JP2019193073A - Sound source separation device, method thereof, and program - Google Patents

Abstract

To provide a sound source separation device with higher separation accuracy than before.SOLUTION: The sound source separation device includes: a diffusive noise removal part configured to remove an estimated diffuse noise signal from observed signal and to obtain the removed signal; a filter design part configured to obtain a filter by combining the probability distribution of modeled removed signal and the probability distribution of modeled transfer function; and a sound source separation part configured to separate noise component estimate value including at least first acoustic signals and coherent noise acoustic signals from observed signal by the filter.SELECTED DRAWING: Figure 4

Description

  The present invention provides an observation signal obtained when a known acoustic signal is applied to a microphone in a noisy environment (for example, when a known acoustic signal is reproduced and a reproduced sound is recorded by the microphone), and a known acoustic signal. The present invention relates to a sound source separation device that separates a speech component and a noise component contained in an observation signal from a signal.

  When evaluating the speech recognition performance of a microphone, there is one that estimates the SN ratio from the observation signal recorded by the microphone and compares the SN ratio estimated value with the speech recognition rate. For example, it is possible to compare the speech recognition rates of the speech recognition apparatus with respect to each SN ratio estimation value by performing speech recognition with one speech recognition apparatus for two or more observation signals having different SN ratio estimation values. it can.

  By using such a method, it is possible to estimate whether or not the observation signal is desired to be recognized by humans (for example, if an observation signal with a high SN ratio is desired to be recognized easily) It is possible to estimate the recognition performance close to the user experience value. In other words, a high SN ratio (easy to hear with less noise component compared to the speech component) results in higher speech recognition recognition accuracy, and a lower SNR (more noise component than the speech component makes it difficult to hear) speech recognition. The recognition performance can be evaluated in consideration of the fact that the recognition accuracy of the image becomes low.

Data for performance evaluation as described above, generally prepared speech signal database, not previously shown to reproduce the target sound s _t from the speaker 71 as shown in FIG. 1, the interference noise n _t from the speaker 72 The SN ratio is estimated by the SN ratio estimator 74 using the observation signal x _t recorded by the microphone 73. Note that the observed signal x _t also include diffuse noise d _t. t is an index indicating time.

Conventional SN ratio estimation techniques, target sound as in FIG. 2 (a original sound signal, a source referred to as a source or source signal) to the speech period information obtained from s _t send speech segment (in Figure 2 The speech component is obtained from T _{s0 to} T _s1 ) and the noise component is obtained from the non-speech interval (T _{n0 to} T _{n1 in} FIG. 2) (see Non-Patent Document 1).

However, if non-stationary noise is present, a difference occurs between the estimated value of the SN ratio and the user experience value (actual feeling value). For example, FIG. 2A is set to a state in which non-stationary noise does not exist, and in FIG. 2B, non-stationary noise exists in a section including a non-speech section (T _{n0 to} T _{n1 in} FIG. 2), and the SN ratio is based on the user experience value. In FIG. 2C, non-stationary noise exists in the section including the speech section (T _{s0 to} T _{s1 in} FIG. 2), and the SN ratio is estimated to be higher than the user experience value.

Therefore, as shown in FIG. 3, it separates the speech and noise components of the observed signal x _t in the sound source separation unit 84, to propose a method for estimating the SN ratio from the signal separated in SN ratio estimation unit 85. Here, the target signal s _{ω, τ} ∈ C (C is the entire set of complex numbers) and the coherent noise n _{ω, τ} ∈ C and the diffusive noise d _{ω, τ} ∈ C are superimposed as follows: _{From ω, τ} ∈ C, the target sound-derived component (speech component) a _ω s _{ω, τ} included in the observation signal x _{ω, τ} and the noise-derived component (noise component) n _{ω, τ} + d _{ω, τ} Dealing with the problem of estimating.
x _{ω, τ} = a _ω s _{ω, τ} + n _{ω, τ} + d _{ω, τ} (1)
Here, x _{ω, τ} , s _{ω, τ} , n _{ω, τ} , d _{ω, τ} are converted from time domain signals x _t , _st , n _t , _dt to frequency domain signals, respectively. , Ω∈ {1,…, Ω} and τ∈ {1,…, Τ} are frequency and (frame) time indices, and a _ω is the transfer characteristic from the target sound position (target sound generation position) to the observation position. (Also called transfer function). Hereinafter, for the sake of simplicity, the absolute value of the complex number is expressed in uppercase letters corresponding to each lowercase letter such as | x _{ω, τ} | = X _{ω, τ} . Unless otherwise noted, lowercase variables are complex and uppercase variables are real. As the diffuse noise, stationary noise such as background noise including air-conditioning sound is assumed. As the coherent noise, non-stationary noise such as an utterance of a person who is not originally recorded, a TV voice, or a sudden sound is assumed.

