JP2009212599A - Method, device and program for removing reverberation, and recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve accuracy of a prediction filter coefficient when an observation signal is short. <P>SOLUTION: A method for removing reverberation is a method for estimating the prediction filter coefficient by using a probability model about the prediction filter coefficient for estimating a reverberation signal. A sound source model estimation section performs a step for estimating the model parameters of a sound source model which receives a time series observation signal and does not contain reverberation, an observation signal covariance estimation section performs a step for estimating the covariance matrix and covariance vector of the observation signal by receiving the model parameters and the observation signal, a prediction filter coefficient estimation section performs a step for estimating the prediction filter coefficient by receiving the covariance matrix and covariance vector of the observation signal and the probability density function of a prediction filter which probabilitistically modeled the factors of a prediction filter for predicting a reverberation signal contained in the observation signal, and a reverberation removing section performs a step for estimating a sound signal which does not contain reverberation by receiving the observation signal and the prediction filter. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  According to the present invention, an acoustic signal generated by a sound source (hereinafter referred to as “sound source signal”) is collected from a signal (hereinafter referred to as “observation signal”) obtained by collecting the sound signal with a microphone in a room with reverberation. The present invention relates to a dereverberation apparatus and method for extracting a removed acoustic signal, a program thereof, and a recording medium.

  When a sound source signal is collected in an environment with reverberation, it is observed as a signal in which reverberation is superimposed on the original sound source signal. For this reason, it becomes difficult to extract the nature of the original sound source signal, and the clarity of the sound source signal decreases. Therefore, a dereverberation method and apparatus for removing the superimposed reverberation have been conventionally used for the purpose of improving intelligibility.

An example of the functional configuration of a conventional dereverberation apparatus 900 disclosed in Non-Patent Document 1 is shown in FIG. The dereverberation apparatus 900 includes a sound source model 90, a prediction filter estimation unit 92, and a dereverberation unit 94. The sound source model 90 is obtained by modeling a speech waveform of a short time section of a sound source signal not including reverberation with a Gaussian distribution. The prediction filter estimation unit 92 receives the observation signal and the sound source model 90 as input, and obtains a prediction filter coefficient that predicts a reverberation signal that maximizes an optimization function that expresses the likelihood of the observation signal. The reverberation removing unit 94 removes the reverberation signal predicted by the prediction filter coefficient from the observed signal and outputs an acoustic signal.
Nakatani, T., Juang, BH, Hikichi, T., Yoshioka, T., Kinoshita, K.Delcroix, M., andMiyoshi, M., ”Study on speech dereverberation with autocorrelation codebook,” Proc.IEEE International Conference on Acoustics , Speech, and Signal Processing (ICASSP-2007), vol.I, pp.193-196, April 2007.

  In the conventional dereverberation method, a prediction filter coefficient for predicting a reverberation signal included in an observation signal is estimated only from the observation signal. Since this estimation requires an observation signal having a length of a certain length or more, if the observation signal is short, it is difficult to predict the reverberation signal with high accuracy, and accurate dereverberation cannot be performed.

  The present invention has been made in view of the above points, and even when the observation signal is short, a dereverberation removal method and apparatus capable of estimating the reverberation signal included in the observation signal with relatively high accuracy, and a program thereof. An object is to provide a recording medium.

  In the dereverberation method according to the present invention, a sound source model estimation unit estimates a model parameter of a sound source model that does not include reverberation using a time-series observation signal as an input, and an observation signal covariance estimation unit And observation signal covariance matrix and covariance vector estimation process, and prediction filter estimator is included in the observation signal covariance matrix and covariance vector A prediction filter coefficient estimation process that estimates a prediction filter coefficient using a prediction filter model that probabilistically models a prediction filter coefficient that predicts a reverberation signal as input, and a dereverberation unit that receives the observation signal and the prediction filter coefficient as input And a dereverberation process for estimating a speech signal not included.

  According to the dereverberation method of the present invention, the prediction filter coefficient is estimated using a probability model related to the prediction filter coefficient for estimating the reverberation signal, in addition to the method of estimating the prediction filter coefficient from the observation signal and the sound source model of the conventional method. Predictive filter coefficients that are more likely to be probabilistic can be estimated by using a probabilistic model for predictive filter coefficients, and even if the observed signal is short, the reverberation signal included in the observed signal can be estimated relatively accurately .

