JP2006243664A - Signal separation device, signal separation method, signal separation program, and recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To separate and extract a target signal with small distortion from a mixed signal obtained by mixing signals emitted from a plurality of signal sources.
First, the window function is multiplied by the mixed signal in each frame set by 1 / S (S> 2) times shift of the length T of the window function, and each calculation result is calculated in the time-frequency domain. Convert to signal. Next, a cluster is generated by clustering the feature amounts extracted from these signals in the time frequency domain, and a mask for each time frequency is generated using the cluster. Then, using the mask and the time-frequency domain signal, a time-frequency domain separation signal is extracted for each time frequency, and the time-frequency domain separation signal is converted into a time-domain separation signal. The separated signals of the areas are added and synthesized.
[Selection] Figure 2

Description

  The present invention relates to a technical field of signal processing, and more particularly, to a technique for estimating and separating and extracting a target signal in a situation where other noises are superimposed on the target signal and observed.

  As a signal extraction technique using a plurality of sensors, a beamformer (also referred to as beamforming) is widely known (for example, see Non-Patent Document 1). However, the beamformer requires information such as the direction of the target signal and a time interval during which the target signal is not active, and if such information cannot be accurately given (or cannot be estimated), the accuracy of signal extraction is low. Another technique is blind signal separation (BSS) by independent component analysis (ICA) (see, for example, Non-Patent Document 2). BSS is superior in that it does not require the information required by the beamformer. However, in both the beamformer and the BSS, signal extraction can be performed accurately only when the number of sensors M is equal to or greater than the number of signals N (number of target signals + number of noises) (N ≦ M).

On the other hand, as a technique when the number of sensors M is smaller than the number of signals N (N> M), a time frequency mask (meaning a mask for each time frequency, for example, a binary mask having a value of 1 or 0 is widely used) (See Non-Patent Document 3, for example). According to this method, signal extraction is possible even when N> M.
BD Van Veen and KM Buckley, "Beamforming: a versatile approach to special filtering," IEEE ASSP Magazine, pp. 2-24, April 1988 A. Hyvaerinen and J. Karhunen and E. Oja, "Independent Component Analysis," John Wiley & Sons, 2001, ISBN 0-471-40540 O. Yilmaz and S. Rickard, "Blind separation of speech mixture via time-frequency masking," IEEE Trans. On SP, vol. 52, no. 7, pp. 1830-1847, 2004.

However, since the process using the time frequency mask is a non-linear process, the signal extracted by the method using the time frequency mask has a drawback that nonlinear distortion occurs. In particular, in the case of an acoustic signal, this nonlinear distortion is called musical noise and is perceived as audible and unpleasant noise.
That is, the processing by the time frequency mask is performed in the time frequency domain, but in the short-time Fourier transform (STFT) used when the signal is converted into the time frequency domain, the length T of the window function used Often a half window shift (T / 2 shift) is used. Then, the time frequency mask is estimated and applied to the time frequency signal of the observation signal obtained by this T / 2 shift. That is, nonlinear processing using a time-frequency mask is performed every half length of the window function, and the extracted signal suddenly rises or falls depending on the granularity. This is a major factor of distortion. T is an integer equal to the number of samples used for the short-time Fourier transform (preferably an even number). The actual time length of the window function is T / f _s seconds where the sampling frequency is f _s .

  The present invention has been made in view of these points, and an object of the present invention is to provide a technique capable of separating and extracting a target signal with a small distortion.

  In the present invention, in order to solve the above-described problem, first, in each frame (window) set with a 1 / S (S> 2) times shift of the length T of the window function, the window function is multiplied by the mixed signal, Each of the calculation results is converted into a time-frequency domain signal. Then, the feature quantities extracted from the signal in the time frequency domain are clustered to generate a cluster, and a mask for each time frequency is generated using the cluster information. Furthermore, using these masks and time-frequency domain signals, the time-frequency domain separation signal is extracted for each time frequency, and the time-frequency domain separation signal is converted to the time-domain separation signal to support each frame. The time domain separation signals to be added are added and combined (superposition addition).

  Here, in the present invention, each frame is set with a 1 / S (S> 2) times shift of the length T of the window function. That is, conventionally, when obtaining the time-frequency representation of the observation signal, each frame is set with a finer window shift (fine shift) than where each frame was set with the T / 2 shift length. By using such a fine window shift (short window shift length), it is possible to increase the overlapping range and overlapping number between frames on the time axis. In the present invention, the calculation result obtained by multiplying the mixed signal by the window function in each frame is converted into a signal in the time frequency domain, and a separated signal in the time frequency domain is extracted using them. It is converted into a separated signal and added and synthesized. Here, since the frames overlap closely on the time axis, the separated signals converted into the time domain also overlap closely on the time axis. Finally, by adding and synthesizing these, each separated signal corresponding to each frame is averaged at each discrete time, and a smoothing effect of the separated signal can be obtained. As a result, the sudden rise or fall of the separated signal is reduced, and nonlinear distortion is reduced.

In the present invention, it is preferable that the window function is a function in which the added composite value of the window function components included in the time domain separation signal corresponding to each frame is constant at each discrete time. Thereby, fluctuations in the intensity of the added and synthesized signal can be prevented, and a high-quality separated signal can be reproduced.
In the present invention, T and S are preferably powers of 2. Thereby, it becomes possible to use a fast Fourier transform (FFT) for the conversion process to the signal of a time frequency domain, and can speed up a process.

  As described above, according to the present invention, the target signal can be separated and extracted with small distortion even when N> M.

でモデル化される。ここで、式（１）におけるｎはサンプリング時刻を、ｐは掃引のための変数を、それぞれ示している。また、すべての信号はあるサンプリング周波数でサンプリングされ、離散的に表現されるものとする。 Hereinafter, embodiments of the present invention will be described with reference to the drawings.
<Principle>
First, the principle of this embodiment will be described.
In the blind signal separation, original signals emitted from a plurality of signal sources are mixed, and a separated signal that is assumed to be an original signal is extracted from only the observed signals under the condition of being observed by a plurality of sensors.
[Model of mixed signal (observation signal)]
Let N be the number of signal sources and M be the number of sensors. In this embodiment, it is assumed that N> M. In addition, s _i is a signal generated from the i (i = 1,..., N) th signal source i, and h _ji is an impulse from the signal source i to the j (j = 1,..., M) th sensor j. A response. In this case, the signal x _j observed by the sensor j is a convolution mixture of the original signal s _i and the impulse response h _ji.