Observed signal x _omega, voice component from _τ a _ω s _{ω, τ} and a noise component n _{ω, τ} + d _ω, a typical method of estimating the _tau, there is a non-linear filtering. In this method, the nonlinear filter is designed with the following equation:

Each signal (component) is estimated as follows.
^ a _ω ^ s _{ω, τ} = G _{ω, τ} x _{ω, τ} (3)
^ n _{ω, τ} + ^ d _{ω, τ} = (1-G _{ω, τ} ) x _{ω, τ} (4)
By estimating each signal (component) in this way, it is possible to estimate sSNR (segmental-SNR), which is the SNR of each time frame as defined by Equation (5), for example.

In Equation (2), in order to estimate the nonlinear filter G _{ω, τ} , it is necessary to estimate the transfer characteristic A _ω , target sound S _{ω, τ} , coherent noise N _{ω, τ} , and diffusive noise D _{ω, τ} There is. In this problem setting, since the target sound S _{ω, τ} is assumed to be known, transfer characteristics A _ω , coherent noise N _{ω, τ} , and diffusive noise D _{ω, τ} are obtained from the observed signal X _{ω, τ.} By estimating, it is possible to estimate the nonlinear filters _{Gω, τ} and SNR.

Many of the conventional methods in the above sound source separation problem assume instantaneous mixing of sound sources in the amplitude region and multiplicative properties in the amplitude region of the transfer characteristics. Assuming that the above assumption holds, the observation signal X _{ω, τ} can be described as follows.
X _{ω, τ} = A _ω S _{ω, τ} + N _{ω, τ} + D _{ω, τ} (6)
There are various methods for estimating each component under this model. Dispersive noise D _omega, typical in a manner of estimating the _tau is to diffuse noise D _{omega, tau} is assumed to be stationary noise, the expected value of the observation signal X _{omega, tau.}

However, only this method can estimate only the diffusive noise _{Dω, τ} among the noise components, and cannot estimate the coherent noise _{Nω, τ} . There is a semi-supervised non-negative matrix factorization (NMF) as a method for estimating the coherent noise N _{ω, τ} . Semi-supervised NMF puts the following models for observed signals X _{ω, τ} .

Where W ^S _{ω, r} and W ^N _{ω, k} are the basis of the amplitude spectrum of the target sound and coherent noise, and H ^S _{r, τ} and H ^N _{k, τ} are the amplitudes of the target sound and coherent noise, respectively. It is the intensity (activation) corresponding to each base of the spectrum, and R and K are the respective base numbers. In this problem setting, since the target sound S _{ω, τ} is known, the base W ^S _{ω, r} and the intensity H ^S _{r, τ} are changed to the target sound S _{ω, τ} and

Learn to minimize the objective function such as generalized KL information amount between, and then minimize the objective function such as generalized KL information amount between observation signal X _{ω, τ} and Equation (7) Thus, the base W ^N _{ω, k} and the intensity H ^N _{k, τ} are learned (see Non-Patent Document 2).

"G.160: Revised Appendix II-Objective measures for the characterization of the basic functioning of noise reduction algorithms", International Telecommunication Union D. Kitamura, N. Ono, H. Saruwatari, Y. Takahashi, and K. Kondo, "DISCRIMINATIVE AND RECONSTRUCTIVE BASIS TRAINING FOR AUDIO SOURCE SEPARATION WITH SEMI-SUPERVISED NONNEGATIVE MATRIX FACTORIZATION", in Proc., IWAENC 2016.

However, since the formula (7) and the transfer characteristic A _omega diffuse noise D _omega, do not consider the _tau, observation signals X _omega, components derived from the target sound from _τ a _ω s _ω, components derived from _tau and noise The separation accuracy of n _{ω, τ} + d _{ω, τ} is low, and it is difficult to estimate the SNR precisely only by applying this.

  An object of the present invention is to provide a sound source separation technology apparatus with higher separation accuracy than the conventional one.

  In order to solve the above problems, according to one aspect of the present invention, a sound source separation device acquires a desired acoustic signal from an observation signal obtained by recording a predetermined acoustic signal emitted from a speaker with a microphone. The observation signal includes a first acoustic signal based on a predetermined acoustic signal and a transfer function that represents a spatial characteristic between the speaker and the microphone, an interference noise acoustic signal that is coherent noise, and diffusive noise. The sound source separation device removes the estimated value of the diffusive noise acoustic signal from the observation signal and obtains the removed signal, and the removed noise signal. A filter design unit that obtains a filter by combining a probability distribution modeled with a probability distribution modeled with a transfer function, and a noise component including at least a first acoustic signal and a coherent noise acoustic signal from an observation signal by the filter A sound source separation unit that separates the estimated value of.

