JP2007235875A - Transmission path estimating method, echo canceling method, sound source separating method, apparatus therefor, program, and recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To estimate a transmission path or input signal without using previous information of the transmission path. <P>SOLUTION: The present invention is based upon processing steps wherein a present signal space matrix and a past/future signal space matrix are determined from a multi-channel signal, a projection residual signal space matrix is determined by projecting the present signal space matrix on an orthogonal base of the past/future signal space matrix, and an optimal response length is directly estimated from the projection residual signal space matrix. A corrected projection residual signal space is determined by correcting a projection residual signal space based on the estimated optimal response length, and an impulse response of the transmission path is estimated from the corrected projection residual signal space. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a transmission path estimation method, apparatus, and the like for estimating a transmission path and an input signal from a multi-channel output signal from a linear transmission path of 1-input multi-output or multi-input multi-output without using prior information of the transmission path, The present invention relates to a program and a recording medium, and a dereverberation method, a sound source separation method, an apparatus, a program, and a recording medium using the same.

The problem of estimating the transmission path and source signal from the observed signal that has passed through the linear transmission path without using any knowledge of the transmission path and source signal is called the blind estimation problem, and is used in the fields of audio-acoustic signal processing and wireless communication. Research is ongoing.
In the field of audio-acoustic signal processing, there is a problem that reverberation caused by reflection from a wall or the like is mixed in the sound recorded indoors, and the clarity of the sound is lowered and is difficult to recognize. Solving the blind estimation problem restores the clarity of speech and makes it easier to recognize. In wireless communication, ghosts are generated by receiving reflections from buildings such as buildings, and transmission errors are likely to occur. Solving the blind estimation problem can reduce transmission errors and improve the transmission rate.

ただしh₁(k)…h_P(k)は伝達経路のインパルス応答であり、Nはインパルス応答の長さになる。
無線通信のブラインド推定問題では、観測信号の周波数特性は伝達経路に由来することがほとんどである。一方、音声音響分野のブラインド推定問題では、観測信号の周波数特性がソース信号と伝達経路の両方に由来し、推定が一層困難になっている。
上記ブラインド推定問題に関して、有色のソース信号から伝達経路を推定する方法として、Least Square Smoothing法（以下LSS法と表記）とその拡張であるJ-LSS法が提案されている（非特許文献１）。 These reverberation and ghost phenomena can be modeled with the source signal as s (k) and the P channel observation signal as y ₁ (k) to y _P (k) as shown in FIG.

However, h ₁ (k)... H _P (k) is the impulse response of the transmission path, and N is the length of the impulse response.
In the wireless communication blind estimation problem, the frequency characteristics of the observed signal are mostly derived from the transmission path. On the other hand, in the blind estimation problem in the audio-acoustic field, the frequency characteristics of the observation signal are derived from both the source signal and the transmission path, making estimation more difficult.
Regarding the blind estimation problem, the Least Square Smoothing method (hereinafter referred to as LSS method) and the J-LSS method, which is an extension thereof, have been proposed as methods for estimating a transmission path from a colored source signal (Non-patent Document 1). .

  The processing of the LSS method will be described with reference to FIGS. In the LSS method, it is assumed that the actual impulse response length N is known, and this is assumed as the assumed length j of the transmission path impulse response, and the current signal space matrix generation unit 111, the past signal space matrix generation unit 112 shown in FIG. The future signal space matrix generation means 113 is set. Here, the multi-channel signal stored in the multi-channel signal storage means 101 shown in FIG. 9 will be described with reference to FIG. The multi-channel signal storage means 101 stores a multi-channel signal having a predetermined time length at a certain time. A space spanned by a vector composed of signals located approximately in the center of the time axis direction of the accumulated multi-channel signal is called a current signal space, and a space spanned by a signal vector that is older in time than the current signal is past signal space, current space. A space in which a signal vector that is temporally newer than the signal space is called a future signal space.

をベースとして説明を進める。
過去信号空間行列生成手段１１２は、多チャネル信号蓄積手段１０１に蓄積された多チャネル信号から、多チャネル過去信号空間として過去信号空間行列Z_P（非特許文献１の“past data matrix”）を生成する。未来信号空間行列生成手段１１３は、同様に多チャネル未来信号空間として未来信号空間行列をZ_F（非特許文献１の“future data matrix”）を生成する。

wの推奨値はjである。またTについては、T>2wに設定する。行列Z_P、Z_Fは（P× w）行（T+1）列の行列になる。 Step SP1 (see FIG. 12): A past signal space, a future signal space, and a current signal space are obtained from the P channel observation signals accumulated in the multichannel signal accumulation means 101.
In the following, a vertical vector consisting of values y ₁ (k) to y _P (k) of time k of the P channel sound pickup signal

The explanation will proceed based on the above.
The past signal space matrix generation unit 112 generates a past signal space matrix Z _P (“past data matrix” in Non-Patent Document 1) as a multi-channel past signal space from the multi-channel signals stored in the multi-channel signal storage unit 101. To do. The future signal space matrix generation unit 113 similarly generates a future signal space matrix Z _F (“future data matrix” of Non-Patent Document 1) as a multi-channel future signal space.