Modeled with Here, n in the expression (1) indicates a sampling time, and p indicates a variable for sweeping. All signals are sampled at a certain sampling frequency and expressed discretely.

と各周波数での単純混合に近似表現できる。ここで、ｆは周波数であり、f=0,f_s/T,...,f_s (T-1)/Tのような離散値である。また、f_sはサンプリング周波数を表す。また、Ｔは短時間フーリエ変換に用いるサンプル数（整数、好ましくは偶数）である。さらに、ｍはSTFTに用いる窓関数の離散時刻（各フレームに対応する整数）を表す。また、H_ji(f)は信号源ｉからセンサｊまでの周波数応答であり、S_i(f,m)は原信号s_iに短時間離散フーリエ変換を施したものである。式（２）を行列とベクトルとを用いて表記すると、
X(f,m)= H(f)・S(f,m) …(3)
となる。ここで、H(f)は、ｊｉ要素に信号源ｉからセンサｊまでの周波数応答H_ji(f)を持つ（N×M）行列（今後これを混合行列と呼ぶ）、S(f,m)= [S₁(f,m),…,S_N(f,m)]^T及びX(f,m)=[X₁(f,m),…,X_M(f,m)]^Tは、それぞれ、原信号s_i及び観測信号x_jのSTFT結果を要素に持つベクトルである。尚、[*]^Tは[*]の転置ベクトルを示す。 [Time-frequency domain representation]
In the blind signal separation, the original signal is separated and extracted from the observation signal modeled by the convolution mixing as described above, but the problem of the convolution mixing is complicated. Therefore, it is effective to handle the problem after applying short-time Fourier transform (STFT) to Equation (1) to convert the observation signal into the time-frequency domain.
In the time frequency domain, equation (1) is

And an approximate expression for simple mixing at each frequency. Here, f is a frequency and is a discrete value such as f = 0, f _s / T,..., F _s (T−1) / T. F _s represents a sampling frequency. T is the number of samples (integer, preferably even) used for short-time Fourier transform. Further, m represents a discrete time (an integer corresponding to each frame) of the window function used for STFT. H _ji (f) is a frequency response from the signal source i to the sensor j, and S _i (f, m) is a signal obtained by subjecting the original signal s _i to a short-time discrete Fourier transform. When Expression (2) is expressed using a matrix and a vector,
X (f, m) = H (f) ・ S (f, m) (3)
It becomes. Here, H (f) is an (N × M) matrix having frequency response H _ji (f) from the signal source i to the sensor j in the ji element (hereinafter referred to as a mixing matrix), S (f, m ) = [S ₁ (f, m), ..., S _N (f, m)] ^T and X (f, m) = [X ₁ (f, m), ..., X _M (f, m)] ^T Are vectors having STFT results of the original signal s _i and the observation signal x _j as elements, respectively. [*] ^T indicates a transposed vector of [*].

[Points of this form]
As described above, the non-linear distortion is caused by the sudden rise and fall of the separated signal. In order to reduce this, in the present embodiment, (a) the window shift length when the observation signal is converted into the time-frequency domain signal is shortened, and (b) the superposition is performed when the time-frequency domain signal is returned to the time-domain signal. An addition method is used, and (c) a window function suitable for this superposition addition is used.
Thus, by shortening the window shift length, it is possible to increase the overlapping range and overlapping number between frames on the time axis. Then, by using the superposition addition method when returning the time-frequency domain signal to the time-domain signal, each separated signal corresponding to each frame can be averaged at each discrete time, and a smoothing effect of the separated signal can be obtained. Can do. As a result, the sudden rise or fall of the separated signal is reduced, and nonlinear distortion is reduced. By using a window function suitable for this superposition addition, a high quality separated signal can be obtained.

<Specific example of this embodiment>
Next, a specific example of this embodiment will be described.
[Hardware configuration]
FIG. 1 is a block diagram illustrating a hardware configuration of a signal separation device 1 according to this embodiment.
As illustrated in FIG. 1, a signal separation device 1 of this example includes a CPU (Central Processing Unit) 10, an input unit 20, an output unit 30, an auxiliary storage device 40, a RAM (Random Access Memory) 50, a ROM (Read Only). Memory) 60 and bus 70.

  The CPU 10 in this example includes a control unit 11, a calculation unit 12, and a register 13, and executes various calculation processes according to various programs read into the register 13. The input unit 20 in this example is an input port for inputting data, a keyboard, a mouse, and the like, and the output unit 30 is an output port for outputting data, a display, and the like. The auxiliary storage device 40 is, for example, a hard disk, an MO (Magneto-Optical disc), a semiconductor memory, or the like, and is observed by a signal separation program area 41 storing a signal separation program for executing the signal separation processing of this embodiment and a sensor. And a data area 42 in which various data such as mixed signals in the time domain are stored. The RAM 50 is, for example, an SRAM (Static Random Access Memory), a DRAM (Dynamic Random Access Memory) or the like, and has a signal separation program area 51 in which a signal separation program is written and a data area 52 in which various data are written. Yes. In addition, the bus 70 in this example connects the CPU 10, the input unit 20, the output unit 30, the auxiliary storage device 40, the RAM 50, and the ROM 60 so that they can communicate with each other.

<Cooperation between hardware and software>
The CPU 10 of this example writes the signal separation program stored in the signal separation program area 41 of the auxiliary storage device 40 in the signal separation program area 51 of the RAM 50 in accordance with the read OS (Operating System) program. Similarly, the CPU 10 writes various data such as a mixed signal in the time domain stored in the data area 42 of the auxiliary storage device 40 in the data area 52 of the RAM 50. Further, the CPU 10 stores the address on the RAM 50 in which the signal separation program and various data are written in the register 13. Then, the control unit 11 of the CPU 10 sequentially reads these addresses stored in the register 13, reads a program and data from the area on the RAM 50 indicated by the read address, and sequentially executes the calculation indicated by the program to the calculation unit 12. The calculation result is stored in the register 13.

  FIG. 2 is an example of a block diagram of the signal separation device 1 configured by reading the signal separation program into the CPU 10 as described above. 3A is a block diagram illustrating details of the time-frequency domain conversion unit 130 in FIG. 2, and FIG. 3B is a configuration of the time-domain conversion unit 160 and the superposition addition unit 170 in FIG. It is the block diagram which showed. In these figures, solid arrows indicate actual data flow, and broken arrows indicate theoretical information flow. Further, for the sake of simplification of description, description of the flow of data entering and exiting the control unit 180 is omitted.

