JP2013044908A - Background sound suppressor, background sound suppression method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a more efficient and highly accurate background sound suppressor capable of reducing a calculation cost and utilizing a probability density function with a more complex form.SOLUTION: In a background sound suppressor 20 of the present invention, a feature quantity extraction unit 100 extracts a high resolution sound source position feature quantity and a high resolution spectral feature quantity from an observation signal, a sound source position occupation degree estimation unit 200 obtains a high resolution sound source position occupation degree, a frequency resolution reduction unit 300 reduces frequency resolution of the high resolution spectral feature quantity and the high resolution sound source position occupation degree, a low resolution occupation degree estimation unit 400 estimates a spectral parameter, a high resolution occupation degree re-estimation unit 510 estimates a high resolution occupation degree, and a target speech estimation unit 600 estimates a target speech.

Description

  The present invention relates to a background sound suppression apparatus, a background sound suppression method, and a program that suppress background sound and estimate / extract target sound from observation signals collected by a plurality of microphones in which target sound and background sound are mixed. .

  Some conventional background sound suppression devices include a high-resolution spectrum model storage unit, a feature extraction unit, a high-resolution occupancy estimation unit, and a target speech estimation unit (see, for example, Non-Patent Document 1 and Non-Patent Document 2). .

  Hereinafter, conventional background sound suppression devices described in Non-Patent Documents 1 and 2 will be described. As described above, the conventional background sound suppression apparatus includes a high-resolution spectrum model storage unit, a feature extraction unit, a high-resolution occupancy estimation unit, and a target speech estimation unit. The high-resolution spectral model storage unit stores the prior probability density function of the spectral parameters representing the overall state of the spectral feature time series for each of the target speech and background sound, and each sound source signal (target purpose) when the spectral parameters are given. Signal or background sound) at each time frequency point, and a spectral feature quantity model that is a posterior probability density function. The feature extraction unit extracts an observation signal obtained by converting a time domain signal collected by a plurality of microphones into a time frequency domain signal, and extracts a high resolution sound source position feature quantity and a high resolution spectrum feature quantity at each time frequency point. . The high-resolution occupancy estimator receives the high-resolution sound source position feature, high-resolution spectral feature, high-resolution spectral feature model, and prior probability density function of the spectral parameters as input. Obtain high resolution occupancy estimates and spectral parameter estimates that are posterior probability density functions of occupying sound source numbers. Further, the target speech estimator includes a high-resolution occupancy estimation value and a spectral parameter estimation value output from the high-resolution occupancy estimation unit, a high-resolution spectral feature value output from the feature extraction unit, and a high-resolution spectral model storage The estimated value of the target speech is extracted using the high-resolution spectral feature model stored in the unit as an input.

Tomohiro Nakatani, Shoko Araki, Takuya Yoshioka, Masakiyo Fujimoto, “Multichannel source separation based on source location cue with log-spectral shaping by hidden Markov source model,” Proc. Of Interspeech-2010, pp. 2766-2769, Sep., 2010 . Tomohiro Nakatani, Akiko Araki, Takuya Yoshioka, Masaki Fujimoto, “Sound source separation based on DOA clustering and logarithmic spectrum HMM of speech” Proc. .

  However, since the conventional background sound suppression apparatus repeatedly performs processing to estimate the spectrum parameter and the high resolution occupancy based on the high resolution sound source position feature quantity and the high resolution spectrum feature quantity, the dimension of each feature quantity is large. There was a problem that the calculation cost increased as the time became. In particular, it is desirable to increase the size of the analysis window in order to properly handle the sound source position information extracted from the sound source position feature in a reverberant environment. The problem is that an increase in cost is inevitable.

  In addition, the conventional background sound suppressor assumes that the sound source position feature model is observed from a point sound source in an environment with relatively little reverberation, so the probability density function of the sound source position feature quantity is a single Gaussian. Only simple things such as distribution could be handled. Therefore, when reverberation longer than the analysis window is included, or when the background sound is not a point sound source or composed of a plurality of sound sources, the target speech cannot be estimated appropriately.

  The present invention has been made in view of such a point, and even when the dimension of each feature amount is large, the calculation cost can be kept low, and long reverberation is included or the background sound is not a point sound source. An object of the present invention is to provide a background sound suppression apparatus that can appropriately estimate a target voice even when it is composed of a plurality of sound sources.

The background sound suppression apparatus of the present invention suppresses the background sound and estimates the target speech from the observation signal x ^(m) _ (n, k) obtained by converting the time domain signal collected by a plurality of microphones into the time frequency domain signal. Extract the value {circumflex over (S ⁾ } ^(j) _ (n, k). First, m represents a microphone number, n represents a frame number, k represents a frequency bin number at high resolution, k￣ represents a frequency bin number at low resolution, and j represents a sound source number. Is represented. The low resolution spectral model storage unit stores the prior probability density function p (q ^(j) ) of the spectral parameters of each sound source signal and the low resolution spectral feature quantity model β￣_ (q ^(j) , n, k of each sound source signal. I) (S) is stored. The high-resolution spectral model storage unit stores a high-resolution spectral feature model β_ (q ^(j) , n, k) (S) of each sound source signal. The feature extraction unit extracts the high-resolution sound source position feature quantity A_ (n, k) and the high-resolution spectral feature quantity X_ (n, k) from the observation signal x ^(m) _ (n, k). The sound source position occupancy estimation unit obtains the sound source position parameter φ ^(j) of each sound source signal from the high resolution sound source position feature amount A_ (n, k), and obtains the high resolution sound source position feature amount A_ (n, k) and the sound source. From the position parameter φ ^(j) , the high-resolution sound source position occupancy Q ^(j) _ (n, k) of each sound source signal is obtained. The frequency resolution reduction unit performs smoothing processing between neighboring frequencies from the high-resolution spectral feature quantity X_ (n, k) and the high-resolution sound source position occupancy Q ^(j) _ (n, k). X￣_ (n, k￣) and low-resolution sound source position occupancy Q￣ ^(j) _ (n, k￣) are obtained. The low-resolution occupancy degree estimation unit includes a low-resolution spectral feature amount X￣_ (n, k￣), a low-resolution sound source position occupancy Q￣ ^(j) _ (n, k￣), and a prior probability density function p (q ^{( j)} ) and the low resolution spectral feature model β￣_ (q ^(j) , n, k￣) (S) to estimate the spectral parameter of each sound source signal so as to maximize the log likelihood function ^ Find q ^(j) . The high-resolution occupancy re-estimation unit includes the high-resolution spectral feature quantity X_ (n, k), an estimated value of spectral parameters ^ q ^(j), and a high-resolution sound source position occupancy Q ^(j) _ (n, k). From the high-resolution spectral feature model β_ (q ^(j) , n, k) (S), an estimated value of high resolution occupancy ^ M ^(j) _ (n, k) is obtained. The target speech estimator is configured to estimate the spectral parameter ^ q ^(j) , the high-resolution occupancy estimate ^ M ^(j) _ (n, k), the high-resolution spectral feature X_ (n, k), and the high-resolution. An estimated value {circumflex over (S)} ^(j) _ (n, k) of the target speech is obtained from the spectral feature model β_ (q ^(j) , n, k) (S).