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

[Basic idea of the invention]
Prior to the description of the embodiments, the basic idea of the dereverberation method of the present invention will be described. In the dereverberation method of the present invention, the maximum likelihood estimation used in the conventional method is replaced with a maximum a posteriori (hereinafter referred to as “MAP”) estimation, and a probability model of a prediction filter coefficient required as a result is changed. It is something to consider. MAP estimation is a method for obtaining an estimated value by obtaining a value that maximizes the posterior probability density function of a target random variable (in the present invention, a prediction filter coefficient) under the condition that an observation signal is obtained. is there. Although the present invention can be composed of one or more microphones, in the following description, in order to simplify the description, the case of two microphones will be described as an example.

The signal name is defined as follows.

s _t ￣ is a vector corresponding to a short-time frame of length N of the target signal from which dereverberation has been removed. ￣ represents a vector, but the notation is correct as shown in the formula and the figure. x _t- ^(v) is a vector corresponding to the short-time frame of the v-th microphone signal. x _t ￣ is a vector that connects short-time frames of all microphone signals. X _t ^(v) is a matrix in which time series of x _t ^(v) ￣ are arranged. _Xt is a matrix in which time-series matrices for the microphones 1 and 2 are arranged. X _{t2: t1} is a matrix in which x _t並 べ is arranged retroactively from t ₂ to t _{1. An} observation signal is modeled in a multichannel autoregressive process as shown in Equation (1).

Expression (1) is a result of prediction by multiplying the past signal sequence X _t-D included in the right side by the prediction filter coefficient c ⁻ for the v-th microphone signal x _t- ^(v) included in the left side at the time t. This means that the prediction error becomes the target signal s _t ^━ .

Here, D is a delay added to the observation signal when predicting the observation signal x _t ^(v)の at time t. It has been reported that the introduction of D> 1 improves the robustness of dereverberation against the estimation error of the prediction coefficient (reference: K. Kinoshita, T. Nakatani, and M. Miyoshi, “Spectral subtraction steered by multi-step forward linear prediction for single channel speech dereverberation, "Proc.ICASSP-2006, vol.1, pp.817-820, May, 2006.). In the following description, the v-th microphone signal is treated as a prediction target signal. Other channels can be predicted in exactly the same way. From the equation (1), the target signal can be written by the equation (2). Therefore, it can be said that c￣ is a value having information equivalent to the inverse filter.

  Next, an optimization function is defined as shown in Expression (3), where a parameter to be estimated as an optimization function is a prediction filter coefficient c￣, and a parameter set including a speech model parameter and a prediction coefficient is θ.

Here, p _x (•), p _s (•), and p _c (•) represent probability density functions related to the observed signal x _t ^(v) ￣, the target signal s _t ￣, and the prediction filter coefficient _cそ れ ぞ れ, respectively. In the development of the above formula, logp _x (X _{T: 1} ; θ), which is a constant unrelated to optimization, is abbreviated. Expressions (4) and (5) are obtained by expanding Expression (3) based on the properties of a general probability density function. Expressions (6) and (7) expand Expression (5) based on Expression (2) and relate to a prediction filter coefficient when a signal other than c￣ is predicted (a signal other than the v-th microphone). Obtained by ignoring the term).

The optimization function of Equation (7) is completely defined if the probability density function p _s (s _t ￣; θ) of the target signal and the probability density function p _c ( _c ￣; θ) of the prediction filter coefficient are given. be able to. The first term of Equation (7) is a function equivalent to the optimization function of the conventional dereverberation method. The second term is a probability model of the prediction filter coefficient. In the present invention, reverberation removal with relatively high accuracy can be realized even when an observation signal having a sufficient length cannot be obtained by newly considering the second term.