  As illustrated in FIG. 2, the signal separation device 1 of the present embodiment mixes the window function in the memory 110 and each frame set with a 1 / S (S> 2) times shift of the length T of the window function. A time frequency domain conversion unit 130 that multiplies the signal and converts each calculation result into a signal in the time frequency domain, a time frequency mask estimation unit 140 that generates a mask for each time frequency, and a signal in the mask and the time frequency domain , A time frequency domain signal extraction unit 150 that extracts a time frequency domain separation signal for each time frequency, a time domain conversion unit 160 that converts a time frequency domain separation signal into a time domain separation signal, and each frame And a superposition adding unit 170 for adding and synthesizing the separated signals in the time domain corresponding to, and a control unit 180 for controlling the entire signal separating apparatus 1.

  Here, the time-frequency mask estimation unit 140 performs clustering of feature amounts extracted from signals in the time-frequency domain, and generates a cluster, and a mask generation that generates a mask for each time frequency using the cluster information. Part 142. The memory 110 includes storage areas 111 to 121, and the control unit 180 includes a temporary memory 181. The memory 110 and the temporary memory 181 in this example are any one of the data area 42 of the auxiliary storage device 40, the data area 51 of the RAM 50, the register 13, or a combination thereof.

Also, as illustrated in FIG. 3A, the time-frequency domain conversion unit 130 includes a window function generation unit 131, a shift adjustment unit 132 that adjusts the shift amount S, counters 133 to 136, and a discrete time r + m. A T / S shift unit 137 that extracts a mixed signal at T / S (r is an integer), and a multiplier 138 that multiplies the mixed signal at the discrete time r + m · T / S by the value of the window function at the discrete time r. And a discrete Fourier transform unit 139 that performs a discrete Fourier transform on the calculation result in the multiplication unit 138.
Further, as illustrated in FIG. 3B, the time domain transforming unit 160 has a function of transforming a time frequency domain signal into a time domain signal by discrete Fourier inverse transform, And a function of superposing and adding the plurality of time domain signals.

<Processing>
Next, processing of the signal separation device 1 in this embodiment will be described. The following processes are performed under the control of the control unit 180, and data in the middle of calculation is used for each calculation process while being read from and written to the temporary memory 181 unless otherwise specified.
[Overall processing]
FIG. 4 is a flowchart for explaining the entire processing of the signal separation device 1 according to this embodiment. Hereinafter, the entire processing of the signal separation device 1 according to the present embodiment will be described with reference to this flowchart.
The input to the signal separation device 1 is an observation signal vector x (n) whose elements are time-domain mixed signals x _j (n) (j = {1,..., M}) observed by M sensors j. ) = [x ₁ (n), ..., x _M (n)] ^T. In this example, these time domain mixed signals x _j (n) are stored in the storage area 111 of the memory 110 in association with the corresponding sensor j and sampling time n. Further, the length T of the window function to be used is stored in the storage area 112 of the memory 110.

When signal separation is started, first, an observation signal vector x (n) = [x ₁ (n), ..., x _M (n)] ^T is input to the time-frequency domain transform unit 130, and the time-frequency domain The conversion unit 130 converts the observation signal vector X (f, m) = [X ₁ (f, m), ..., X _M (f, m) in the time-frequency domain by short-time Fourier transform (STFT). ] Convert to ^T. In the case of this example, the time frequency domain transform unit 130 first reads the mixed signal x _j from the storage area 111 of the memory 110. The time-frequency domain transforming unit 130, to which the time-domain mixed signal _xj is input, performs the window function in each frame set by a 1 / S (S> 2) times shift of the length T of the window function w (n). Multiplying the mixed signal x _j (n) by w (n) and converting each calculation result into a signal X _j (f, m) in the time-frequency domain, and storing them in the storage area 115 of the memory 110 ( Step S1). Note that m is an integer corresponding to each frame. Details of the processing of the time-frequency domain transform unit 130 will be described later.

Next, the observation signal vector X (f, m) = [X ₁ (f, m), ..., X _M (f, m)] ^T is input to the time-frequency mask estimation unit 140, and the time-frequency mask estimation is performed. The unit 140 uses this observation signal vector X (f, m) = [X ₁ (f, m), ..., X _M (f, m)] ^T to separate and extract a mask M (f, m) = [M ₁ (f, m), ..., M _K (f, m)] ^T is estimated for each time frequency. Here, M _k (f, m) (k = {1,..., M}) is a mask for extracting the k-th separated signal, and K is the number of signals to be separated and extracted (≦ N). is there. In this example, first, the clustering unit 141 reads the signal X _j (f, m) (j∈ {1,..., M}) in the time frequency domain from the storage area 115 of the memory 110, and then the feature quantity. θ (f, m) is extracted and stored in the storage area 116 of the memory 110 (step S2). Next, in the clustering unit 141, the feature quantity θ (f, m) is read from the storage area 115 of the memory 110, the feature quantity θ (f, m) is clustered to generate a cluster, and each cluster is specified. Information θ _k ^˜ (k = {1,..., K}) for storing is stored in the storage area 117 of the memory 110 (step S3). Next, the mask generator 142 generates a mask M _k (f, m) for extracting the k-th separated signal for each time frequency, and stores these in the storage area 118 of the memory 110 (step S4). Details of the processing of the time-frequency mask estimation unit 140 will be described later.

Next, in the time-frequency domain signal extraction unit 150, the separated signal in the time-frequency domain
Y (f, m) = M (f, m) X _J (f, m)
Is calculated. Here, Y (f, m) is a separated signal vector Y (f, m) = [Y ₁ (f, m), ... with the separated signal Y _k (f, m) in the time-frequency domain as an element. , Y _K (f, m)] ^T. X _J (f, m) is one of the signals X _j (f, m) in the time frequency domain, and J∈ {1,..., M}. M (f, m) is [M ₁ (f, m), ..., M _K (f, m)] ^T. In this example, first, in the time-frequency domain signal extraction unit 150, the mask M _k (f, m) is stored from the storage area 118 of the memory 110, and one of the time-frequency domain signals X _J (f, m) from the storage area 115. Read m). And the time frequency domain signal extraction part 150 uses these for every time frequency.
Y _k (f, m) = M _k (f, m) X _J (f, m)
The time frequency domain separation signal Y _k (f, m) is extracted, and the extracted time frequency domain separation signal Y _k (f, m) is stored in the storage area 119 of the memory 110 (step S5). ).