  The background sound suppression apparatus of the present invention can reduce the calculation cost even when the dimension of each feature amount is large, and includes a long reverberation, a background sound that is not a point sound source, or a plurality of sound sources. In this case, the target speech can be estimated appropriately.

The block diagram which shows the structure of the conventional background sound suppression apparatus. The flowchart which shows operation | movement of the conventional background sound suppression apparatus. 1 is a block diagram illustrating a configuration of a background sound suppression device according to Embodiment 1. FIG. 5 is a flowchart showing the operation of the background sound suppression apparatus according to the first embodiment. The example of the filter coefficient which a frequency resolution reduction part uses. FIG. 6 is a block diagram illustrating a configuration of a background sound suppression device according to a second embodiment. 10 is a flowchart illustrating the operation of the background sound suppression device according to the second embodiment. FIG. 9 is a block diagram illustrating a configuration of a background sound suppression device according to a third embodiment. 10 is a flowchart illustrating the operation of the background sound suppression device according to the third embodiment.

  Hereinafter, embodiments of the present invention will be described in detail. In addition, the same number is attached | subjected to the structure part which has the same function, and duplication description is abbreviate | omitted.

First, symbols used for description will be described. The observed signal is superimposed target speech and the background sound, picking up the sound signal in N _m the microphones. An observation signal obtained by converting an acoustic signal collected from the m-th microphone into a frequency domain signal using a short-time Fourier transform or the like is represented as x ^(m) _ (n, k). n is the nth time, that is, the frame number, k is the kth frequency, that is, the bin number, and when referring to the time frequency point corresponding to the nth time and the kth frequency, the time frequency point (n, k ). The total number of frequency bins for each frame is denoted as _Nk . j is the number of the j-th sound source signal, j = 1 represents the target sound, and j = 2 represents the background sound. There is the following correspondence between the expression in the mathematical expression and the expression in the text.

<Description of conventional example>
First, the outline of the operation of the conventional background sound suppression apparatus 10 will be described with reference to FIGS. FIG. 1 is a block diagram showing the configuration of a conventional background sound suppression apparatus 10. FIG. 2 is a flowchart showing the operation of the conventional background sound suppression apparatus 10.

In the conventional background sound suppression apparatus 10, the simultaneous probability density function of the entire spectrum time series {S ^(j) _ (n, k)} of the j th sound source signal is modeled as shown in the following equation.

Here, q ^(j) represents a spectrum parameter representing the state of the entire spectrum time series of the j-th sound source signal. Hereinafter, q ^(j) of all sound source signals is collectively expressed as q = [q ⁽¹⁾ , q ⁽²⁾ ].

Further, β_ (q ^(j) , n, k) (S) is a model of the spectral feature quantity, and as expressed in the equation (3), each time is given under the spectral parameter q ^(j). This is a probability density function in which the spectrum value of the sound source signal at the frequency point (n, k) is S.

Equation (2) introduces the assumption that the spectral values S ^(j) _ (n, k) at different time frequency points are independent of each other when the spectral parameters are known.

In the conventional example, as shown in Expression (4), the spectrum value S ^(j) of a sound source signal having the largest energy at each time frequency point (n, k) (hereinafter referred to as an exclusive sound source signal ^). It is assumed that _ (n, k) matches the spectrum value of the observation signal.

It is assumed that the non-occupying sound source j has a relationship of S ^(j) _ (n, k) ≦ X_ (n, k). Then, it is known that the posterior probability density function of the high-resolution spectral feature quantity X_ (n, k) of the observation signal can be expressed as follows under the condition that the spectral parameters of each sound source signal are known (in detail) “See SJ Rennie, JR Hershey, and PA Olsen,“ Hierarchical variational loopy belief propagation for multi-talker speech recognition, ”Proc. ASRU-2009, pp. 176-181 2009.”).

  In the conventional example, it is further assumed that the above equation can be decomposed as follows.

  Z_ (n, k) is a random variable representing the number of the sound source that is occupied at the time frequency point (n, k), and Z_ (n, k) = j is a sound source in which the jth sound source is occupied. Show the case.

Further, in the conventional background sound suppressing apparatus 10, since the sound source position parameter φ ^(j) is estimated from the high resolution sound source position feature quantity A_ (n, k), the high resolution sound source position feature quantity model p (A_ (n, k); φ) is introduced. The sound source position feature quantity model p (A_ (n, k); φ) corresponding to each sound source j assumes that the energy of each sound source signal is sparsely distributed over different time frequency points, and the time frequency points. , It is assumed that it depends only on the sound source position of the exclusive sound source. When the sound source position parameters φ ^(j) of all sound sources are collectively expressed as φ = [φ ⁽¹⁾ , φ ⁽²⁾ ], a model p (A_ (n, k); φ), that is, the probability density function of the high-resolution sound source position feature quantity of the observation signal can be developed as a mixture distribution as shown in Expression (8).

  In Equation (8), p (Z_ (n, k) = j) represents a prior probability density function in which the jth sound source becomes an exclusive sound source at the time frequency point (n, k). Further, the following notation is used in the following description.

γ_ (φ ^(j) , n, k) (A) is a high-resolution sound source position feature quantity A_ (n, k) when the number of the sound source occupied at the time frequency point (n, k) is j. Represents the resulting probability density function. This depends only on the sound source position parameter φ ^(j) of the j-th sound source. Specific definitions of γ_ (φ ^(j) , k) (A) and φ ^(j) will be described later.

Original expression ^{(8), γ_ (φ (} j), k) if (A) is defined, the sound source position parameter phi ^(j) and exclusively pre probability for the number of sound source density function p (Z_ ( If n, k) = j) is given, the model p (A_ (n, k); φ) of the sound source position feature quantity can be uniquely determined. Conversely, when the sound source position feature A_ (n, k) is observed, the prior probability density function p (Z_ (n, k) relating to the sound source position parameter and the number of the occupied sound source according to a method such as maximum likelihood estimation. ) = J) and its posterior probability density function.

  According to the above definition, the probability density function of complete data is derived as shown in equation (10).

