JP2013097176A - Sound source separation device, sound source separation method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a sound source separation technique which provides a method for efficiently estimating an arrival time difference δ and eliminates the need of overall search operation to increase the speed of estimation.SOLUTION: An arrival time difference δ, a power spectrum σof noise, a sound source spectrum s, and a sound source existence probability p(k|θ) are estimated from an observation signal x=[x,x]in a frequency domain. The estimation is performed in accordance with formula where: mis a posterior probability indicative of an expected value of attribution of the observation signal xto a sound source k; D is an interval between two microphones; c is a speed of an original signal; φ=sinc(2πfD/c) is true; ξ=[x-φ(x-s)] is true; ψand ψare phases of spectra sand ξrespectively; and δis the arrival time difference. The uncertainty of ±π which the estimated arrival time difference δincludes is corrected.

Description

  The present invention belongs to the technical field of signal processing. In particular, the present invention relates to a sound source separation technique for estimating and separating and extracting each original signal from an observed signal in a situation where a plurality of original signals are mixed with noise and observed with two microphones. In particular, the present invention belongs to a blind sound source separation technique for estimating each original signal from only the observation signal in which a plurality of original signals and noise are mixed without using the information on the original signals and how they are mixed.

Non-Patent Document 1 is known as a conventional technique of sound source separation. In Non-Patent Document 1, the arrival time difference δ _k between the two microphones of the original signal emitted from the sound source k is calculated.

As estimated.

Yosuke Izumi, Junki Ono, Shigeki Hatakeyama, “Underdetermined Blind Source Separation under Noise / Reverberation Environment Based on Sparse Mixture Model”, Proceedings of the IEICE General Conference, March 2008

However, in the prior art, since an analytical update formula for the arrival time difference δ _k is not given, it is necessary to estimate the arrival time difference δ _k by a full search operation that requires a large calculation cost.

  The present invention focuses on the nature of the arrival time difference δ, provides a method for efficiently estimating the arrival time difference δ, eliminates the full search operation required in the prior art, and provides a high-speed sound source separation technique. Objective.

In order to solve the above problems, according to the first aspect of the present invention, in a situation where a plurality of original signals are mixed together with noise and observed by two microphones, each original signal is separated from the observed signal. Extract. The source signal source index is k, and the arrival time difference δ of the original signal to the two microphones from the frequency domain observation signal x _{f, t} = [x _{f, t, L} , x _{f, t, R} ] ^T _{f, k} , noise power spectrum σ _f ² , original signal spectrum s _{f, t, k} and sound source existence probability p (k | θ) are estimated. Generate separated signals y _{f, t, k} from observed signals x _{f, t} and estimated parameters θ = {δ _{f, k} , σ _f ² , s _{f, t, k} , p (k | θ)}. . The posterior probability that the observed signal x _{f, t} indicates the expected value attributed to the sound source k is m _{f, t, k} , the distance between the two microphones is D, the speed of the original signal is c, φ _f = sinc ( 2πfD / c), and ξ _{f, t, k} = [x _{f, t, R} −φ _f (x _{t, t, L} −s _{f, t, k} )], and spectra s _{f, t, k} and ξ The phases of _{f, t, k} are ψ _sk and ψ _ξk , respectively, and the arrival time difference δ _{f, k} is

Estimate as The uncertainty of ± π included in the estimated arrival time difference δ _{f, k} is corrected.

  The present invention provides a method for efficiently estimating the arrival time difference δ, and has the effect of enabling high-speed sound source separation.

The functional block diagram of the sound source separation apparatus which concerns on 1st embodiment. The figure which shows the processing flow of the sound source separation apparatus which concerns on 1st embodiment. The functional block diagram of an E step calculation part. The functional block diagram of an M step calculation part. The figure which shows the processing flow of a parameter estimation part. The figure which shows the environment which performed the simulation. FIG. 7A shows a simulation result when the number of sound sources is two, and FIG. 7B shows a simulation result when the number of sound sources is three. The functional block diagram of the M step calculation part in the case of performing the process of a sound source spectrum estimation part before the process of a time difference estimation part.

  Hereinafter, embodiments of the present invention will be described. In the drawings used for the following description, constituent parts having the same function and steps for performing the same process are denoted by the same reference numerals, and redundant description is omitted. Further, the processing performed for each element of a vector or matrix is applied to all elements of the vector or matrix unless otherwise specified.

