JP4113169B2 - Method for estimating the number of signal sources, estimation apparatus, estimation program, and recording medium - Google Patents

Description

  The present invention relates to a technique for estimating the number of signals using observation values obtained by observing a mixed unknown number of signals with a plurality of sensors, and more particularly to a technique for correctly estimating the number of signal sources in an actual environment.

となる。ここでPはインパルス応答h_jk(l)の持続時間を、n_j(t)はセンサでのノイズを表す。具体的な例としては、音信号が室内で混合される場合、音源からマイクまでの距離により音が減衰・遅延し、また壁などの反射により残響が発生し、さらにマイクに背景ノイズが付加される。 There has been proposed a method for estimating the number of signal sources by performing short-time Fourier transform on an observation signal mixed with a plurality of sound sources and examining eigenvalues of a spatial correlation matrix in each frequency bin (for example, see Non-Patent Document 1). .)
[Problem formulation]
First, the problem handled by this method is formulated. All signals are sampled at a certain sampling frequency and expressed discretely. Assume that N signals are mixed and observed by M sensors. The following deals with the situation where there is a distance from the signal source to the sensor, the signal is attenuated / delayed, and reaches the sensor via a plurality of paths. Mixing in this situation is convolutional mixing with impulse response h _jk (l) from signal source k to sensor j.

It becomes. Here, P represents the duration of the impulse response h _jk (l), and n _j (t) represents the noise at the sensor. As a specific example, when sound signals are mixed indoors, the sound is attenuated / delayed depending on the distance from the sound source to the microphone, reverberation occurs due to reflection from the wall, etc., and background noise is added to the microphone. The

を求める。ここでfは周波数であり、f=0, (1/L)f_s,...,{(L-1)/L}f_sと離散化されている（f_sはサンプリング周波数)。g(l)は窓関数であり、ハニング窓g(l)=(1/2)(1+cos(2πl/L))などのg(0)にパワーの中心を持つ窓関数を用いることで、X_j(f,τ)は時刻t=τを中心とする観測信号x_j(t)の周波数特性を表現する。X_j(f,τ)はLサンプルにわたる情報を含んでいるため、すべての時間τに対してX_j(f,τ)を求める必要はなく、適当な間隔（例えばL/2やL/4）の時間τ毎にX_j(f,τ)を求める。 [Method based on eigenvalues]
Next, the method for estimating the number of signal sources proposed in Non-Patent Document 1 will be described step by step. It is assumed that the number of sensors is equal to or more than the number of signal sources, that is, N ≦ M.
First, a time series for each frequency is applied to the observation signal x _j (t) at sensor j by applying a short-time discrete Fourier transform of L points.

Ask for. Here, f is a frequency, and is discretized as f = 0, (1 / L) f _s ,..., {(L-1) / L} f _s (f _s is a sampling frequency). g (l) is a window function. By using a window function having a center of power at g (0) such as Hanning window g (l) = (1/2) (1 + cos (2πl / L)) , X _j (f, τ) represents the frequency characteristic of the observation signal x _j (t) centered at time t = τ. Since X _j (f, τ) contains information over L samples, it is not necessary to obtain X _j (f, τ) for every time τ, and an appropriate interval (for example, L / 2 or L / 4) ) X _j (f, τ) is obtained every time τ.

と各周波数での単純混合に近似表現できるからである。ここで、H_jk(f)は信号源kからセンサｊまでの周波数応答、S_k(f,τ)やN_j(f,τ)は式(2)と同様の式に従って源信号s_k(t)やノイズn_j(t)に短時間離散フーリエ変換を施したものである。 Operation in the frequency domain is effective for convolution mixed signals. The convolutional mixing in the time domain expressed by Equation (1) is

This is because it can be approximated to simple mixing at each frequency. Here, H _jk (f) is the frequency response from the signal source k to the sensor j, and S _k (f, τ) and N _j (f, τ) are the source signals s _k ( t) and noise n _j (t) are subjected to short-time discrete Fourier transform.

Next, for X (f, τ) = [X ₁ (f, τ), ..., X _M (f, τ)] ^T , the correlation matrix R (f) = <X (f, τ) X (f, τ) ^H > _τ is calculated, and this is subjected to eigenvalue decomposition as R (f) = V (f) · Λ (f) · V (f) ^H. Note that V (f) = [v ₁ (f), v ₂ (f), ..., v _M (f)], and Λ (f) is λ ₁ (f), λ ₂ (f), ..., a diagonal matrix of M rows and M columns with λ _M (f) as diagonal elements.・^H is the operation to find the conjugate transpose of the matrix, 〈〉 _τ is the average over time τ, v _j (f) is the eigenvector (M-dimensional vertical vector), λ _j (f) is the corresponding eigenvalue Yes, they are sorted in the order of λ ₁ (f) ≧ λ ₂ (f) ≧ ... ≧ λ _M (f). Each eigenvalue λ _j (f) is expressed as [Y ₁ (f, τ), ..., Y _M (f, τ)] ^T ← V (f) ^H・ [X ₁ (f, τ),. .., X _M (f, τ)] indicates the power value of the j-th signal Y _j (f, τ) when ^T.