  According to the present invention, there is an effect that the separation accuracy is higher than the conventional one. Further, the use of each separated component has the effect that the SN ratio estimation accuracy is higher than in the prior art.

The figure for demonstrating the prior art which estimates SN ratio. 2A is a diagram showing a state where non-stationary noise does not exist, FIG. 2B is a diagram showing a state where non-stationary noise exists in a section including a non-speech section, and FIG. 2C is a non-stationary noise in a section including a utterance section. The figure which shows the state which exists. The figure for demonstrating the prior art which estimates SN ratio. The functional block diagram of the SN ratio estimation apparatus which concerns on 1st embodiment. The figure which shows the example of the processing flow of the SN ratio estimation apparatus which concerns on 1st embodiment.

  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. In the following explanation, the symbol “^” etc. used in the text should be described immediately above the character immediately after it, but it is described immediately before the character due to restrictions on the text notation. In addition, the symbol “_” or the like used in the text should be described immediately below the character immediately after it, but it is described immediately before the character due to restrictions on the text notation. In the formula, these symbols are written in their original positions. Further, the processing performed for each element of the vector or matrix is applied to all elements of the vector or matrix unless otherwise specified.

<Points of first embodiment>
In this embodiment, by extending the semi-supervised NMF, the transfer characteristics A _ω , coherent noise N _{ω, τ} , and diffusive noise D _{ω, τ} are estimated from the observed signals X _{ω, τ} , and the SN ratio is calculated. Provide an estimation method. The point of this embodiment is
・ Semi-supervised NMF is preliminarily estimated for diffusive noise to be applied to the observation model in the real environment such as Equation (1) and Equation (6), and is removed from the observation signal.
-Optimum based on posterior probability maximization (MAP: maximum-a-posteriori) estimation by incorporating a term related to the transfer characteristic A _ω into a semi-supervised NMF (see Non-Patent Document 2) that is probability-modeled based on the signal after removal It is to provide an algorithm that performs the conversion. With such a configuration, the target sound-derived component and the noise-derived component can be separated from the observation signal with high accuracy in the actual environment, and the SN ratio can be estimated.

  First, how the observation signal is modeled will be described.

<Modeling of observation signal>
In order to model the observation signal in accordance with Equation (6), first, the observation signal _{Xω, τ} is approximated as follows.

It is assumed that diffuse noise D _omega, extending the existing technology related to the estimation of _tau, diffuse noise D _{omega, tau} is a stationary noise in during a certain time frame. Further, assuming that the target sound S _{ω, τ} and the coherent noise N _{ω, τ} are sparse signals in time, the diffusive noise D _{ω, τ} is estimated as follows.
^ D _{ω, τ} ← Υ ・ min [X _{ω, τ-F_wd} , X _{ω, τ-F_wd + 1} ,…, X _{ω, τ + B_wd} ] (8)
Here, F_wd and B_wd are parameters that define the number of time frames in which D _{ω and τ} are stationary, and can be obtained by tuning. For example, each may be set to about 20. Moreover, Υ is a predetermined value. Then, the observation signal (hereinafter, also referred to as “removed signal”) Y _{ω, τ} from which the diffusive noise D _{ω, τ} has been removed can be described as follows.

Here, the base W ^S _{ω, r} and the intensity H ^S _{r, τ} of the amplitude spectrum of the target sound can be estimated by using a conventional semi-supervised NMF framework (see Non-Patent Document 2). Hereinafter, a method for estimating the base W ^N _{ω, k} , the intensity H ^N _{k, τ} and the transfer characteristic A _ω of the amplitude spectrum of the coherent noise from the removed signal Y _{ω, τ} will be described. The estimated values of W ^S _{ω, r} , H ^S _{r, τ} , W ^N _{ω, k} , H ^N _{k, τ} , A _ω are ^ W ^S _{ω, r} , ^ H ^S _{r, τ} , ^ W ^N _{It is written as ω, k} , ^ H ^N _{k, τ} , ^ A _ω .

