JP2013054258A - Sound source separation device and method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a sound source separation device capable of effectively suppressing an additive noise included in an input signal.SOLUTION: In a sound source separation device, a feature vector ψ(t) featuring each time frequency bin of a multichannel observation signal y(t) is classified into N+1 components caused by each of N target sound sources and N additive noises, so that posterior probability of each of the target sound sources and additive noises is estimated by the method of maximum likelihood, and then a covariance matrix ^Rof the n-th target signal and a covariance matrix ^Rof a multichannel observation signal y(t) included in an observation signal are calculated, so that a generalized multichannel Wiener filter wfor recovering the target signal is calculated by determining unnecessary components other than the n-th target signal included in the observation signal on the basis of the covariance matrix ^Rof the n-th target signal and the covariance matrix ^Rof the multichannel observation signal y(t). Finally, an estimate ^Sof the n-th target signal is outputted on the basis of the multichannel observation signal y(t), the generalized multichannel Wiener filter w, and the posterior probability of each of the target sound sources.

Description

  The present invention relates to a sound source separation apparatus that extracts each target signal with high accuracy when the input signal includes a plurality of target signals and additive noise, and a method and program thereof.

  When an acoustic signal is collected in an environment where a plurality of target sound sources exist, a mixed signal in which the target signals overlap each other is often observed. At this time, when the target sound source of interest is an audio signal, the clarity of the target sound is greatly reduced due to the influence of other sound source signals superimposed on the target signal. As a result, it becomes difficult to extract the nature of the original target speech signal (hereinafter referred to as the target signal), and the recognition rate of the automatic speech recognition (hereinafter referred to as speech recognition) system is significantly reduced. Furthermore, when additive noise is present in addition to the target signal, the clarity and the reduction in the recognition rate of the speech recognition system also increase. In order to prevent this reduction in the recognition rate, it is necessary to devise a method (method) for recovering the clarity of the target signal by separating a plurality of target signals.

  The elemental technology for separating the plurality of target signals can be used for various acoustic signal processing systems. For example, a hearing aid that extracts the target signal from the sound collected in the real environment to improve ease of hearing, a TV conference system that improves the intelligibility of the voice by extracting the target signal, and audio used in the real environment It can be used for a recognition system, a machine-human interaction device in a machine control interface, a music information processing system for searching and recording music, and the like.

  FIG. 9 shows a functional configuration example of a conventional sound source separation device 900 disclosed in, for example, Non-Patent Documents 1 and 2, and its operation will be briefly described. The sound source separation device 900 includes a feature vector calculation unit 90, a speech presence probability calculation unit 91, and a 1ch filtering unit 92.

  The feature vector calculator 90 calculates a feature vector that characterizes each time frequency bin of the multi-channel input signal. The speech existence probability calculation unit 91 receives the feature vector as an input, and calculates the existence probability of each of the N target sound sources included in the input signal in each time frequency bin. The existence probability is calculated by performing maximum likelihood estimation of the parameters of the mixture model with N mixture. The 1ch filtering unit 92 sets the existence probability calculated by the speech existence probability calculating unit 91 to 0 (meaning that no signal exists) or 1 (meaning that a signal exists) in each time frequency bin value of the input signal. ) Is multiplied by the converted value to calculate the estimated value of the target signal of the target sound source. By using this method, a plurality of target signals included in the input signal can be recovered.

H. Sawada, S. Araki, and S. Makino, “Underdetermined convolutive blind source separation via frequency bin-wise clustering and permutation alignement,” IEEE Trans. Audio, Speech and Lang. Process., Vol. 19, pp.516- 527, March 2011. H. Sawada, S. Araki, and S. Makino, “A two-stage frequency domain blind source separation method for underdetermined convolutive mictures,” in Proc. IEEE WASPAA, 2007, pp. 139-142.

  However, in the conventional method, it has not been assumed that additive noise is included in the input signal. Therefore, if additive noise is included in the input signal, it cannot be suppressed, and the target signal cannot be effectively recovered.