Here, the example in which the separation signal Y _k (f, m) in the time-frequency domain is extracted by the calculation of Y _k (f, m) = M _k (f, m) X _J (f, m) has been described. However, the separation signal Y _k (f, m) in the time-frequency domain may be extracted using a method in which another signal separation extraction method and the mask M _k (f, m) are combined. For example, methods combining masks for each time frequency and independent component analysis (ICA) (for example, “S. Araki, S. Makino, A. Blin, R. Mukai and H. Sawada,“ Blind separation of more speech than sensors with less distortion by combining sparseness and ICA, "Proc. IWAENC2003, pp. 271-274, 2003.") and methods that combine masks and beamformers for each time frequency (for example, "N. Roman and D. Wang" , "Binaural sound segregation for multisource reverberant environments," Proc. ICASSP2004, pp. 373-386, 2004.) etc. may be used to extract the separation signal Y _k (f, m) in the time-frequency domain .

に変換し、各時間領域の分離信号y_k ^m(n)をメモリ１１０の記憶領域１２０に格納する（図３（ｂ），ステップＳ６）。尚、vは虚数単位であり、g=n−m・R、R=T/Sである。また、時間領域変換部１６０の処理の詳細については後述する。
最後に、重畳加算部１７０において、メモリ１１０の記憶領域１２０から各サンプリング時刻ｎに対応する時間領域の分離信号y_k ^m(n),y_k ^m+1(n),y_k ^m+2(n),...,y_k ^m+S-1(n)を読み出して加算合成し（図３（ｂ））、時間領域の分離信号y_k(n)を算出し、記憶領域１２１に格納する（ステップＳ７）。尚、重畳加算部１７０の処理の詳細については後述する。 Next, in the time domain transforming unit 160, the separated signals Y _k (f, m), Y _k (f, m + 1), Y _k (f, m + 2) in the time frequency domain from the storage area 119 of the memory 110. , ..., Y _k (f, m + S−1) are read out, and these are separated in the time domain by short-time inverse Fourier transform (STIFT).

And the separated signal y _k ^m (n) of each time domain is stored in the storage area 120 of the memory 110 (FIG. 3B, step S6). Note that v is an imaginary unit, and g = n−m · R and R = T / S. Details of the processing of the time domain conversion unit 160 will be described later.
Finally, in the superposition adding unit 170, the separated signals y _k ^m (n), y _k ^{m + 1} (n), y _k ^{m + 2} (in the time domain corresponding to each sampling time n from the storage area 120 of the memory 110. n), ..., y _k ^{m + S-1} (n) are read out and synthesized (FIG. 3 (b)), a time domain separation signal y _k (n) is calculated and stored in the storage area 121 (Step S7). Details of the processing of the superposition adding unit 170 will be described later.

[Details of Processing of Time Frequency Domain Transformer 130]
The time-frequency domain transforming unit 130 extracts a signal from the observed time-domain mixed signal x _j (n) using a window function w (n) having a length T, and performs a discrete Fourier transform on the signal to thereby obtain a time frequency. The region signal X _j (f, m) is calculated. Then, while shifting this window function by a T / S (S> 2) shift that is finer than the conventional (T / 2 shift), the signal X _j (f, m) in the time frequency domain is sequentially calculated. Details of the processing by the time-frequency domain transforming unit 130 will be described below.

First, preprocessing performed by the time-frequency domain transform unit 130 will be described.
First, the time frequency domain transform unit 130 reads the length T from the storage area 112 of the memory 110 in the window function generation unit 131, generates a window function w (n) of length T, and stores it in the storage area 113 of the memory 110. Store. A window function w (n) suitable for this embodiment will be described later. Further, the window function w (n) to be used may be stored in the storage area 113 of the memory 110 in advance without providing the window function generation unit 131.

  Next, the shift adjustment unit 132 determines the shift rate 1 / S (S> 2) and stores the parameter S in the storage area 112 of the memory 110. Here, the shift adjustment unit 132 compares and considers the quality required for the separated signal (how much nonlinear distortion is allowed) and the calculation time allowed for the system (the larger the S, the longer the calculation time), and the appropriate parameter. Determine S and output. Specifically, for example, a table (lookup table) in which the parameter S is specified is stored in the memory 110 using the quality required for the separated signal and the calculation time allowed for the system as keys, and the shift adjustment unit 132 sets the table. With reference to the table, a parameter S suitable for the quality required for the separately given separated signal and the calculation time allowed for the system is determined. Alternatively, the shift adjustment unit 132 may not be provided, and a human may select an appropriate parameter S and store it in the storage area 112 of the memory 110. Note that the above-described length T and parameter S are preferably powers of 2. This is because fast Fourier transform can be used for processing in the discrete Fourier transform unit 139 described later, and the processing speed can be increased.

の演算によって、時間領域の混合信号x_jを時間周波数領域の信号Ｘ_j(f,m)に変換する。尚、Ｒは窓シフト長を示しR=T/Sを満たす。またｖは虚数単位（v²=−１）を示す。以下、この処理をフローチャートを用いて説明する。 Next, the conversion process from the time domain to the time frequency domain by the time frequency domain conversion unit 130 of this embodiment will be described.
The time-frequency domain transforming unit 130

Thus, the mixed signal x _j in the time domain is converted into a signal X _j (f, m) in the time frequency domain. R represents the window shift length and satisfies R = T / S. V represents an imaginary unit (v ² = −1). Hereinafter, this process will be described with reference to a flowchart.

FIG. 5 is a flowchart for explaining the conversion process from the time domain to the time frequency domain by the time frequency domain conversion unit 130.
First, j is initialized to 1 in the counter 133 of the time-frequency domain converter 130 (step S11), m is initialized to 0 in the counter 134 (step S12), and f is initialized to 0 in the counter 136 (step S13). The counter 135 initializes r to 0, and the control unit 180 stores X into which 0 is substituted in the temporary memory 181 (step S14).