The recommended value for w is j. For T, set T> 2w. The matrices Z _P and Z _F are (P × w) rows (T + 1) columns.

ステップSP2：Z行列生成手段１１５にて、多チャネル過去信号空間行列Z_Pと多チャネル未来信号空間行列Z_Fから過去未来信号空間行列Z（非特許文献１の“future-past data matrix”）を生成する。

The current signal space matrix generating unit 111 generates a current signal space matrix Y (“current data matrix” in Non-Patent Document 1) of (P × w) rows (T + 1) columns as a multi-channel current signal space.

Step SP2: In the Z matrix generation means 115, the past future signal space matrix Z (“future-past data matrix” in Non-Patent Document 1) is obtained from the multi-channel past signal space matrix Z _P and the multi-channel future signal space matrix Z _F. Generate.

Step SP3: The multi-channel current signal space is projected onto the multi-channel past signal space and the multi-channel future signal space to obtain a projected residual signal space. For this purpose, the past future signal space orthogonal basis calculation means 121 obtains the orthogonal basis of the matrix Z corresponding to the past future signal space. This orthogonal basis V is the matrix Z
Z = UΣV ^T
It is obtained by singular value decomposition as follows.

In the projection residual signal matrix extracting means 131 calculates the projection of the past future signal space of the current signal space YVV ^T. And projective residual signal space
E = Y−YVV ^T
Is obtained as a projection residual signal space matrix E (“projection error matrix” in Non-Patent Document 1).

とし、h^₁(0)とh₁(0)の推定値とした場合、前記の左特異ベクトルの各成分は、

のようにPチャネル収音経路インパルス応答の推定値と対応する。 Step SP4: In the transmission path estimation means 161,
The impulse response of the P channel sound collection path that extracts the left singular vector corresponding to the maximum singular value by singular value decomposition of the projection residual signal space matrix E

And assuming that h ^ ₁ (0) and h ₁ (0) are estimated values, each component of the left singular vector is

This corresponds to the estimated value of the P channel sound collection path impulse response.

を得ることができる。このP本のベクトルにより伝達経路推定手段１６１は伝達経路を推定することができ、この伝達経路推定値を用いて逆フィルタ算出手段２０１は逆フィルタ係数を算出し、逆フィルタ手段２０２のフィルタ特性を設定する。以上がLSS法の処理となる。 Therefore, by extracting the components of the left singular vector for every P and vectorizing them, P vectors that respond to the estimated value of the P channel sound collection path impulse response

Can be obtained. The transmission path estimation means 161 can estimate the transmission path from the P vectors, and using this transmission path estimation value, the inverse filter calculation means 201 calculates the inverse filter coefficient, and the filter characteristic of the inverse filter means 202 is obtained. Set. The above is the processing of the LSS method.

In the LSS method, it is assumed that the actual impulse response length N is known, but since it is actually unknown, processing for determining the response length is required. A method including this processing is the J-LSS method described in Non-Patent Document 2.
The J-LSS method will be described with reference to FIGS. The characteristic configuration of the J-LSS method shown in FIG. 13 is that projection residual signal matrix extraction means 131 and response length estimation means 141 of transmission path estimation means 161 ′ are provided, and accompanying this, transmission path estimation means. The procedure of 161 is slightly changed to 161 ′. As shown in FIG. 14, the processing flow of steps SP1 to SP3 is common to the LSS method and the J-LSS method as shown in FIG. 14. Therefore, in FIG. I will explain.

Step SP6: The response length estimation means 141 performs singular value decomposition on the projection residual signal space matrix E.
E = U _E Σ _E V _E ^T
And the response length is assumed. This is assumed to be an assumed length k (1 ≦ k ≦ j). As an example, setting to j can be considered.

を生成する。 Step SP7: The left singular vector corresponding to P (j + 1) −j + k−1 singular values is extracted from the matrix U _E from the smaller of the matrix E. Generate a matrix D with the left singular vector in the row,
D = | D ₁ … D _j |
, And divide equally into j and block Hankel matrix Γ _k (D)

Is generated.

Step SP8: In order to check the validity of the assumed length k, a linear expression for the block Hankel matrix
Γ _k (D) f = 0
Check if has a solution other than 0 vector. When this linear expression has a solution other than the 0 vector, it is determined that the assumed length k matches the impulse response length N, and the process goes to step SP10. When there is no solution other than the zero vector, the process goes to step SP9.