  The present invention has been made in view of such problems, and even when additive noise is included in an input signal, the additive noise is appropriately suppressed and each target signal of a plurality of target sound sources is recovered. It is an object of the present invention to provide a sound source separation device, a method and a program thereof.

  The sound source separation apparatus according to the present invention includes a feature vector calculation unit, a speech / noise existence probability calculation unit, a speech / noise feature calculation unit, a speech estimation filter calculation unit, and a multi-channel filtering unit. The feature vector calculation unit calculates a feature vector characterizing each time frequency bin of the multi-channel observation signal by normalizing the observation signal in the complex region with its norm. The speech / noise existence probability calculation unit receives the feature vector, classifies the feature vector into N + 1 components each resulting from N target sound sources and additive noise, and relates to each target sound source and additive noise. Maximum likelihood estimation of posterior probability. The speech / noise feature calculation unit receives the posterior probability of the target signal of each target sound source, the posterior probability of additive noise, and the multi-channel observation signal, and is included in the covariance matrix of the nth target signal and the observation signal. Calculate the covariance matrix of the multichannel observation signal. The speech estimation filter calculation unit receives the covariance matrix of the nth target signal and the covariance matrix of the multi-channel observation signal, obtains an unnecessary component other than the nth target signal included in the observation signal, and obtains the target signal Compute a generalized multi-channel Wiener filter to recover. The multi-channel filtering unit receives the multi-channel observation signal, the generalized multi-channel Wiener filter, and the posterior probability relating to each target sound source, and outputs an estimated value of the n-th target signal.

  According to the sound source separation device of the present invention, the multi-channel observation signal is classified and processed into the component caused by each of the N target sound sources and the component caused by the additive noise, so that the additive noise can be effectively processed. Can be suppressed. Specific effects confirmed in the evaluation experiment will be described later.

The figure which shows the function structural example of the sound source separation apparatus 100 of this invention. The figure which shows the operation | movement flow of the sound source separation apparatus. The figure which shows the function structure of the audio | voice and noise presence probability calculation part 20. FIG. The figure which shows the operation | movement flow of the audio | voice and noise presence probability calculation part 20. FIG. The figure which shows the function structural example of the audio | voice / noise characteristic calculation part 30. The figure which shows the operation | movement flow of the audio | voice and noise characteristic calculation part 30 It is a figure which shows the signal waveform before an audio | voice separation process, (a) is the clean audio | voice of the speaker 1, (b) is the clean audio | voice of the speaker 2, (c) is a figure which shows the audio | voice waveform of a mixed signal. It is a figure which shows the speech waveform after a speech separation process, (a) is the speech waveform of the speaker 1 separated by the conventional method, (b) is the speech waveform of the speaker 2 separated by the conventional method, and (c) is this The speech waveform of the speaker 1 separated by the method of the invention is shown in (d), and the speech waveform of the speaker 2 separated by the method of the invention is shown. The figure which shows the function structural example of the conventional audio | voice separation apparatus 900.

  Embodiments of the present invention will be described below with reference to the drawings. The same reference numerals are given to the same components in a plurality of drawings, and the description will not be repeated. Before the description of the embodiment, the observation signal is modeled.

ここで、ｔは時間フレームのインデックス、ｋは周波数インデックスを表す。観測信号ｙ（ｋ，ｔ）は、Ｍ個の混合信号であるｙ（ｋ，ｔ）＝[Ｙ_１（ｋ，ｔ）…Ｙ_Ｍ（ｋ，ｔ）]^Ｔであり、ｘ_ｎ（ｋ，ｔ）はｎ番目のチャネル応答を伴った信号成分ｘ_ｎ（ｋ，ｔ）＝ｈ_ｎ（ｋ）Ｓ_ｎ（ｋ，ｔ）である。Ｓ_ｎ（ｋ，ｔ）はｎ番目の目的信号である。 [Modeling of observed signals]
It is assumed that the observation signal includes both the target signal caused by N (N ≧ 1) point sound sources and additive noise. In this case, the multi-channel observation signal y (k, t) observed using M microphones is expressed by the following equation (1) in the complex spectral region through a clipping process in a short time window and a short time Fourier transform. It is expressed as follows.