  FIG. 1 shows a functional configuration example of a dereverberation apparatus 100 using the dereverberation method of the present invention as a first embodiment. The operation flow is shown in FIG. The dereverberation apparatus 100 includes a prediction filter model recording unit 10, a sound source model estimation unit 11, an observation signal covariance estimation unit 12, a prediction filter estimation unit 13, and a dereverberation unit 44. The dereverberation apparatus 100 replaces the sound source model 90 of the conventional dereverberation apparatus 900 with the sound source model estimation unit 11, adds a prediction filter model recording unit 10 and an observation signal covariance estimation unit 12, and provides a prediction filter estimation unit 13. The content of the processing in is changed. In addition, the dereverberation apparatus 100 of this example is realized by reading a predetermined program into a computer including, for example, a ROM, a RAM, a CPU, and the like, and executing the program by the CPU.

The sound source model estimation unit 11 receives the time series signal of the observation signal and estimates model parameters of the sound source model that does not include reverberation (step S11). In this method, for example, as in the case where the observation signal is subjected to the prewhitening process, an autoregressive coefficient relating to the observed signal is obtained and used as an approximate value of the autoregressive coefficient of the speech signal. Hereinafter, a sound source model, that is, a model of a target signal will be described as a probability density function p _s and a model parameter as an autocorrelation matrix r. Each is defined as shown in Equation (8) and Equation (9).

Here, a is an upper triangular Toeplitz matrix (N × N) defined by the equation (10) from the autoregressive coefficient α = [α ₁ α ₂ ... Α _p ] of the target signal s _t ￣.

From the autocorrelation matrix r and the above equation (8), the added term included in the first term of the equation (7) defining the optimization function can be expanded as shown in the equation (11).

  Here, equation (12) is used, and constant terms unrelated to optimization are omitted in the development of equation (11).

The observation signal covariance estimation unit 12 receives the autocorrelation matrix r and the observation signal x _t ^(v) ￣ as inputs, and obtains the observation signal covariance matrix Φ and the covariance vector φ based on the equations (13) and (14). Estimate (step S12).

  Here, the reason for obtaining the covariance matrix Φ and the covariance vector φ of the observation signal will be described. The above optimization function can be efficiently maximized using an expected value maximization method (hereinafter referred to as “EM”) algorithm based on the definition of the above probability density function. With the state i of the prediction coefficient as a hidden variable, the Q function in the EM algorithm is defined by Equation (15).

Here, p _c (•) is a mixed Gaussian distribution obtained by stochastically modeling the prediction filter coefficient for predicting the reverberation signal included in the observation signal, and is defined by Expressions (16) to (18).

Here, i is an integer (1 ≦ i ≦ K) representing the state of the prediction filter coefficient, and g _i represents the mixing ratio. The Gaussian distribution in each state can be expressed by equation (19).

The mixed Gaussian distribution is learned in advance in a specific room, and the model parameters {g ₁ , μ ₁ , Σ ₁ ,...} Are obtained in advance and recorded in the prediction filter model recording unit 10. . As a modification of equation (19), it is possible to increase the calculation efficiency by setting μ _i = 0 for all i or by setting the non-diagonal element of Σ _i to 0. The parameters in that case can also be determined in advance by using a general mixed Gaussian parameter learning algorithm. The learning method may be a general method and will not be described.

Here, the right side of the equation (15) represents a conditional expected value function with the prediction filter coefficient c ￣ ^“n” given. If only the terms related to the prediction filter coefficient c ￣ ^“n” are left and rearranged, the Q function can be expressed by Expression (20).

However,

  Here, the first term of the equation (21) is the observation signal covariance matrix Φ shown in the above equation (13). Further, the first term of the equation (22) is the observation signal covariance vector φ shown in the above equation (14). Therefore, assuming that the estimated values of the covariance matrix Φ and the covariance vector φ of the observation signal are given by the equations (13) and (14), in the E step of the EM algorithm, the equations (21) and (22) are used. It is only necessary to find the second term. This second term can be obtained by the following equations (23) to (25).

In the M step of the EM algorithm, the updated value of the prediction filter coefficient c is determined as the value of Expression (22) that maximizes Expression (20) (this corresponds to the expected value of the prediction coefficient). As described above, the prediction filter estimation unit 13 receives the covariance matrix Φ and the covariance vector φ of the observation signal and the model parameters {g ₁ , μ ₁ , Σ ₁ ,. Estimate habit.

The prediction filter estimation unit 13 includes an initial value setting unit 131, a posterior probability calculation unit 132, an expected value calculation unit 133, and a conditional expected value function calculation unit 134. The initial value setting unit 131 determines the initial value c￣ ^“0” of the prediction filter coefficient by, for example, multistep linear prediction shown in the reference (step S131). At this time, the repeated counter n is set to n = 0.

