signal Y(ω,k), The following expression (3) holds.

式（３）のｗ（ω）は、M個の要素からなる列ベクトルである。

W(ω) in Expression (3) is a column vector composed of M elements.

The output signal Y(ω,k) after the beam forming process is converted into a time waveform by, for example, the overlap-add method.

There are various beamforming methods depending on the design method of the filter w(ω). The minimum dispersion beamformer, which is one of the various methods, is a method of determining the filter w(ω) by the optimization problem of the following expression (4).

Here, a(ω, θ) is a vector expressing a transfer characteristic (phase difference) of a sound wave coming from the direction θ to each microphone, and is called a steering vector. When a sound wave arrives at a microphone array from a certain direction, the acoustic signal observed by each microphone has a phase difference according to the path of the sound wave. The steering vector expresses the phase difference.

When the delay time until the sound wave is observed by the m-th microphone is τ _m (θ), the steering vector a(ω,θ) is expressed by the following equation (5).

The constraint condition a ^H (ω,θ)w(ω)=1 in the equation (4) means that the gain of the sound wave coming from the direction θ is kept at 1. By this constraint condition, the all-pass characteristic of the target sound is passed. Is guaranteed.

The objective function w ^H (ω)R(ω)w(ω) of the equation (4) represents the variance of the output signal Y(ω,k) represented by the equation (3). The name minimum dispersion beamformer comes from minimizing the dispersion of the output signal Y(ω,k).

式（６）のＲ^−１（ω）は、Ｒ（ω）の逆行列を表す。Now, the optimum solution of the optimization problem of the equation (4) can be expressed in the closed form of the following equation (6).

R ⁻¹ (ω) in Expression (6) represents an inverse matrix of R(ω).

Therefore, according to the equation (6), in order to realize the minimum dispersion beamformer, it suffices to know the steering vector a(ω, θ) and the covariance matrix R(ω).

The theoretical value of the steering vector a(ω, θ) can be calculated if the arrangement of the microphones and the arrival direction of the target sound are determined. Alternatively, it is possible to use actual measurement values measured in advance for some arrival directions, or to use values calculated from them by arithmetic processing such as interpolation.

The covariance matrix R(ω) is estimated from the observed signal in a practical situation. For example, when the covariance matrix R(ω) is obtained by the maximum likelihood estimation using the observation signal in the time section from the frame k _s to the frame k _e (k _s <k _e ), it is calculated by the following equation (7). be able to.

The least-dispersion beamformer is theoretically very simple, but there are problems in practical use. If the observed signal used for learning the covariance matrix R(ω) contains the target sound, there is a problem that the target sound is distorted. This is one of the main reasons that the steering vector a(ω, θ) contains an error. If the steering vector a(ω,θ) contains an error, the all-pass characteristic with respect to the true arrival direction of the target sound is not guaranteed, and rather the target sound is not minimized in order to minimize the objective function of equation (4). It will be suppressed.

In order to avoid this problem, for example, there is a method of estimating the covariance matrix R(ω) from an observation signal in a time section in which the target sound is not sounding (for example, immediately before the target sound starts sounding). This method expects that noise other than the target sound is stationary. Since the covariance matrix R(ω) does not include information about the target sound, it is possible to prevent the target sound from being suppressed by the optimization problem of Expression (4).

However, if such a method of using the observation signal in the time section in which the target sound is not sounding is adopted, it is possible to suppress the noise when there is noise that is heard only while the target sound is sounding. I can't.

That is, in the minimum variance beamformer, if the observation signal used for estimation of the covariance matrix R(ω) contains the target sound, the target sound will be distorted, while the observation signal used for estimation will have the target sound. In the case where is not included, there is a problem that the noise that occurs only while the target sound is sounding cannot be suppressed.

Therefore, the inventors of the present invention have devised a method of suppressing the influence of the target sound while using a signal containing the target sound as an observation signal used for estimating the covariance matrix R(ω). The method will be described below.