The transfer characteristic A _ω is originally a physical parameter, and estimation accuracy can be improved by incorporating acoustic prior knowledge such as the shape of the room and the observation environment. In order to realize this, in this embodiment, each parameter is estimated by MAP estimation. Specifically, the likelihood function p (_A, _N | _S, _Y) for the removed signal Y _{ω, τ} and the prior distribution p (_A | _α) for the transfer characteristic A _ω are designed, and the following equation (11 ) Are estimated so as to maximize the joint probability L of).
L = p (_A, _N | _S, _Y) p (_A | _α) (11)
_A: = [^ A _ω ] ∈R ^Ω
_N: = [^ N _{ω, τ} ] ∈R ^{Ω × Τ}
_S: = [S _{ω, τ} ] ∈R ^{Ω × Τ}
_Y: = [Y _{ω, τ} ] ∈R ^{Ω × Τ}
_α: = [α _ω ] ∈R ^Ω
_α is a set of parameters used to model the prior distribution for the transfer characteristic ^ A _ω . Here, a Poisson distribution, which is a probability distribution obtained by probabilistic interpretation of the generalized KL information amount, is applied to the likelihood function. As for the transfer characteristic A _ω , since the transfer characteristic A _ω is a non-negative variable, the Poisson distribution is applied. Each distribution can then be described as follows:

Here, since each distribution is an exponential family, it is more efficient in numerical calculation to maximize the simultaneous probability L by maximizing the logarithmic simultaneous distribution with logarithms on both sides. Here, when logarithm is taken for each distribution, it can be described as follows.

Therefore, the objective function to be maximized is

It becomes. Maximizing this objective function J (Θ) means maximizing the joint probability L.

<Derivation of update formula>
An algorithm for estimating the estimated value of the base ^ W ^N _{ω, k} , the estimated value of the intensity ^ H ^N _{k, τ} and the estimated value of the transfer characteristic ^ A _ω so as to maximize Equation (18) is described. Since it is difficult to directly maximize Equation (18), an update algorithm using the auxiliary function method will be described in this embodiment. For simplicity of the problem, R = K. From the log-sum inequality, λ _{r, ω, τ} ≧ 0 and

Then, the following inequality holds.

Then, the objective function J (Θ) can be suppressed from below by the following J ′ (Θ).

According to the auxiliary function method, first, J ′ (Θ) is maximized with respect to λ _{r, ω, τ} , and by repeating the process of maximizing each variable under the λ _{r, ω, τ} , the objective function J ( The parameter can be estimated to monotonically increase Θ). The update algorithm based on the auxiliary function method is as follows.

Note that when calculating using the matrix calculation library, Equations (22) and (23) may be changed to the following update rule as an approximation of the above algorithm.

T is a transpose, _E is a matrix of Ω × Τ and all elements are 1, and division of the matrix represents division for each element. Also _Z = [_ Z ^(S) , _W ^(N) ], _H = [(_ H ^(S) ) ^T , (_ H ^(N) ) ^T ] ^T , _Z ^(S) : = {^ A _ω ^ W _{ω , r} ^S } ∈R ^{Ω × R} , _W ^(N) : = {^ W _{ω, k} ^N } ∈R ^{Ω × K} , _H ^(S) : = {^ H _{r, τ} ^S } ∈R ^{R × Τ} , _H ^(N) : = {^ H _{k, τ} ^N } ∈R ^{K × Τ} .

Also in order not to update the _Z ^(S) and _H ^(S), replaced with a _H ^(S) and _Z ^(S) for each update to the pre-learned values.

<SNR ratio estimation apparatus according to the first embodiment>
FIG. 4 is a functional block diagram of the SN ratio estimation apparatus according to the first embodiment, and FIG. 5 shows an example of the processing flow.

  The SN ratio estimation apparatus 100 includes an initialization unit 102, a diffusive noise removal unit 103, a filter design unit 104, a sound source separation unit 105, and a signal-to-noise ratio estimation unit 106.

SN ratio estimation apparatus 100, the target sound s _omega obtained by converting the target sound s _t in the time domain to be reproduced by the loudspeaker 71 into a frequency domain _{signal, tau,} of the observed signal x _t the frequency domain of the time domain was recorded by the microphone 73 The observation signal x _{ω, τ} converted to a signal and various parameters are input. The various parameters here are, for example, Υ in formula (8), bases R and K (for example, R = K = can be set to about 10), initial value of transfer characteristic estimation value ^ A (for example, ^ A _ω = 1) etc. In this embodiment, the target sound s _{ω, τ} in the frequency domain and the observation signal x _{ω, τ} are described as input. However, the target sound s _{t in} the time domain and the observation signal x _t are input. It is good also as a structure to be. Where t is a time index. In this case, the signal-to-noise ratio estimation apparatus 100 performs processing for conversion to a frequency domain signal. For example, fast Fourier transform or the like may be used for frequency conversion, and the Fourier transform length may be 256 points and the number of shift points may be 128 points.