<Sound source separation apparatus 1 according to the first embodiment>
FIG. 1 is a functional block diagram of a sound source separation device 1 according to the present embodiment, and FIG. 2 shows a processing flow thereof.

  The sound source separation device 1 includes a frequency domain conversion unit 11, a parameter estimation unit 12, a separated signal generation unit 13, and a time domain conversion unit 14. The parameter estimation unit 12 further includes an E step calculation unit 121 and an M step calculation unit 122. Including.

First, the observation signal will be described. There are K signal sources (hereinafter also referred to as “sound sources”), k (k = 1, 2,..., K) is an index of a sound source, and a signal (hereinafter referred to as “original signal”) emitted from the sound source k is s _k. (T). In a situation where a plurality of original signals s ₁ (t),..., S _K (t) are observed with two microphones L and R together with noise, a time domain observation signal observed with the microphone L is expressed as x _L. (T), a time domain observation signal observed by the microphone R is x _R (t), and a time domain observation signal observed by the two microphones L and R is x (t) = [x _L ( _t), and x R ^{(t)] T.} “ ^T ” represents transposition. t represents a frame number and a time corresponding to the frame number. When T is the total number of time frames, t = 0, 1,..., T−1. Here, the observation signal in the frequency domain is _{expressed as xf, t} = [ _{xf, t, L} , _{xf, t, R} ] ^T. Note that f is a discrete point obtained by equally dividing the sampling frequency f _s by F, and f∈ {0, f _s / F,..., (F−1) f _s / F}. Hereinafter, when there is no notice, the observation signal means the observation signal x _{f, t} = [x _{f, t, L} , x _{f, t, R} ] ^T in the frequency domain, and in the case of the observation signal in the time domain, Specify clearly.
Here, the observed signal is

Suppose that Here, s _{f, t, k} is the spectrum of the original signal s _k (t) (hereinafter also referred to as “sound source spectrum”), and n _{f, k} = [n _{f, k, L} , n _{f, k, R} ] Represents additive noise in the microphones L and R. B _{f, k} = [b _{f, k, L} , b _{f, k, R} ] ^T is a steering vector related to the sound source k (a vector for specifying the direction of the sound source k, also referred to as a direction vector), and the original signal If the arrival time difference between s _k (t) and the microphones L and R is δ _k ,

(See Non-Patent Document 1). In the present embodiment, it is assumed that the sound source and the microphone are fixed within the observation time of the original signal, and that all the K sound sources are arranged at different positions. That is, it is assumed that the steering vector b _{f, k} takes a different value depending on the value of k regardless of the time t. The purpose of sound source separation is to estimate all sound source spectra s _{f, t, k} using only the observation signals x _{f, t} .

The sound source separation apparatus 1 according to the present embodiment receives a time domain observation signal x (t) = [x _L (t), x _R (t)] ^T as an input, and uses a time domain separation signal (estimated original signals). Signal) y _k (t) is output. Hereinafter, the processing content of each part is demonstrated.

<Frequency domain converter 11>
First, the frequency domain transform unit 11 receives the time domain observation signal x (t) = [x _L (t), x _R (t)] ^T collected by the microphones L and R, and uses this as a short-time Fourier transform. The frequency domain observation signal x _{f, t} = [x _{f, t, L} , x _{f, t, R} ] is converted to ^T by conversion or the like (s 1), and is output to the parameter estimation unit 12 and the separated signal generation unit 13.

<Parameter estimation unit 12>
Next, the parameter estimation unit 12 estimates a parameter θ necessary for sound source separation from the observation signals _{xf, t} (s2). Note that θ = {δ _{f, k} , σ _f ² , s _{f, t, k} , p (k | θ)}, where δ _{f, k} represents the arrival time difference of the original signal to the two microphones. σ _f ² represents the noise power spectrum, s _{f, t, k} represents the sound source spectrum, and p (k | θ) represents the sound source existence probability (contribution rate of the sound source k in the mixed signal).

  In the present embodiment, the following two assumptions are used to estimate the above parameters.