  In Expression (10), q is a spectral parameter, and φ is a sound source position parameter. In the conventional example, the spectrum parameter q and the sound source position parameter φ are estimated as values that maximize the next log likelihood function.

In Expression (12), the random variable Z_ (n, k) is treated as a hidden variable. For example, an expectation maximization algorithm can be used to maximize the log likelihood function including hidden variables. In the expected value maximization algorithm, the a posteriori probability density function ^ M ^(j) _ (n, k) = p of the number of the occupied sound source under the observation signal obtained based on the estimated value ^ q of the spectral parameter (Z_ (n, k) | A_ (n, k), X_ (n, k) ^ q; ^ φ) also needs to be estimated at the same time. In the conventional example, the value of this function is referred to as high resolution occupancy, and this value is also included in the parameter to be estimated.

  In the following, description will be made in the order of procedures actually performed. The conventional background sound suppression apparatus 10 includes a feature extraction unit 100, a high resolution occupancy estimation unit 500, a target speech estimation unit 600, and a high resolution spectrum model storage unit 800.

The high-resolution spectral model storage unit 800 includes a prior probability density function p (q ^(j) ) of a spectral parameter q ^(j) representing the state of the entire spectrum time series of the target speech and background sound, and the spectral parameter q ^{(j )} Is stored, the model β_ (q ^(j) , n, k) (S) of the high resolution spectral feature quantity at each time frequency point of each sound source signal is stored. (S) is a variable representing the sound source power feature amount X_ (n, k). The prior probability density function p (q ^(j) ) and the high resolution spectral feature quantity model β_ (q ^(j) , n, k) (S) are given by prior learning for each of the target speech and the background sound. It shall be.

The feature extraction unit 100 receives an observation signal x ^(m) _ (n, k) obtained by converting a time domain signal collected by a plurality (N _m ) of microphones into a time frequency domain signal, and inputs each time frequency point ( The high-resolution sound source position feature quantity A_ (n, k) and the high-resolution spectral feature quantity X_ (n, k) at n, k) are extracted (S101, S102).

  The high-resolution spectral feature amount X_ (n, k) is extracted as, for example, a logarithmic power spectrum of a signal picked up by the first microphone. This is calculated as shown in equation (13).

  The high-resolution sound source position feature quantity A_ (n, k) generally appears in a signal phase difference or intensity ratio between different microphones at each time frequency point. Therefore, the high-resolution sound source position feature amount A_ (n, k) is a vector that can be obtained by grouping different signal phase differences and intensity ratios for different microphone pairs, or a value obtained as a result of performing some feature extraction therefrom. It is extracted as there is. For example, when a phase difference between signals collected by two microphones is extracted as a high-resolution sound source position feature amount A_ (n, k), the calculation is performed as shown in Expression (14).

  In addition to the above, for example, a normalized complex spectrum vector calculated as shown in Expression (14 ′) can also be used as the sound source position feature amount (for details, see “Hiroshi Sawada, Shoko Araki, and Shoji Makino, “Underdetermined convolutive blind source separation via frequency bin-wise clustering and permutation alignment,” IEEE Trans. Audio, Speech, and Language Processing, vol. 19, no. 3, pp. 516-527, 2011. 1)).

  Hereinafter, in the present specification, the configuration of the invention will be described using the high-resolution sound source position feature quantity A_ (n, k) according to the equation (14 ′). Refer to Non-Patent Document 1 and Non-Patent Document 2 for the configuration of the invention using Expression (14).

The high-resolution occupancy estimation unit 500 includes a high-resolution sound source position feature amount A_ (n, k) and a high-resolution spectral feature amount X_ (n, k) output from the feature extraction unit 100, and a high-resolution spectral model storage unit 800. Input of the spectral parameter prior probability density function p (q ^(j) ) and the spectral feature quantity model β_ (q ^(j) , n, k) (S) Degree ^ M ^(j) _ (n, k) is estimated.

First, the high-resolution occupancy estimation unit 500 initializes the high-resolution occupancy ^ M ^(j) _ (n, k) for each sound source j, for example, with random numbers so that Σ _j ^ M ^(j) = 1. Turn into. Thereafter, the following processes (1) to (3) are repeated until convergence.

(1) Update of spectrum parameters (S501)
High-resolution spectral feature quantity X_ (n, k), high-resolution occupancy ^ M ^(j) _ (n, k), spectral parameter prior probability density function p (q ^(j) ), and high-resolution spectral feature quantity model Using β_ (q ^(j) , n, k) (S), the estimated value ^ q ^(j) of the spectral parameter is updated (M-step) as shown in the equation (15).

(2) Update of sound source position parameter (S502)
Using the high resolution occupancy ^ M ^(j) _ (n, k) and the high resolution sound source position feature A_ (n, k), as shown in the equation (17), the sound source position parameter ^ φ ^(j) Is updated (M-step).

(3) Update of high resolution occupancy (S503)
Spectral parameter ^ q ^(j) , high-resolution sound source position feature A_ (n, k), high-resolution spectral feature X_ (n, k), and high-resolution spectral feature model β_ (q ^(j) , n, k ) (S) and the high resolution occupancy ^ M ^(j) _ (n, k) is updated (E-step) as shown in the equation (18).

The target speech estimation unit 600 includes a high-resolution spectral feature quantity X_ (n, k), a high-resolution occupancy ^ M ^(j) _ (n, k), a spectral parameter ^ q ^(j) _n, and a high-resolution spectral feature. Using the quantity model β_ (q ^(j) , n, k) (S) as an input, an estimated value {circumflex over (S)} ^(j) _ (n, k) of the target speech is obtained by least square error estimation (S600). The estimation method is as follows.

<Problem of conventional example>
In the conventional background sound suppression apparatus 10, in the high resolution occupancy estimation unit 500, the spectrum parameter ^ q ^(j) , the sound source position parameter ^ φ ^(j) , and the high resolution occupancy ^ M ^(j) _ (n, k (15), (17), and (18) are repeatedly executed for the update of. At this time, the calculation cost increases as the dimensions of the high-resolution spectral feature quantity X_ (n, k) and the high-resolution sound source position feature quantity A_ (n, k) increase, that is, as the total number N _k of frequency bins of each frame increases. There was a problem that became larger.

In the conventional example, in the probability density function of the sound source position feature quantity, the probability density function of the high resolution sound source position feature quantity of each sound source represented by γ_ (φ ^(j) , k) (A) is a single Only simple things such as Gaussian distribution could be handled. Therefore, only the statistical properties of the high-resolution sound source position feature of a point sound source with relatively little reverberation can be expressed, the reverberation is long, the background sound includes multiple point sound sources, and sound sources other than point sound sources are included. In this case, the high-resolution sound source position feature quantity of the target voice and the background sound could not be expressed appropriately.