Then, the number N of eigenvalues having a dominant value among the decomposed eigenvalues is estimated as the number of signal sources, and the size of the remaining MN eigenvalues is estimated as the noise power value σ _n (f) ² (Λ _{N + 1} (f) =... = Λ _M (f) = σ _n (f) ² ).
Kiyoshi Yamamoto, WFG van Rooijen, EY Ling, Tai Asano, Takeshi Yamada, Nobuhiko Kitawaki, “Application of Sound Source Number Estimation Method Using SVM to Sound Source Separation System”, Acoustical Society of Japan 2002 Autumn Meeting, 2-5 -10, pp. 537-538, September 2002 "

However, the conventional method based on eigenvalues has a problem that the actual number of signal sources may not be estimated correctly.
For example, when the above-described method based on eigenvalues is used in a realistic situation, the following two problems must be considered.
The first problem is the effect of reverberation. In general, since the reverberation length is longer than the frame length L of the short-time discrete Fourier transform, a certain time component of a certain signal affects a plurality of frames. As a result, the number of dominant eigenvalues may be estimated more than the actual number of signals.

FIG. 11A shows normalized power values of eigenvalues at each frequency when only one sound source is played under the conditions shown in FIG. As shown in this figure, the power value of the eigenvalue taking the second largest value due to the effect of reverberation is about −20 dB. In the case of the above-described method based on the eigenvalue, the number of eigenvalues having a value larger than a predetermined threshold value is determined as the number of signal sources. When this threshold value is smaller than −20 dB, the above-described 2 The eigenvalue with the second largest value is counted as one of the “dominant eigenvalues”, and the number of sound sources is estimated to be two. That is, due to the effects of reverberation, the number of sound sources cannot be estimated accurately unless this threshold is set to a certain large value.
The second problem is that the power of each signal may not appear properly in the eigenvalue. This problem is particularly noticeable at low frequencies where the phase difference is small.

  FIG. 11B shows normalized power values of eigenvalues at each frequency when all three sound sources are played under the conditions shown in FIG. In this example, since the number of sound sources is 3, there should be three dominant eigenvalues. However, as shown in this figure, although the power values of the sound sources are set to be equal, the power value of the eigenvalue gradually decreases as it becomes the second and third. This tendency becomes more prominent as the frequency becomes lower. Therefore, in order to estimate that there are three sound sources at many frequencies in this situation, the above threshold value must be set to about −30 dB or less. However, if the threshold is set small, the eigenvalue corresponding to reverberation is also counted as a “dominant eigenvalue”, and for example, in the case of one sound source in FIG.

As explained above, the method based on the eigenvalues of the prior art cannot accurately estimate the actual number of signal sources due to the effect of reverberation and the problem that the power of each signal does not appear properly in the eigenvalues. There is.
The present invention has been made in view of these points, and an object thereof is to provide a technique capable of correctly estimating the number of signal sources even in an actual environment.

In the present invention, in order to solve the above-described problem, first, observation signals x _j (t) (j = {1,..., M}) from M sensors are converted to time-series data X _j (f , τ), and generates a separated signal Y _i (f, τ) (i = {1, ..., M}) from the time series data X _j (f, τ) and stores it in the storage unit . Next, the power value of each separated signal Y _i (f, τ) is obtained and stored in the storage unit, and the correlation value of the envelope with respect to the time difference Δτ between the different separated signals Y _i (f, τ) is calculated. Store in the storage. Note that the envelope of separated signals Y _i (f, τ), the envelope of the absolute value of the separated signal _{| Y i (f, τ)} | means. Then, the power value and envelope correlation value of each separated signal Y _i (f, τ) are compared with a plurality of parameters stored in the storage unit, and the separated signal Y _i (f, τ) is a source signal component. It is determined whether or not.

  Here, since the source signal and the reverberation signal have a correlation, the envelope correlation value between them is high. Also, the reverberation signal has less power than the corresponding source signal. That is, the reverberation signal has a high envelope correlation value and a relatively small power value. In the present invention, paying attention to this feature, the envelope correlation value or power value of the separated signal is compared with a plurality of parameters indicating each threshold value to determine whether or not the separated signal is a source signal. This makes it possible to estimate the number of signal sources in consideration of the actual environment, compared to the case where the source signal is determined using only the power value of the eigenvalue as an index.

  As described above, in the present invention, the envelope correlation value and the power value of the separated signal are calculated, and the source signal is determined by comparing them with a plurality of parameters. It becomes possible to estimate correctly.

Embodiments of the present invention will be described below with reference to the drawings.
[First Embodiment]
First, a first embodiment of the present invention will be described.
<Overall configuration>
FIG. 1 is a block diagram showing an entire estimation apparatus 1 in this embodiment.
The estimation apparatus 1 causes a Neumann computer, to which a CPU (central processing unit), a RAM (Random Access Memory), a ROM (Read Only Memory), a hard disk, and the like are connected by a bus to execute a predetermined program (estimation program), for example. It is constructed by

  As illustrated in FIG. 1, the estimation apparatus 1 according to the present embodiment includes a memory 10, a frequency domain conversion unit 20, a signal source number estimation unit 30, a result integration unit 40, and a control unit 50. Here, the signal source number estimation unit 30 includes a signal separation unit 31, a power calculation unit 32, an envelope correlation calculation unit 33, and a determination unit 34. The memory 10 includes an observation signal region 11, a time-series data region for each frequency. 12, a separation signal region 13, a power value region 14, an envelope region 15, a parameter region 16, and a signal source number region 17. The control unit 50 includes a register 51 and controls the estimation device 1 as a whole. Also, the broken-line arrows in this figure indicate the theoretical information flow, and the solid-line arrows indicate the actual data flow (at the same time, the electrical or information connection relationship). However, the notation of input / output data in the control unit 50 is omitted.

