JP5172536B2 - Reverberation removal apparatus, dereverberation method, computer program, and recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To flexibly remove reverberation even when reverberation environment is changed according to sound source movement and room temperature variation. <P>SOLUTION: A observation power time sequence creation section 1 creates an observation power time sequence. An initial setting section 2 sets an indoor impulse response estimation value and an original sound power estimation value time sequence. A reverberation power estimation value time sequence calculating section 3 calculates a reverberation power estimation value time sequence. An indoor impulse response updating section 4 updates the indoor impulse response estimation value. An original sound power estimation value time sequence updating section 5 updates the original sound power estimation value time sequence. A parameter normalizing section 6 normalizes the indoor impulse response estimation value, and normalizes the original sound power estimation value time sequence. A convergence determination section 7 determines whether or not, the indoor impulse response estimation value and the original sound power estimation value time sequence satisfy predetermined criteria. A parameter output section 8 outputs the indoor impulse response estimation value and the original sound power estimation value time sequence. <P>COPYRIGHT: (C)2010,JPO&amp;INPIT

Description

  The present invention relates to a dereverberation apparatus, a dereverberation method, a computer program, and a recording medium that remove a reverberation component from an acoustic signal.

  Various approaches have been proposed for removing reverberation components from an acoustic signal. For example, an approach that operates on an acoustic signal input to a single microphone is based on clean speech assumptions and models (such as harmonics, sparsity, autoregressive models, autocorrelation function codebooks) There is known one that estimates an inverse filter of a room impulse response so that the restored sound has as clean a sound as possible (for example, see Non-Patent Document 1). In general, the indoor impulse response may change remarkably every moment depending on the sound source position. Therefore, in the technique described in Non-Patent Document 1, how robustly the inverse filter can be estimated from a short observation signal. Is an important issue.

As a method for removing a reverberation component from an acoustic signal, a technique of performing inverse filtering on a power time envelope for each subband is also known (see, for example, Non-Patent Document 2 and Non-Patent Document 3). According to the techniques described in Non-Patent Document 2 and Non-Patent Document 3, it is the phase spectrum that changes significantly depending on the sound source position in the room impulse response, and the amplitude spectrum or the power spectrum is relatively unaffected. Based on a hypothesis. The convolution model of the power envelope can only be approximated, so it is expected that there will be a certain limit to the accuracy of dereverberation, but it is not possible to estimate the inverse filter of the room impulse response based on the clean speech quality. Compared to the technique described in Patent Document 1, there is a possibility that the operation may be more robust against changes in the sound source position and the like.
T.A. Nakatani, B .; Jung, T .; Hikichi, T .; Yoshioka, K .; Kinoshita, M .; Delcroix, M .; Miyoshi, “Study on Speech Develberation with Automation Codebook,” in Proc. ICASSP 2007, pp. 193-197, 2007. Shigeki Hirobayashi, Hiroaki Nomura, Tsunehiko Koike, Mikio Higashiyama, “Recovering reverberant speech by inverse filtering of power envelope transfer function,” IEICE Transactions, Vol. J81-A, no. 10, pp. 1323-1330, 1998. M.M. Unoki, M .; Furukawa, K .; Sakata, M .; Akagi, “A Method Based on the MTF Concept for Developing the Power Environment from the Reverse Signal,” in Proc. ICASSP 2003, Vol. 1, pp. 888-891, 2003.

  In the techniques of Non-Patent Document 2 and Non-Patent Document 3, the power envelope of the reverberation component is modeled by a parametric function, and in the technique of Non-Patent Document 3, the parameter is estimated based on a scale called modulation degree. ing. However, in an actual environment, since the reverberation environment changes as the sound source moves or changes in room temperature, it is extremely rare for an actual reverberation component to ideally follow these function classes. Therefore, these techniques have a problem that it is not always guaranteed that the reverberation is satisfactorily removed.

  The present invention has been made in view of such circumstances, and a dereverberation apparatus and a dereverberation method capable of removing reverberation while flexibly responding to changes in a reverberation environment accompanying movement of a sound source or changes in room temperature. An object of the present invention is to provide a computer program and a recording medium.