<Outline of the present invention>
In the first embodiment, the iterative estimation of the spectrum parameter ^ q ^(j) , which has increased the calculation cost in the conventional example, is obtained by the iterative estimation in a space with a reduced frequency resolution. As a result, the spectral parameter ^ q ^(j) can be estimated efficiently. Further, the high resolution occupancy can be estimated by using the estimated spectral parameter ^ q ^(j) , the high resolution spectral model, and the high resolution sound source position occupancy. At this time, since iterative processing is not necessary for estimating the high resolution occupancy, the calculation cost does not increase. As a result, the spectral parameter ^ q ^(j) and the high resolution occupancy can be obtained efficiently. Even in this configuration, iterative processing is necessary for estimating the high-resolution sound source position occupancy, but this processing is smaller than the iterative processing for estimating the high-resolution occupancy in the conventional example. Can be realized at a cost.

  In the second embodiment, a high-resolution sound source position feature quantity model learned in advance is also provided, so that the high-resolution sound source position occupancy can be estimated without performing iterative processing. As a result, the background sound can be suppressed more efficiently. Furthermore, when learning a model of high-resolution sound source position feature quantity in advance, it becomes possible to use a more complicated probability density function of high-resolution sound source position feature quantity. Even when it is not composed of only a single point sound source, the sound source position feature quantities of the target sound and the background sound can be more appropriately distinguished.

  In the third embodiment, the frequency resolution is not reduced, and a model of the high-resolution sound source position feature quantity learned in advance is provided. Since the model of the sound source position stored in the high-resolution sound source position model storage unit can be used, it is not necessary to estimate the model parameters of the sound source positions of the target sound and the background sound, and the calculation cost can be kept low. . In addition, as a model of the sound source position stored in the high-resolution sound source position model storage unit, a model having a more complicated distribution shape such as a mixture distribution can be used. Even in an environment included in sound, background sound can be appropriately suppressed.

  Next, the operation of the background sound suppression apparatus 20 according to the first embodiment of the present invention will be described in detail with reference to FIGS. FIG. 3 is a block diagram showing the configuration of the background sound suppression apparatus 20 according to the first embodiment of the present invention. FIG. 4 is a flowchart showing the operation of the background sound suppression apparatus 20 according to the first embodiment of the present invention.

  In the following, description will be made in the order of procedures actually performed. The background sound suppression apparatus 20 of this embodiment includes a feature extraction unit 100, a sound source position occupancy estimation unit 200, a frequency resolution reduction unit 300, a low resolution occupancy estimation unit 400, a high resolution occupancy re-estimation unit 510, and target speech estimation. Unit 600, low-resolution spectral model storage unit 700, and high-resolution spectral model storage unit 810.

The low-resolution spectral model storage unit 700 includes a prior probability density function p (q ^(j) ) of a spectral parameter q ^(j) representing the state of the entire spectrum time series of the target speech and background sound, and the spectral parameter q ^{(j )} Is stored, the model β ス ペ ク ト ル _ (q ^(j) , n, k￣) (S) of the low resolution spectral feature quantity at each time frequency point of each sound source signal is stored. (S) is a variable representing the low-resolution spectral feature amount X￣_ (n, k￣). The simultaneous probability density function of the entire time series {S ￣ ^(j) _ (n, k ￣)} of the low-resolution spectral feature quantity of the j-th sound source signal is expressed by the following equations (1 ′), (2 ′), and (3 ′). Model as shown.

Furthermore, the spectrum parameter q ^(j) is decomposed into a state series representing the state at each time as q ^(j) = {q ^(j) _ (0), q ^(j) _ (1),. Suppose that state transitions occur at each time according to a first-order Markov process. However, the spectrum parameter q ^(j) _ (0) represents the initial state of the hidden Markov model. The posterior probability density function of S￣ ^(j) _ (n, k￣) at each time frequency point (n, k￣) defined by the equation (3 ′) is the state q ^(j) _ (n ) Is assumed to follow a Gaussian distribution that depends only on. This can be expressed by equations (20) and (21).

Here, π ^(j) _ (i) = p (q ^(j) _ (0) = i) is a prior probability that the initial state of the hidden Markov model is i, α ^(j) _ (i, h) = P (q ^(j) _ (n) = h | q ^(j) _ (n-1) = i) is the state transition probability that the hidden Markov model moves from state i to state h, β 、 _ (i, n, k￣) (S) = p (S￣ ^(j) _ (n, k￣) = S | q ^(j) _ (n) = i) = N (S￣ ^(j) _ (n, k ￣); μ￣ ^(j) _ (i, k￣), σ￣ ^(j) _ (i, k￣)) is the probability density function of the output in state i of the hidden Markov model, and μ￣ ^{(j )} _ (I, k￣) and σ￣ ^(j) _ (i, k￣) are their mean and variance. Π ^(j) _ (i), α ^(j) _ (i, h), μ￣ ^(j) _ (i, k￣), σ￣ ^(j) _ ( ^{) for} all h, i, j, k In this embodiment, i, k￣) are all obtained in advance by learning from a speech database or the like.

The high-resolution spectral model storage unit 810 is a model β_ (q ^(j) , n, k) (S) of the high-resolution spectral feature quantity at each time frequency point of each sound source signal when the spectral parameter q ^(j) is given. ) Is stored.

The feature extraction unit 100 receives the observation signal x ^(m) _ (n, k) as an input, and extracts a logarithmic power spectrum as a high resolution spectral feature quantity X_ (n, k) based on the equation (13) (S101). . Further, based on the equation (14 ′), the normalized complex spectrum is extracted as the high-resolution sound source position feature amount A_ (n, k) (S102).

The sound source position occupancy estimation unit 200 receives the high-resolution sound source position feature quantity A_ (n, k) as an input and estimates the sound source position parameter φ ^(j) (S201). For this estimation, reference 1 or “Tomohiro Nakatani, Shoko Araki, Takuya Fujimoto, Masakiyo Fujimoto,“ Joint unsupervised learning of hidden Markov source models and source location models for multi-channel source separation, ”Proc. Of IEEE ICASSP-2011 , pp. 237-240, 2011. (hereinafter referred to as Reference Document 2) and the like. For this reason, in this embodiment, the normalized complex spectrum of the observation signal derived from each sound source signal has an average value μ ^(j) _ (k) and variance σ ^(j) _ (k) that differ for each frequency. Assume that the following distribution follows.