Next, the T / S shift unit 137 reads S and T from the storage area 112 of the memory 110, receives m from the counter 134, receives r from the counter 135,
r + m ・ R (R = T / S)
And the result of the calculation is stored in the temporary memory 181 (step S15). Next, the T / S shift unit 137 reads r + m · R from the temporary memory 181, extracts the time-domain mixed signal x _j (r + m · R) from the storage area 111 of the memory 110, and a multiplication unit 138 (step S16). The multiplier 138 further reads the window function w (n) from the storage area 113 of the memory 110, receives r from the counter 135,
w (r) ・ x _j (r + m ・ R)
Is stored in the storage area 114 of the memory 110 (step S17).

Next, the discrete Fourier transform unit 139 reads w (r) · x _j (r + m · R) from the storage area 114 of the memory 110, receives r from the counter 135, receives f from the counter 136, and stores the temporary memory. Read X from 181;
X + w (r) ・ x _j (r + m ・ R) ・ e ^-v2πft
The calculation result is stored in the temporary memory 181 as a new X (step S18).

Next, the control unit 180 determines whether r of the counter 135 is T-1 (step S19). Here, if r = T−1 is not satisfied, r + 1 is set as a new r in the counter 135 (r is counted up), and the process returns to step S15 (step S20). On the other hand, if r = T−1, the discrete Fourier transform unit 139 uses the latest X stored in the temporary memory 181 as the signal X _j (f, m) in the time frequency domain, and the storage area 115 of the memory 110. (Step S21).

Next, the control unit 180 reads a T from the storage area 112 of the memory 110, f of the counter 136 determines whether a _{(T-1) f s /} T ( step S22). If not f = (T−1) f _s / T, the counter 136 sets f + f _s / T as a new f (counts up), and returns to step S14 (step S23). On the other hand, if f = (T−1) f _s / T, the control unit 180 determines whether m of the counter 134 is m _max (the maximum value of m) (step S24). In this example, the value of m _max is determined in advance. If m = m _{max is} not satisfied, m + 1 is set as a new m in the counter 134 (m is counted up), and the process returns to step S13 (step S25). On the other hand, if m = m _max , the control unit 180 determines whether j of the counter 133 is M (step S26). If j = M is not satisfied, j + 1 is set as a new j (counted up) in the counter 133, and the process returns to step S12. On the other hand, if j = M, the process of step S1 is terminated.

[Details of processing of time-frequency mask estimation unit 140]
Next, details of the processing of the time-frequency mask estimation unit 140 will be described.
In the method of this embodiment, signal sparsity is assumed. Here, sparse means that the signal is 0 at most sampling times n. The sparsity of the signal is confirmed by an audio signal, for example. By assuming the sparseness of the signal, it can be assumed that even if there are a plurality of signals, the probability that they are observed overlapping each other at each time frequency point (f, m) is low. Therefore, the observed signal at each sensor at each time frequency point (f, m) is the frequency response H _i (f from the signal S _i (f, m) active at that time frequency point (f, m) to each sensor. ) = [H _1i (f), ..., H _Mi (f)] ^T is observed. Therefore, the observed signal vectors can be clustered by the reflected frequency response H _i (f). Each signal can be separated and extracted by using a mask M (f, m) for each time frequency for extracting only the observation signal X (f, m) corresponding to the time frequency of the member belonging to each class. Hereinafter, a method for generating a mask for each time frequency will be exemplified.

を算出し、それから信号の推定到来方向（Direction of Arrival : DOA）

を算出し、これを特徴量としてメモリ１１０の記憶領域１１６に格納する（ステップＳ２）。尚、Ｊ'はセンサｊの何れか１つ（Ｊ'∈{1,...,M}）である。また、dはセンサｊとセンサＪ'との距離、ｃは信号速度を示す。 First, the clustering unit 141 of the time-frequency mask estimation unit 140 reads the time-frequency domain signal X _j (f, m) from the storage area 115 of the memory 110, and uses these to extract the feature quantity used for clustering. In the case of this example, the clustering unit 141 sequentially reads out the signal X _j (f, m) in the time frequency domain from the storage area 115 of the memory 110 and observes the two sensors (each sensor j and the reference sensor J ′). Phase difference between signals X _j (f, m) and X _{J '} (f, m)

And then the estimated direction of arrival (DOA) of the signal

Is stored in the storage area 116 of the memory 110 as a feature amount (step S2). J ′ is any one of the sensors j (J′ε {1,..., M}). Further, d is a distance between the sensor j and the sensor J ′, and c is a signal speed.

という時間周波数毎のマスクを生成し、これらをメモリ１１０の記憶領域１１８に格納する（ステップＳ４）。このマスクM_k(f,m)は、θ_k ^〜−Δ≦θ(f,m)≦θ_k ^〜＋Δの範囲に属する信号の推定到来方向θ(f,m)に対して１の値をとり、それ以外の範囲に属する信号の推定到来方向θ(f,m)に対して０の値をとるバイナリマスクである。ここで、ΔはマスクM_k(f,m)を用いて抽出できる分離信号の範囲を与える。すなわち、Δを十分小さくすると良い分離抽出性能が得られるが非線型歪みが大きくなる。また、Δを大きくすると、非線型歪みは減少するが分離性能が劣化する。 Next, the clustering unit 141 reads the estimated arrival direction θ (f, m) (corresponding to “feature”) of the signal from the storage area 116 of the memory 110 and uses, for example, the k-means method (for example, “RO Duda, Clustering using PE Hart, and DG Stork (translated by Morio Onoe), Pattern Recognition, John Wiley & Sons, New Technology Communications, ISBN 0-471-05669-3, 2003. . Then, the clustering unit 141 obtains an average value θ _k ^˜ (k = {1,..., K}) (corresponding to “information for identifying each cluster”) of each generated cluster, and stores them in the memory 110 is stored in the storage area 117 (step S3). Then, the mask generating unit 142 reads out the average value theta _k ^~ of each cluster from the storage area 117 of the memory 110,

A mask for each time frequency is generated and stored in the storage area 118 of the memory 110 (step S4). This mask M _k (f, m) has a value of 1 for the estimated arrival direction θ (f, m) of signals belonging to the range of θ _k ^to −Δ ≦ θ (f, m) ≦ θ _k ^to + Δ. The binary mask takes a value of 0 with respect to the estimated arrival direction θ (f, m) of signals belonging to other ranges. Here, Δ gives the range of the separated signal that can be extracted using the mask M _k (f, m). That is, if Δ is sufficiently small, good separation and extraction performance can be obtained, but nonlinear distortion increases. When Δ is increased, nonlinear distortion is reduced, but the separation performance is deteriorated.