<Outline of processing>
In this embodiment, the number of source signals is estimated from observation signals x ₁ (t),..., X _M (t) obtained by observing mixed signals obtained by mixing source signals with M sensors.
In this embodiment, as a pre-process, a plurality of parameters (a first parameter th _noise indicating a noise level threshold, a second parameter th _rev indicating a reverberation level threshold, and an envelope correlation value threshold are shown. Data specifying the third parameter th _cor ) is stored in the memory 10. The input time domain observation signals x _j (t) (j = {1,..., M}) are respectively converted into time-series data X _j (f, τ) for each frequency by the frequency domain converter 20. The signal is converted and sent to the signal separation unit 31 of the signal source number estimation unit 30. The signal separation unit 31 generates a separation signal Y _i (f, τ) (i = {1,..., M}) from the time series data X _j (f, τ) for each frequency f. Then, the power calculation unit 32 calculates the power value of each separated signal Y _i (f, τ), and the envelope correlation calculation unit 33 uses the envelope correlation value for the time difference Δτ between different separated signals Y _i (f, τ). Is calculated. When these are calculated, the determination unit 34 compares the power value and envelope correlation value of each separated signal Y _i (f, τ) with each parameter in the memory, and the separated signal Y _i (f, τ). Is the source signal component, and the number of signal sources EN (f) for the frequency f is estimated.

Thereafter, for a wideband signal such as voice, finally, the estimated value for each frequency is integrated in the result integrating unit 40 to obtain an estimated value en of the number of signal sources as a whole. On the other hand, for narrowband signals used in the communication field or the like, it is not necessary to integrate estimated values for each frequency, and it is only necessary to obtain an estimated value EN (f) at the frequency f of interest.
<Details of this embodiment>
2A shows the functional configuration of the signal separation unit 31 illustrated in FIG. 1, FIG. 2B shows the functional configuration of the power calculation unit 32, and FIG. 3A shows the functional configuration of the envelope correlation calculation unit 33. FIG. 3B is a block diagram illustrating the functional configuration of the determination unit. 4 and 5 are flowcharts for explaining a method of estimating the number of signal sources in the present embodiment.

The details of the configuration and processing in this embodiment will be described below with reference to FIGS.
[Preprocessing]
First, as preprocessing, a first parameter th _noise indicating a noise level threshold, a second parameter th _rev indicating a reverberation level threshold, and a third parameter th _cor indicating an envelope correlation value threshold are specified. The data to be stored is stored in the parameter area 16 of the memory 10 (corresponding to the “storage unit”). Examples of parameters include th _noise = 0.01, th _rev = 0.2, and th _cor = 0.5. However, if there is a sample that shows the number of sound sources that are sounding during actual measurement, each parameter may be adjusted based on the observation data. Specifically, for example, the first parameter th _noise is adjusted to be larger than the normalized power value of the noise signal and smaller than the normalized power value of the reverberation signal, and the second parameter th _{rev is set} to Adjustment is made to be smaller than the normalized power value and larger than the normalized power value of the reverberation signal, and the third parameter th _cor is adjusted to be smaller than the envelope correlation value between the source signal and the reverberation signal. The meaning of the normalized power value and the envelope correlation value will be described later.
Further, the observation signal x _j (t) (j = {1,..., M}) in the time domain that is the target of estimating the number of signal sources is written in the observation signal area 11 of the memory 10. This observation signal x _j (t) is an observation signal from M sensors (microphones, etc.), and the subscript j is the observation signal x _j (t) observed by the jth sensor. It shows that.

[Conversion to frequency domain]
First, the control unit 50 (FIG. 1) assigns 1 to the variable j and stores it in the register 51 (step S1). Next, the frequency domain transform unit 20 accesses the observation signal region 11 of the memory 10 and reads the observation signal x _j (t) (step S2). The frequency domain conversion unit 20 that has read the observation signal x _j (t) converts it into time series data X _j (f, τ) for each frequency and stores it in the time series data area 12 for each frequency in the memory 10. (Step S3). In this example, this conversion is performed using a sampling frequency f _s and a short-time discrete Fourier transform of L points.
Next, the control unit 50 determines whether or not the variable j stored in the register 51 is M (step S4). If it is determined that j = M is not satisfied, the control unit 50 sets a value obtained by adding 1 to j as a new j (step S5), stores it in the register 51, and returns to the process of step S2. On the other hand, if it is determined that j = M, the process proceeds to the following signal source estimation process.

[Signal source estimation processing]
First, the control unit 50 (FIG. 1) assigns 0 to the variable f and stores it in the register 51 (step S6).
Independent component analysis (ICA) processing:
Next, the independent component analysis (ICA) unit 31a (FIG. 2A) extracts the time series data X _j (f, τ) from the time series data region 12 for each frequency in the memory 10, and the independent component analysis ( ICA), and X (f, τ) = [X ₁ (f, τ), ..., X _M (f, τ)] ^T , the separation matrix W (f) of the M × M matrix and the ICA separation A signal Z (f, τ) = [Z ₁ (f, τ),..., Z _M (f, τ)] ^T is generated and stored in the register 31b (corresponding to the “storage unit”) (step S7). Here, the signal separation by ICA is such that W (f, τ) = W (f) · X (f, τ) so that each element of the ICA separation signal Z (f, τ) becomes independent from each other. ). Various ICA algorithms are shown in A. Hyvarinen and J. Karhunen and E. Oja, “Independent Component Analysis,” John Wiley & Sons, 2001, ISBN 0-471-40540, and the like.