Step SP9: Change the assumed length k and return to step SP7. As an example, when the assumed length k is initially set to j, it is conceivable to change the assumed length k as k ← k−1.
Step SP10: In the transmission path estimation means 161 ′,
Γ _k (D) f = 0
An impulse response is obtained as a solution f other than the zero vector of. The correspondence between f and the impulse response is the same as the correspondence between the singular vector and the impulse response in step SP4.
L.Tong and Q. Zhao, "Joint Order Detection and Blind Channel Estimation by Least Squares Smoothing," IEEE Transactions on signal processing, Vol. 47, No.9 1999.

In the J-LSS method (prior art), the optimal impulse response length is obtained by repeatedly verifying the validity of the assumed impulse response length k while changing the assumption little by little. There is a problem that requires enormous operations.
An object of the present invention is to estimate a transmission path or an input signal with a small amount of computation without changing the assumption of the impulse response length and repeatedly performing verification processing.

In the present invention, the optimum response length is directly estimated from the initially obtained projection residual signal space. Based on this optimum response length, the projection residual signal space is corrected to obtain a corrected projection residual signal space, and the processing procedure for estimating the impulse response of the transmission path from the corrected projection residual signal space is a basic operation.
Hereinafter, specific procedures of the transmission path estimation method according to the present invention, the dereverberation method that operates using the transmission path estimation method, and the sound source separation method will be described.
A transmission path estimation method according to the present invention includes a multi-channel signal accumulation processing means for accumulating multi-channel observation signals in a transmission path estimation method for estimating a transmission path from observation signals observed from a source signal through a plurality of linear transmission paths. Current signal space matrix generation processing for obtaining a current signal space matrix from accumulated multi-channel observation signals, past future signal space matrix generation processing for generating past future signal space matrices from accumulated multi-channel observation signals, and past future Projected residual when projecting current signal space matrix to past future signal space from past future signal space orthogonal basis calculation processing to obtain orthogonal basis of signal space matrix and orthogonal base of current signal space matrix and said past future signal space Projection residual signal matrix extraction processing to obtain the signal space matrix, and the redundant projection residual signal space matrix Eb from the rank r of the projection residual signal matrix E to obtain the projection Projection residual signal matrix correction processing that takes out a component orthogonal to the redundant projection residual signal space matrix Eb from the residual signal space matrix E to obtain a corrected projection residual signal space matrix, and the corrected projection residual signal matrix as a singular value A transmission path estimation process for decomposing and extracting a left singular vector corresponding to the maximum singular value to obtain an impulse response of the multi-channel transmission path.

  The dereverberation method according to the present invention applies an inverse filter calculation process for obtaining an inverse filter from the estimated transmission path impulse response, and an inverse filter applied to the multichannel observation signal, in addition to each process applied in the above transmission path estimation method. And an inverse filter process for outputting a source signal estimation result.

  Furthermore, in the sound source separation method according to the present invention, in addition to the processes used in the transmission path estimation method, a dereverberation filter calculation process for obtaining a dereverberation filter from the estimated transmission path impulse response, and reverberation for a multichannel observation signal. The present invention includes a dereverberation filtering process that applies a dereverberation filter obtained by a removal filter calculation process, and a sound source separation process that performs a sound source separation process using a signal after the dereverberation process as an input.

  The present invention directly estimates the optimum response length from the projected residual signal space obtained by projecting the current signal space matrix onto the future past signal space matrix. Then, based on the optimum response length, the projection residual signal space is corrected to obtain a corrected projection residual signal space, and the impulse response of the transmission path is estimated from the corrected projection residual signal space. As a result, the present invention eliminates the need for iterative processing as in the conventional method (J-LSS method), and therefore estimates the impulse response of the transmission path with a smaller amount of computation than in the prior art (J-LSS method). Can do. Therefore, the present invention is particularly effective when the estimated impulse response length may be large.

  The transmission path estimation method, the dereverberation method, the sound source separation method, and these apparatuses according to the present invention can all be realized by hardware. However, in order to realize the simplest method, the transmission path estimation program according to the present invention and the dereverberation are implemented in a computer. The best mode is one in which a program and a sound source separation program are installed and the computer functions as these devices.

  When the computer functions as a transmission path estimation device, a transmission path estimation program is installed in the computer, and by this program, multi-channel signal storage means, current signal space matrix generation means, past signal space matrix generation means, The past future signal space orthogonal basis calculation means, the projection residual signal matrix extraction means, the projection residual signal matrix correction means, and the transmission path estimation means are constructed and function as a transmission path estimation device.

When making a computer function as an dereverberation device, a dereverberation program is installed in the computer, and a configuration in which an inverse filter calculating unit and an inverse filter unit are added to the configuration of the transmission path estimation device in the computer by this program And make it function as a dereverberation device.
When making a computer function as a sound source separation device, a sound source separation program is installed in the computer, and by this program, the dereverberation filter calculating means, the dereverberation filtering means, the sound source separation means, Is added to make it function as a sound source separation device.