The SN ratio estimation apparatus 100 uses the target sound s _{ω, τ} and the observation signal x _{ω, τ} to separate the speech component and the noise component contained in the observation signal x _{ω, τ} to obtain the signal-to-noise ratio. ,Output.

  The SN ratio estimation device is, for example, a special configuration configured by reading a special program into a known or dedicated computer having a central processing unit (CPU), a main memory (RAM), and the like. Device. For example, the SN ratio estimation apparatus executes each process under the control of the central processing unit. Data input to the SN ratio estimation device and data obtained in each process are stored in, for example, a main storage device, and the data stored in the main storage device is read out to the central processing unit as necessary. Used for other processing. At least a part of each processing unit of the SN ratio estimation apparatus may be configured by hardware such as an integrated circuit. Each storage unit included in the SN ratio estimation device can be configured by, for example, a main storage device such as a RAM (Random Access Memory), or middleware such as a relational database or a key-value store. However, each storage unit is not necessarily provided in the SN ratio estimation device, and is configured by an auxiliary storage device configured by a semiconductor memory element such as a hard disk, an optical disk, or a flash memory, and the SN ratio is determined. It is good also as a structure provided in the exterior of an estimation apparatus.

  Hereinafter, each part will be described.

<Initialization unit 102>
The initialization unit 102 receives the target sound _{sω, τ} , the observation signal _{xω, τ,} and various parameters.

The initialization unit 102 estimates the diffusive noise D _{ω, τ according} to the equation (8) using the observation signal x _{ω, τ} and Υ, and outputs the estimated value ^ D _{ω, τ} .
^ D _{ω, τ} ← Υ ・ min [X _{ω, τ-F_wd} , X _{ω, τ-F_wd + 1} ,…, X _{ω, τ + B_wd} ] (8)
The initialization unit 102 uses, for example, the target sound s _{ω, τ} and the basis number R to estimate the basis value in an existing NMF framework (see Non-Patent Document 2) based on generalized KL information minimization. ^ W ^S _{ω, r} and intensity estimate ^ H ^S _{r, τ} are obtained and output. For example, since the target sound S _{ω, τ} is known, the base estimate ^ W ^S _{ω, r} and the intensity estimate ^ H ^S _{r, τ} are changed to the target sound S _{ω, τ}

Learning is performed so as to minimize an objective function such as a generalized KL information amount between (see Non-Patent Document 2). The base estimation value ^ W ^N _{ω, k} and the intensity estimation value ^ H ^N _{k, τ} are initialized with a non-negative random number or the like.

The initialization unit 102 performs, for example, the estimation value ^ D _{ω, τ} , the basis estimation value ^ W ^S _{ω, r} , the intensity estimation value ^ H ^S _{r, τ} , and the basis estimation value ^ W ^N by the above-described method. Initial values of _{ω, k} and estimated intensity ^ H ^N _{k, τ} are obtained (S102) and output. Note that the estimated transfer value ^ A _{ω, k} , the estimated base value ^ W ^N _{ω, k} , and the estimated strength value ^ H ^N _{k, τ} are values that are repeatedly updated in this embodiment, but the estimated value ^ D _The initial values of _{ω, τ} , the estimated base value ^ W ^S _{ω, r} and the estimated intensity value ^ H ^S _{r, τ} may be used as they are once set for one usage environment.

<Diffusion noise removing unit 103>
The diffusive noise removing unit 103 receives the observation signal x _{ω, τ} and the estimated value ^ D _{ω, τ of the} diffusive noise D _{ω, τ} as inputs, and from the observation signal x _{ω, τ} by the expression (9) The estimated value of D _{ω, τ} is removed, and the removed signal Y _{ω, τ} is obtained (S103) and output.

<Filter design unit 104>
The filter design unit 104 calculates a basis estimate ^ W ^S _{ω, r} , an intensity estimate ^ H ^S _{r, τ} , a basis estimate ^ W ^N _{ω, k} and an intensity estimate ^ H ^N _{k, τ} initial value, and removing spent signal Y _{omega, tau,} diffuse noise D _omega, estimate of _tau ^ D _{omega, tau,} observed signal x _{omega, tau,} base number K, and input various parameters including R. Filter design unit 104-removed signal Y _omega, give a probability distribution that models the _tau, and the probability distribution that models the transfer characteristics A _omega, the nonlinear filter G _omega by combining _{the tau} (S104), the output To do. For example, the joint probability L of Equation (11) that combines the likelihood function p (_A, _N | _S, _Y) for the removed signal Y _{ω, τ} and the prior distribution p (_A | _α) for the transfer characteristic A _ω Each parameter _A, _N, and _α is estimated so as to maximize.
L = p (_A, _N | _S, _Y) p (_A | _α) (11)
This process maximizes the following objective function J (Θ) with each parameter (base estimate ^ W ^N _{ω, k} , strength estimate ^ H ^N _{k, τ} , transfer characteristic estimate ^ A _This corresponds to the process of estimating _ω ).