  Note that the ICA solution has arbitrary scaling. This is because the independence between elements does not change even if a scalar value in an element of Z (f, τ) is multiplied. Therefore, at this stage, there is a high possibility that the power of the source signal observed by the sensor is not correctly reflected in the ICA separation signal. If the number N of source signals is less than the number M of sensors, N elements of the ICA separation signal Z (f, τ) correspond to the source signal, and the remaining MN elements correspond to noise and reverberation components. However, the size of the element corresponding to the noise and reverberation at this stage is generally amplified. Therefore, next, the scaling unit 31c (FIG. 2A) solves the problem of the arbitraryness of the scaling.

Scaling problem solving process:
In the scaling unit 31c, the following operation is performed in order to solve the problem of the arbitraryness of scaling. First, the diagonal matrix generation unit 31ca reads the separation matrix W (f) from the register 31b, and generates a diagonal matrix Λ (f) for solving the scaling problem from the separation matrix W (f) (step S8). ). As this diagonal matrix Λ (f), for example,
Λ (f) = sqrt (diag [(W (f) ・ W (f) ^H ) ^-1 ]) (4)
Can be illustrated. Here, • ⁻¹ is an inverse matrix, • ^H is a conjugate transpose matrix, diag is an operation for setting values other than diagonal components to 0, and sqrt is an operation for calculating the square root of each element.

が成り立つ。すなわち、分離信号Y_i(f,τ)のパワーの総和とセンサでの観測信号X_j(f,τ)のパワーの総和が等しくなる。さらに、分離信号Y_i(f,τ)が互いに独立であれば、

The generated diagonal matrix Λ (f) is sent to the product calculation unit 31cb, and the product calculation unit 31cb uses this and the ICA separation signal Z (f, τ) read from the register 31b, and [Y ₁ ( f, τ), ..., Y M (f, τ)] T ← Λ (f) · [Z 1 (f, τ), ..., the calculation of _{Z M (f, τ)]} T, A separation signal Y _i (f, τ) (i = {1,..., M}) that solves the scaling problem (recovers power) is generated and stored in the separation signal area 13 (FIG. 1) of the memory 10. (Step S9).
Here, the separation signal Y _i (f, τ) has the following two properties by the series of operations including the expression (4). First, if the separated signals Y _i (f, τ) are uncorrelated with each other,

Holds. That is, the sum of the powers of the separated signals Y _i (f, τ) is equal to the sum of the powers of the observation signals X _j (f, τ) at the sensor. Furthermore, if the separated signals Y _i (f, τ) are independent of each other,

Holds. Note that S _k (f, τ) (k = {1,..., N}) represents a source signal component. That is, the power of a certain separated signal Y _i (f, τ) is equal to the sum of the power when the corresponding source signal S _k (f, τ) is observed by all sensors. That the separated signals Y _i (f, τ) are uncorrelated with each other and further independent of each other is the purpose of the independent component analysis, and in many cases, this condition is almost satisfied. Therefore, the power of each separated signal Y _i (f, τ) becomes close to the sum of the power when the corresponding source signal S _k (f, τ) is observed by the sensor by the above series of operations.
Note that the diagonal matrix Λ (f) = diag [W (f) ^-1 ] may be used instead of the diagonal matrix Λ (f) in equation (4), and more generally W (f) It is also possible to use a diagonal matrix [Lambda] having a ^-1 column i-th row and j-th row as a diagonal component of the j-th row. In this case, the power of each separated signal Y _i (f, τ) is the power _obtained by observing the corresponding source signal S _k (f, τ) with a certain sensor j, that is, | H _jk (f) · S _k (f, τ) | approximates to ^2.

[Determination process]
In the determination process, the number of source signals is estimated from the separated signal Y _i (f, τ) that has solved the scaling problem (recovered power). First, the control unit 50 (FIG. 1) assigns 1 to the variable i and stores it in the register 51 (step S10).
Next, the average power calculation unit 32a (FIG. 2B) of the power calculation unit 32 sequentially extracts the separation signal Y _i (f, τ) for each τ from the separation signal region 13 (FIG. 1) of the memory 10, for example. Then, the power value | Y _i (f, τ) | ² is sequentially calculated and stored in the register 32b. Then, the average power calculation unit 32a reads the power value | Y _i (f, τ) | ² stored in the register 32b, and calculates the average power value σ _i ² (for the time τ of the separated signal Y _i (f, τ). f) ← 〈│Y _i (f, τ) ｜ ² 〉 _τ
Is calculated and stored in the register 32b (corresponding to the “storage unit”) (step S11).

Next, the envelope calculation unit 33a (FIG. 3A) of the envelope correlation calculation unit 33 sequentially extracts the separation signal Y _i (f, τ) for each τ from the separation signal region 13 (FIG. 1) of the memory 10, for example. Then, the absolute value | Y _i (f, τ) | is sequentially calculated and stored in the register 33b. Next, the envelope calculation unit 33a reads the absolute value | Y _i (f, τ) | stored in the register 32b, and the absolute value of the separated signal Y _i (f, τ) so that the average with respect to the time τ becomes zero. Value | Y _i (f, τ) | normalized envelope v _i (f, τ) ← | Y _i (f, τ) |-<| Y _i (f, τ) |> _τ
Is calculated and stored in the register 33b (corresponding to the “storage unit”) (step S12).