Here, t represents a time frame index, and k represents a frequency index. The observed signal y (k, t) is M mixed signals y (k, t) = [Y ₁ (k, t)... Y _M (k, t)] ^T , and x _n (k, t) t) is the signal component x _n (k, t) = h _n (k) S _n (k, t) with the _nth channel response. S _n (k, t) is the nth target signal.

The channel response h _n (k) between the n-th sound source and each microphone is expressed as h _n (k) = [H _1n (k)... H _Mn (k)] ^T. The additive noise component v (k, t) is v (k, t) = [V _1n (k)... V _M (k)] ^T. In the present invention, it is assumed that the target additive noise changes sufficiently slowly compared to other sound sources. In the present invention, the channel response is time invariant.

Since each processing of the present invention is performed individually for each frequency k, in the following description, the frequency index k is appropriately omitted for simplicity.
In addition, the observation signal introduces a sparse assumption that there is only sound due to at least one point sound source in a certain time frequency bin, and no sound due to other point sound sources. Model the observed signal as shown in (2).

That is, it is assumed that only the sound caused by additive noise and the nth target sound source exists in the time frequency bin. Alternatively, as shown in Equation (3), it is assumed that there is no sound due to the point sound source, and there is only noise.

  If the sparseness assumption is introduced in this way, each time frequency bin can be roughly divided into characteristics due to any of the N target sound sources or characteristics due to noise alone. . The following example will be described on the assumption that the observation signal y (t) is modeled as described above.

  FIG. 1 shows a functional configuration example of a sound source separation device 100 of the present invention. The operation flow is shown in FIG. The sound source separation apparatus 100 includes a feature vector calculation unit 10, a voice / noise existence probability calculation unit 20, a voice / noise feature calculation unit 30, a voice estimation filter calculation unit 40, and a multi-channel filtering unit 50. To do. The function of each unit of the sound source separation device 100 is realized by reading a predetermined program into a computer configured by, for example, a ROM, a RAM, and a CPU, and executing the program by the CPU.

  The feature vector calculation unit 10 calculates a feature vector ψ (t) characterizing each time frequency bin of the multichannel observation signal y (t) by normalizing the observation signal in the complex region with its norm (step S10). The speech / noise existence probability calculation unit 20 receives the feature vector ψ (t) as an input, classifies the feature vector ψ (t) into N + 1 components each resulting from N target sound sources and additive noise, Maximum likelihood estimation is performed on posterior probabilities regarding each target sound source and additive noise (step S20).

The speech / noise feature calculation unit 30 receives the posterior probability of the target signal of each target sound source, the posterior probability of additive noise, and the multichannel observation signal y (t) as inputs, and the covariance matrix of the nth target signal. ^ covariance matrix of a multichannel observed signal y included in _{R Xnxn} the observation signal (t) _^ to calculate the _{R yy} (step S30). Speech estimation filter calculation unit 40, n-th and covariance matrix _^ R xnxn object signals, a covariance matrix ^ R _yy multichannel observed signal y (t) as an input, n-th included in the observation signal An unnecessary component other than the target signal is obtained, and a generalized multi-channel Wiener filter w _n ^(β) for recovering the target signal is calculated (step S40). Multichannel filtering unit 50 as inputs and posterior probability multichannel observed signal y (t) and generalized multi-channel Wiener filter w _{n ^(beta)} for each target source, estimates of the n-th target signal ^ ^~ S _n ^(β) is output (step S50). The control unit 60 controls time-series operations between the above-described units. It should be noted that the notations such as ^ ^to are the correct notations that are located immediately above the variables as described in the figures and equations.

  In the conventional binary mask processing for dividing the voice existence probability described in the background art by binary of 1/0, additive noise in the time frequency bin where the target signal exists cannot be removed. According to the method, it is possible to suppress the additive noise included in the observation signal and separate and extract each of the N target signals.