A first embodiment of the present invention will be described with reference to FIGS. In addition, the same code | symbol is attached | subjected and shown to the block which performs the same process as the prior art shown in FIG.
First, the assumed transmission path impulse response length j is set to a value expected to be equal to or longer than the impulse response length N in advance in the current signal space matrix generation unit 111, the past signal space matrix generation unit 112, and the future signal space matrix generation unit 113 of FIG. Set. In the first embodiment shown in FIG. 1, a past signal space matrix generation unit 112 and a future signal space matrix generation unit 113 are provided, and a Z matrix is generated from the past signal space matrix and the future signal space matrix generated by them. However, since there is also a method for generating the Z matrix without going through the past signal space matrix generation means 112 and the future signal space matrix generation means 113, here the past signal space matrix generation means 112, the future signal space matrix generation means 113, The Z matrix generation unit 115 including the Z matrix generation unit 115 is referred to as a Z matrix generation unit 102.

Step 1: The current signal space matrix generation unit 111 obtains the current signal space matrix Y from the multichannel signals stored in the multichannel signal storage unit 101. Similarly, the past signal space matrix generation means 112 obtains the past signal space matrix Z _P and the future signal space matrix generation means 113 obtains the future signal space matrix Z _F. The detailed processing content is the same as that in Step SP1 of “Background Technology”, and the description thereof is omitted here.

により行列Zを生成する。詳細な処理内容は「背景技術」のステップSP2と同一なのでここでは説明を省略する。 Step 2: The Z matrix generation unit 102 calculates the past signal space matrix Z _P and the future signal space matrix Z _F from

To generate a matrix Z. Detailed processing contents are the same as in step SP2 of “Background Technology”, and thus description thereof is omitted here.

Step 3: In the past future signal space orthogonal basis calculation means 121, the orthogonal basis V of the matrix Z is obtained, and the projection residual signal space matrix E is obtained.
E = Y−YVV ^T
Ask for. Detailed processing contents are the same as in step SP2 of “Background Technology”, and thus description thereof is omitted here.

  Step 4: In the projection residual signal matrix correction means 151, the rank r of the projection residual signal space matrix E is obtained, and j−r + 1 is set as the optimum response length. P × (r−1) rows are extracted from the bottom of the projection residual signal space matrix E and set as a redundant projection residual signal space matrix Eb. Its size is P × (r−1) rows (T + 1) columns.

次に、取り出した冗長射影残差信号空間行列Ebの直交基底Vbを、後述のように特異値分解等を経由して求め、
E₂＝E_a−E_aV_bV_b ^T
により補正された射影残差信号行列E₂を求める。このステップ４が背景技術のLSS法にもJ−LSS法にも無い本願特有の処理である。 Here, it is assumed that the projection residual signal space matrix E is divided into the matrix Ea and the redundant projection residual signal space matrix Eb as follows, and only the redundant projection residual signal space matrix Eb from the projection residual signal space matrix E Take out.

Next, the orthogonal base Vb of the extracted redundant projection residual signal space matrix Eb is obtained via singular value decomposition or the like as described later,
E ₂ = E _a −E _a V _b V _b ^T
The projection residual signal matrix E ₂ corrected by is obtained. This step 4 is a process unique to the present application which is neither in the background art LSS method nor in the J-LSS method.

Step 5: In pathways estimating means 161, and singular value decomposition corrected projection residual matrix E _2, the left singular vector corresponding to the largest singular value, to obtain an impulse response estimate of the multi-channel sound path. Since the detailed processing contents other than the input being the corrected residual matrix E ₂ instead of the projected residual matrix E are the same as those in step SP4 of the “background art” LSS method, the description thereof is omitted here.

Here, supplementary explanation will be given with reference to FIG. 2A indicates each matrix generated by the current signal space matrix generation unit 111, the past signal space matrix generation unit 112, and the future signal space matrix generation unit 113. It is assumed that the matrix Y includes a redundant part X. Projection residual signal matrix E ₂ obtained in step 4 is indexing the amount of redundant portions X as shown in FIG. 2B, to remove the redundant portion X, a projection residual signal matrix corrected to a proper amount. By the estimation of the transmission path by using the corrected projected residual signal matrix E _2, so that it is possible to estimate the transmission path at an appropriate impulse response length.

Therefore, according to the present invention, the corrected residual signal matrix E ₂ corrected by performing the past future signal space orthogonal base extraction process, the projected residual signal matrix extraction process, and the projected residual signal matrix correction process once is obtained. Therefore, the amount of calculation can be reduced.
In the prior art J-LSS method, it is necessary to change the assumption of the response length little by little and repeatedly perform the process of verifying the validity of the assumption. On the other hand, in the present invention, iterative processing (loop of steps SP7-SP8-SP9 shown in FIG. 14) in the J-LSS method is unnecessary.