を算出して、メモリ１０のパワー値領域１４に格納する（ステップＳ１６）。 Next, the control unit 50 (FIG. 1) determines whether or not the variable i stored in the register 51 is M (step 13). If i = M is not satisfied, the control unit 50 adds 1 to i, sets the value as a new i, stores the value in the register 51 (step S14), and returns to step S11. On the other hand, if i = M, the controller 50 assigns 1 to the variable i and stores it in the register 51 (step S15), and executes the following processing.
First, the power normalization unit 32c (FIG. 2B) of the power calculation unit 32 extracts the average power values σ ₁ ² (f),..., Σ _M ² (f) from the register 32b, and calculates the average power. Normalized power value obtained by normalizing the value σ _i ² (f)

Is calculated and stored in the power value area 14 of the memory 10 (step S16).

を算出し、その演算結果Cor_i,k(f)をレジスタ３３ｂに格納する（ステップＳ１８）。なお、この例のΔτは、例えばプログラムコードに組み込まれた定数である。 Next, the control unit 50 (FIG. 1) assigns 1 to the variable k and stores it in the register 51 (step S17). Next, the correlation calculation unit 33c (FIG. 3A) of the envelope correlation calculation unit 33 extracts the envelopes v _i (f, τ) and v _k (f, τ) from the register 33b. Then, the correlation calculation unit 33c uses the envelopes v _i (f, τ) and v _k (f, τ), and uses the separated signal Y _i (f, τ) based on the time difference Δτ (for example, L / 2 or L / 4). ) Separated signal Y _k (f, τ)

And the calculation result Cor _{i, k} (f) is stored in the register 33b (step S18). In this example, Δτ is a constant incorporated in the program code, for example.

Next, the control unit 50 (FIG. 1) determines whether or not the variable k stored in the register 51 is M (step S19). If k = M is not satisfied, the control unit 50 adds 1 to k, stores the value as a new k in the register 51, and returns to step S18 (step S20). On the other hand, when k = M, the maximum value calculation unit 33d (FIG. 3A) of the envelope correlation calculation unit 33 receives the envelope correlation value Cor _{i, 1} (f from the register 33b (FIG. 3A). ), ..., Cori _{, M} (f) is extracted. Then, the maximum value calculation unit 33d calculates the maximum value maxCor _i (f) for each _{i of} the envelope correlation values Cor _{i, k} (f) using these, and stores them in the envelope area 15 (FIG. 1) of the memory 10. (Step S21).

Next, the comparison unit 34 a (FIG. 3B) of the determination unit 34 uses the average power value normalized value NP _i (f), the envelope from the power value region 14, the envelope region 15, and the parameter region 16 of the memory 10. The maximum value maxCor _i (f) of the correlation value, the first parameter th _noise , the second parameter th _rev and the third parameter th _cor are read out. Then, the comparison unit 34a determines whether the separated signal Y _i (f, τ) corresponds to the source signal, noise, or reverberation component according to the following logical expression (step S22).

That is, here, three types of parameters th _noise , th _rev , and th _cor are used. If the normalized value NP _i (f) of the average power value is less than the first parameter th _noise , it is determined as a noise component, and the normalized value NP _i (f) of the average power value is less than the second parameter th _rev In addition, if the maximum value maxCor _i (f) of the envelope correlation value exceeds the third parameter th _cor , it is determined as a reverberation component. Eventually, when sig _i (f) becomes 0, it is determined that the separated signal Y _i (f, τ) corresponds to noise and reverberation components (not source signal components), and sig _i (f) = When it is 1, it is determined that the separated signal Y _i (f, τ) corresponds to the source signal. The determination result sig _i generated in this way is sent to and stored in the register 34b (FIG. 3B). Note that at least a part of “<” in the above logical expression may be “≦”, and “>” may be “≧”.

Next, the control unit 50 determines whether or not the variable i stored in the register 51 (FIG. 1) is M (step S23). If i = M is not satisfied, the controller 50 adds 1 to i, stores the value as a new i in the register 51, and returns to step S16 (step S24). On the other hand, if i = M, the signal source number calculation unit 34c (FIG. 3B) of the determination unit 34 extracts the determination results sig ₁ (f),..., Sig _M (f) from the register 34b. The number of signal sources
EN (f) = Σ _i sig _i (f)
Is stored in the signal source number area 17 (FIG. 1) of the memory 10 (step S25).

Next, the control unit 50 determines whether or not the variable f stored in the register 51 is {(L-1) / L} f _s (f _s is a sampling frequency) (step S26). Storing Here, when the variable f is not a {(L-1) / L } f s, adds f _s / L to the control unit 50 is variable f, and its value as a new variable f, the register 51 Then, the process returns to step S7 (step S27). On the other hand, when the variable f is a {(L-1) / L } f s, it performs the following result integration processing.
[Result integration processing]
First, the result integrating unit 40 estimates the number of signal sources EN (0) _,. ), And based on this, an estimated value en of the number of signal sources as a whole is calculated and output (step S28). In this example, the overall estimated value en is simply determined by majority vote. A large weight may be given to a reliable frequency (for example, a high frequency), and the overall estimated value en may be determined by weighting majority.

[Application result]
The result of having applied the estimation method of the number of signal sources of this form to estimation of the number of sound sources is shown.
FIG. 10 illustrates general experimental conditions. The experimental conditions are as follows.
・ Signal source: 7 seconds of sound ・ Reverberation time: T _R = 200 ms
-Background noise power: -21.8 dB
・ Sampling frequency: f _s = 8000 Hz
-Room size: 4.45 m x 3.55 m x 2.50 m
・ Number of sound sources: 1 to 3 ・ Sound source arrangement / interval: 4 cm in a straight line ・ Number of sensors: 3 ・ Distance between the central sound source and each sensor: 1.1 m
・ An angle between a straight line connecting the central sound source and each sensor and a straight line where each sensor is arranged: 45 °, 90 °, 120 °
1 to 3 sound sources were sounded under the conditions shown in FIG. 10, and the number of sound sources sounding was estimated using observation signals from three microphones.