For example, using Equations (21) to (24) or Equations (21), (25), (26), and (24), the estimated base value ^ W ^N _{ω, k} and the estimated intensity value ^ H ^N _{k, τ} , Updating the estimated value of transfer characteristic ^ A _ω (S104-1) means maximizing the joint probability L and estimating each parameter _A, _N, _α.

However, _Z = [_ Z ^(S) , _W ^(N) ], _H = [(_ H ^(S) ) ^T , (_ H ^(N) ) ^T ] ^T , _Z ^(S) : = {^ A _ω ^ W _{ω , r} ^S } ∈R ^{Ω × R} , _W ^(N) : = {^ W _{ω, k} ^N } ∈R ^{Ω × K} , _H ^(S) : = {^ H _{r, τ} ^S } ∈R ^{R × Τ} , _H ^(N) : = {^ H _{k, τ} ^N } ∈R ^{K × Τ} , and when updating by equations (21), (25), (26), (24), _Z ^(S) and In order not to update _H ^(S) , _Z ^(S) and _H ^(S) are replaced with pre-learned values for each update.

When the predetermined condition is satisfied (S104-2), the filter design unit 104 ends the update, and the estimated value of the base ^ W ^N _{ω, k} and the estimated value of the intensity ^ H ^N _{k, τ} Using the estimated value ^ A _ω of the characteristic, a nonlinear filter G _{ω, τ} represented by the following equation is obtained (S104-3) and output.

The filter design unit 104 repeats the update process S104-1 until a predetermined condition is satisfied. As the predetermined condition, (i) S104-1 is repeated a predetermined number of times (for example, 100 times), and (ii) the update amount is smaller than a predetermined value. In short, it is only necessary that the update amount of the base estimated value ^ W ^N _{ω, k} , the intensity estimated value ^ H ^N _{k, τ} , and the transfer characteristic estimated value ^ A _ω converge to a desired level.

<Sound source separation unit 105>
The sound source separation unit 105 receives the observation signal x _{ω, τ} and the filter G _{ω, τ} as input, and uses the filter G _{ω, τ} to at least estimate the speech component ^ a _ω ^ s _{ω, τ} from the observation signal x _{ω, τ . τ} and the estimated noise component including coherent noise n _{ω, τ} are separated. For example, the estimated value ^ a _ω ^ s _{ω, τ} of the speech component and the estimated value ^ n _{ω, τ} + ^ d _{ω, τ of the} noise component are separated by the following equation (S105) and output.
^ a _ω ^ s _{ω, τ} = G _{ω, τ} x _{ω, τ} (3)
^ n _{ω, τ} + ^ d _{ω, τ} = (1-G _{ω, τ} ) x _{ω, τ} (4)

<Signal to Noise Ratio Estimator 106>
The signal-to-noise ratio estimation unit 106 receives a speech component estimate value ^ a _ω ^ s _{ω, τ} and a noise component estimate value ^ n _{ω, τ} + ^ d _{ω, τ} as inputs, and obtains a signal-to-noise ratio ( S106) and output. For example, sSNR is obtained by the following equation.

<Effect>
With such a configuration, the speech component and noise component can be separated from the observation signal recorded by the microphone in the noisy environment, so the SN ratio in the utterance interval can be highly accurate even in the presence of non-stationary noise. Can be estimated. By using the obtained SN ratio estimated value, application to the following applications becomes possible.
・ Comparison of noise suppression performance between microphones: For example, by comparing the noise suppression performance of microphones by obtaining SNR estimates from observation signals recorded with two or more microphones with a noise cancellation function. it can.
・ Comparison of speech recognition performance between speech recognition systems connected to microphones: For example, an SN ratio estimate is obtained from observation signals recorded with microphones in a noisy environment, and speech recognition processing is performed with two or more speech recognition systems. And the speech recognition performance with respect to the SN ratio estimated value for each speech recognition system can be compared from the SN ratio estimated value and the speech recognition result.
・ Comparison of microphone observation signal and user sensation recognition rate: For example, the SN ratio estimated value is obtained from the observation signal recorded with the microphone in the noisy environment, and the user's sensation recognition rate for the observation signal is obtained. The ratio estimated value and the user's bodily sensation recognition rate can be compared.
Comparison of microphone observation signal and speech recognition engine recognition performance: For example, each speech signal recognition engine performs speech recognition on two or more observation signals having different SN ratio estimates, thereby estimating each signal-to-noise ratio. The speech recognition performance of the speech recognition engine against the value can be compared.