The posterior probability calculation unit 132 obtains the posterior probability of each state i with the prediction filter coefficient c￣ ^“n” given by the above equation (23) (step S132). The conditional expected value function calculation unit 133 calculates the conditional expected value using the above-described formula (24) and formula (25) (step S133). The expected value calculation unit 134 obtains an updated value of the expected value of the prediction filter coefficient by using the above formula (21) and formula (22) (step S134). If the updated value has not converged, the counter n is set to n = n + 1 (step S136), and the process returns to step S131.

The reverberation removing unit 44 removes the reverberation signal estimated from the observation signal based on the above equation (2) using the updated prediction filter coefficient c 係数^“n” (step S44). The operation of the dereverberation unit 44 is the same as in the conventional method.

  As described above, according to the dereverberation method of the present invention, it is possible to estimate a predictive filter coefficient that is more likely by using a probabilistic model related to a predictive filter coefficient. The reverberation signal included in the observation signal can be estimated with high accuracy.

[Modification]
In Example 1, the sound source model was defined on the assumption that the sound source follows a steady autoregressive process. On the other hand, by introducing a sound source model with higher accuracy, it is possible to realize dereverberation with higher accuracy. For example, a method of introducing a sound source model modeled by a finite state machine can be considered. FIG. 3 shows a functional configuration example of the dereverberation apparatus 300 according to the method, and FIG. 4 shows an operation flow thereof.

The dereverberation apparatus 300 selects an autocorrelation matrix r that most closely matches the observation signal of each short time frame t of the observation signal x _t ^(v) ￣. Therefore, a sound source model recording unit 30 that records a plurality of autocorrelation matrices is newly provided. The sound source model estimation unit 31 selects one from a plurality of autocorrelation matrices r recorded in the sound source model recording unit 30 with reference to the observation signal x _t ^(v) 、. Is different from the dereverberation apparatus 100 of the first embodiment in that it includes a convergence determination unit 321 that repeatedly operates from the selection of the autocorrelation matrix until the target signal s _tし た from which the derever has been removed converges.

The dereverberation apparatus 300 also uses the above equation (7) as an optimization function for dereverberation. In this example, only the definition of the portion related to the sound source model of the first term of Expression (7) is modified, and the same term is used for the second term.

Parameters of the speech models for the target signal s _t ¯ at each time t is the index of the autocorrelation codebook is denoted to as i _t. i _t, when the index m assumes a value of code words contained in the autocorrelation codebook 1 ≦ m ≦ M, take one of the values that any. A self-correlation matrix corresponding to each m written as r _m, the self-correlation matrix corresponding to the r _it to the index i _t of the autocorrelation codebook of time t. Further, the entire model parameter of the speech time series is assumed to be the entire time series I = {i ₁ , i ₂ ,..., I _T } of the index of the autocorrelation codebook.

The voice model at time t can be written by equation (26).

However, the parameter to be estimated by the dereverberation method is θ = {c￣, I}. With the above,
Expression (7) of the optimization function can be defined based on the above expressions (18), (19), and (26). In this example, the optimization function is maximized alternately with respect to the prediction filter coefficient c￣ and the entire time series I of the autocorrelation codebook index.

The sound source model estimation unit 31 sets the observation signal x _t ￣ ^(v) itself as the initial estimated value s _t ￣ ^[0] (step S31). At the same time, the repeat counter n _{1 is set} to n ₁ = 0. Then, by selecting the observed signal x _{t ^(v)} one of the autocorrelation matrix r _it from a plurality of autocorrelation matrix recorded in the sound source model recording unit 30 with reference to ^¯ defining a i _t by the equation (27).

The observation signal covariance estimation unit 12 receives the observation signal x _t Φ ^(v) and the autocorrelation matrix r _it and estimates the covariance matrix Φ and covariance vector φ of the observation signal. The process up to step S44 for removing the reverberation from the observed signal and estimating the target signal s _t ￣ is the same as that in the first embodiment. In this example, the convergence determination unit 321 in the dereverberation unit 32 repeatedly counts up the counter n ₁ (step S322) until the target signal s _t収束 converges (step S321), and the sound source model estimation unit 31. estimating the prediction filter coefficients c¯ change the autocorrelation matrix r _it in.