FIG. 6 shows a comparison of estimation results between the conventional method and the method of this embodiment. Here, FIG. 6A shows the estimation result of the number of signal sources by the conventional method based on the eigenvalue, and FIG. 6B shows the estimation result of the number of signal sources by the method of the present embodiment. The horizontal axis represents the number of true sound sources, and the vertical axis represents the number of frequency bins estimated as the number of sound sources 0, 1, 2, and 3, respectively.
As shown in this figure, in the conventional method, even in the case of one sound source or three sound sources, it is estimated that there are two sound sources according to the majority vote. As described above with reference to FIG. 11, the reason why the conventional method makes a mistake in estimation is that the power of each sound source or noise does not appear properly in the eigenvalue, and the influence of reverberation is not taken into consideration. is there. On the other hand, according to the method of this embodiment, it is correctly estimated in all cases.

Next, the reason why the estimation by the method of this embodiment is accurate will be described.
First, consider power recovery (whether each separated signal appropriately reflects the power of each signal). FIG. 7 shows the true power value (FIG. 7 (a)) observed by the sensor in the case of three sound sources and the power of the separated signal (FIG. 7 (b)) obtained by separating these mixed sounds by the method of this embodiment. The comparison with the value is shown. In addition, the observation result of Fig.7 (a) is measured by sounding each sound source one by one, and normalizing the result. As shown in these figures, the power value of each separated signal according to the method of the present embodiment approximates the true power value of the observed value of each sound source, and the power is restored correctly enough to estimate the number of sound sources. I understand.

  Next, let us consider how to deal with the effects of reverberation. FIG. 8 shows the power values (FIG. 8 (a)) of the first and second (i = 1, 2) separated signals in the case of one sound source and the correlation values (FIG. 8 (b)) of their envelopes. Is. Looking only at the power value in FIG. 8A, the power value of the second separated signal is never sufficiently small, so it is difficult to determine whether it is a signal source or noise including reverberation. However, when the correlation value of the envelopes of the first and second separated signals shown on the right side is seen, the value is sufficiently large, so the second separated signal is a noise containing a large amount of reverberation components of the first separated signal. I understand that. That is, the correlation value of the envelope takes a value of −1 to +1, and approaches 0 as the correlation between signals is lower. In the example of FIG. 8B, the correlation values of the envelope are concentrated between 0.6 and 1, and it can be seen that the correlation between the first separated signal and the second separated signal is high. It can be estimated that the second separated signal having a weak power value is the reverberation component of the first signal.

The first parameter th _noise indicating the noise level threshold necessary for these determinations, the second parameter th _rev indicating the reverberation level threshold, and the third parameter th indicating the envelope correlation value threshold. By appropriately setting _cor , the number of active source signals can be accurately estimated in an actual environment where the influence of noise and reverberation cannot be ignored.
[Second Embodiment]
Next, a second embodiment of the present invention will be described.
The present embodiment is a modification of the first embodiment, and is a form in which signal separation based on eigenvalues is performed instead of signal separation using ICA. Below, it demonstrates centering around difference with 1st Embodiment, and abbreviate | omits description about the matter which is common in 1st Embodiment.

FIG. 9A is a block diagram illustrating the configuration of the signal separation unit 131 in this embodiment.
Note that the only difference between the estimation apparatus of the present embodiment and the estimation apparatus 1 of the first embodiment is that the signal separation unit 31 becomes the signal separation unit 131. Further, the difference between the process of the present embodiment and the process of the first embodiment is only the signal separation process (FIG. 4: steps S7 to 9) and the average power calculation process (FIG. 4: step S11).

[Signal separation processing]
FIG. 9B is a flowchart for explaining the signal separation processing of the present embodiment.
First, the correlation matrix generation unit 131a (FIG. 9A) of the signal separation unit 131 sequentially extracts time-series data X _j (f, τ) from the time-series data region 12 (FIG. 1) for each frequency in the memory 10. And the time series vector X (f, τ) = [X ₁ (f, τ), ..., X _M (f, τ)] ^T correlation matrix R (f) = <X (f, τ) X (f, τ) ^H > _τ is generated (step S31).
The generated correlation matrix R (f) is sent to the eigenvalue decomposition unit 131b (FIG. 9A), and the eigenvalue decomposition unit 131b converts the correlation matrix R (f) into R (f) = V (f) · Λ. It is decomposed into a product of (f) · V (f) ^H (step S32). Note that V (f) = [v ₁ (f), v ₂ (f), ..., v _M (f)], and Λ (f) is λ ₁ (f), λ ₂ (f),. .., λ _M (f) is a diagonal matrix of M rows and M columns, v _j (f) is an eigenvector, and λ _j (f) is a corresponding eigenvalue. The generated eigenvalues λ ₁ (f), λ ₂ (f),..., Λ _M (f) are stored in the memory 10 (FIG. 1) in association with the corresponding τ (step S32), and V (f ) Is sent to the product operation unit 131c (FIG. 9A).