  FIG. 3 shows the result of the simulation performed to show the effectiveness of the method of the first embodiment. The simulation shown in FIG. 3 uses a sound signal of 8 kHz sampling as a source signal, performs estimation of a transmission path and estimation of an inverse filter from an observation signal observed through a transmission path of one input and two outputs. The impulse response length of the 1-input 2-output transmission path is set to N = 500, and the initial impulse response length is assumed to be J = 550. The transmission path uses a room impulse response with a reverberation time of 200 ms, measured at a sampling frequency of 8 ”kHz, cut to 500 taps.

A1 and A2 shown in FIG. 3 are impulse responses (true values) from the signal source to the latter stage of the transmission path. Also, C1 and C2 shown in FIG. 3 are impulse response estimation results by the method of the first embodiment. When A1 and C1 and A2 and C2 are compared, both have substantially the same waveform, and it can be seen that the impulse response is well estimated by the method of the first embodiment.
Modification of the first embodiment: After obtaining an optimum response length via rank from the projection residual signal matrix E, as a method for obtaining the impulse response of the transmission path, the following implementation method for recalculating the matrices Y, Z, and E Is also possible.
(1) The assumed length is reset to j−r + 1 (r + 1 is redundant), and the matrix Y is obtained by applying step 1 of the first embodiment.
(2) The assumed length is reset to j−r + 1, and the matrix Z is obtained by applying step 2 of the first embodiment.
(3) Applying Step 3 of Example 1 to recalculate the projection residual matrix.
(4) Step 5 of Embodiment 1 is applied with the recalculated projection residual matrix as an input.
In the method according to the first embodiment, the orthogonal basis of the matrix E _{2 having} P × (r−1) rows (T + 1) columns may be obtained.

However, in the implementation method based on the above recalculation, it is necessary to obtain the orthogonal basis of the matrix Z of (2P × w) rows (T + 1) columns with w = j−r + 1. Compared to the method of Example 1, in the exemplary method according to recalculation, the amount of computation extra required according to the size difference of the matrix E ₂ and the matrix Z.

A second embodiment of the present invention will be described with reference to FIG. In the second embodiment, the source signal from which reverberation is removed is estimated using the estimation result of the transmission path impulse response according to the first embodiment. It should be noted that blocks that perform the same processing as in the prior art in FIG.
Step 1: An impulse response estimation value of a multi-channel sound pickup path is obtained using the method of the first embodiment.
Step 2: Inverse filter calculation means 201 obtains an inverse filter from the estimated transmission path impulse response using the method described in Japanese Patent Laid-Open No. 62-190935.
Step 3: The inverse filter means 202 applies the inverse filter to the multichannel observation signal to estimate the original source signal.
In order to show the effectiveness of the technique of the second embodiment, a simulation was performed using the same settings as the simulation of the first embodiment. FIG. 5 shows an impulse response from the signal source to the latter stage of the inverse filter when the inverse filter obtained by the method of the second embodiment is used. This impulse response is an impulse response of only a direct wave, and it can be seen that the reverberation is well removed.

Next, a third embodiment of the present invention will be described with reference to FIG.
The purpose of this embodiment is to eliminate and reverberate signals observed through a multi-input multi-output transmission path as shown in FIG. The impulse response length of the multi-input multi-output transmission path is set to N, and the assumed length is set to a value j that is expected to be N or more. Here, the number of sound sources is known as Q, and the number of sound collection channels is P. However, Q <P.

をベースとして説明を進める。
同様に多チャネル信号蓄積手段１０１に蓄積された多チャネル信号から、過去信号空間行列生成手段１１２により、多チャネル過去信号空間として行列Z_Fを生成する。

但し、音源数Qをもちいて、w≧Q×jに設定する。またTについては、T>2wに設定する。行列Z_P、Z_Fのサイズは（P×w）行（T＋1）列である。
現信号空間行列生成手段１１１は、多チャネル信号蓄積手段１０１に蓄積された多チャネル観測信号から、多チャネル現信号空間として、行列Yを生成する。このとき行列Yのサイズは（P×j）行（T＋1）列になる。

Step 1: A current signal space, a past signal space, and a future signal space are obtained from the multichannel observation signals stored in the multichannel signal storage means 101.
In the following, a vertical vector consisting of values y ₁ (k) to y _P (k) of time k of the P channel sound pickup signal

The explanation will proceed based on the above.
Similarly, a matrix Z _F is generated as a multi-channel past signal space by the past signal space matrix generation unit 112 from the multi-channel signals stored in the multi-channel signal storage unit 101.

However, w ≧ Q × j is set using the number of sound sources Q. For T, set T> 2w. The sizes of the matrices Z _P and Z _F are (P × w) rows (T + 1) columns.
The current signal space matrix generation unit 111 generates a matrix Y as a multichannel current signal space from the multichannel observation signals stored in the multichannel signal storage unit 101. At this time, the size of the matrix Y is (P × j) rows (T + 1) columns.

により行列Zを生成する。 Step 2: In the Z matrix generation means 115, from the past signal space matrix Z _P and the future signal space matrix Z _F

To generate a matrix Z.