(Assumption 1) is an assumption that the noise n can be modeled by stationary noise according to a normal distribution of mean 0 and covariance matrix σ _f ² V _f . Here, σ _f ² is the noise power at the frequency f, and V _f is, for example, diffuse noise.

(See Non-Patent Document 1). Here, c is the speed of the original signal (sound speed, etc.), and D is the interval between the two microphones.

  (Assumption 2) is a signal in which the original signal is sparse (that is, a signal in which non-zero components (non-zero components) are sparse, in other words, a signal in which most (or many) components are zero). This is an assumption. That is, it is assumed that only one original signal is dominant at a certain time frequency (f, t). According to (Assumption 2), Equation (2) is

(See Non-Patent Document 1).
Based on the above (Assumption 1) and (Assumption 2), the likelihood that the original signal emitted from the sound source k existing in the direction of the arrival time difference δ _{f, k} arrives and x _{f, t} is observed is

(See Non-Patent Document 1). “ ^H ” represents Hermitian transpose. Here, b _{f, k} is expressed by Equation (3). Thus, the arrival time difference δ _{f, k} , the sound source spectrum s _{f, t, k} and the noise power spectrum σ _f ² are expressed by the log likelihood function.

Is obtained by maximum likelihood estimation as a parameter that maximizes (see Non-Patent Document 1). However, in the above, p (k | δ _{f, k} , σ _f ² , s _{f, t, k} ) represents a sound source existence probability and is a contribution rate of the sound source k in the mixed signal,

(See Non-Patent Document 1).
Specifically, an expected value maximization method (hereinafter also referred to as “EM algorithm”) is applied, and the Q function

Parameter θ = {δ _{f, k} , σ _f ² , s _{f, t, k} , p (k | θ)} that maximizes is obtained by repeating the following E step and M step (see Non-Patent Document 1). ). Here, m _{f, t, k} is given by the following equation (11), p (x _{f, t} | k, θ ′) is given by the above-mentioned equation (6), and θ ′ is the previous iteration. Means the parameter being obtained.

  This iterative calculation is performed by the parameter estimation unit 12. Details of the processing of the parameter estimation unit 12 will be described below.

  3 is a functional block diagram of the E step calculation unit 121, FIG. 4 is a functional block diagram of the M step calculation unit 122, and FIG. 5 shows a processing flow of the parameter estimation unit 12.

First, the parameter estimation unit 12 initializes each parameter (s20). Among the parameters, σ _f ² , δ _{f, k} , and p (k | θ) are initialized. For example, σ _f ² = | x _{f, t = 0, L} | ² , δ _{f, k} = (D / c) cos α _k (c is the speed of the original signal (sound speed, etc.), D is the microphone interval, and α _k is An initial value in the direction of the sound source k (appropriate value between −π / 2 to π / 2)), p (k | θ) = 1 / K. Further, using the initialized parameters δ _{f, k} and x _{f, t = 0} , s _{f, t, k} is initialized based on Expression (3) and Expression (19) described later. The number of updates n = 0. Note that the maximum number of updates N and the convergence determination threshold value Δ are set in advance by the designer or user of the apparatus.

  The E step calculation unit 121 includes a posterior probability estimation unit 1211 and a Q function calculation unit 1212 (see FIG. 3). The M step calculation unit 122 includes a time difference estimation unit 1221, a sound source spectrum estimation unit 1222, a noise power estimation unit 1223, and a sound source existence probability estimation unit 1224. The time difference estimation unit 1221 further includes an arctangent calculation unit 12211, a time correction unit 12212, and the like. (See FIG. 4). The processes in the posterior probability estimation unit 1211 and the Q function calculation unit 1212 are collectively referred to as E step, and the processes in the time difference estimation unit 1221, the sound source spectrum estimation unit 1222, the noise power estimation unit 1223, and the sound source existence probability estimation unit 1224 are combined. This is called M step.

(A posteriori probability estimation unit 1211)
The posterior probability estimation unit 1211 of the E step calculation unit 121 uses the observed signal x _{f, t} and the parameter θ ′ obtained in the previous iteration (however, the parameter θ ′ obtained in the previous iteration). Is present, that is, in the first posterior probability estimation, the above-described initialized parameter) is used as an input, and these values are used to determine the posterior probability m _{f, t, k} = p (k | x _{f, t} , θ ′) is obtained from the following equation (11) (s22), and is output to the Q function calculator 1212 and the M step calculator 122.