As described above, for example, by using a sound source model modeled by a finite state machine, a more accurate sound source model can be obtained, and as a result, highly accurate dereverberation can be realized. The dereverberation method described in the first embodiment and the modified example is a method based on the premise that all signals have been acquired in advance and can be batch-processed. Next, a dereverberation method according to the present invention for sequentially estimating the latest prediction filter coefficient for observed signals obtained sequentially will be described as a second embodiment.

FIG. 5 shows a functional configuration example of the dereverberation apparatus 500 that sequentially estimates the latest prediction filter coefficients, and FIG. 6 shows an operation flow thereof. The dereverberation apparatus 500 estimates / updates the prediction filter coefficient c で at predetermined time intervals. At the time of each update, the prediction filter coefficient c で is estimated by applying the above-described maximization algorithm to all or part of the observation signal obtained before that time, and sequentially at each time. This is a configuration in which the latest prediction filter coefficient c￣ obtained so far is applied to the observation signal at that time for the obtained observation signal.

The dereverberation apparatus 500 also includes an updating unit 50 that repeatedly operates the observation signal covariance estimation unit at predetermined time intervals and updates the prediction filter coefficient c￣. The observation signal covariance estimation unit includes the latest covariance matrix Φ _{n. −1} and the covariance vector φΦ _n−1 are different from the dereverberation apparatus 100 in that the observation signal covariance estimation unit 51 includes a covariance recording unit 511 that records the covariance vector φΦ _n−1 .

The operation of the dereverberation apparatus 500 until the first prediction filter coefficient c￣ is estimated is basically the same as that of the dereverberation apparatus 100, but the second and subsequent operations are repeated at predetermined time intervals by the update unit 50 ( Step S50) is different. In addition, the observation signal covariance estimation unit 51
When the covariance matrix Φ and covariance vector φ of the observed signal are estimated, the latest covariance matrix Φ and covariance vector φ are covariance as covariance matrix Φ _n-1 and covariance vector φ _n-1 The point of recording in the recording unit 511 is different. Further, the processing (step S131 ′) in which the initial value setting unit 131 ′ of the prediction filter estimation unit 13 ′ sets the initial value of the prediction filter coefficient is different only in the first time. The prediction filter coefficient c 係数 is updated at predetermined time intervals, but dereverberation is continuously performed with the latest prediction filter coefficient c￣.

In the dereverberation apparatus 500, the dereverberation process of the dereverberation unit 44 is performed by the prediction filter estimation unit 13.
The prediction filter estimation process of ′ is performed in parallel and asynchronously. As a result, the dereverberation unit 44 applies the prediction filter estimation unit 13 ′ to the observed signals input sequentially.
The dereverberation can be sequentially performed based on the latest prediction filter estimated value estimated by (without waiting for the completion of the next prediction filter update process of the prediction filter estimation unit). Note that the estimated value of the prediction filter is, for example, 0 at the time until the prediction filter estimation unit 13 ′ obtains the first estimated value. Alternatively, a value calculated based on an observation signal measured in advance may be used.

The observation signal covariance estimation unit 51 performs estimation of the covariance matrix Φ and the covariance vector φ using Expression (28) and Expression (29).

Here, T _i represents all the time indexes of the observation signal corresponding to the predetermined time interval before each update. α and β are forgetting factors, and are constants of 0 <α and β <1.

By doing so, dereverberation can be performed using the latest prediction filter coefficient obtained at each time. Next, Embodiment 3 of the present invention in which dereverberation is performed in the frequency domain.
Will be explained.

FIG. 7 shows a functional configuration example of a dereverberation apparatus 700 that performs dereverberation in the frequency domain. The dereverberation apparatus 700 is different from the dereverberation apparatuses 100 and 500 that perform the dereverberation in the time domain in that the dereverberation apparatus 700 includes a frequency domain dereverberation unit 70 that removes reverberation in the frequency domain.

The dereverberation process, which subtracts the energy of the reverberant signal from the energy of the observed signal by spectral subtraction, makes the estimation error of the prediction filter coefficient more robust against the difference in the sound source position, etc. Has been.