<2. Block diagram of a signal processing device according to the present embodiment>
FIG. 1 is a block diagram showing a configuration example of a signal processing device that suppresses the influence of a target sound while using a signal containing the target sound as an observation signal in a minimum dispersion beamformer.

The signal processing device 1 includes a covariance matrix calculation unit 11, a steering vector generation unit 12, a projection matrix generation unit 13, a filter generation unit 14, and a beamforming processing unit 15.

In the signal processing device 1, a short time Fourier transform (STFT) is applied to the acoustic signals x ₁ (n) to x _M (n) observed by a microphone array including M microphones as an input signal. , The narrowband observation signals X ₁ (ω) to X _M (ω) converted into the time-frequency domain are input. Note that the observation signals X ₁ (ω) to X _M (ω) are obtained by omitting the frame numbers of the observation signals X ₁ (ω,k) to X _M (ω,k) described above. Then, the beam forming process is sequentially performed on a plurality of frames.

Further, by performing a short-time Fourier transform on the acoustic signals x ₁ (n) to x _M (n), narrowband observation signals in a plurality of time frequency regions ω ₁ , ω ₂ , ω ₃ ,... X ₁ (ω ₁ ) to X _M (ω ₁ ), X ₁ (ω ₂ ) to X _M (ω ₂ ), X ₁ (ω ₃ ) to X _M (ω ₃ ),. To simplify the description, the signal processing device 1 will be described as a process for the observation signals X ₁ (ω) to X ₄ (ω) in the predetermined time-frequency domain ω, but in reality, the signal processing device 1 does not , The processing described below is performed for each of the time-frequency regions ω ₁ , ω ₂ , ω ₃ ,....

The M observation signals X ₁ (ω) to X _M (ω) input to the signal processing device 1 are supplied to the covariance matrix calculation unit 11 and the beamforming processing unit 15.

The covariance matrix calculation unit 11 calculates the covariance matrix R(ω) from the observation signals X ₁ (ω) to X _M (ω) obtained from the time interval in which the desired target sound is produced, according to the above-described equation (7). ) Is calculated and supplied to the filter generation unit 14.

Incidentally, observation signals X ₁ time interval ringing of the desired target sound _(omega) to X _{M (omega)} is the observed signal X ₁ for a predetermined period that includes both the presence or absence of the target sound _(omega) to X _M ( ω) may be input, and the observation signals X ₁ (ω) to X _M (ω) in which the target sound is generated may be extracted, or obtained from the time interval in which the desired target sound is generated. Only the observed signals X ₁ (ω) to X _M (ω) may be input to the signal processing device 1.

In the steering vector generation unit 12, the arrival direction θ _s of the target sound and the arrival directions θ _n1 , θ _n2 ,..., θ _nN of the target sound, which are detected by using a predetermined method such as a sound source direction estimation technique, are _displayed. Supplied. That is, in this technique, the arrival of the arrival direction theta _s and noise of the target sound direction _{θ n1, θ n2, ··,} θ nN is, both are either detected by any method, target sound as a detection result arrival direction theta _s and noise arrival direction theta _n1 of, theta _n2, · ·, theta _nN is supplied to the steering vector generation unit 12. N represents the number of known noise sources, and the number N of noise sources is smaller than the number M of microphones (0<N<M). Further, the arrival direction θ _s of the target sound is different from the arrival directions θ _n1 , θ _n2 ,..., θ _nN of the noise.

The steering vector generation unit 12 calculates the steering vector a(ω, θ) of each of the target sound and the noise from the information of the arrival direction θ _s of the target sound and the arrival directions θ _n1 , θ _n2 ,..., θ _nN of the noise. To do. That is, the steering vector generation unit 12 calculates the steering vector a(ω, θ _s ) of the target sound and the steering vector a(ω, θ _n1 ), a(ω, θ _n2 ),... , a(ω, θ _nN ) is calculated.