In the following, the function of each part of the sound source separation device 100 will be described in more detail.
[Feature vector calculator]
Each of the multi-channel observation signals y (t) observed by M microphones is converted into a signal in the complex spectral region by a short-time Fourier transform process. The vector of the complex spectrum Y ₁ (k, t) of the first microphone to the complex spectrum Y _M (k, t) of the _Mth microphone is y (t) = [Y ₁ (k, t)... Y _M ( k, t)] ^T.
The multi-channel observation signal y (t) in the complex region is normalized by its norm, and the feature vector ψ (t) is calculated by Expression (4).

[Speech / Noise Presence Probability Calculator]
FIG. 3 shows a more specific functional configuration example of the voice / noise existence probability calculation unit 20. The operation flow is shown in FIG. The voice / noise existence probability calculation unit 20 includes a cluster classification unit 201, an initialization unit 202, an expected value calculation unit 203, a maximization unit 204, and a convergence determination unit 205. The speech / noise existence probability calculation unit 20 receives the feature vector ψ (t) as an input, and converts the feature vector ψ (t) into N components of “target signal + additive noise” and “additive noise”. Are automatically classified into clusters resulting from the above, and the posterior probability p [C _n | ψ (t), θ] relating to each cluster is estimated using an expected value maximization method (EM algorithm) and output as a speech existence probability.

  The cluster classification unit 201 models the feature vector ψ (t) with a probability density function using equation (5) (step S201). That is, the cluster classification unit 201 classifies the feature vector ψ (t) into a component caused by each of the N target sound sources and a component caused by additive noise, and is modeled by N + 1 probability density functions. To do.

ここで、混合分布のパラメータθはθ＝{ａ_１，σ_１，…，ａ_Ｎ＋１，σ_Ｎ＋１}であり、ｎ番目の分布の重みパラメータα_ｎは、Σ_ｎα_ｎ＝１，０≦α_ｎ≦１の制約を満たす。 A parameter θ _n characterizing the density function is represented by θ _n = {a _n , σ _n }. a _n is the average of the n-th cluster C _n , and σ _n ² is its variance.

Here, the parameter θ of the mixed distribution is θ = {a ₁ , σ ₁ ,..., A _{N + 1} , σ _{N + 1} }, and the weight parameter α _n of the _nth distribution is Σ _n α _n = 1, 0 ≦ α Satisfies the constraint of _n ≦ 1.

ここで（ｑ）は、ＥＭアルゴリズムの繰り返し回数を表す。 The initialization unit 202 initializes each mixture distribution parameter θ _n with a random number (step S202).
The expected value calculation means 203 calculates an expected value (E step) using equation (7) (step S203).

Here, (q) represents the number of repetitions of the EM algorithm.

The maximizing means 204 calculates a covariance matrix R related to the feature vector ψ (t) using Expression (8), and performs eigenvalue decomposition on R.

Then, the eigenvector corresponding to the maximum eigenvalue is substituted into the average parameter a _n ^(q) , the dispersion parameter σ _n ² is updated by equation (9), and the mixture weight parameter α _n is updated by equation (10) (M Step) (step S204).

The convergence determination unit 205 repeats the processing of step S203 and step S204 until the update width of the dispersion parameter σ _n ² and the mixture weight parameter α _n becomes sufficiently small (convergence of step S205). By repeating the calculation by the EM algorithm until convergence, it is possible to calculate the existence probability of each signal component in each time frequency bin. Note that the posterior probability p [C _n | ψ (t), θ] or simply p [C _n | ψ (t)] (n = 1,..., N + 1) regarding the cluster C _n obtained by this processing is The characteristic shown in Formula (11) shall be satisfy | filled.

  Equation (11) shows that the probability that the nth signal exists in a certain time frequency bin is completely defined by the feature vector ψ (t). Since the calculation of the speech existence probability is performed independently at each frequency k, a permutation (replacement) problem occurs that a certain n-th signal component has different cluster indexes at different frequencies. A conventional method (for example, Non-Patent Document 1) can be used to solve the permutation problem for bundling the same signal between frequencies.