Step 3: In the past future signal space orthogonal basis calculation means 121, the orthogonal basis V of the matrix Z is obtained, and the projection residual signal space matrix E is obtained.
E = Y−YVV ^T
Ask for.

冗長射影残差信号空間行列Ebの直交基底Vbを、上述のように特異値分解等を経由して求め、
E₂＝E_a−E_aV_bV_b ^T
により補正された射影残差信号空間行列を求める。 Step 4: The projection residual signal matrix correction means 151 obtains the rank r of the projection residual signal matrix E. Using this rank r, P × (r−Q) rows are extracted from the bottom of the projection residual signal space matrix E and set as a redundant projection residual signal space matrix Eb. The size of the matrix Eb is P × (r−Q) rows (T + 1) columns. The matrix E is divided into a matrix Ea and a redundant projection residual signal space matrix Eb as follows.

Obtain the orthogonal basis Vb of the redundant projection residual signal space matrix Eb via singular value decomposition as described above,
E ₂ = E _a −E _a V _b V _b ^T
The projection residual signal space matrix corrected by is obtained.

とすると、

はq番目の音源に関するPチャネル収音経路のインパルス応答推定値になっている。 Step 5: left singular vectors in the transmission path estimator 161, the singular value decomposition of the corrected projection residual signal matrix E _2, the size of left singular vector corresponding to the largest singular value corresponds to the Q-th singular value Take out. The qth singular vector

Then,

Is the impulse response estimate of the P channel sound collection path for the qth sound source.

  Step 6: The dereverberation filter calculating means 203 obtains an dereverberation filter from the transmission path impulse response estimated in Step 5 by using the method described in the above-mentioned Japanese Patent Application Laid-Open No. 62-190935.

  Step 7: In the dereverberation filtering means 204, this dereverberation filter is applied to the multichannel observation signal to estimate the multichannel signal from which the reverberation component has been removed.

  Step 8: The sound source separation unit 301 applies sound source separation processing to the multichannel signal from which the reverberation component has been removed, and extracts the sound source signal.

For the sound source separation processing, a blind separation algorithm based on independent component analysis (ICA) described in Reference 1 can be used. (Reference 1: “JF Cardoso, Blind Signal Separation: Statistical Principles, Proceedings of the IEEE, VOL.86, NO.10, pp.2009-2025, 1998.”)
The transmission path estimation method included in the above dereverberation process is based on the LSS method (Non-patent Document 1). However, the LSS method deals only with the case where the sound source is a single colored signal and the transmission path is a one-input multi-output system.
In the LSS method, an assumed impulse response length j is recommended as w when generating a past signal space matrix Z _P and a future signal space matrix Z _F from a multi-channel signal. However, the desired result cannot be obtained even if the LSS method is applied as it is to the observation signal of a multi-input multi-output system.

Therefore, in the present invention, in order to extend the LSS method to a multi-input multi-output system, w ≧ Q × j is set. However, Q is the number of sound sources Q.
Further, a technique called frequency domain blind sound source separation (frequency domain BSS) described in Japanese Patent Application Laid-Open No. 2003-333682 is often used for the problem of separating the convolution mixed signal of the third embodiment.

However, in the frequency domain BSS, the direct sound component of the sound to be removed can be almost completely removed. It is known that the separation performance is significantly reduced at around 10 dB.
Japanese Patent Laid-Open No. 2003-99093 also proposes a method of performing noise suppression processing (crosstalk component suppression processing) as reverberation processing after the frequency domain blind sound source separation processing with the aim of improving separation performance. The structure described in this patent document is shown in FIG. In the noise suppression processing of the noise suppression unit 302, the reverberation component is modeled as a signal in which the delay and gain of the estimated direct sound component are adjusted. Based on this model, the crosstalk component included in the signal after the sound source separation processing of the sound source separation means 301 is estimated and subtracted. However, since the reverberation component is simply modeled as a reflected wave of several orders, the improvement of the separation performance is only 3 to 4 dB on average.

On the other hand, by using the method of the third embodiment, it is possible to extract a desired sound without a reverberation component while suppressing the crosstalk component to around −30 dB.
The transmission path estimation apparatus described in the first embodiment, the dereverberation apparatus described in the second embodiment, and the sound source separation apparatus described in the third embodiment are all transmission path estimation that causes a computer to operate according to the procedure described in each embodiment. It can be realized by causing a computer to function by a program, a dereverberation program, and a sound source separation program.
Each program is written in a computer-readable program language, and is recorded on a computer-readable recording medium such as a magnetic disk, CD-ROM, or semiconductor memory. It is installed in a computer from these recording media or through a communication line, and is decrypted and executed by a CPU provided in the computer.

  The transmission path estimation apparatus, dereverberation apparatus, and sound source separation apparatus according to the present invention are each utilized in the field of hands-free voice communication systems and the like.