The steering vector generation unit 12 generates the generated target sound steering vector a(ω, θ _s ) and noise steering vector a(ω, θ _n1 ), a(ω, θ _n2 ),..., A(ω, θ _nN ) is supplied to the projection matrix generation unit 13. The steering vector generation unit 12 also supplies the steering vector a(ω, θ _s ) of the target sound to the filter generation unit 14.

The projection matrix generation unit 13 converts the target sound steering vector a(ω, θ _s ) into N noise steering vectors a(ω, θ _n1 ), a(ω, θ _n2 ),..., A(ω , calculates the component q _para on subspace W where theta _nN) is put, in the case of orthogonal decomposed into components q _perp on the orthogonal complement W ^perp, the component q _perp on orthogonal complement W ^perp.

Further, the projection matrix generation unit 13 generates a vector p representing the projection direction from the calculated component q _perp on the orthogonal complementary space W ^perp , and calculates the projection matrix P using the vector p.

Specifically, first, in order to calculate the component q _perp on orthogonal complement W ^perp, projection matrix generating unit 13, for example, by Gram-Schmidt process (Gram-Schmidt), subspace W An orthonormal basis {e ₁ , e ₂ ,..., E _N } is calculated.

The orthonormal basis {e ₁ , e ₂ ,..., E _N } of the subspace W by the Gram-Schmidt orthonormalization method can be calculated by the following equation (8).

According to the equation (8), e ₁ is calculated by a(ω, θ _n1 ) when l=1, and when 2≦l≦N, a(ω, θ _nl ) and e ₁ ,... _,, calculated by using _el-1 .

Next, the projection matrix generation unit 13 uses the calculated orthonormal basis {e ₁ , e ₂ ,..., E _N } of the subspace W to calculate the component q _perp on the orthogonal complementary space W ^perp as follows. It is calculated by the equation (9).

Then, the projection matrix generation unit 13 calculates the vector p expressing the projection direction by normalizing the magnitude of the component q _perp to 1 by the expression (10).

式（１１）のIは、単位行列を表す。Finally, the projection matrix generation unit 13 calculates the projection matrix P by the formula (11) from the vector p obtained by the formula (10).

I in Expression (11) represents an identity matrix.

The calculated projection matrix P is supplied to the filter generation unit 14. In addition, the covariance matrix R(ω) is also supplied from the covariance matrix calculation unit 11 to the filter generation unit 14 as described above.

式（１２）のλは、正の値を持つ正則化パラメータである。正則化パラメータは、物理的にはノイズレベルを意味し、その値は、マイクロホンアレイを含む電気回路で生じる雑音や演算誤差等のノイズフロアに基づいて決定される。The filter generation unit 14 uses the projection matrix P generated by the projection matrix generation unit 13 to calculate a processing covariance matrix S(ω) obtained by processing the covariance matrix R(ω) by the following equation (12). ..

Λ in the equation (12) is a regularization parameter having a positive value. The regularization parameter physically means a noise level, and its value is determined based on a noise floor such as noise or calculation error generated in an electric circuit including a microphone array.

Next, the filter generation unit 14 processes the covariance matrix R(ω) and obtains the processed covariance matrix S(ω) and the steering vector a(ω) of the target sound supplied from the steering vector generation unit 12. , θ _s ) is used to calculate the filter w(ω) of the minimum dispersion beamformer. The filter w(ω) of the minimum-dispersion beamformer is calculated by the following equation (13) in which the covariance matrix R(ω) of equation (6) is replaced with the processed covariance matrix S(ω).

The filter generation unit 14 supplies the calculated filter w(ω) to the beamforming processing unit 15.

The beamforming processing unit 15 uses the filter w(ω) supplied from the filter generation unit 14 to perform beamforming processing on the column vector X(ω) of the observation signal. That is, the beamforming processing unit 15 executes the equation (3) that multiplies the _M observation signals X ₁ (ω) to X _M (ω) input to the signal processing device 1 by the filter w(ω).