However, φ ^(j) _ (k) = [μ ^(j) _ (k), σ ^(j) _ (k)] is obtained by extracting a portion related to only the frequency k from the sound source position parameter φ ^(j). Φ ^(j) is a collection of φ ^(l) _ (k) for all frequencies k φ ^(j) = [φ ^(j) _ (1),..., Φ ^(j) _ (N _k ) ]. Based on this assumption, in this embodiment, the probability density function of the high-resolution sound source position feature quantity of the observation signal x ^(j) _ (n, k) is modeled by equations (8), (9), and (19). And

Subsequently, the sound source position occupancy estimation unit 200 estimates the high-resolution sound source position occupancy Q ^(j) _ (n, k) as follows based on the estimated sound source position parameter φ ^(j) (S202). ).

  The frequency resolution reduction unit 300 uses the high-resolution spectral feature amount X_ (n, k) output from the feature extraction unit 100 and the high-resolution sound source position occupancy Q_ (n, k) output from the sound source position occupancy estimation unit 200. By applying a smoothing process between neighboring frequencies as an input, it is converted into a low resolution spectral feature quantity X￣_ (n, k￣) and a low resolution sound source position occupancy Q￣_ (n, k￣).

In order to reduce the frequency resolution of the high-resolution spectral feature value X_ (n, k), for example, filter bank processing or the like often used in the feature value extraction for speech recognition is used. Now, F_ (k￣) = [F_ (k￣, 1), F_ (k￣, 2),..., F_ (k￣, N _k )] is obtained to obtain the k￣th output of the filter bank processing. Filter coefficients. The conversion from the high-resolution spectral feature quantity X_ (n, k) to the low-resolution spectral feature quantity X （_ (n, k￣) is obtained as follows using the filter coefficient F_ (k￣) (S301). ).

Here, k￣ represents the frequency number of the low-resolution spectral feature quantity X￣_ (n, k￣), and k￣ ≦ k.

  Next, the frequency resolution reduction unit 300 converts the high-resolution sound source position occupancy Q_ (n, k) to the low-resolution sound source position occupancy Q￣_ (n, k￣) with the same filter coefficient F_ (k￣ ) Is performed as follows (S302).

  FIG. 5 shows an example of the filter coefficient F_ (k￣).

The low-resolution occupancy estimation unit 400 receives the low-resolution spectral feature amount X￣_ (n, k￣) and the low-resolution sound source position occupancy Q￣_ (n, k￣) as inputs, according to an expected value maximization algorithm. obtaining the estimated value of the spectrum parameter ^{^ q} and ^(j) a low-resolution occupancy estimate ^{^ M¯ (j) _ (n} , k¯). For this reason, the following processes (1) and (2) are repeated until convergence.

(1) Update of estimated values of spectral parameters (S401)
For each sound source j, an estimated value of spectral parameters ^ q (j) = [^ q ^(j) _ (0),..., ^ Q ^(j) _ (N _s )] satisfying Expression (22) is represented by the Viterbi algorithm. Update using.

(2) Update of low resolution occupancy (S402)
The low resolution occupancy M￣ ^(j) _ (n, k￣) is updated as shown in the equation (32) (E-step).

The spectral parameter estimation value ^{ circumflex over ⁽ q ⁾ } ^(j) obtained as a result of the repetition of the above (1) and (2) is the output of the low resolution occupancy estimation unit 400.

The high-resolution occupancy re-estimation unit 510 includes a high-resolution spectral feature quantity X_ (n, k), a high-resolution sound source position occupancy Q_ (n, k), a spectral parameter estimate value ^ q ^(j), and a spectrum The feature value model β_ (q ^(j) , n, k) (S) is used as an input, and an estimated value of high resolution occupancy ^ M ^(j) _ (n, k) is obtained according to equation (32 ′). (S510).

The target speech estimation unit 600 includes a high-resolution spectral feature amount X_ (n, k), a high-resolution occupancy estimation value ^ M ^(j) _ (n, k), and a spectral parameter estimation value ^ q ^(j). And a high-resolution spectral feature model β_ (q ^(j) , n, k) (S) as an input, and based on the same equation (36) as in the conventional example, the target speech in which the background sound is suppressed from the observation signal determination of the estimated value ^{^ S (j) _ (n} , k) (S600).

As described above, the background sound suppression apparatus 20 according to the present embodiment obtains the iterative estimation of the spectrum parameter ^ q ^(j) , which has increased the calculation cost in the conventional example, by the iterative estimation in the space where the frequency resolution is reduced. . As a result, the spectral parameter ^ q ^(j) can be estimated efficiently. Further, by using the estimated spectral parameter ^ q ^(j) , the high resolution spectral feature model β_ (q ^(j) , n, k) (S) and the high resolution sound source position occupancy Q_ (n, k). The high resolution occupancy M ^(j) _ (n, k) can be estimated. At this time, since iterative processing is not necessary for estimating the high resolution occupancy, the calculation cost does not increase. As a result, the spectral parameter ^ q ^(j) and the high resolution occupancy M ^(j) _ (n, k) can be obtained efficiently. In this configuration as well, iterative processing is necessary for estimating the high-resolution sound source position occupancy Q_ (n, k). This processing is similar to the iterative processing for estimating the high-resolution occupancy in the conventional example. In comparison, it can be realized with a small calculation cost.

  Next, the operation of the background sound suppression device 30 according to the second embodiment of the present invention will be described in detail with reference to FIGS. FIG. 6 is a block diagram showing the configuration of the background sound suppression apparatus 30 according to the second embodiment of the present invention. FIG. 7 is a flowchart showing the operation of the background sound suppression apparatus 30 according to the second embodiment of the present invention. Below, it demonstrates centering on difference with Example 1, and abbreviate | omits description about the matter which is common in Example 1. FIG.

  The background sound suppression apparatus 30 of this embodiment includes a feature extraction unit 100, a sound source position occupancy estimation unit 210, a frequency resolution reduction unit 300, a low resolution occupancy estimation unit 400, a high resolution occupancy re-estimation unit 510, and target speech estimation. Unit 600, low-resolution spectral model storage unit 700, high-resolution spectral model storage unit 810, and high-resolution sound source position model storage unit 900.

The high-resolution sound source position model storage unit 900 stores a probability density function γ ^(j) _ (k) (A) of a high-resolution sound source position feature quantity for each sound source signal (target sound or background sound). The shape of the probability density function γ ^(j) _ (k) (A) is fixed by prior learning and does not need to be estimated from the observed signal. Further, as in the equation (19), it is not necessary to easily estimate the parameters from the observation signal, and a more complicated format can be obtained.