The product calculation unit 131c extracts the time series data X _j (f, τ) from the time series data area 12 (FIG. 1) for each frequency in the memory 10, and [Y ₁ (f, τ),. ^{_{M (f, τ)] T}} = V (f) H · [X 1 (f, τ), ..., by the calculation of _{X M (f, τ)]} T, separated signal Y _i (f, τ) (i = {1,..., M}) is generated and stored in the separated signal area 13 of the memory 10 (step S33). Note that <| Y _i (f, τ) | ² > _τ = λ _i (f) holds.

[Average power calculation processing]
In this embodiment, in step S11 of the first embodiment, the average power calculation section 32a of the power calculation section 32 (FIG. 2 (b)), the separation signal Y _i (f, tau) from the power value | Y _i ( f, τ) | ² is calculated, and instead of calculating the average power value σ _i ² (f) ← <| Y _i (f, τ) | ² > _τ and storing it in the register 32b, the average power is calculated. The unit 32a sequentially extracts the eigenvalue λ _i (f) from the memory 10 (FIG. 1), and the average power value σ _i ² (f) ← λ _i (f) regarding the time τ of the separated signal Y _i (f, τ).
Is calculated and stored in the register 32b.
Other processes are the same as those in the first embodiment.

In the case of the configuration described above, the problem that the power is not properly recovered (the problem that the power of each signal does not appear properly in the eigenvalue) is not solved, but the problem of the effect of reverberation is the same as in the first embodiment. can be solved. Therefore, the number of signal sources can be accurately estimated even in this embodiment in the frequency region where the influence of the problem that power is not properly recovered is small.
The present invention is not limited to the embodiments described above. For example, the various processes described above are not only executed in time series according to the description, but may also be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. Needless to say, other modifications are possible without departing from the spirit of the present invention.

Further, when the above-described configuration is realized by a computer, processing contents of functions that each device should have are described by a program. The processing functions are realized on the computer by executing the program on the computer.
The program describing the 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 program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.
A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

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

  As an application example of the sound signal of the present invention, for example, a process of estimating the number of active sound sources in a certain section in pre-processing of adaptive beamformer or blind sound source separation can be exemplified.

The block diagram which shows the whole estimation apparatus in 1st Embodiment. FIG. 2A is a block diagram illustrating a functional configuration of a signal separation unit illustrated in FIG. 1. FIG. 6B is a block diagram illustrating a functional configuration of a power calculation unit. FIG. 4A is a block diagram illustrating a functional configuration of an envelope correlation calculation unit. FIG. 6B is a block diagram illustrating a functional configuration of a determination unit. The flowchart for demonstrating the estimation method of the number of signal sources in 1st Embodiment. The flowchart for demonstrating the estimation method of the number of signal sources in 1st Embodiment. (A) is the graph which showed the estimation result of the number of signal sources by the conventional method based on an eigenvalue. (B) is the graph which showed the estimation result of the number of signal sources by the method of this form. (A) is a diagram showing the true power value observed by the sensor in the case of three sound sources. (B) is a diagram showing the power value of the separated signal obtained by separating the mixed sound by the method of the first embodiment. (A) is the figure which showed the power value of the 1st and 2nd separated signal in the case of 1 sound source. (B) is the figure which showed the correlation value of those envelopes. (A) is the block diagram which illustrated the composition of the signal separation part in a 2nd embodiment. FIG. 6B is a flowchart for explaining signal separation processing according to the second embodiment. The figure which showed experiment conditions. The figure which showed the power estimated value by the method based on an eigenvalue. (A) is the normalized power value of the eigenvalue at each frequency when only one sound source is played. (B) is the normalized power value of the eigenvalue at each frequency when all three sound sources are played.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Estimation apparatus 31 Signal separation part 10 Memory 32 Power calculation part 20 Frequency domain conversion part 33 Envelope correlation calculation part 30 Signal reduction estimation part 34 Determination part

Claims (9)