The signal Y(ω,k) after the beam forming processing, which is the calculation result of the equation (3), is output from the signal processing device 1 as an output signal.

<3. Beamforming process flow chart>
The beamforming process executed by the signal processing device 1 will be described with reference to the flowchart of FIG.

In the signal processing device 1, the acoustic signals x ₁ (n) to x _M (n) observed by the M microphones are observed signals X ₁ (ω) to X ₁ (ω) to which short-time Fourier transform (STFT) is applied. X _M (ω) is input.

In step S1, the covariance matrix calculation unit 11 calculates the covariance matrix R(ω) from the observation signals X ₁ (ω) to X _M (ω) in the time period in which the desired target sound is sounding. More specifically, the covariance matrix calculation unit 11 uses the column vector X(ω) of the observation signals X ₁ (ω) to X _M (ω) in the time section in which the desired target sound is generated, According to (7), the covariance matrix R(ω) is calculated and supplied to the filter generation unit 14.

In step S2, the steering vector generation unit 12 determines, based on the information of the arrival direction θ _s of the target sound and the arrival directions θ _n1 , θ _n2 ,..., θ _nN of the noise, the steering vector a(ω, θ) is calculated. That is, the steering vector generation unit 12 calculates the steering vector a(ω, θ _s ) of the target sound and the steering vector a(ω, θ _n1 ), a(ω, θ _n2 ),... , a(ω, θ _nN ) is calculated.

The steering vector a(ω, θ _s ) of the target sound is supplied to the projection matrix generation unit 13 and the filter generation unit 14, and the noise steering vector a(ω, θ _n1 ), a(ω, θ _n2 ),... , a(ω, θ _nN ) is supplied to the projection matrix generation unit 13.

In step S3, the projection matrix generation unit 13 sets the steering vector a(ω, θ _s ) of the target sound to the steering vector a(ω, θ _n1 ), a(ω, θ _n2 ),... , a (ω, θ _nN) calculates the component q _para on the subspace W spanned by, in the case of orthogonal decomposed into components q _perp on the orthogonal complement W ^perp, the component q _perp on orthogonal complement W ^perp To do.

In step S4, the projection matrix generation unit 13 generates a vector p expressing a projection direction from the calculated component q _perp on the orthogonal complementary space W ^perp , and further uses the vector p to calculate the projection by the equation (11). Calculate the matrix P. The calculated projection matrix P is supplied to the filter generation unit 14.

In step S5, the filter generation unit 14 uses the projection matrix P generated by the projection matrix generation unit 13 to process the covariance matrix R(ω) according to the equation (12), and the processed covariance matrix S(ω). To calculate.

In step S6, the filter generation unit 14 uses the calculated processing covariance matrix S(ω) and the steering vector a(ω, θ _s ) of the target sound supplied from the steering vector generation unit 12 to calculate the minimum variance. The beam former filter w(ω) is calculated. The calculated filter w(ω) is supplied to the beamforming processing unit 15.

In step S7, the beam forming processing unit 15 uses the filter w(ω) supplied from the filter generation unit 14 to perform the beam forming processing. That is, the beamforming processing unit 15 executes the equation (3) that multiplies the _M observation signals X ₁ (ω) to X _M (ω) input to the signal processing device 1 by the filter w(ω).

The signal Y(ω,k) after the beamforming processing is output from the signal processing device 1 as an output signal, and the output signal Y(ω,k) is output in the subsequent stage of the signal processing device 1 by, for example, a superposition addition method or the like. Is converted into a time waveform.

The beam forming process of FIG. 2 is executed for each of the plurality of time-frequency regions ω ₁ , ω ₂ , ω ₃ ,... After the short-time Fourier transform. In the time direction, the beamforming process of FIG. 2 is repeatedly executed in units of predetermined frames.