  The present invention receives an input of an acoustic signal and generates an observation power time series that is a time series of the amplitude or power of a subband signal for each frequency channel by short-time frequency analysis, and for each frequency channel An initial setting unit that sets a room impulse response estimated value having a non-negative constraint and an original sound power estimated time series that is a power estimated value time series for each frequency channel of the original sound, the indoor impulse response estimated value, and the original sound A reverberant sound power estimated value time series calculating unit that calculates a reverberant sound power estimated value time series that is a power time series of a reverberant sound model for each frequency channel, and the observed power time series, Based on the reverberant power estimate time series, the indoor impulse response estimate, and the original sound power estimate time series, non-negative constraints are set. Thus, the room impulse response update unit for updating the room impulse response estimated value, the observation power time series, the reverberation sound power estimated time series, the room impulse response estimated value, and the original sound power estimated value An original sound power estimated value time series updating unit that updates the original sound power estimated value time series while satisfying a non-negative constraint, and the indoor impulse response estimated value updated by the indoor impulse response updating unit. The original sound power estimated value time series normalized by the sum of the element values of the impulse response estimated values to be a constant value and updated by the original sound power estimated value time series update unit is the element value of the original sound power estimated value time series. A parameter normalization unit that normalizes so that the sum of the values becomes a constant value, the indoor impulse response estimated value normalized by the parameter normalization unit, and the original sound A convergence determination unit for determining whether or not the time estimation value time series satisfies a predetermined criterion, and the convergence determination unit, the indoor impulse response estimation value and the original sound power normalized by the parameter normalization unit A parameter output unit that outputs the room impulse response estimated value and the original sound power estimated value time series when it is determined that the estimated time series satisfies a predetermined criterion, the convergence determining unit When the room impulse response estimated value and the original sound power estimated value time series normalized by the parameter normalization unit do not satisfy a predetermined criterion, the room normalized by the parameter normalization unit Based on the impulse response estimated value and the original sound power estimated value time series, the reverberant power estimated value time series calculating unit calculates the reverberant power estimated value time series, and An inner impulse response update unit updates the indoor impulse response estimated value, the original sound power estimated value time series update unit updates the original sound power estimated value time series, and the original sound power estimated value time series update unit updates the original sound power. An dereverberation apparatus that updates an estimated time series.

Further, in the dereverberation apparatus of the present invention, the indoor impulse response update unit includes, for each frequency channel, a correlation function between the observed sound and the estimated original sound, which is a correlation function between the observed power time series and the original sound power estimated value time series. and the observed sound-estimated original correlation function calculation section for calculating a, for each frequency channel, between the estimated reverberation Probable original sound is a correlation function of the reverberation power estimate time series and the original sound power estimate time series An estimated reverberation sound / estimated original sound correlation function calculation unit for calculating a correlation function, and a time series element value of the observed sound / estimated original sound correlation function for each frequency channel, the estimated reverberant sound / estimated original sound correlation function An indoor impulse response estimated value update coefficient calculating unit that calculates an indoor impulse response estimated value update coefficient that is a value divided by a time-series element value, and for each frequency channel, the indoor impulse Integrating the response estimation value and the room impulse response estimate update coefficient for each element, and the room impulse response estimate update value output unit that calculates the room impulse response estimate update value, comprising: a.

Further, in the dereverberation apparatus of the present invention, the room impulse response update unit includes, for each frequency channel, a time series obtained by dividing the observation power time series by the reverberation sound power estimated value time series, and the original sound power. A modeling error ratio sequence / estimated original sound correlation function calculating unit that calculates a correlation function between a modeled error ratio sequence / estimated original sound, which is a correlation function with an estimated value time series, and for each frequency channel, Estimated original sound partial sum series calculation unit that calculates an estimated original sound partial sum series that is a series with element values of the partial sum of element values of each specific range of the sequence, and between the modeling error ratio sequence and the estimated original sound for each frequency channel the correlation function, and the room impulse response estimate update coefficient calculation unit for calculating a room impulse response estimate update coefficient is a value obtained by dividing the element values of the estimated original partial sum sequence, frequency An indoor impulse response estimated value update value output unit that calculates the indoor impulse response estimated value update value by integrating the indoor impulse response estimated value and the indoor impulse response estimated value update coefficient for each element for each channel. It is characterized by that.