Note that p (x _{f, t} | k, θ ′) is given by equation (6). The posterior probability m _{f, t, k} represents the posterior probability that the observed signal x _{f, t} belongs to the sound source k.

(Inverse tangent calculation unit 12211)
The arc tangent calculation unit 12211 of the time difference estimation unit 1221 of the M step calculation unit 122 has an observation signal x _{f, t} , a posteriori probability m _{f, t, k,} and a parameter θ ′ ( More specifically, it is the sound source spectrum s _{f, t, k} of θ ′, provided that there is no parameter θ ′ obtained in the previous iteration, that is, in the first arc tangent calculation. And the above-described initialized parameter) as inputs, and using these values, an arrival time difference δ _{f, k} (more specifically _, a value 2πfδ _{f, k obtained} by multiplying the arrival time difference δ _{f, k} by 2πf) is as follows: (S25) and output to the time correction unit 12212.

[psi _sk (where subscript sk represents _{s k)} and ψ _ξk (although representing a subscript Kushik is xi] _k), respectively _{s f, t, k} and xi] _{f, t,} represents the phase of the _k. The derivation of equation (13) will be described later.

In the prior art, δ _k is estimated by a discrete full search based on Equation (1), which requires a lot of calculation cost. In this embodiment, δ _{f, k} is calculated based on Equation (13). The calculation cost can be reduced without requiring a full search. Moreover, the point which calculates | requires arrival time difference (delta) _{f, k} for every frequency f based on Formula (13) differs from a prior art.

(Time correction unit 12212)
The time correction unit 12212 of the time difference estimation unit 1221 of the M step calculation unit 122 includes an observation signal x _{f, t} , a posteriori probability m _{f, t, k,} and a parameter θ ′ obtained by the previous iteration (however, When the parameter θ ′ obtained in the previous iteration does not exist, that is, in the first time correction, the above-described initialized parameter) is used as an input, and using these values, the equation ( The uncertainty of ± π included in the arrival time difference δ _{f, k} estimated in 13) is corrected (s26). This correction is such that the left side 2πfδ _{f, k} of equation (13) takes a value from −π to π, whereas the arctangent of the right side of equation (13) returns only a value from −π / 2 to π / 2. Therefore, it is necessary to correctly obtain a value when 2πfδf _{, k} is in the range of −π to −π / 2 and π / 2 to π. When the value obtained by Expression (13) is described as 2πfδ ′ _{f, k} , correction is performed as follows.

Here ξ _{f, t, k,} respectively formula φ _f (14), the given by (15), ψ _sk (where subscript sk represents _{s k), ψ ξk} (where subscript Kushik is xi] _k Represents the phases of s _{f, t, k} and ξ _{f, t, k} , respectively. The derivation of equations (17) and (18) will be described later.

Further, the time correcting unit 12212 is corrected value 2Paiefuderuta _f, the _k divided by 2 [pi] f, calculated arrival time difference [delta] _{f, k,} outputted to the sound source spectrum estimation section 1222 and the noise power estimation section 1223 and the Q function calculating unit 1212 To do.

(Sound source spectrum estimation unit 1222)
The sound source spectrum estimation unit 1222 of the M step calculation unit 122 receives the arrival time difference δ _{f, k} and the observation signal x _{f, t} and inputs the sound source spectrum s _{f, t, k} using the following formula ( 19) (s27) and output to the noise power estimation unit 1223 and the Q function calculation unit 1212.

Here, b _{f, k} and V _f are expressed by equations (3) and (4), respectively.

(Noise power estimation unit 1223)
The noise power estimation unit 1223 of the M step calculation unit 122 receives the arrival time difference δ _{f, k} , the sound source spectrum s _{f, t, k} , the observation signal x _{f, t} and the posterior probability m _{f, t, k} , Is used to estimate the power spectrum σ _f ² of noise by the following equation (20) (s28) and output to the Q function calculator 1212.

T is the total number of time frames.

(Sound source existence probability estimation unit 1224)
The sound source existence probability estimation unit 1224 of the M step calculation unit 122 receives the posterior probabilities m _{f, t, and k} , estimates the sound source existence probability by the following equation (21) (s29), and outputs it to the Q function calculation unit 1212 To do.