In the beam forming process described with reference to FIG. 2, the process of each step is described to be executed in order, but the process order of some steps may be reversed or executed in parallel. It is also possible. In other words, if the filter generation unit 14 can obtain the projection matrix P, the covariance matrix R(ω), and the target sound steering vector a(ω, θ _s ) for the calculation of the filter w(ω), The order of calculating the projection matrix P, the covariance matrix R(ω), and the target sound steering vector a(ω, θ _s ) up to that point is not limited.

<4. Features of this beamforming process>
FIG. 3 is a diagram illustrating the concept of the vector p in the signal processing device 1 when the number M of microphones is 3 and the number N of noise sources is 2.

The vector p expressing the projection direction calculated by the equation (10) is the two noise steering vectors a(ω, θ _n1 ) and a( among the target sound steering vectors a(ω, θ _s ). represents a component orthogonal to the subspace W spanned by ω, θ _n2 ).

The operation of multiplying P=(I-pp ^H ) in the above equation (12) means projection onto a subspace (subspace V described later) orthogonal to the vector p. That is, the matrix S(ω) of the equation (12) is projected onto a subspace orthogonal to the vector p, so that information about noise is extracted from the information on the spatial property of the covariance matrix R(ω). It means that only the information about the target sound is excluded while saving.

Therefore, in the beam forming process of the signal processing device 1, the influence of the target sound is suppressed by removing the component derived from the target sound from the covariance matrix R(ω) using the arrival direction θ _s of the target sound. A processed covariance matrix S(ω) is generated, and beamforming is performed by using it as a covariance matrix. This makes it possible to suppress distortion with respect to the target sound even when using an observation signal in a time section including the target sound.

That is, it is possible to provide a method of suppressing the influence of the target sound while using the observation signal of the time section including the target sound when learning the spatial property of noise in the beamforming process.

Further, regarding the noise source, in the beamforming processing of the signal processing device 1, the processing is performed so that the components due to the noise are retained by using the arrival directions θ _n1 , θ _n2 ,..., θ _nN of the noise. Therefore, the amount of noise suppression is maintained.

<5. Other simple projection method 1>
As described above, the effect of emphasizing the target sound is that, of the steering vector a(ω, θ _s ) of the target sound, N noise steering vectors a(ω, θ _n1 ), a(ω, θ _n2 ), .., which is inferior to the method of generating the projection matrix P that projects onto the subspace orthogonal to the component (vector p) orthogonal to the subspace W spanned by a(ω, θ _nN ) (hereinafter referred to as the present projection method). However, two simpler projection methods will be described.

The first of the simple projection method, N number of noise of the steering vector _{a (ω, θ n1),} a (ω, θ n2), ··, a (ω, θ nN) how to projection to the subspace W that spanned Is.

Specifically, the projection matrix generation unit 13 uses the formula (8) to calculate steering vectors a(ω, θ _n1 ), a(ω, θ _n2 ),..., A(ω, θ _nN ) of N noises. ), the orthonormal basis {e ₁ , e ₂ ,..., E _N } of the subspace W is calculated.

Next, the projection matrix generation unit 13 calculates the projection matrix P by the following equation (14) using the orthonormal basis {e ₁ , e ₂ ,..., E _N } of the obtained subspace W. To do.

Therefore, the first simple projection method does not calculate the vector p, but the steering vectors a(ω, θ _n1 ), a(ω, θ _n2 ),..., A(ω, θ _{nN of} N noises. ) Will be projected on the subspace W.

Also by this method, a blind spot is formed in the arrival direction of noise, and it is possible to suppress the noise given in the arrival direction. However, it is not possible to suppress diffuse noise and reverberation that have no directivity.

On the other hand, the projection method has a mechanism capable of suppressing diffuse noise and reverberation to some extent, as described below.

In this projection method, the subspace orthogonal to the vector p is projected, but the subspace of the projection destination and N noise steering vectors a(ω, θ _n1 ), a(ω, θ _n2 ), · _Does not match the subspace W spanned by a(ω, θ _nN ). That is, assuming that the subspace to which the vector p is projected is V, there is a subspace U that satisfies the following expression (15).