を特徴量としてクラスタリングを行うこととしてもよい。また、観測信号ベクトルX(f,m)をクラスタリングすることもできる。 Here, clustering is performed using the estimated arrival direction of the signal as a feature quantity. For example, the observation signals X _j (f, m) and X _{J in} two sensors (each sensor j and a reference sensor J ′) are used. _' Phase difference between (f, m) (Equation (4)) itself and gain ratio of both

It is good also as performing clustering by using as a feature-value. Further, the observation signal vector X (f, m) can be clustered.

[Details of processing of time domain conversion unit 160]
Next, details of the processing of the time domain conversion unit 160 will be described.
The time domain transforming unit 160 transforms the time frequency domain separated signal Y _k (f, m) into a time domain separated signal y _k ^m (n) by inverse discrete Fourier transform. In the case of this example, the time domain conversion unit 160 reads S and T from the storage area 112 of the memory 110, and separates signals Y _k (f, m) and Y _k (f , m + 1), Y _k (f, m + 2),..., Y _k (f, m + S−1) are read out. Then, the time domain conversion unit 160 calculates the following time domain separation signal y _k ^m (n) for each frame m for each _k , and stores it in the storage area 120 of the memory 110 (step S6). ). In the following, v is an imaginary unit, and g = n−m · R and R = T / S.

FIG. 6 is a conceptual diagram illustrating a state in which a time-domain separated signal y _k ^m (n) is generated from a time-frequency domain separated signal Y _k (f, m) and further added and synthesized.
As shown in this figure, by the processing of the time domain transforming unit 160, the separation signal Y _k (in the time frequency domain of each frame {m, m + 1, m + 2,..., M + S−1}). f, m), Y _k (f, m + 1), Y _k (f, m + 2), ..., Y _k (f, m + S−1) are long from n = m It is converted into a time domain separation signal y _k ^m (n) that has a value of T and the rest is zero.

の関係を満たす。すなわち、ここで算出された時間領域の分離信号y_k ^m(n)は、本来の時間領域の分離信号y_k(n)に窓関数成分w(n−m・R)を乗じたものとなっている。これは、時間周波数領域変換部１３０の乗算部１３８において混合信号に窓関数w(n)を乗じていたことに起因するものである。 The right side (m · R ≦ n ≦ m · R + T−1) of the equation (6) is

Satisfy the relationship. In other words, the time domain separation signal y _k ^m (n) calculated here is obtained by multiplying the original time domain separation signal y _k (n) by the window function component w (n−m · R). ing. This is because the multiplication unit 138 of the time-frequency domain conversion unit 130 multiplies the mixed signal by the window function w (n).

この処理により、図６に示すような加算合成された分離信号y_k(n)が生成される。尚、ここでは分離信号y_k(n)の順序は規定されない。すなわちy_k(n)が原信号s_k(n)の推定値とは限らない。 [Details of processing of superposition adding unit 170]
Next, the details of the processing of the superposition adding unit 170 will be described.
In the case of this example, in the superposition addition unit 170, the separation signal y _k ^m of the time domain of each frame {m, m + 1, m + 2,..., M + S−1} from the storage area 120 of the memory 110. (n), y _k ^{m + 1} (n), y _k ^{m + 2} (n), ..., y _k ^{m + S-1} (n) are read out, and these are added and synthesized according to the following equation. The combined result is stored in the storage area 121 of the memory 110 as a time domain separation signal y _k (n) (step 7). In this example, C is a constant determined by the window function w (n) and the shift rate 1 / S (described later).

As a result of this processing, the separated combined signal y _k (n) as shown in FIG. 6 is generated. Here, the order of the separated signals y _k (n) is not defined. That is, y _k (n) is not necessarily an estimated value of the original signal s _k (n).

を満たすものであることが望ましい。 [Window function w (n) suitable for this form]
Next, the window function w (n) suitable for this embodiment will be described.
As described above, the time domain separation signal y _k ^m (n calculated by the time domain transform unit 160 is obtained because the mixed signal is multiplied by the window function w (n) in the multiplication unit 138 of the time frequency domain transform unit 130. ) Is obtained by multiplying the original time domain separated signal y _k (n) by the window function component w (n−m · R) (formula (7)). Therefore, an appropriate window function is provided so that the added and combined portion of the window function component w (n−m · R) when the separated signals y _k ^m (n) in the time domain are added and combined according to Equation (8) does not cause distortion. w (n) must be selected. Specifically, it is assumed that the added composite value of the window function component w (n−m · R) included in the time domain separation signal y _k ^m (n) corresponding to each frame m is constant in a predetermined range on the time axis. It is desirable to use the window function w (n) That is, the window function w (n) is

It is desirable to satisfy.

Examples of window functions that satisfy such conditions include the following.
Hanning window
w (n) = 0.5−0.5 ・ cos (2πn / T) (n = 0, ..., T−1)
Hamming window
w (n) = 0.54−0.46 ・ cos (2πn / T) (n = 0, ..., T−1)
Bartlett window

  FIGS. 7A to 7C show how the Hanning window, the Hamming window, and the Bartlett window are added and synthesized according to the equation (9). FIG. As shown in this figure, the added composite waveforms 202, 212, and 222 obtained by adding and synthesizing the window function waveforms 201, 211, and 221 of each window function according to the equation (9) are 1 in the range of n = 500 to 1300. Yes.

When the Hanning window or Bartlett window described above is used, the constant C in Equation (8) is preferably 0.5S, and when a Hamming window is used, it is preferably 0.54S. Further, when a window function obtained by generalizing a Hanning window or Bartlett window is multiplied by b, the constant C is preferably 0.5 S · b, and when a Hamming window is used, it is 0.54 S · b. It is desirable. Further, if the output level of the separated signal y _k (n) calculated by the equation (8) is not a problem, the constant C is set to a value other than this (for example, the window function w (n) or the shift ratio 1 / S). It may be a value independent of.