(Q function calculator 1212)
The Q function calculation unit 1212 of the E step calculation unit 121 includes an observation signal x _{f, t} and parameters θ = {δ _{f, k} , σ _f ² , s _{f, t, k} , p (k | θ)}, The posterior probabilities m _{f, t, and k} are used as inputs, and the Q function is obtained by the above equation (10) using these values (s30).

When the processing from s20 to s30 is completed, the parameter estimation unit 12 determines that (condition 1) the amount of change in the value of the Q function (| Q (θ | θ ^n-1 ) −Q (θ | θ ⁿ ) |) is predetermined. (Condition 2) Whether the number of updates n is equal to or greater than a predetermined maximum number of updates N (for example, N = 20) is determined (s31).

When the parameter estimation unit 12 satisfies any one of (Condition 1) and (Condition 2), the parameter estimation unit 12 obtains the latest posterior probability m _{f, t, k} and arrival time difference δ acquired at that time. _{f and k} are output to the separated signal generator 13.

When neither of (Condition 1) and (Condition 2) is satisfied, the parameter estimation unit 12 repeats the E step and the M step (s31, s21). Note that the parameter θ and the value Q (θ | θ ⁿ ) of the Q function are stored in a storage unit (not shown), and are used in the next iteration.

<Separated signal generator 13>
The separated signal generation unit 13 receives the posterior probability m _{f, t, k} , the arrival time difference δ _{f, k} and the observation signal x _{f, t} and inputs the separated signal y _{f, t, k according} to the following equation (22). Is generated (s3) and output to the time domain conversion unit 14.

The sound source spectrum s _{f, t, k} is an estimation of the spectrum of the original signal emitted by the sound source. However, when this sound source spectrum is simply converted into a signal in the time domain, the source sound emitted by another sound source is obtained. A signal may remain. Only the original signal emitted from the sound source k can be extracted and separated by the above equation (22).

<Time domain conversion unit 14>
The time domain transform unit 14 receives the separated signals y _{f, t, and k} as input, and uses a time domain transform method (for example, short-time Fourier inverse transform) corresponding to the frequency domain transform method performed in the frequency domain transform unit 11 to separate the separated signals. y _{f, t, k} is converted into a separation signal y _k (t) in the time domain (s4) and output as an output value of the sound source separation device 1.

<Points of this embodiment>
Hereinafter, the points of the present embodiment will be described, and the derivation methods of equations (13), (17), and (18) will be described.

In this embodiment,
(Property 1) The arrival time difference δ _{f, k} is a value that affects the phase difference between the R channel and the L channel. (Property 2) Uses two properties that the phase difference is an amount that takes a periodic value. Then, the arrival time difference δ _{f, k} is estimated. Here, (property 1) is obvious when the case where the noise n _{f, t} is sufficiently small in the expression (2) is considered. (Property 2) is not only a value in the range of −π ≦ Θ <π, but also a periodic indefiniteness of Θ ± 2πM (M is an arbitrary integer). It means that it has the property of taking the value to be.

  (Property 2) will be further described below. Formula (6)

The vectors in the parentheses of exp and the matrices x, b, and V on the right side of

(C is a constant independent of δ _{f, k} ). Expression (32) is a Von Mises distribution that is a distribution for a variable that takes a periodic value such as a phase or angle distribution.

(See Reference 1).
[Reference Document 1] C.I. M.M. Bishop, Motoda et al., “Pattern Recognition and Machine Learning (Part 1)”, Springer Japan, 2006.

Here, −π <x ≦ π, μ is the distribution average (−π <μ ≦ π), κ> 0 is the distribution concentration parameter (corresponding to (1 / dispersion) in the normal distribution), I ₀ (x ) Is a zeroth-order first-type Bessel function.

That is, the variable x in the equation (33) corresponds to ψ _ξk −ψ _sk in the equation (32), the average μ in the equation (33) corresponds to 2πfδ _{f, k} in the equation (32), and the equation (33) The degree of concentration κ corresponds to (| ξ _{f, t, k} || s _{f, t, k} |) / (σ _f ² (1−φ _f ² )) in Expression (32).