BRIEF DESCRIPTION OF THE DRAWINGS The block diagram for demonstrating Example 1 of this invention. FIG. 5 is a diagram for explaining the operation of the first embodiment. FIG. 6 is a waveform diagram for explaining the effect of the first embodiment. The block diagram for demonstrating Example 2 of this invention. FIG. 6 is a waveform diagram for explaining the effect of the second embodiment. The block diagram for demonstrating Example 3 of this invention. FIG. 6 is a layout diagram for explaining a usage situation of the third embodiment. The block diagram for demonstrating the connection relation of the sound source separation means described in the well-known literature, and a noise suppression means. The block diagram for demonstrating a prior art. The layout for demonstrating the utilization condition of a prior art. The figure for demonstrating the mode inside a multichannel signal storage means. The flowchart for demonstrating background art. The block diagram for demonstrating the other example of background art. 14 is a flowchart for explaining another background art shown in FIG. 13.

Explanation of symbols

101 Multi-channel signal accumulating unit 131 Projected residual signal matrix extracting unit 102 Z matrix generating unit 141 Response length estimating unit 111 Current signal space matrix generating unit 151 Projected residual signal matrix correcting unit 112 Past signal space matrix generating unit 161, 161 ′ Transmission path estimation means 113 Future signal space matrix generation means 201 Inverse filter calculation means 115 Z matrix generation means 202 Inverse filter means 121 Past future signal space orthogonal basis calculation means

Claims (8)