式（１５）のプラスを丸で囲んだ記号は、直和を意味する。

The symbol in which the plus in the formula (15) is circled means a direct sum.

The subspace W mainly contains the information of the direct wave component of the noise to which the direction is given, while the subspace U mainly contains the diffuse noise and the reverberation component which are not given the direction. It is included. In this projection method, the projection is performed on the subspace V including the subspace U, and the filter is optimized so that the variance is minimized. Therefore, not only the noise given in the arrival direction but also the diffuse noise and the reverberation component are to some extent. It is possible to suppress.

On the other hand, since the first simple projection method does not include the subspace U, it is not possible to suppress even diffuse noise and reverberation that have no directivity.

<6. Other simple projection method 2>
The second simple projection method is a method of projecting the target sound in the direction of the steering vector a(ω, θ _s ).

Specifically, the projection matrix generation unit 13 sets a vector obtained by normalizing the magnitude of the steering vector a(ω, θ _s ) of the target sound to 1 by the equation (16) as a vector p.

式（１７）のIは、単位行列を表す。Next, the projection matrix generation unit 13 uses this vector p to calculate the projection matrix P by Expression (17).

I in Expression (17) represents an identity matrix.

The steering vector a(ω, θ _s ) of the target sound is stretched with N noise steering vectors a(ω, θ _n1 ), a(ω, θ _n2 ),..., A(ω, θ _nN ). Since the component on the subspace W is also included, the orthogonal projection using the steering vector a(ω, θ _s ) of the target sound distorts the information regarding noise included in the covariance matrix R(ω). The second simple projection method is a fixed beamformer and is equivalent to a delay-sum beamformer, which is the most basic method.

As described above, the effect of emphasizing the target sound is greatest in the main projection method, and then in the order of the first simple projection method and the second simple projection method.

The signal processing device 1 is configured to be capable of executing any of the projection method, the first simple projection method, and the second simple projection method, and to select the projection method to be executed by user setting or the like as necessary. be able to. Alternatively, any one or two of the main projection method, the first simple projection method, and the second simple projection method may be configured to be executable.

<7. Example of computer configuration>
The series of processes described above can be executed by hardware or software. When the series of processes is executed by software, the programs constituting the software are installed in the computer. Here, the computer includes a computer incorporated in dedicated hardware and, for example, a general-purpose personal computer capable of executing various functions by installing various programs.

FIG. 4 is a block diagram showing a configuration example of hardware of a computer that executes the series of processes described above by a program.

In a computer, a CPU (Central Processing Unit) 101, a ROM (Read Only Memory) 102, and a RAM (Random Access Memory) 103 are connected to each other by a bus 104.

An input/output interface 105 is further connected to the bus 104. An input unit 106, an output unit 107, a storage unit 108, a communication unit 109, and a drive 110 are connected to the input/output interface 105.

The input unit 106 includes a keyboard, a mouse, a microphone and the like. The output unit 107 includes a display, a speaker and the like. The storage unit 108 includes a hard disk, a non-volatile memory, or the like. The communication unit 109 includes a network interface or the like. The drive 110 drives a removable recording medium 111 such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory.

In the computer configured as described above, the CPU 101 loads the program stored in the storage unit 108 into the RAM 103 via the input/output interface 105 and the bus 104 and executes the program, thereby executing the above-described beam. Forming processing is performed.

Further, the CPU 101 performs a process of performing a short-time Fourier transform on the acoustic signals x ₁ (n) to x _M (n) observed by the microphone array, and the output signal Y(ω,k) by the superposition addition method. For example, the processing of converting to a time waveform can also be executed by the above.

In the computer, the program can be installed in the storage unit 108 via the input/output interface 105 by mounting the removable recording medium 111 in the drive 110. Further, the program can be received by the communication unit 109 via a wired or wireless transmission medium such as a local area network, the Internet, or digital satellite broadcasting, and installed in the storage unit 108. In addition, the program can be installed in the ROM 102 or the storage unit 108 in advance.