The dereverberation apparatus of the present invention also uses a power subtraction technique that obtains the predicted value e _t ￣ ^(v) of the reverberation signal from the observation signal and the prediction filter coefficient by Equation (30) and subtracts it from the short-time power spectrum of the observation signal. It is possible to remove dereverberation.

Frequency domain dereverberation unit 70, the observed signal x _t ¯ ^(v) and the respective predicted values e _t ¯ of the reverb signal ^(v), for example, a power spectrum using the common frequency conversion technology short time Fourier transform, etc. After conversion to, the dereverberation is performed between the power spectra, and can be configured by conventional techniques. For example, the observed signal x _t ￣ ^(v) , the predicted value of the reverberation signal e _t ￣ ^(v) , the target signal s _t ￣, and the short-time Fourier transform of each of them, X _{n, m} ^{to (V)} , En _{, m} ^{to (V)} , S _{n, m} ^˜ (where n and m are indices of time frames and frequency bins) and spectral subtraction can be expressed by equations (31) to (33).

Here, ε is a positive constant sufficiently smaller than 1. Furthermore, short-time Fourier transform S _n of the resulting target _signal, applying an inverse Fourier transform short time _m ^~, it is possible to obtain a time domain signal of the target signal by performing such overlap-add. As described above, according to the third embodiment, more accurate dereverberation can be performed.

〔simulation result〕
The performance of the dereverberation method of the present invention was evaluated. First, the estimation condition of the model parameter of the prior probability distribution of the prediction filter of the room to be dereverberation will be described. FIG. 8 schematically shows a room to be analyzed. The object of analysis was a reverberation chamber with a reverberation time of 0.5 seconds and a size of 3.5 m × 4.5 m × 2.5 m. An area of 1 square meter near the center of the room with a height of 1.5 m was equally divided by 30 × 30 points, and each point was set as a sound source position. Two microphones at the receiving point were prepared, and the distance between the microphones was 0.2 m. Two microphones Mic. 1 and Mic. The impulse response to 2 is obtained by mirror image method, the prediction filter coefficient of each transmission path is estimated by multi-channel multi-step linear prediction, and the model parameter {g ₁ , μ ₁ , Σ ₁ , ...} were estimated. The number of mixtures in the mixed Gaussian distribution was 4.

  The dereverberation of speech was performed using the obtained prior probability distribution of the prediction filter coefficients. The simulation conditions were as follows. Sound source positions were set at two locations indicated by speakers sp1 and sp2 in FIG. The speaker sp1 is located at the center in the learning range. The speaker sp2 is located outside the learning range. The order of multi-channel multi-step linear prediction was 2800, and step size D = 400. The sound source was an ATR audio database. The sampling frequency was 8 kHz and the number of quantization bits was 16. An impulse response from the sound source positions of the speakers sp1 and sp2 corresponding to a reverberation time of 0.5 seconds created by a mirror image method is convoluted with an audio signal that does not include reverberation, and is 0.5 to 1.0. Stereo observation signals with different lengths of 2.0 and 4.0 seconds were created. Using these observation signals of different lengths, the dereverberation method of the present invention and the microphone Mic. 1 compared the dereverberation performance of the observed signal.

The comparison was made by the cepstrum distortion of the original speech signal without reverberation and the dereverberation speech. Cepstrum distortion (CD) is defined by equation (34).

Here, c _k ^ and c _k are the cepstrum coefficients of the dereverberation voice and the original voice signal, respectively, and D ₀ and D ₁ are the dimensions of the cepstrum coefficients. In this simulation, cepstrum distortion was defined using cepstrum coefficients from the 0th order to the 12th order. The cepstrum coefficient for each dimension was obtained by subtracting the average value of the time series.

  FIG. 9 shows the time average value of the cepstrum distortion for each observation signal length and the time average value of the cepstrum distortion of the observation signal according to the conventional method and the method of the present invention. FIG. 9A shows the case where the sound source position is the speaker sp1, and FIG. Each horizontal axis represents the observation time length [second], and the vertical axis represents the time average value of cepstrum distortion in [dB].