<Modification>
In this embodiment, the signal-to-noise ratio is used as the output of the apparatus, but the estimated value ^ a _ω ^ s _{ω, τ} of the speech component that is the output value of the sound source separation unit 105 and the estimated noise component ^ n _{ω, τ} + ^ d _{ω, τ} may be output from the apparatus, and the signal-to-noise ratio estimation unit 106 may not be provided. In this case, it is called a sound source separation device. It can be said that the SN ratio estimation apparatus includes a sound source separation apparatus.

In the present embodiment, the sound source separation unit 105, the filter G _omega, the observed signal x _{omega, tau} by _tau, an estimate of at least speech component ^ a _ω ^ s _ω, the estimate of _tau and a noise component ^ n _{omega, tau} + ^ d _{ω, τ} is separated, but it is not always necessary to separate the diffusive noise d _{ω, τ} from the observed signal when estimating the signal-to-noise ratio _. Only _τ may be separated. In this case, the filter may be designed without taking diffuse noise into consideration.

<Other variations>
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>
In addition, various processing functions in each device described in the above embodiments and modifications may be realized by a computer. In that case, the processing contents of the functions that each device should have are described by a program. Then, by executing this program on a computer, various processing functions in each of the above devices are 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, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may 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 storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its storage unit. When executing the process, this computer reads the program stored in its own storage unit and executes the process according to the read program. As another embodiment of this program, a computer may read a program directly from a portable recording medium and execute processing according to the program. Further, each time a program is transferred from the server computer to the computer, processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program includes information provided for processing by the electronic computer and equivalent to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In addition, although each device is configured by executing a predetermined program on a computer, at least a part of these processing contents may be realized by hardware.

Claims (8)