The program executed by the computer may be a program in which processing is performed in time series in the order described in this specification, or in parallel, or at a necessary timing such as when a call is made. It may be a program in which processing is performed.

The embodiments of the present disclosure are not limited to the above-described embodiments, and various modifications can be made without departing from the scope of the present disclosure.

For example, as the signal processing device 1, it is possible to adopt a form in which all or some of the plurality of embodiments described above are combined.

For example, the processing performed by the signal processing device 1 can be configured as cloud computing in which one device shares and jointly processes one function via a network.

In addition, each step described in the above-described flowcharts can be executed by one device or shared by a plurality of devices.

Further, when one step includes a plurality of processes, the plurality of processes included in the one step can be executed by one device or shared by a plurality of devices.

It should be noted that the effects described in the present specification are merely examples and are not limited, and effects other than those described in the present specification may be present.

Note that the present technology may also be configured as below.
(1)
A covariance matrix calculation unit that calculates a covariance matrix used in the minimum-dispersion beamformer from an observation signal in a time section including the target sound,
A steering vector generation unit that calculates at least one of the steering vector of the target sound or the steering vector of noise,
Using at least one of the calculated steering vector of the target sound or the steering vector of the noise, a projection matrix generation unit that calculates a projection matrix that processes the covariance matrix,
And a filter generation unit that calculates a filter of the minimum dispersion beamformer using a processed covariance matrix obtained by processing the covariance matrix with the projection matrix.
(2)
The steering vector generation unit calculates both the steering vector of the target sound and the steering vector of the noise,
The projection matrix generation unit obtains a component of the target sound steering vector that is orthogonal to the subspace spanned by the noise steering vector, and calculates the projection matrix that is projected to the subspace orthogonal to the component. The signal processing device according to 1).
(3)
The steering vector generation unit calculates both the steering vector of the target sound and the steering vector of the noise,
The signal processing device according to (1) or (2), wherein the projection matrix generation unit obtains a subspace spanned by the steering vector of the noise, and calculates the projection matrix projected onto the subspace.
(4)
The steering vector generation unit calculates a steering vector of the target sound,
The signal processing device according to any one of (1) to (3), wherein the projection matrix generation unit calculates the projection matrix projected in the direction of the steering vector of the target sound.
(5)
Any one of (1) to (4) above, further comprising: a beamforming processing unit that performs beamforming processing of the minimum dispersion beamformer on the observation signal using the filter calculated by the filter generation unit. The signal processing device according to.
(6)
The projection matrix generation unit calculates the projection matrix that excludes only the information about the target sound, while preserving the information about the noise among the information on the spatial properties of the covariance matrix. The signal processing device according to any one of (5).
(7)
The signal processor
Calculate the covariance matrix used by the minimum variance beamformer from the observed signal in the time interval containing the target sound,
Calculate at least one of the steering vector of the target sound or the steering vector of noise,
At least one of the calculated steering vector of the target sound or the steering vector of the noise is used to calculate a projection matrix for processing the covariance matrix,
A signal processing method comprising: calculating a filter of the minimum dispersion beamformer using a processed covariance matrix obtained by processing the covariance matrix by the projection matrix.
(8)
Computer,
A covariance matrix operation unit that calculates a covariance matrix used in the minimum variance beamformer from an observation signal in a time section including the target sound,
A steering vector generation unit that calculates at least one of the steering vector of the target sound or the steering vector of noise,
Using at least one of the calculated steering vector of the target sound or the steering vector of the noise, a projection matrix generation unit that calculates a projection matrix that processes the covariance matrix,
A program for functioning as a filter generation unit that calculates a filter of the minimum dispersion beamformer using a processed covariance matrix obtained by processing the covariance matrix with the projection matrix.