To make the description more intuitive, consider the case where the noise parameter is φ _f = 0 (this is not the diffusive noise shown in Equation (4), but the Gaussian noise with variance σ _f ² is observed x _L, means that the ride each to _{x R).} At this time, since ξ _{f, t, k} = x _{f, t, R according} to the equation (14), in the equation (32), ψ _ξk is ψ _{xf, t, R} (however, the subscripts xf, t, R Represents x _{f, t, R} and ψ _{xf, t, R} represents the phase of the observation signal x _{f, t, R} of the microphone R). Also, according to equations (3) and (5) (where n _{f, t} = 0 in equation (5)), s _{f, t, k} = x _{f, t, L} , so ψ _sk in equation (32) is [psi _{xf, t, L} (where subscripts xf, t, L represents a _{x f, t, L, ψ} xf, t, L represents an observed signal _{x f, t, L} of the phase of the microphone L) It becomes. Further, as described above, ξ _{f, t, k} = x _{f, t, R} , and from equations (3) and (5), x _{f, t, R} = e ^{j2πfδk, f} · s _{f, t, k} (however, , E subscripts δk, f represent δ _{k, f} ), so in equation (32), | ξ _{f, t, k} | = | e ^{j2πfδk, f} · s _{f, t, k} | = | s _{f, t, k} | Therefore, equation (32) becomes

It becomes. Combine this with the Von Mises distribution,
(1) The distribution of variables taking periodic values of phase differences ψ _{xf, t, R} −ψ _{xf, t, L} between the R channel and the L channel takes an average μ = 2πfδ _{f, k} .
(2) The degree of concentration κ of the Von Mises distribution corresponds to the SN ratio | s _{f, k, t} | ² / σ _f ² . That is, when the S / N ratio is low (the condition is bad), the degree of concentration of the R channel and L channel phase difference values decreases (= dispersion increases). This corresponds to a phenomenon in which the measured value of the phase difference varies under a low SN ratio.
Two points can be said.

In this embodiment, (property 1) and (property 2) are used, and as an implementation procedure, an amount related to the phase difference between the R channel and the L channel is a distribution for a variable that takes a periodic value. Focusing on the fact that it can be expressed by the Mises distribution, the signal arrival time difference parameter δ _{f, k} is estimated by estimating the average value μ = 2πfδ _{f, k} of the distribution by maximum likelihood estimation with respect to the parameter of the Von Mises distribution.

Substituting p (x _{f, t} | k, θ) of equation (31) into equation (10) of the Q function,

(13) is obtained by solving.
Further, the expressions (17) and (18) of the function F in the time correction unit 12212 are the second-order derivatives of the Q function

It is. In the time correction unit 12212, whether the solution δ ′ _{f, k} obtained by the equation (13) gives the maximum value or the minimum value of the Q function is determined by whether the value of F (2πfδ ′ _{f, k} ) is positive or negative. When δ ′ _{f, k} gives a minimum value, the above correction of + π or −π is performed.

<Effect>
In the conventional method, since an analytical update formula is not given in the time difference estimation unit, a full search operation requiring a large calculation cost is required. Therefore, it is a problem to reduce the calculation cost of the time difference estimation unit and to provide a high-speed sound source separation unit. In the present embodiment, focusing on the fact that the arrival time difference δ _{f, k} is a value that affects the phase difference between the microphone R and the microphone L and that the phase difference has a periodic value, Estimate δ _{f, k} . This eliminates the need for a full search operation that has been required in the past, and thus provides a high-speed sound source separation unit.

<Simulation results>
An experiment was conducted to show the effect of the invention. In the room shown in FIG. 6, the number of sound sources was two (70 degrees and 150 degrees) or three (30, 70, 150 degrees). As the sound source, an English speaker voice was used, and when the number of sound sources was 2 and 3, the experiment was performed with 10 different sound source combinations.

As the noise, Gaussian noise having an average of 0 and a covariance matrix σ _f ² V _f (V _f is given by Equation (4)) was superimposed at an SN ratio of about 25 dB. The reverberation time of the room was 130 ms, the sampling frequency was 8 kHz, and the short Fourier transform window length and shift length were 64 ms and 16 ms, respectively.

In the conventional method, when the arrival time difference δ _k is estimated, the sound source position Θ (see FIG. 6) is changed in increments of 1 degree from 0 degrees to 180 degrees, and a total search is performed for the corresponding 181 types of arrival time differences δ _k. Was done.