The sound source position occupancy estimation unit 210 outputs the high-resolution sound source position feature quantity A_ (n, k) output from the feature extraction unit 100 and the probability density function γ ^{(j) —} stored in the high-resolution sound source position model storage unit 900. (K) Using (A) as an input, the high-resolution sound source position occupancy Q ^(j) _ (n, k) is estimated according to the following equation (S210).

  Other components and the processing flow are the same as those of the background sound suppression apparatus 20 of the first embodiment.

Next, a prior learning method of the probability density function γ ^(j) _ (k) (A) will be described. Now, an observation signal including only the sound source j (target sound or background sound) is obtained as data for prior learning, and a high-resolution sound source position feature A_ (n, k) is extracted from the observation signal. Suppose that n = 1 to N. At this time, any function can be used as the probability density function γ ^(j) _ (k) (A) as long as it represents the probability density function of this feature quantity at each frequency k. As an example, a case where a mixed distribution having a distribution F (A; μ ^(j) _ (k), σ ^(j) _ (k)) defined by Expression (19) as an element will be described. At this time, the probability density function γ ^(j) _ (k) (A) is modeled as follows.

Here, r is an element number of the mixture distribution, u ^(j) _ (r) is a mixture ratio of the element, and F (A; μ ^(j) _ (r, k), σ ^{( j)} _ (r, k)) represents the distribution of the elements. One difference between Equation (19) and Equation (19 ′) is that the probability density function for each sound source j is modeled with only one element in Equation (19), whereas Equation (19 ′) is It is a mixed distribution consisting of multiple elements. For each sound source j, parameters to be determined in advance learning are u ^(j) _ (r), μ ^(j) _ (r, k), and σ ^(j) _ (r, k) for all r, k. It is. Using the high-resolution sound source position feature quantity A_ (n, k) extracted from the pre-learning data, these parameters can be obtained by the following procedure using an expected value maximization algorithm.

(1) Initialize μ ^(j) _ (r, k) and σ ^(j) _ (r, k) for all r and k. For example, μ ^{(j) —} (r, k) is initialized with a random number, and σ ^{(j) —} (r, k) is initialized to σ ^{(j) —} (r, k) = 1.

(2) u ^(j) _ (r) (> 0) is initialized with a random number, for example, so that Σ _r u ^(j) _ (r) = 1.

(3) The following (3-1) to (3-4) are repeated until convergence.
(3-1) K ^(j) _ (n, r, k) is updated as follows.

(3-2) σ ^(j) _ (r, k) is updated as follows.

(3-3) An eigenvalue for the maximum eigenvalue of the matrix R_ (r, k) obtained as follows is obtained, and is substituted by μ ^(j) _ (r, k) to be updated.

(3-4) u ^(j) _ (r, k) is updated as follows.

As a result of the above iteration, u ^(j) _ (r), μ ^(j) _ (r, k), and σ ^(j) _ (r, k) finally obtained are parameters obtained by prior learning. In accordance with these parameters, the probability density function γ ^(j) _ (k) (A) is defined by equation (19 ′).

As described above, the background sound suppression apparatus 30 according to the present embodiment includes the model of the high-resolution sound source position feature value learned in advance, so that the high-resolution sound source position occupancy Q ^(j) can be obtained without iterative processing. Since _ (n, k) can be estimated, the calculation cost can be kept low. As a result, the background sound can be suppressed more efficiently.

In addition, when learning a model of a high resolution sound source position feature quantity in advance, a more complicated probability density function γ ^(j) _ (k) (A) of a high resolution sound source position feature quantity can be used. Therefore, even when a long reverberation is included or the background sound is not composed of only a single point sound source, the sound source position feature quantities of the target sound and the background sound can be more appropriately distinguished. As a result, background sound suppression can be performed more appropriately.

  Next, the operation of the background sound suppression device 40 according to the third embodiment of the present invention will be described in detail with reference to FIGS. FIG. 8 is a block diagram showing the configuration of the background sound suppression apparatus 40 according to the third embodiment of the present invention. FIG. 9 is a flowchart showing the operation of the background sound suppression apparatus 40 according to the third embodiment of the present invention. Below, it demonstrates centering around difference with Example 2, and abbreviate | omits description about the matter which is common in Example 2. FIG.

  The background sound suppression apparatus 40 of this embodiment includes a feature extraction unit 100, a sound source position occupancy estimation unit 210, a high resolution occupancy estimation unit 520, a target speech estimation unit 600, a high resolution spectrum model storage unit 800, a high resolution sound source position. A model storage unit 900 is provided.

The high-resolution occupancy estimation unit 520 includes the high-resolution spectral feature amount X_ (n, k) output from the feature extraction unit 100 and the high-resolution sound source position occupancy Q ^(j) _ output from the sound source position occupancy estimation unit 210. Using (n, k) as an input, according to an expectation maximization algorithm, an estimated value of spectrum parameter ^ q ^(j) and an estimated value of high resolution occupancy ^ M ^(j) _ (n, k) are obtained. For this reason, the following processes (1) and (2) are repeated until convergence.
(1) Update of estimated values of spectral parameters (S521)
For each sound source j, an estimated value of spectral parameters ^ q ^(j) = [^ q ^(j) _ (0), ..., ^ q ^(j) _ (N _s )] satisfying Expression (22 ′) is expressed as Viterbi. Update using algorithm.

(2) Update of high resolution occupancy (S522)
The high resolution occupancy M ^(j) _ (n, k) is updated (E-step) as shown in Expression (32 ′).

Spectral parameter estimates ^{ circumflex over ⁽ q ⁾ } ^(j) and high resolution occupancy estimates {circumflex over (M ⁾ } ^(j, _ (n, k) obtained as a result of repetition of (1) and (2) above This is the output of the degree estimation unit 520.

  Note that the high-resolution occupancy estimation unit 520 differs from the low-resolution occupancy estimation unit 400 of the second embodiment only in the frequency resolution of the feature amount, and the processing contents are the same.

  Other components and processing flow are the same as those of the background sound suppression device 30 of the second embodiment.

The background sound suppression apparatus 40 of the present embodiment pays attention only to the result of the entire processing. In the second embodiment, the filter coefficient F_ (k￣) = [F_ (k￣, 1), F_ ( k ￣, 2), ..., F_ ( _k Ｎ, N _k )] is N _k and each element is set to F_ (k ￣, k) = 1 when k ￣ = k, otherwise This corresponds to the case where F_ (k￣, k) = 0. In this case, the input / output of the frequency resolution reduction unit 300 of the second embodiment is the same. That is, the frequency resolution reduction unit 300 is equivalent to performing no processing. In addition, the model β￣_ (q ^(j) , n, k￣) (S) of the low resolution spectral model feature quantity and the model β_ (q ^(j) , n, k) (S) of the high resolution spectral model feature quantity are used. Are the same, and the estimated value of low resolution occupancy ^ M￣ ^(j) _ (n, k￣) and the estimated value of high resolution occupancy ^ M ^(j) _ (n, k) are also the same. Become.