1 signal processing device, 11 covariance matrix calculation unit, 12 steering vector generation unit, 13 projection matrix generation unit, 14 filter generation unit, 15 beamforming processing unit, 101 CPU, 102 ROM, 103 RAM, 106 input unit, 107 output Section, 108 storage section, 109 communication section, 110 drive

Claims (8)

A covariance matrix operation unit that calculates a covariance matrix used in the minimum variance beamformer from an observation signal in a time section including the target sound,
A steering vector generation unit that calculates both the steering vector of the target sound and the steering vector of noise,
Of the steering vector of the target sound, a component orthogonal to the subspace spanned by the noise steering vector is obtained, and a projection matrix projected onto the subspace orthogonal to the component is calculated as a projection matrix for processing the covariance matrix. A projection matrix generator that
And a filter generation unit that calculates a filter of the minimum dispersion beamformer using a processed covariance matrix obtained by processing the covariance matrix with the projection matrix. The projection matrix generation unit, of the steering vector of the target sound, obtains a component orthogonal to the subspace stretched by the steering vector of the noise, to calculate a projection matrix to project in the subspace orthogonal to the component first A second mode for calculating the projection matrix projected in the direction of the steering vector of the target sound,
The filter generation unit, by using a processing covariance matrix obtained by processing the covariance matrix by the projection matrix generated by the first mode or the second mode selected alternatively The signal processing apparatus according to claim 1 , wherein a filter of the minimum dispersion beamformer is calculated . A covariance matrix operation unit that calculates a covariance matrix used in the minimum variance beamformer from an observation signal in a time section including the target sound,
A steering vector generation unit that calculates both the steering vector of the target sound and the steering vector of noise,
A submatrix stretched by the steering vector of the noise is obtained, and the projection matrix projected onto the subspace is calculated as a projection matrix processing unit that processes the covariance matrix,
A filter generation unit that calculates a filter of the minimum dispersion beamformer using a processed covariance matrix obtained by processing the covariance matrix by the projection matrix;
A signal processing device comprising: The projection matrix generation unit obtains a subspace stretched by the steering vector of the noise, a first mode for calculating the projection matrix projected on the subspace, and the projection projected in the direction of the steering vector of the target sound. A second mode for computing a matrix,
The filter generation unit, by using a processing covariance matrix obtained by processing the covariance matrix by the projection matrix generated by the first mode or the second mode selected alternatively , Calculate the filter of the minimum variance beamformer
The signal processing device according to claim 3 . The beam forming processing unit that executes the beam forming process of the minimum dispersion beam former on the observation signal using the filter calculated by the filter generating unit. Signal processing device. The projection matrix generation unit, among the information spatial nature of the covariance matrix, while preserving the information relating to the noise, according to claim 1 for calculating the projection matrix to eliminate only the information about the target sound Signal processing equipment. The signal processing device
Calculate the covariance matrix used by the minimum variance beamformer from the observed signal in the time interval containing the target sound,
Calculate both the steering vector of the target sound and the steering vector of noise,
Of the steering vector of the target sound, a component orthogonal to the subspace spanned by the noise steering vector is obtained, and a projection matrix projected onto the subspace orthogonal to the component is calculated as a projection matrix for processing the covariance matrix. Then
A signal processing method comprising: calculating a filter of the minimum dispersion beamformer using a processed covariance matrix obtained by processing the covariance matrix by the projection matrix. Computer,
A covariance matrix operation unit that calculates a covariance matrix used in the minimum variance beamformer from an observation signal in a time section including the target sound,
A steering vector generation unit that calculates both the steering vector of the target sound and the steering vector of noise,
Of the steering vector of the target sound, a component orthogonal to the subspace spanned by the noise steering vector is obtained, and a projection matrix projected onto the subspace orthogonal to the component is calculated as a projection matrix for processing the covariance matrix. A projection matrix generator that
A program for functioning as a filter generation unit that calculates a filter of the minimum dispersion beamformer using a processed covariance matrix obtained by processing the covariance matrix with the projection matrix.

Images