  Figure 7 shows the calculation time in SIR (signal-to-noise ratio, noise includes other speaker's voice), SDR (signal-to-distortion ratio), and personal computer (IntelXeon (registered trademark) X5650 2.67 GHz (6 Core) x 2 CPU). Show. FIG. 7A shows the case where there are two sound sources (70 degrees and 150 degrees), and FIG. 7B shows the case where there are three sound sources (30 degrees, 70 degrees and 150 degrees). A larger value of SIR and SDR indicates better performance. The numbers are average values of 10 different sound source combinations. As shown in FIGS. 7A and 7B, it can be seen that the present embodiment can achieve substantially the same separation performance as the conventional method with a calculation time of about 1/10.

<Other variations>
As described above, the point of this embodiment is a method for estimating the arrival time difference. Therefore, other processes and parameter estimation methods are not limited to the above-described embodiment, and other conventional techniques may be used.

  In this embodiment, data is directly transferred between the respective units, but each data may be read and written via a storage unit (not shown).

In this embodiment, the noise power estimation unit 1223 obtains the sound source spectrum s _{f, t, k} from the sound source spectrum estimation unit 1222, but uses the arrival time difference δ _{f, k} and the observation signal x _{f, t.} The noise power estimator 1223 may be configured to calculate based on the equation (19).

  In this embodiment, the case of diffuse noise is assumed, but noise having other characteristics may be used. In that case, the sinc (2πfD / c) in the equations (4) and (15) may be appropriately changed according to the noise characteristics.

The separated signal generation unit 13 receives the posterior probabilities m _{f, t, k} and the sound source spectra s _{f, t, k} as inputs, and instead of the equation (22), the separated signals y _{f, t, k are expressed} by the following equations. May be generated.

In this case, when the amount of change in the value of the Q function (| Q (θ | θ ⁿ⁻¹ ) −Q (θ | θ ⁿ ) |) is smaller than the predetermined convergence determination threshold Δ, Alternatively, the latest signal posterior probabilities m _{f, t, k} and the sound source spectrum s _{f, t, k} acquired when the number of updates n is equal to or greater than the predetermined maximum number of updates N (s31) are separated signal generation unit 13 Output to. With such a configuration, the separated signals y _{f, t, and k} can be generated in the same manner as Expression (22) (see Expression (19)).

  The calculation order of M steps is not limited to the calculation order of this embodiment. For example, whichever of the calculation order of the time difference estimation unit 1221 and the sound source spectrum estimation unit 1222 may be performed first. FIG. 8 is a functional block diagram when the arrival time difference estimation processing (s25, s26) of the time difference estimation unit 1221 is performed after the sound source spectrum estimation processing (s27) of the sound source spectrum estimation unit 1222 is performed. Hereinafter, a description will be given with reference to FIG.

The sound source spectrum estimation unit 1222 of the M step calculation unit 122 is the parameter θ ′ obtained in the previous iteration (more specifically, the arrival time difference δ _{f, k} of θ ′. If the parameter θ ′ obtained by repeating the above is not present, that is, in the first sound source spectrum calculation, the above-described initialized parameter) and the observation signal x _{f, t} are input, and these values are used. The sound source spectrum s _{f, t, k} is estimated by the equation (19) and output to the noise power estimation unit 1223, the Q function calculation unit 1212, the time difference estimation unit 1221, and the time correction unit 12212.

The arc tangent calculation unit 12211 of the time difference estimation unit 1221 of the M step calculation unit 122 receives the observation signal x _{f, t} , the posterior probability m _{f, t, k,} and the sound source spectrum s _{f, t, k,} and these Is used to estimate the arrival time difference δ _{f, k} by the equation (13) and output it to the time correction unit 12212.

The time correction unit 12212 of the time difference estimation unit 1221 of the M step calculation unit 122 repeats the observation signal xf _{, t} , the posterior probability _{mf, t, k} , the sound source spectrum sf _{, t, k,} and the previous iteration. Is the noise power spectrum σ _f ² of θ ′. More specifically, when the parameter θ ′ obtained in the previous iteration does not exist, that is, In the first time correction, the above-described initialized parameter) is used as an input, and by using these values, the uncertainty of ± π included in the arrival time difference δ _{f, k} estimated by Expression (13) is included. Correction is made based on equation (16). Further, the time correction unit 12212, the correction value 2Paiefuderuta _f, the _k divided by 2 [pi] f, calculated arrival time difference [delta] _{f, k,} and outputs to the noise power estimation section 1223 and the Q function calculating unit 1212.