A method for estimating the number of signal sources, which estimates the number of signal sources from observed signals,
Data for specifying a plurality of parameters is stored in the storage unit,
The frequency domain converter converts the observation signal x _j (t) (j = {1, ..., M}) from M sensors into time-series data X _j (f, τ) for each frequency. The procedure of storing in the storage unit;
A procedure in which the signal separation unit generates a separation signal Y _i (f, τ) (i = {1, ..., M}) from the time series data X _j (f, τ) and stores it in the storage unit; ,
The power calculation unit calculates the power value of each of the separated signals Y _i (f, τ) and stores it in the storage unit,
A procedure in which an envelope correlation calculating unit calculates an envelope correlation value for a time difference Δτ between different separated signals Y _i (f, τ) and stores it in a storage unit;
The determination unit compares the power value and envelope correlation value of each of the separated signals Y _i (f, τ) with the parameters, and the separated signal Y _i (f, τ) is a source signal component. A procedure for determining whether or not
A method for estimating the number of signal sources. A method for estimating the number of signal sources according to claim 1,
The procedure for generating the separated signal Y _i (f, τ) from the time series data X _j (f, τ) and storing it in the storage unit is as follows.
The independent component analysis unit uses the independent component analysis to separate the M × M matrix separation matrix W (f) and ICA from the time series data X ₁ (f, τ),..., X _M (f, τ). Generating a signal [Z ₁ (f, τ), ..., Z _M (f, τ)] ^T and storing it in the storage unit;
A diagonal matrix generation unit generating a diagonal matrix Λ (f) for solving the scaling problem from the separation matrix W (f);
The product operation unit is [Y ₁ (f, τ), ..., Y _M (f, τ)] ^T ← Λ (f) ・ [Z ₁ (f, τ), ..., Z _M (f , τ)] ^T to generate the separated signal Y _i (f, τ) and store it in the storage unit by calculating ^T ;
A method for estimating the number of signal sources. A method for estimating the number of signal sources according to claim 1,
The procedure for generating the separated signal Y _i (f, τ) from the time series data X _j (f, τ) and storing it in the storage unit is as follows.
The correlation matrix generator generates a correlation matrix R (f) ← <X (f for time series vector X (f, τ) = [X ₁ (f, τ), ..., X _M (f, τ)] ^T , τ) ・ X (f, τ) ^H 〉 _τ ,
The eigenvalue decomposition unit sets the correlation matrix R (f) to V (f) = [v ₁ (f), v ₂ (f), ..., v _M (f)], and sets Λ (f) to λ ₁ (f), λ ₂ (f),..., Λ _M (f) are diagonal elements of M rows and M columns, v _j (f) is an eigenvector, and λ _j (f) Is a product of R (f) = V (f) · Λ (f) · V (f) ^H , where E is the corresponding eigenvalue,
The product operation unit is [Y ₁ (f, τ), ..., Y _M (f, τ)] ^T ← V (f) ^H・ [X ₁ (f, τ), ..., X _M ( f, τ)] ^T to generate the separated signal Y _i (f, τ) and store it in the storage unit by calculating ^T ,
A method for estimating the number of signal sources. A method for estimating the number of signal sources according to claim 1,
The procedure for calculating the power value of each separated signal Y _i (f, τ) and storing it in the storage unit is as follows:
The average power calculation unit calculates the average power value for the time τ of each separated signal Y _i (f, τ),
A power normalization unit normalizing the average power value,
The procedure for determining whether the separated signal Y _i (f, τ) is a source signal component is as follows:
It is a procedure for comparing the normalized value of the average power value and the envelope correlation value with each of the above parameters and determining whether or not the separated signal Y _i (f, τ) is a source signal component.
A method for estimating the number of signal sources. A method for estimating the number of signal sources according to claim 1,
The procedure for calculating the envelope correlation value and storing it in the storage unit is as follows.
Envelope calculation unit, the isolated signals in the so average with respect to time tau becomes 0 Y _i (f, τ) the absolute value of _{| Y i (f, τ)} | envelope was normalized v _{i (f,} τ ) And storing it in the storage unit;
The correlation calculator

A procedure for calculating an envelope correlation value Cor _{i, k} (f) by calculation of and storing it in the storage unit,
A maximum value calculating unit calculating a maximum value maxCor _i (f) of the envelope correlation value Cor _{i, k} (f) for each _i , and
The procedure for determining whether the separated signal Y _i (f, τ) is a source signal component is as follows:
The power value of each of the separated signals Y _i (f, τ) and the maximum value maxCor _i (f) are compared with the parameters, and the separated signal Y _i (f, τ) is a source signal component. It is a procedure to determine whether there is,
A method for estimating the number of signal sources. A method for estimating the number of signal sources according to claim 1,
The above parameters are
A first parameter th _noise indicating a threshold of a noise level, a second parameter th _rev indicating a threshold of a reverberation level, and a third parameter th _cor indicating a threshold of an envelope correlation value,
The procedure for determining whether the separated signal Y _i (f, τ) is a source signal component is as follows:
When the power value of the separated signal Y _i (f, τ) is less than or less than the first parameter th _noise , or the power value of the separated signal Y _i (f, τ) is the second parameter th _rev Or less, and when the envelope correlation value is greater than or equal to the third parameter th _cor , the separated signal Y _i (f, τ) has a procedure for determining that it is not a source signal component.
A method for estimating the number of signal sources. An estimation device for estimating the number of signal sources from observed signals,
A storage unit storing data for specifying a plurality of parameters;
Frequency to convert observation signal x _j (t) (j = {1, ..., M}) from M sensors into time-series data X _j (f, τ) for each frequency and store in memory An area conversion unit;
A signal separation unit that generates a separation signal Y _i (f, τ) (i = {1, ..., M}) from the time series data X _j (f, τ) and stores it in a storage unit;
A power calculation unit that calculates the power value of each of the separated signals Y _i (f, τ) and stores it in the storage unit;
Calculating an envelope correlation value for a time difference Δτ between the different separated signals Y _i (f, τ) and storing the envelope correlation value in a storage unit;
The power value and envelope correlation value of each separated signal Y _i (f, τ) is compared with each of the above parameters, and whether or not the separated signal Y _i (f, τ) is a source signal component is determined. A determination unit for determining;
The estimation apparatus characterized by having. A procedure for converting observation signals x _j (t) (j = {1, ..., M}) from M sensors into time-series data X _j (f, τ) for each frequency;
A procedure for generating a separation signal Y _i (f, τ) (i = {1, ..., M}) from the time series data X _j (f, τ) and storing it in a storage unit,
A procedure for calculating the power value of each separated signal Y _i (f, τ) and storing it in the storage unit,
A procedure for calculating an envelope correlation value for a time difference Δτ between different separated signals Y _i (f, τ) and storing it in a storage unit;
The power value and envelope correlation value of each separated signal Y _i (f, τ) described above are compared with each parameter stored in the storage unit, and the separated signal Y _i (f, τ) is the source signal component. A procedure for determining whether or not
An estimation program for causing a computer to execute.   A computer-readable recording medium on which the estimation program according to claim 8 is recorded.

Images