[Voice / Noise Feature Calculator]
FIG. 5 shows a more specific functional configuration example of the voice / noise feature calculation unit 30. The operation flow is shown in FIG. The speech / noise feature calculation unit 30 includes observation signal covariance matrix calculation means 301, additive noise covariance matrix calculation means 302, and target signal covariance matrix calculation means 303.
The observation signal covariance matrix calculation unit 301 calculates a covariance matrix R _yy of the multichannel observation signal y (t). The covariance matrix R _yy of the multichannel observation signal y (t) is given by Expression (12).

In actual calculation, a value obtained by multiplying the vector y (t) of the multi-channel observation signal and its Hermitian transposition y ^H (t) is averaged by the total number of observation frames T (formula (13), step S301).

Next, a method for calculating a covariance matrix related to a target signal that does not include an additive noise component will be described. Considering that the multichannel observation signal y (t) is classified into N + 1 clusters in the speech / noise existence probability calculation unit 20, the covariance matrix R _yy of the observation signal shown in Expression (12) is as follows: Can be decomposed into the sum of each cluster.

The nth integral term is given by equation (15).

covariance matrix for the n-th cluster, and covariance matrix R _vv relates additive noise, expressed by n-th form of the sum of the covariance matrix R _Xnxn the target signal. Since the (N + 1) -th cluster has N target sound sources, the feature regarding additive noise is captured. That is, the covariance matrix R _{N + 1} related to the ( _{N + 1} ) th cluster represents a covariance matrix related to additive noise (R _{N + 1} = R _vv ).

The additive noise targeted in this embodiment is assumed to be noise that changes sufficiently slowly as compared with the target sound source. Therefore, the additive noise component included in the 1st to Nth clusters and the N + 1th cluster are included. The observed additive noise component can be considered to have sufficiently close characteristics. Accordingly, the covariance matrix ^ R _vv relates additive noise, the covariance matrix _^ R xnxn about the target signal can be calculated as follows.

The additive noise covariance matrix calculation means 302 receives the multi-channel observation signal y (t) and the posterior probability p [C _n | ψ (t), θ] of the target sound source as input. A value obtained by multiplying the vector, its Hermitian transposition y ^H (t), and the posterior probability p [C _{N + 1} | y (t)] relating to additive noise over the total number of observation frames T, and adding the covariance matrix of additive noise ^ R _vv is calculated (formula (18), step S302).

The target signal covariance matrix calculation means 303 receives the multi-channel observation signal y (t), the posterior probability p [C _n | ψ (t), θ] of the target sound source, and the additive noise covariance matrix ^ R _vv as inputs. , A value obtained by multiplying the vector of the multi-channel observation signal y (t), its Hermitian transposition y ^H (t), and the posterior probability p [C _n | y (t)] for each target sound source with the total number of observation frames T from the values, additive noise covariance matrix _^ by subtracting _{R vv} to calculate the covariance matrix _{^ R Xnxn} for each signal of interest (the formula (19), step S303).

[Sound estimation filter calculator]
The speech estimation filter calculation unit 40 receives the covariance matrix R _{yy of the} multichannel observation signal y (t) and the covariance matrix ^ R _{xnxn related} to the target signal, and minimizes the signal component caused by the nth target signal. Estimate the square error.

The least square error estimate of the nth target signal component is given as follows:

Equation (20) is derived by introducing the sparseness assumption. Posteriori probability for the n-th cluster in the above equation, the minimum square error estimation value _{^{E {~ S n (t)}} | y (t), C n} has the effect of smoothly mask. The second term on the right side of the above equation is equivalent to obtaining the multi-channel Wiener filter w that minimizes the following square error ε _n (w).

The filter w that minimizes ε _n (w) is generally derived by solving the following Yule-walker equation.

Here, u ₁ becomes u ₁ = [10... 0] ^T when the n-th target signal in the first microphone is to be recovered. Furthermore, the filter of Expression (22 ⁾ is generalized to a filter w _n ^(β) that can adjust by β how much the component other than the nth target sound source is suppressed, as in the following expression. can do.

Here, _Run , which is an unnecessary component other than the nth target signal, is calculated as follows.

The speech estimation filter calculation unit 40 inputs an unnecessary component R _un other than the n-th target signal as a covariance matrix R _{yy of the} multichannel observation signal y (t) and a covariance matrix ^ R _xnxn related to the target signal. And a generalized multi-channel Wiener filter that recovers the target signal is calculated by Equation (24).