By setting it as the above-mentioned structure, the effect similar to 1st embodiment can be acquired. Similarly, after performing the noise power estimation process (s28) of the noise power estimation unit 1223 based on the arrival time difference δ _{f, k} in the parameter θ ′ obtained in the previous iteration, the time difference estimation unit The arrival time difference estimation process 1221 (s25, s26) may be performed.

  The present invention is not limited to the above-described embodiments and modifications. 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. In addition, it can change suitably in the range which does not deviate from the meaning of this invention.

<Program and recording medium>
The sound source separation device described above can also be functioned by a computer. In this case, each process of a program for causing a computer to function as a target device (a device having the functional configuration shown in the drawings in various embodiments) or a processing procedure (shown in each embodiment) is processed by the computer. A program to be executed by the computer may be downloaded from a recording medium such as a CD-ROM, a magnetic disk, or a semiconductor storage device or via a communication line into the computer, and the program may be executed.

Claims (3)

In a situation where a plurality of original signals are mixed with noise and observed with two microphones, a sound source separation device that separates and extracts each original signal from the observed signal,
The sound source index of the original signal is k, and the original signal to the two microphones from the observed signal x _{f, t} = [x _{f, t, L} , x _{f, t, R} ] ^T in the frequency domain. Parameter estimation means for estimating the arrival time difference δ _{f, k} , noise power spectrum σ _f ² , original signal spectrum s _{f, t, k} and sound source existence probability p (k | θ);
A separated signal y _{f, t, k} is generated from the observed signal x _{f, t} and the estimated parameter θ = {δ _{f, k} , σ _f ² , s _{f, t, k} , p (k | θ)}. Separating signal generating means for
The parameter estimation means includes
The posterior probability that the observed signal x _{f, t} indicates the expected value attributed to the sound source k is m _{f, t, k} , the distance between the two microphones is D, the speed of the original signal is c, φ _f = sinc (2πfD / c) and ξ _{f, t, k} = [x _{f, t, R} −φ _f (x _{t, t, L} −s _{f, t, k} )], and the spectrum s _{f, t,} The phases of _k and ξ _{f, t, k} are ψ _sk and ψ _ξk , respectively, and the arrival time difference δ _{f, k} is

Arc tangent calculation means to estimate as
Correction means for correcting the indefiniteness of ± π included in the estimated arrival time difference δ _{f, k} .
Sound source separation device. In a situation where a plurality of original signals are mixed with noise and observed by two microphones, a sound source separation method for separating and extracting each original signal from the observed signal,
The sound source index of the original signal is k, and the original signal to the two microphones from the observed signal x _{f, t} = [x _{f, t, L} , x _{f, t, R} ] ^T in the frequency domain. Parameter estimation step for estimating the arrival time difference δ _{f, k} , noise power spectrum σ _f ² , original signal spectrum s _{f, t, k,} and sound source existence probability p (k | θ);
A separated signal y _{f, t, k} is generated from the observed signal x _{f, t} and the estimated parameter θ = {δ _{f, k} , σ _f ² , s _{f, t, k} , p (k | θ)}. And a separated signal generating step
The parameter estimation step includes:
The posterior probability that the observed signal x _{f, t} indicates the expected value attributed to the sound source k is m _{f, t, k} , the distance between the two microphones is D, the speed of the original signal is c, φ _f = sinc (2πfD / c) and ξ _{f, t, k} = [x _{f, t, R} −φ _f (x _{t, t, L} −s _{f, t, k} )], and the spectrum s _{f, t,} The phases of _k and ξ _{f, t, k} are ψ _sk and ψ _ξk , respectively, and the arrival time difference δ _{f, k} is

Arc tangent calculation step to estimate as
A correction step of correcting the indefiniteness of ± π included in the estimated arrival time difference δ _{f, k} .
Sound source separation method.   A program for causing a computer to function as the sound source separation device according to claim 1.

Images