<Features of this embodiment>
[Hearing characteristics]
As described above, the separation signal Y _k (f, m) in the time frequency domain is extracted by nonlinear processing using the mask M _k (f, m) for each time frequency, and the time domain of the length T is obtained in the time domain conversion unit 160. The signal y _k ^m (n) is converted into the signal y _k ^m (n) and added and synthesized by the equation (8) to calculate the separated signal y _k (n). Here, the shape of a mask that is superimposed and added in order to confirm qualitatively whether the separation signal y _k (n) is extracted smoothly is shown. This is _obtained by superimposing and adding the mask M _k (f, m) for each time frequency determined at each frequency f and each discrete time m in Expression (5) as in Expression (8). It is expressed.

FIG. 8 is a graph in which the vertical axis represents M _f (n) and the horizontal axis represents sampling time n. Here, FIG. 8 (a) is a graph of M _f (n) in the case of S = 2 (rough shift), FIG. 8 (b), M in the case of S = 8 (fine shift) _f ( It is a graph of n). Also, T = 512 (equivalent to 64 ms for sampling frequency = 8000 Hz), and shift amount R = T / S when S = 2 is 256 (32 ms for sampling frequency = 8000 Hz) The shift amount R = T / S when S = 8 is 64 (corresponding to 8 ms for the sampling frequency = 8000 Hz experiment). Further, M _f (n) in FIG. 8 is _obtained by applying Equation (10) to a mask M _k (f, m) for extracting a signal emitted from the signal source S _{1 under} the experimental conditions in FIG.

From FIG. 8, it can be seen that there is a gap of 32 ms (> 22 ms) in M _f (n) when S = 2 (position (A) and position (C) in FIG. 8A). Here, with regard to the smoothness of the acoustic signal in the time direction, it is said that “a human can recognize a gap of 22 ms or more (a silent section with a sudden rise / fall in a continuous signal)” (for example, “ BCJ Moore, An introduction to the psychology of hearing, 3rd Ed., Academic Press, 1989. "). Therefore, it is quite possible that this gap contributes to musical noise.
On the other hand, when S = 8, although a gap is seen (position (B) in FIG. 8B), its length is 8 ms and is not recognized by humans. Furthermore, the value of M _f (n) gradually changes near the position (D) in FIG. Therefore, the gap near the position (D) is unlikely to cause musical noise. That is, it is considered that the generation of musical noise can be reduced by shortening the shift length.

Also, when looking at the change in the amplitude of M _f (n), one step is as small as 0.5 when S = 2 (coarse shift), and one step is as small as 0.125 when S = 8 (fine shift). . This is because M _k (f, m) ∈ {1, 0}, so that the size of one stage is always 1 / S. That is, when a fine shift is used, the amplitude of M _f (n) does not change abruptly. Therefore, it can be seen that the fine shift and addition synthesis used in this embodiment do not cause a sudden change in the amplitude of the separated signal, which is a cause of musical noise.
As described above, the same effect is obtained in the case of S = 4 as compared to the case of S = 2.

[Experimental results (objective evaluation by separation performance and subjective evaluation by MOS]
FIG. 10 is a table showing a performance comparison of the separated signal y _k (n) when the shift amount S is changed in the present embodiment. Here, SIR (signal-to-interference signal ratio) represents separation performance, SDR (signal-to-distortion ratio) represents the degree of signal distortion, and MOS (Mean Opinion Score) represents auditory quality. In particular, MOS is an average value of points obtained by listening to output signals (sounds) actually extracted by 20 subjects. For any evaluation value, a larger value indicates better performance. FIG. 10A shows the result in an environment without reverberation, and FIG. 10B shows the result when the reverberation time RT ₆₀ = 130 ms.

As the original signal, speech signals from three speakers (two men and one woman) were used (Fig. 9). In FIG. 10A, an observation signal with two omnidirectional microphones in an environment without reverberation was simulated. In FIG. 10 (b), an impulse response observed in a room with 130 ms of reverberation was convolved with the original signal, and an observation signal was created according to equation (1). Also, clustering is performed with DOA, and Hanning window w (n) = 0.5−0.5 · cos (2πn / T) (n = 0,..., T−1) is used as the window function, and in equation (8) C was set to 0.5S.
From the result of FIG. 10, it can be seen that by decreasing the shift length (= increasing S), the values of SDR and 'MOS can be improved without degrading the separation performance (SIR). Here, not only the MOS but also the SDR value has risen. Therefore, the method of this embodiment is expected to be effective not only for removing the acoustic signal musical noise but also for removing non-linear distortion of general signals. In addition, about MOS value, the significance was recognized by analysis of variance.

Furthermore, SIR is improved by making the shift finer. This is presumably because the use of a fine window shift length (fine shift) increases the number of sample points at each frequency and improves the accuracy of clustering. This is one of the effects of fine shift.
FIG. 11 shows the result of comparing the performance depending on the window shift amount when other signal separation and extraction methods are used. Here, FIG. 11A shows a performance comparison when using a method combining the above-described mask M _k (f, m) and independent component analysis (ICA), and FIG. The performance comparison when using the method combining the mask M _k (f, m) and the beam former described above is shown. Both show the results under the above-mentioned reverberant environment. Here again, it can be seen that by making the shift finer (increasing S), the SDR value can be improved without degrading the separation performance (SIR). In addition, it is shown on the homepage of <URL> http://www.kecl.ntt.co.jp/icl/signal/araki/fineshift.html that this embodiment can obtain a good sound.

  The present invention is not limited to the embodiment described above. For example, the various processes shown in FIG. 4 and FIG. 5 are not only executed in time series according to the description, but may be executed by changing the processing order as necessary. In addition, each of these processes may be executed in parallel or individually as required by the processing capability of the apparatus that executes the process.

  Also, in this embodiment, a binary pattern in which the transition from a high level value (“1” in the above example) to a low level value (“0” in the above example) as shown in Expression (5) is a discontinuous mask. It was decided to use a mask. However, it is also possible to use a mask having a smooth shape in which the transition from the high level value to the low level value is continuous (for example, “S. Araki, S. Makino, H. Sawada and R. Mukai”). , "Underdetermined Blind Speech Separation with Directivity Pattern based Continuous Mask and ICA," EUSIPCO2004, pp.1991-1994, Sept. 2004. "). Here, the “smooth-shaped mask” is a mask that tries to smoothly extract a signal that is estimated to have reached the sensor from a physically limited direction without being spatially cut by 1/0. In other words, it can be said to be a spatially smooth mask. On the other hand, the present embodiment described above has an effect of suppressing a rapid change in a signal in the time direction, and it can be said that a mask that is smooth in time is generated. Therefore, it can be said that the distortion of the separated signal can be further reduced by combining a spatially smooth mask with the configuration of this embodiment.