[Multi-channel filtering section]
The multi-channel filtering unit 50 includes a multi-channel observation signal y (t), a generalized multi-channel Wiener filter w _n ^(β) , a posterior probability p [C _n | ψ (t), θ] for each target signal, As an input, the estimated value of the nth target signal is filtered by Expression (26) and output.

[Evaluation experiment]
An evaluation experiment was conducted for the purpose of evaluating the performance of the sound source separation device 100 of the present invention. The experimental conditions were as follows.
Two target signals (N = 2) were used, and data of 12 male and female speakers extracted at random from the TIMIT database were used. As the mixing conditions, three conditions were mixed: mixing two female voices, mixing two male speakers, and mixing one female speaker and one male speaker. The positions of the two speakers were 2 m away from the microphone array and 160 degrees apart from each other, and were mixed at the same volume level (SIR: Signal-to-Interference Ratio = 0 dB).

  As additive noise, bubble noise extracted from the noisex database was used and added so that the SNR (Signal-to-Noise Ratio) of each microphone signal was 5 to 20 dB. As the present invention, a multi-channel Wiener filter (β = 1 in equation (24)) and MVDR (Minimimum Variance Distortionless Responds, β = 0 in equation (24)) were created and shown in Non-Patent Documents 1 and 2. Comparison was made with the prior art. As the number of microphones, two conditions of 8 and 16 were prepared.

Table 1 shows the SNR comparison results, and Table 2 shows the SIR comparison results.

表１と２の比較結果から明らかなように、この発明の音源分離方法の方が、マイクロホンの数によらず高い性能を示した。

As is clear from the comparison results in Tables 1 and 2, the sound source separation method of the present invention showed higher performance regardless of the number of microphones.

  7 and 8 show the evaluation results as signal waveforms. FIG. 7 shows a waveform before processing, where (a) is a clean voice of speaker 1, (b) is a clean voice of speaker 2, and (c) is a voice obtained by mixing each speaker voice and noise. . FIG. 8 shows signal waveforms after sound source separation. (A) and (b) are signal waveforms of speaker 1 and speaker 2 separated by the conventional method, and (c) and (d) are of speaker 1 and speaker 2 separated by the sound source separation method of the present invention. It is a signal waveform. Comparing waveforms around 4 seconds when the speaker's signal is interrupted, it can be seen that the SNR is better when the sound source is separated by the method of the present invention. As described above, the sound source separation method of the present invention enables extraction of a target signal in which additive noise is effectively suppressed.

  When the processing means in the speech separation apparatus 100 described above is realized by a computer, the processing contents of functions that each apparatus should have are described by a program. Then, by executing this program on the computer, the processing means in each apparatus is realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape or the like, and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only). Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording medium, MO (Magneto Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. 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. Further, the program may be distributed by storing the program in a recording device of a server computer and transferring the program from the server computer to another computer via a network.

  Each means may be configured by executing a predetermined program on a computer, or at least a part of these processing contents may be realized by hardware.

Claims (5)