Therefore, in this embodiment, the frequency resolution reduction unit 300 is omitted, and the high-resolution spectral feature amount X_ (n, k) output from the feature extraction unit 100 and Q ^(j) _ ( ⁾ output from the sound source position occupancy estimation unit 210 are omitted. n, k) is input to the high resolution occupancy estimation unit 520. Further, the low resolution occupancy estimation unit 400 and the low resolution spectrum model storage unit 700 are omitted, and the high resolution occupancy estimation value ^ M (j) _ (n, k) output from the high resolution occupancy estimation unit 520 The estimated value ^ q ^{(j) of the} spectral parameter and the model β_ (q ^(j) , n, k) (S) of the high resolution spectral feature quantity stored in the high resolution spectral model storage unit 800 are used to estimate the target speech. Shall be input to the department.

As described above, the background sound suppression apparatus 40 according to the present embodiment includes the high-resolution sound source position feature quantity model learned in advance, so that the high-resolution sound source position occupancy Q ^(j) can be obtained without performing the iterative process. Since _ (n, k) can be estimated, the calculation cost can be kept low. As a result, the background sound can be suppressed more efficiently.

In addition, when learning a model of a high resolution sound source position feature quantity in advance, a more complicated probability density function γ ^(j) _ (k) (A) of a high resolution sound source position feature quantity can be used. Therefore, even when a long reverberation is included or the background sound is not composed of only a single point sound source, the sound source position feature quantities of the target sound and the background sound can be more appropriately distinguished. As a result, background sound suppression can be performed more appropriately.

<Confirmation experiment>
A confirmation experiment was conducted for the purpose of evaluating the background sound suppressor of the present invention.

  The experimental conditions will be described. In a room with reverberation, we used two microphones and used the sound of the speaker in front of the microphone as well as various surrounding background sounds. This observation signal includes a relatively long reverberation, and the background sound includes a plurality of point sound sources or a sound source that is not a point sound source. In order to appropriately handle such observation signals, the model of the high-resolution sound source position feature amount shown in Example 2 of the present invention was prepared by prior learning. In this confirmation experiment, in the invention of Example 2, a comparison was made between the case where the frequency resolution was reduced (the present invention) and the case where the frequency resolution was not performed (the conventional example). In both cases, the analysis window length of the short-time Fourier transform is set to 100 milliseconds so that the sound source position information of the signal including reverberation can be appropriately handled. Since the sampling frequency was 16 kHz, the dimension of the high-resolution spectral feature amount was 801. On the other hand, the dimension of the low-resolution spectral feature is 40.

  First, the real-time factor was measured as a comparison of calculation costs. The real time factor is the ratio of the time (seconds) required for background sound suppression processing to the observed signal length (seconds). When the real time factor is 1 or less, it means that the processing is completed in a time shorter than the length of the observation signal. In our experiments, the real-time factors of the conventional example and the present invention were 4.52 and 0.69, respectively. Thereby, it was confirmed that the present invention can greatly reduce the calculation cost.

  Next, the result of applying automatic speech recognition to the observed signal and the signal with the background sound suppressed is shown. The word correct rate when the observed signal was speech recognized as it was was 69.4%, whereas the word correct rate when the sound with the background sound suppressed in the conventional example and the present invention was recognized as speech was 82, respectively. 7% and 81.6%. Since the speech recognition rate was greatly improved in both the conventional example and the present invention, it can be seen that the high-resolution sound source position feature quantity model of Example 2 functioned effectively. In addition, the speech recognition performance was slightly reduced by the present invention as compared with the conventional example, but the difference was extremely small.

  From the above results, it was confirmed that the present invention can realize the effect of greatly reducing the calculation cost of the conventional example without substantially deteriorating the background sound suppression performance.

<Program, recording medium>
The various processes described above are not only executed in time series according to the description, but may also be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. Needless to say, other modifications are possible without departing from the spirit of the present invention.

  Further, when the above-described configuration is realized by a computer, processing contents of functions that each device should have are described by a program. The processing functions are realized on the computer by executing the program on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. 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. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

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

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

  INDUSTRIAL APPLICABILITY The present invention can be used for estimating and extracting a target sound by suppressing the background sound from observation signals collected by a plurality of microphones in which the target sound and the background sound are mixed.

10, 20, 30, 40 Background sound suppression device 100 Feature extraction unit 200, 210 Sound source position occupancy estimation unit 300 Frequency resolution reduction unit 400 Low resolution occupancy estimation unit 500, 520 High resolution occupancy estimation unit 510 High resolution occupancy Re-estimation unit 600 Target speech estimation unit 700 Low-resolution spectral model storage unit 800, 810 High-resolution spectral model storage unit 900 High-resolution sound source position model storage unit

Claims (7)