In addition, as a signal separation method of this embodiment, a method in which such a smooth mask and another signal separation extraction method, for example, independent component analysis (ICA) are combined (for example, “S. Araki, S. Makino”). , H. Sawada and R. Mukai, "Underdetermined Blind Speech Separation with Directivity Pattern based Continuous Mask and ICA," EUSIPCO 2004, pp. 1991-1994, Sept. 2004.) may be used.
In this embodiment, N> M. However, the present invention may be applied when N ≦ M. Needless to say, other modifications are possible without departing from the spirit of the present invention.

  Further, the signal separation program describing the above-described processing contents can be recorded on a computer-readable recording medium. The computer-readable recording medium may be any medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, or a semiconductor memory. Specifically, for example, the magnetic recording device may be a hard disk device or a flexible Discs, magnetic tapes, etc. as optical disks, DVD (Digital Versatile Disc), DVD-RAM (Random Access Memory), CD-ROM (Compact Disc Read Only Memory), CD-R (Recordable) / RW (ReWritable), etc. As the magneto-optical recording medium, MO (Magneto-Optical disc) or the like can be used, and as the semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory) or the like can be used.

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

  As another execution form of the signal separation program described above, the computer may directly read the signal separation program from a portable recording medium and execute processing according to the program. Each time the signal separation program is transferred, the processing according to the received program may be executed sequentially. In addition, the above-mentioned processing is executed by a so-called ASP (Application Service Provider) type service that realizes processing functions only by the execution instruction and result acquisition without transferring the signal separation program from the server computer to the computer. It is good also as composition to do. 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.

  According to the present invention, it is possible to accurately separate and extract a target signal even in an environment where various noises, interference signals, and interference signals exist. For example, when the present invention is applied to the audio field, the target speech is separated even in a situation where the input microphone of the speech recognizer is far away from the speaker and the microphone picks up sounds other than the target speaker. By extracting, a speech recognition system with a high recognition rate can be constructed.

The block diagram which illustrated the hardware constitutions of the signal separation apparatus. The block diagram of the signal separation apparatus comprised when a signal separation program is read by CPU. (A) is the block diagram which illustrated the detail of the time frequency domain conversion part in FIG. FIG. 3B is a block diagram illustrating a configuration of a time domain conversion unit and a superposition addition unit in FIG. 2. The flowchart for demonstrating the whole process of a signal separation apparatus. The flowchart for demonstrating the conversion process from the time domain to a time frequency domain by a time frequency domain conversion part. Separation signal Y _k (f, m) in the time-frequency domain separated signals from the time domain y _k ^m (n) is generated, further conceptual diagram thereof is illustrated the manner in which the additive synthesis. (A) is the figure which showed a mode that the Hanning window (Hanning window) was added and synthesize | combined according to Formula (9). FIG. 6B is a diagram illustrating a state in which a Hamming window is added and synthesized according to Equation (9). (C) is the figure which showed a mode that addition synthesis | combination of the Bartlett (triangle) window (Bartlett window) according to Formula (9). A graph in which the vertical axis represents M _f (n) and the horizontal axis represents sampling time n. Here, FIG. 8A is a graph of M _f (n) in the case of S = 2 (coarse shift). (B) is a graph of M _f (n) in the case of S = 8 (fine shift). The figure which showed experiment conditions. The table | surface which showed the performance comparison of isolation | separation signal _yk (n) at the time of changing the shift amount S. FIG. (A) is the table | surface which showed the performance comparison when the method which combined mask _Mk (f, m) and independent component analysis (ICA) was used. (B) is a table showing a performance comparison when using the method combining the mask M _k (f, m) and the beam former described above.

Explanation of symbols

1 Signal Separator 130 Time Frequency Domain Transformer 140 Time Frequency Mask Estimator 150 Time Frequency Domain Signal Extractor 160 Time Domain Transformer 170 Superposition Adder

Claims (7)

A signal separation device that separates a mixed signal obtained by mixing signals emitted from a plurality of signal sources into the signal,
Time for multiplying the mixed signal by the window function in each frame set with 1 / S (S> 2) times shift of the length T of the window function, and converting each operation result into a signal in the time-frequency domain A frequency domain transforming means;
Clustering the feature quantities extracted from the signals in the time-frequency domain, and clustering means for generating a cluster;
Mask generating means for generating a mask for each time frequency using the cluster information;
Using the mask and the signal in the time frequency domain, time frequency domain signal extraction means for extracting a separation signal in the time frequency domain for each time frequency;
Time domain transforming means for transforming the time frequency domain separation signal into a time domain separation signal;
Superimposing and adding means for adding and synthesizing the time domain separation signals corresponding to each frame;
A signal separation device comprising: The signal separation device according to claim 1,
The window function is
The added composite value of the window function components of the time domain separation signal corresponding to each frame is a function that is constant in a predetermined range on the time axis.
A signal separation device. The signal separation device according to claim 1,
T and S are powers of 2,
A signal separation device. The signal separation device according to claim 1,
Time frequency domain transforming means
T / S shift means for extracting the mixed signal at discrete time r + m · T / S (r and m are integers);
Multiplication means for multiplying the mixed signal at the discrete time r + m · T / S by the value of the window function at the discrete time r;
Discrete Fourier transform means for performing a discrete Fourier transform on the operation result in the multiplication means;
A signal separation device comprising: A signal separation method for separating a mixed signal obtained by mixing signals emitted from a plurality of signal sources into the signal,
The time-frequency domain transforming means to which the mixed signal is input multiplies the mixed signal by the window function in each frame set with a 1 / S (S> 2) times shift of the length T of the window function, A procedure for converting each of the calculation results into a signal in the time-frequency domain,
Clustering means clustering the feature quantities extracted from the signal in the time frequency domain, and generating a cluster;
A mask generating means uses the cluster information to generate a mask for each time frequency; and
The time frequency domain signal extraction means uses the mask and the signal of the time frequency domain to extract a separation signal of the time frequency domain for each time frequency,
A time domain converting means for converting the time-frequency domain separation signal into a time-domain separation signal;
A step of superimposing and adding and synthesizing the time domain separated signals corresponding to each frame;
A signal separation method comprising:   A signal separation program for causing a computer to function as the signal separation device according to claim 1.   A computer-readable recording medium storing the signal separation program according to claim 6.

Images