A feature vector calculation unit that calculates a feature vector that characterizes each time-frequency bin of the multi-channel observation signal by normalizing the observation signal in the complex region with its norm;
Using the feature vector as an input, the feature vector is classified into N + 1 components each resulting from N target sound sources and additive noise, and the maximum likelihood estimation of posterior probabilities for each target sound source and additive noise is performed・ Noise existence probability calculator,
Using the posterior probability for the target signal of each target sound source, the posterior probability for additive noise, and the multichannel observation signal as input, the covariance matrix of the nth target signal and the multichannel observation signal included in the observation signal A speech / noise feature calculator that calculates the covariance matrix of
Generalization of recovering the target signal by obtaining an unnecessary component other than the nth target signal included in the observation signal by using the covariance matrix of the nth target signal and the covariance matrix of the multi-channel observation signal as inputs. A speech estimation filter calculator for calculating a multi-channel Wiener filter;
A multi-channel filtering unit that receives the multi-channel observation signal, the generalized multi-channel Wiener filter, and the posterior probability associated with each target sound source, and outputs an estimated value of the n-th target signal;
A sound source separation apparatus comprising: The sound source separation device according to claim 1,
The voice / noise feature calculator is
A value obtained by multiplying the vector of the multi-channel observation signal y (t) by the Hermitian transposition y ^H (t) is averaged over the total number of observation frames T to obtain a covariance matrix R _yy of the multi-channel observation signal y (t). An observation signal covariance matrix calculation means for calculating,
The multi-channel observation signal y (t) and the posterior probability p [C _n | ψ (t), θ] of the target sound source are input, and the vector of the multi-channel observation signal y (t) and its Hermitian transposition y ^H (t) The additive noise covariance is calculated by averaging the value obtained by multiplying the posterior probability p [C _{N + 1} | y (t)] with respect to the additive noise by the total number of observed frames T and calculating the additive noise covariance matrix ^ _Rvv. Matrix calculation means;
The multi-channel observation signal y (t) is input using the multi-channel observation signal y (t), the posterior probability p [C _n | ψ (t), θ] of the target sound source and the covariance matrix ^ R _vv of the additive noise. Of the additive noise from a value obtained by averaging the observed signal T and the value obtained by multiplying the vector of the vector and its Hermitian transposition y ^H (t) and the posterior probability p [C _n | y (t)] for each target sound source. A target signal covariance matrix calculating means for calculating a covariance matrix ^ R _xnxn for each target signal by subtracting the matrix ^ R _vv ;
A sound source separation device comprising: A feature vector calculation process for calculating a feature vector that characterizes each time-frequency bin of a multi-channel observation signal by normalizing the observation signal in the complex region by its norm,
Using the feature vector as an input, the feature vector is classified into N + 1 components each resulting from N target sound sources and additive noise, and the maximum likelihood estimation of posterior probabilities for each target sound source and additive noise is performed・ Noise existence probability calculation process,
Using the posterior probability for the target signal of each target sound source, the posterior probability for additive noise, and the multichannel observation signal as input, the covariance matrix of the nth target signal and the multichannel observation signal included in the observation signal Voice / noise feature calculation process to calculate the covariance matrix of
Generalization of recovering the target signal by obtaining an unnecessary component other than the nth target signal included in the observation signal by using the covariance matrix of the nth target signal and the covariance matrix of the multi-channel observation signal as inputs. A voice estimation filter calculation process for calculating a multi-channel Wiener filter;
A multi-channel filtering process for outputting an estimated value of the n-th target signal, with the multi-channel observation signal, the generalized multi-channel Wiener filter, and the posterior probability for each target sound source as inputs;
A sound source separation method comprising: In the sound source separation method according to claim 3,
The voice / noise feature calculation process is as follows:
A value obtained by multiplying the vector of the multi-channel observation signal y (t) by the Hermitian transposition y ^H (t) is averaged over the total number of observation frames T to obtain a covariance matrix R _yy of the multi-channel observation signal y (t). An observation signal covariance matrix calculation step to be calculated;
The multi-channel observation signal y (t) and the posterior probability p [C _n | ψ (t), θ] of the target sound source are input, and the vector of the multi-channel observation signal y (t) and its Hermitian transposition y ^H (t) The additive noise covariance is calculated by averaging the value obtained by multiplying the posterior probability p [C _{N + 1} | y (t)] with respect to the additive noise by the total number of observed frames T and calculating the additive noise covariance matrix ^ _Rvv. Matrix calculation step;
The multi-channel observation signal y (t) is input using the multi-channel observation signal y (t), the posterior probability p [C _n | ψ (t), θ] of the target sound source and the covariance matrix ^ R _vv of the additive noise. And the Hermitian transposition y ^H (t) and the value obtained by multiplying the posterior probability p [C _n | y (t)] for each target sound source by the total number of observed frames T, An objective signal covariance matrix calculating step of calculating a covariance matrix ^ R _xnxn for each target signal by subtracting the covariance matrix ^ R _vv ;
A sound source separation method comprising:   A program for causing a computer to function as the sound source separation device according to claim 1.

Images