In a transfer path estimation method for estimating a transfer path from an observation signal observed through a plurality of linear transfer paths from a source signal,
Multi-channel signal accumulation processing for accumulating multi-channel observation signals;
Current signal space matrix generation processing for obtaining a current signal space matrix from the accumulated multi-channel observation signals;
Past future signal space matrix generation processing for generating a past future signal space matrix from accumulated multi-channel observation signals;
A past future signal space orthogonal basis calculation process for obtaining an orthogonal basis of the past future signal space matrix;
From the orthogonal base of the current signal space matrix and the past future signal space, a projection residual signal matrix extraction process for obtaining a projection residual signal space matrix when the current signal space matrix is projected onto the past future signal space;
A redundant residual signal space matrix Eb is obtained from the rank r of the projected residual signal matrix E, and a component that is orthogonal to the redundant projected residual signal space matrix Eb is extracted from the projected residual signal space matrix E. A projection residual signal matrix correction process as a matrix;
Singular value decomposition of the corrected projected residual signal matrix, taking a left singular vector corresponding to the maximum singular value, and obtaining a transmission path estimation process for obtaining an impulse response of a multi-channel transmission path;
Including a transmission path estimation method. In the dereverberation method for estimating the transmission path and the source signal from the observation signal observed through the plurality of linear transmission paths from the source signal,
Multi-channel signal accumulation processing for accumulating multi-channel observation signals;
Current signal space matrix generation processing for obtaining a current signal space matrix from the accumulated multi-channel observation signals;
Past future signal space matrix generation processing for generating a past future signal space matrix from accumulated multi-channel observation signals;
A past future signal space orthogonal basis calculation process for obtaining an orthogonal basis of the past future signal space matrix;
A projection residual signal matrix extraction process for obtaining a projection residual signal space matrix when the current signal space matrix is projected onto the past future signal space from orthogonal bases of the current signal space matrix and the past future signal space;
A redundant residual signal space matrix Eb is obtained from the rank r of the projected residual signal matrix E, and a component orthogonal to the redundant projected residual signal space matrix Eb is extracted from the projected residual signal space matrix E to obtain a corrected projected residual. A projection residual signal matrix correction process as a signal matrix;
Singular value decomposition of the corrected projection residual signal matrix, taking out the left singular vector corresponding to the maximum singular value and obtaining the impulse response of the multi-channel transmission route,
An inverse filter calculation process for obtaining an inverse filter from the estimated transmission path impulse response;
Applying an inverse filter to the multi-channel observation signal and outputting a source signal estimation result;
A dereverberation method characterized by comprising: In a sound source separation method for estimating a source signal from an observation signal obtained by mixing a plurality of source signals through a plurality of linear transmission paths,
Multi-channel signal accumulation processing for accumulating multi-channel observation signals;
Current signal space matrix generation processing for obtaining a current signal space matrix from the accumulated multi-channel observation signals;
Past future signal space matrix generation processing for generating a past future signal space matrix from accumulated multi-channel observation signals;
A past future signal space orthogonal basis calculation process for obtaining an orthogonal basis of a past future signal space matrix;
A projection residual signal matrix extraction process for obtaining a projection residual signal space matrix when the current signal space matrix is projected onto the past future signal space from orthogonal bases of the current signal space matrix and the past future signal space;
A redundant residual signal space matrix Eb is obtained from the projected residual signal matrix E rank r, and a component orthogonal to the redundant projected residual signal space matrix Eb is extracted from the projected residual signal space matrix E to obtain a corrected projected residual signal. A projection residual signal matrix correction process as a matrix;
Singular value decomposition of the corrected projection residual signal matrix, taking out the left singular vector corresponding to the maximum singular value and obtaining the impulse response of the multi-channel transmission route,
A dereverberation filter calculation process for obtaining a dereverberation filter from the estimated transfer path impulse response;
A dereverberation filtering process that applies the dereverberation filter obtained in the dereverberation filter calculation process to a multi-channel observation signal;
Sound source separation processing that performs sound source separation processing using the signal after the dereverberation processing as input,
A sound source separation method comprising: In a transmission path estimation device that estimates a transmission path from an observation signal observed from a source signal through a plurality of linear transmission paths,
Multi-channel signal storage means for storing multi-channel observation signals;
Current signal space matrix generating means for obtaining a current signal space matrix from the accumulated multi-channel observation signals;
A past future signal space matrix generating means for generating a past future signal space matrix from the accumulated multi-channel observation signals;
Past future signal space orthogonal basis calculating means for obtaining an orthogonal basis of the past future signal space matrix;
Projected residual signal matrix extraction means for obtaining a projected residual signal space matrix when the current signal space matrix is projected onto the past future signal space from orthogonal bases of the current signal space matrix and the past future signal space;
A redundant residual signal space matrix Eb is obtained from the rank r of the projected residual signal matrix E, and a component orthogonal to the redundant projected residual signal space matrix Eb is extracted from the projected residual signal space matrix E to obtain a corrected projected residual. A projection residual signal matrix correction means as a signal matrix;
Singular value decomposition of the corrected projected residual signal matrix, taking out the left singular vector corresponding to the maximum singular value, and obtaining the impulse response of the multi-channel transmission path, transmission path estimation means,
A transmission path estimation apparatus comprising: In the dereverberation device that estimates the transmission path and the source signal from the observation signal observed through the plurality of linear transmission paths from the source signal,
Multi-channel signal storage means for storing multi-channel observation signals;
Current signal space matrix generating means for obtaining a current signal space matrix from the accumulated multi-channel observation signals;
A past future signal space matrix generating means for generating a past future signal space matrix from the accumulated multi-channel observation signals;
Past future signal space orthogonal basis calculating means for obtaining an orthogonal basis of the past future signal space matrix;
Projected residual signal matrix extraction means for obtaining a projected residual signal space matrix when the current signal space matrix is projected onto a past future signal space from orthogonal bases of the current signal space matrix and the past future signal space;
A redundant residual signal space matrix Eb is obtained from the projected residual signal matrix E rank r, and a component orthogonal to the redundant projected residual signal space matrix Eb is extracted from the projected residual signal space matrix E to obtain a corrected projected residual signal. A projection residual signal matrix correction means as a matrix;
Singular value decomposition of the corrected projected residual signal matrix, taking out the left singular vector corresponding to the maximum singular value, and obtaining the impulse response of the multi-channel transmission route,
An inverse filter calculating means for obtaining an inverse filter from the estimated transmission path impulse response;
Means for applying an inverse filter to the multi-channel observation signal and outputting a source signal estimation result;
A dereverberation apparatus comprising: In a sound source separation device that estimates a source signal from an observation signal in which a plurality of source signals are mixed through a plurality of linear transmission paths,
Multi-channel signal storage means for storing multi-channel observation signals;
Current signal space matrix generating means for obtaining a current signal space matrix from the accumulated multi-channel observation signals;
A past future signal space matrix generating means for generating a past future signal space matrix from the accumulated multi-channel observation signals;
Past future signal space orthogonal basis calculating means for obtaining an orthogonal basis of the past future signal space matrix;
Projected residual signal extraction means for obtaining a projected residual signal space matrix when the current signal space matrix is projected onto the past future signal space from orthogonal bases of the current signal space matrix and the past future signal space;
A redundant residual signal space matrix Eb is obtained from the rank r of the projected residual signal matrix E, and a component orthogonal to the redundant projected residual signal space matrix Eb is extracted from the projected residual signal space matrix E to obtain a corrected projected residual. A projection residual signal matrix correction means for making a difference signal matrix;
Singular value decomposition of the corrected projected residual signal matrix, taking out the left singular vector corresponding to the maximum singular value, and obtaining the impulse response of the multi-channel transmission path, transmission path estimation means,
Dereverberation filter calculating means for obtaining an dereverberation filter from the estimated transfer path impulse response;
Dereverberation filtering means for applying the dereverberation filter obtained by the dereverberation filter calculating means to the multi-channel observation signal;
Sound source separation means for performing sound source separation processing using the signal after the dereverberation processing as input,
A sound source separation device comprising:   A program written in a program language that can be read by a computer, and causing the computer to function as the device according to claims 4 to 6.   A recording medium comprising a computer-readable recording medium, wherein the program according to claim 7 is recorded on the recording medium.

Images