From the observation signal x ^(m) _ (n, k) obtained by converting the time domain signals collected by a plurality of microphones into a time frequency domain signal, the background sound is suppressed and the estimated value of the target speech ^ S ^(j) _ (n , K) to extract a background sound,
m represents a microphone number, n represents a frame number, k represents a frequency bin number at high resolution, k￣ represents a frequency bin number at low resolution, and j represents a sound source number. As
The prior probability density function p (q ^(j) ) of the spectral parameters of each sound source signal and the low resolution spectral feature quantity model β￣_ (q ^(j) , n, k￣) (S) of each sound source signal are stored. A low-resolution spectral model storage unit;
A high-resolution spectral model storage unit storing high-resolution spectral feature models β_ (q ^(j) , n, k) (S) of each sound source signal;
A feature extraction unit for extracting a high-resolution sound source position feature quantity A_ (n, k) and a high-resolution spectral feature quantity X_ (n, k) from the observed signal x ^(m) _ (n, k);
A sound source position parameter φ ^(j) of each sound source signal is obtained from the high resolution sound source position feature A_ (n, k), and the high resolution sound source position feature A_ (n, k) and the sound source position parameter φ ^{(j ) To determine} the high-resolution sound source position occupancy Q ^(j) _ (n, k) of each sound source signal;
From the high-resolution spectral feature quantity X_ (n, k) and the high-resolution sound source position occupancy Q ^(j) _ (n, k), the low-resolution spectral feature quantity X￣_ ( n, k￣) and a low resolution sound source position occupancy Q￣ ^(j) _ (n, k￣),
The low-resolution spectral feature amount X￣_ (n, k￣), the low-resolution sound source position occupancy Q￣ ^(j) _ (n, k￣), and the prior probability density function p (q ^(j) ) From the low-resolution spectral feature model β￣_ (q ^(j) , n, k￣) (S), an estimated value ^ q ^{(j of} each sound source signal so as to maximize the log likelihood function ⁾ To obtain a low-resolution occupancy estimation unit;
The high-resolution spectral feature quantity X_ (n, k), the estimated spectral parameter q ^(j) , the high-resolution sound source position occupancy Q ^(j) _ (n, k), and the high-resolution spectral feature quantity model a high-resolution occupancy re-estimation unit for obtaining an estimated value of high-resolution occupancy ^ M ^(j) _ (n, k) from β_ (q ^(j) , n, k) (S);
Estimated value ^ q ^{(j) of the} spectral parameter, estimated value ^ M ^(j) _ (n, k) of the high resolution occupancy, the high resolution spectral feature quantity X_ (n, k), and the high resolution spectral feature A background sound characterized by comprising a target speech estimator that obtains an estimate value {circumflex over (S)} ^(j) _ (n, k) of the target speech from the quantity model β_ (q ^(j) , n, k) (S) Suppressor. The background sound suppression device according to claim 1,
The frequency resolution reduction unit uses the high resolution spectral feature value X_ (n, k) and a filter coefficient F_ (k￣) to perform the low resolution spectral feature value X￣_ (n, k￣) by a filter bank process. An average obtained by weighting the high-resolution sound source position occupancy Q ^(j) _ (n, k) with a function value having the filter coefficient F_ (k￣) and the high-resolution spectral feature X_ (n, k) as arguments. The background sound suppression device, wherein the low-resolution sound source position occupancy Q￣ ^(j) _ (n, k￣) is obtained by calculating a value. The background sound suppression device according to claim 1 or 2,
A high-resolution sound source position model storage unit storing a probability density function γ ^(j) _ (k) (A) of a high-resolution sound source position feature amount of each sound source signal;
The sound source position occupancy estimation unit calculates the high-resolution sound source position occupancy of each sound source signal from the high-resolution sound source position feature quantity A_ (n, k) and the probability density function γ ^(j) _ (k) (A). A background sound suppression apparatus characterized by obtaining a degree Q ^(j) _ (n, k). From the observation signal x (m) _ (n, k) obtained by converting the time domain signals collected by a plurality of microphones into a time frequency domain signal, the background sound is suppressed and the estimated value of the target speech ^ S ^(j) _ (n , K) to extract the background sound,
m represents a microphone number, n represents a frame number, k represents a frequency bin number at high resolution, k￣ represents a frequency bin number at low resolution, and j represents a sound source number. As
The low-resolution spectral model storage unit stores the prior probability density function p (q ^(j) ) of the spectral parameters of each sound source signal and the low-resolution spectral feature model β モ デ ル _ (q ^(j) , n, k￣ of each sound source signal. ) (S) is stored,
The high-resolution spectral model storage unit stores a high-resolution spectral feature model β_ (q ^(j) , n, k) (S) of each sound source signal,
A feature extraction step in which a feature extraction unit extracts a high-resolution sound source position feature quantity A_ (n, k) and a high-resolution spectral feature quantity X_ (n, k) from the observed signal x ^(m) _ (n, k). When,
A sound source position occupancy estimation unit obtains a sound source position parameter φ ^(j) of each sound source signal from the high resolution sound source position feature A_ (n, k), and the high resolution sound source position feature A_ (n, k). And a sound source position occupancy estimation step for obtaining a high resolution sound source position occupancy Q ^(j) _ (n, k) of each sound source signal from the sound source position parameter φ ^(j) ;
The frequency resolution reduction unit performs a smoothing process between neighboring frequencies from the high-resolution spectrum feature amount X_ (n, k) and the high-resolution sound source position occupancy Q ^(j) _ (n, k). A frequency resolution reduction step for obtaining a feature amount X￣_ (n, k￣) and a low-resolution sound source position occupancy Q￣ ^(j) _ (n, k￣);
The low-resolution occupancy estimation unit includes the low-resolution spectral feature amount X￣_ (n, k￣), the low-resolution sound source position occupancy Q￣ ^(j) _ (n, k￣), and the prior probability density function. From p (q ^(j) ) and the low-resolution spectral feature model β￣_ (q ^(j) , n, k￣) (S), the spectrum of each sound source signal is maximized so that the log-likelihood function is maximized. A low-resolution occupancy estimation step for obtaining an estimated value ^ q ^(j) of the parameter;
The high-resolution occupancy re-estimation unit performs the high-resolution spectral feature quantity X_ (n, k), the estimated value of the spectral parameter ^ q ^(j), and the high-resolution sound source position occupancy Q ^(j) _ (n, k ) And the high-resolution spectral feature model β_ (q ^(j) , n, k) (S), the high-resolution occupancy re-establishment for obtaining an estimated value of high resolution occupancy ^ M ^(j) _ (n, k) An estimation step;
The target speech estimator is configured to estimate the spectral parameter ^ q ^(j) , the high resolution occupancy ^ M ^(j) _ (n, k), and the high resolution spectral feature X_ (n, k). And a target speech estimation step for obtaining an estimated value of the target speech ^ S ^(j) _ (n, k) from the high-resolution spectral feature model β_ (q ^(j) , n, k) (S). The background sound suppression method characterized by this. The background sound suppression method according to claim 4,
In the frequency resolution reduction step, the low resolution spectral feature value X￣_ (n, k￣) is obtained by filter bank processing using the high resolution spectral feature value X_ (n, k) and a filter coefficient F_ (k￣). An average obtained by weighting the high-resolution sound source position occupancy Q ^(j) _ (n, k) with a function value having the filter coefficient F_ (k￣) and the high-resolution spectral feature X_ (n, k) as arguments. The background sound suppression method, wherein the low-resolution sound source position occupancy Q 度^(j) _ (n, k￣) is obtained by calculating a value. The background sound suppression method according to claim 4 or 5,
The high-resolution sound source position model storage unit stores a probability density function γ ^(j) _ (k) (A) of the high-resolution sound source position feature amount of each sound source signal,
In the sound source position occupancy estimation step, the high-resolution sound source position occupancy of each sound source signal is calculated from the high-resolution sound source position feature quantity A_ (n, k) and the probability density function γ ^(j) _ (k) (A). Determining the degree Q ^(j) _ (n, k), a background sound suppression method.   The program which makes a computer perform the background sound suppression method in any one of Claims 4-6.

Images