A sound source separation device for obtaining a desired acoustic signal from an observation signal obtained by recording a predetermined acoustic signal emitted from a speaker with a microphone,
The observation signal includes a first acoustic signal based on the predetermined acoustic signal and a transfer function that is a function expressing a spatial characteristic between the speaker and the microphone, and a coherent noise acoustic signal that is coherent noise. A diffusive noise acoustic signal that is diffusive noise, and
Removing an estimate of the diffusive noise acoustic signal from the observed signal and obtaining a removed signal;
A filter design unit that obtains a filter by combining a probability distribution modeling the removed signal and a probability distribution modeling the transfer function;
A sound source separation unit that separates at least the first acoustic signal and an estimated value of a noise component including the coherent noise acoustic signal from the observation signal by the filter;
Sound source separation device. The sound source separation device according to claim 1,
ω = {1,2, ..., Ω} and τ = {1,2, ..., Τ} are frequency and time indices, respectively, and the estimated transfer function is ^ A _ω. Is assumed to be ^ N _{ω, τ} , the predetermined acoustic signal is S _{ω, τ} , the removed signal is Y _{ω, τ} , _A: = [^ A _ω ] ∈R ^Ω , _N: = [ ^ N _{ω, τ} ] ∈R ^{Ω × Τ} , _S: = [S _{ω, τ} ] ∈R ^{Ω × Τ} , _Y: = [Y _{ω, τ} ] ∈R ^{Ω × Τ} , _α: = [α _ω ] ∈ R ^Ω , the probability distribution modeled on the removed signal is a likelihood function p (_A_N | _S, _Y) related to the removed signal, and the probability distribution modeled on the transfer function is a prior distribution related to the transfer function p (_A | _α), and the filter design unit has the joint probability
L = p (_A, _N | _S, _Y) p (_A | _α)
Parameter is estimated to maximize the filter, and the filter is obtained from the estimated parameter.
Sound source separation device. The sound source separation device according to claim 2,
The observed signal is X _{ω, τ} , the estimated values of the amplitude spectra of the predetermined acoustic signal and the coherent noise acoustic signal are ^ W _{ω, r} ^S and ^ W _{ω, k} ^N _, respectively. Assume that the estimated values of the intensity corresponding to the basis of the amplitude spectrum of the acoustic signal and the coherent noise acoustic signal are ^ H _{r, τ} ^S and ^ H _{k, τ} ^N , respectively, and the predetermined acoustic signal and the coherent noise acoustic signal Let R and K be the basis numbers of the amplitude spectrum of
The filter design unit includes:
By
Or
T is transpose, _E is a matrix with Ω x Τ and all elements are 1. Matrix division is element-by-element division, _Z = [_ Z ^(S) , _W ^(N) ], _H = [(_ H ^{( S)} ) ^T , (_ H ^(N) ) ^T ] ^T , _Z ^(S) : = {^ A _ω ^ W _{ω, r} ^S } ∈R ^{Ω × R} , _W ^(N) : = {^ W _{ω, k} ^N } ∈R ^{Ω × K} , _H ^(S) : = {^ H _{r, τ} ^S } ∈R ^{R × Τ} , _H ^(N) : = {^ H _{k, τ} ^N } ∈R ^{K × Τ}
By updating λ _{r, ω, τ} , ^ W _{ω, τ} ^N , ^ H _{ω, τ} ^N , ^ A _ω , parameters are estimated so as to maximize the joint probability.
Sound source separation device. The sound source separation device according to claim 3,
The estimated value of the diffusive noise acoustic signal is set to ^ D _{ω, τ} , and the filter design unit repeats the update process until a predetermined condition is satisfied, and uses the parameters at the end of the update to filter the filter.
Get as,
Sound source separation device. A sound source separation method for obtaining a desired acoustic signal from an observation signal obtained by recording a predetermined acoustic signal emitted from a speaker with a microphone,
The observation signal includes a first acoustic signal based on the predetermined acoustic signal and a transfer function that is a function expressing a spatial characteristic between the speaker and the microphone, and a coherent noise acoustic signal that is coherent noise. A diffusive noise acoustic signal that is diffusive noise, and
Removing the estimated value of the diffusive noise acoustic signal from the observed signal and obtaining a removed signal;
A filter design step of obtaining a filter by combining a probability distribution modeling the removed signal and a probability distribution modeling the transfer function;
A sound source separation step of separating at least the first acoustic signal and an estimated value of a noise component including the coherent noise acoustic signal from the observation signal by the filter,
Sound source separation method. The sound source separation method according to claim 5,
ω = {1,2, ..., Ω} and τ = {1,2, ..., Τ} are frequency and time indices, respectively, and the estimated transfer function is ^ A _ω. Is assumed to be ^ N _{ω, τ} , the predetermined acoustic signal is S _{ω, τ} , the removed signal is Y _{ω, τ} , _A: = [^ A _ω ] ∈R ^Ω , _N: = [ ^ N _{ω, τ} ] ∈R ^{Ω × Τ} , _S: = [S _{ω, τ} ] ∈R ^{Ω × Τ} , _Y: = [Y _{ω, τ} ] ∈R ^{Ω × Τ} , _α: = [α _ω ] ∈ R ^Ω , the probability distribution modeled on the removed signal is a likelihood function p (_A_N | _S, _Y) related to the removed signal, and the probability distribution modeled on the transfer function is a prior distribution related to the transfer function p (_A | _α), and the filter design step has the joint probability
L = p (_A, _N | _S, _Y) p (_A | _α)
Parameter is estimated to maximize the filter, and the filter is obtained from the estimated parameter.
Sound source separation method. The sound source separation method according to claim 6,
The observed signal is X _{ω, τ} , the estimated values of the amplitude spectra of the predetermined acoustic signal and the coherent noise acoustic signal are ^ W _{ω, r} ^S and ^ W _{ω, k} ^N _, respectively. Assume that the estimated values of the intensity corresponding to the basis of the amplitude spectrum of the acoustic signal and the coherent noise acoustic signal are ^ H _{r, τ} ^S and ^ H _{k, τ} ^N , respectively, and the predetermined acoustic signal and the coherent noise acoustic signal Let R and K be the basis numbers of the amplitude spectrum of
The filter design step includes
By
Or
T is transpose, _E is a matrix with Ω x Τ and all elements are 1. Matrix division is element-by-element division, _Z = [_ Z ^(S) , _W ^(N) ], _H = [(_ H ^{( S)} ) ^T , (_ H ^(N) ) ^T ] ^T , _Z ^(S) : = {^ A _ω ^ W _{ω, r} ^S } ∈R ^{Ω × R} , _W ^(N) : = {^ W _{ω, k} ^N } ∈R ^{Ω × K} , _H ^(S) : = {^ H _{r, τ} ^S } ∈R ^{R × Τ} , _H ^(N) : = {^ H _{k, τ} ^N } ∈R ^{K × Τ}
By updating λ _{r, ω, τ} , ^ W _{ω, τ} ^N , ^ H _{ω, τ} ^N , ^ A _ω , parameters are estimated so as to maximize the joint probability.
Sound source separation method.   A program for causing a computer to function as the sound source separation device according to any one of claims 1 to 4.