  Compared with the time average value of the cepstrum distortion of the conventional method (plotted with ●), the method of the present invention (plotted with □) shows a characteristic with less distortion when the observation time length is within 2 seconds. In particular, when the observation time length is 0.5 seconds, the conventional method shows distortion that is worse than that of the observation signal, whereas the method of the present invention is large (2.3 to 3 dB) regardless of the location of the sound source. The time-averaged value of improved cepstrum distortion is shown.

  As described above, according to the dereverberation method of the present invention, when the observation time length that can be used for estimating the prediction filter coefficient is short, the dereverberation method is more robust against the difference in sound source position than the conventional method and can perform good dereverberation. I was able to confirm.

  In addition, the method and apparatus of this invention are not limited to the above-mentioned embodiment, In the range which does not deviate from the meaning of this invention, it can change suitably. Further, the processes described in the above method and apparatus are not only executed in time series according to the order of description, but also may be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. Good.

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

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, 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. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape or the like, and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only). Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording medium, MO (Magneto Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Further, the program may be distributed by storing the program in a recording device of a server computer and transferring the program from the server computer to another computer via a network.

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

The figure which shows the function structural example of the dereverberation apparatus 100 using the dereverberation method of this invention. The figure which shows the operation | movement flow of the dereverberation apparatus. The figure which shows the function structural example of the dereverberation apparatus 300 of this invention. The figure which shows the operation | movement flow of the dereverberation apparatus. The figure which shows the function structural example of the dereverberation apparatus 500 of this invention. The figure which shows the operation | movement flow of the dereverberation apparatus. The figure which shows the function structural example of the dereverberation apparatus 700 of this invention. The figure which shows typically the room made into the analysis object used for simulation. It is a figure which shows a simulation result, (a) shows the case where a sound source position is speaker sp1, (b) shows the case where a sound source position is speaker sp2. The figure which shows the function structural example of the conventional dereverberation apparatus 900 disclosed by the nonpatent literature 1. FIG.

Claims (8)

A sound source model estimation unit for estimating a model parameter of a sound source model that does not include reverberation using a time-series observation signal as an input;
An observation signal covariance estimation unit that estimates the covariance matrix and covariance vector of the observation signal by using the model parameter and the observation signal as inputs;
A prediction filter model recording unit for recording a prediction filter model that probabilistically models a prediction filter coefficient for predicting a reverberation signal included in the observation signal;
A prediction filter estimation unit that estimates a prediction filter coefficient using the covariance matrix and covariance vector of the observed signal as input, and the prediction filter model;
A dereverberation unit that estimates a speech signal that does not include reverberation using the observed signal and the prediction filter coefficient as inputs,
A dereverberation apparatus comprising: The dereverberation apparatus according to claim 1, wherein
An update unit that operates the observation signal covariance estimation unit at predetermined time intervals,
The observation signal covariance estimation unit updates the covariance matrix and the covariance vector at the predetermined time interval from the model parameters and the time series of the observation signals sequentially input,
The dereverberation apparatus, wherein the prediction filter estimation unit and the dereverberation unit operate corresponding to the updated covariance matrix and the covariance vector. In the dereverberation apparatus according to claim 1 or 2,
The dereverberation apparatus, wherein the dereverberation unit operates in a frequency domain. A sound source model estimation unit for estimating a model parameter of a sound source model that does not include reverberation using a time-series observation signal as an input;
An observation signal covariance estimation unit that estimates the covariance matrix and covariance vector of the observation signal by using the model parameter and the observation signal as inputs; and
The prediction filter estimator estimates the prediction filter coefficient using as input the covariance matrix and covariance vector of the observed signal and the prediction filter model that stochastically models the prediction filter coefficient that predicts the reverberation signal included in the observed signal Predictive filter estimation process,
A dereverberation unit, which receives the observation signal and the prediction filter coefficient as input and estimates a speech signal that does not include reverberation;
A dereverberation method comprising: In the dereverberation method according to claim 4,
An dereverberation method comprising: an update process in which the observed signal covariance estimation process, the prediction filter estimation process, and the dereverberation process are repeatedly performed at predetermined time intervals. In the dereverberation method according to claim 4 or 5,
The dereverberation method, wherein the dereverberation process operates in a frequency domain.   An apparatus program for causing a computer to function as the dereverberation apparatus according to claim 1.   A computer-readable recording medium on which any of the apparatus programs according to claim 7 is recorded.