Further, in the dereverberation apparatus of the present invention, the original sound power estimated value time-series updating unit is an observation sound / estimated impulse response which is a correlation function between the observed power time series and the indoor impulse response estimated value for each frequency channel. A correlation function between the observed sound / estimated impulse response for calculating an inter-correlation function, and a correlation function between the reverberation power estimate time series and the indoor impulse response estimate for each frequency channel, and the correlation function And a sparse correction term for calculating a correlation function between the estimated reverberant sound and the estimated impulse response with a sparse correction term, which is a time series obtained by adding the time series obtained by multiplying the original sound power estimated value time series by a constant for each element and further multiplying by a constant and regarding the estimated reverberation-estimated impulse response correlation function calculation unit, for each frequency channel, when the observed sound-estimated impulse response correlation function Original sound power estimated value time series update for calculating an original sound power estimated value time series update coefficient, which is a value obtained by dividing the elements of the column by the time series element value of the estimated reverberant sound / estimated impulse response correlation function with the sparse correction term The original sound power estimated value for calculating the original sound power estimated value time series update value by adding the original sound power estimated value time series and the original sound power estimated value time series update coefficient for each element for the coefficient calculating unit and each frequency channel And a time-series update value output unit.

  Further, in the dereverberation apparatus of the present invention, the original sound power estimated value time series update unit, for each frequency channel, a time series obtained by dividing the observed power time series by the reverberant power estimated value time series, and A modeled error ratio sequence / estimated impulse response correlation function calculating unit that calculates a correlation function between a modeled error ratio sequence / estimated impulse response, which is a correlation function with the indoor impulse response estimated value, and for each frequency channel, A series having an element value as a partial sum of the element values of each specific range of the impulse response estimated value is calculated, and the series and a time series obtained by multiplying the original sound power estimated value time series by a constant and multiplying by a constant An estimated impulse response partial sum sequence calculation unit with a sparse correction term that calculates an estimated impulse response partial sum sequence with a sparse correction term that is an added time series; For each channel, the original sound power estimated value time series is a value obtained by dividing the time series elements of the modeled error ratio series / estimated impulse response correlation function by the element values of the estimated impulse response partial sum series with sparse correction terms. An original sound power estimated value time series update coefficient calculation unit for calculating an update coefficient, and for each frequency channel, the original sound power estimated value time series and the original sound power estimated value time series update coefficient are integrated element by element to estimate the original sound power An original sound power estimated value time series update value output unit for calculating a value time series update value.

  The present invention also provides an observation power time series generation unit that receives an input of an acoustic signal and generates an observation power time series that is a time series of amplitude or power of a subband signal for each frequency channel by short-time frequency analysis. The power time series generation step and the initial setting unit set a room impulse response estimation value having a non-negative constraint for each frequency channel and an original sound power estimation value time series that is a power estimation time series for each frequency channel of the original sound. A reverberation sound that is a power time series of a reverberation sound model for each frequency channel by initial setting step and a reverberation sound power estimate time series calculation unit convolves the room impulse response estimation value and the original sound power estimation value time series. A reverberation sound power estimate time series calculating step for calculating a power estimate time series, and an indoor impulse response updating unit include the observation power Based on a time series, the reverberation sound power estimated value time series, the room impulse response estimated value, and the original sound power estimated value time series, a room impulse that satisfies the non-negative constraint and updates the room impulse response estimated value A response update step; and an original sound power estimated value time series update unit based on the observed power time series, the reverberant sound power estimated value time series, the indoor impulse response estimated value, and the original sound power estimated value time series. An original sound power estimated value time series updating step that updates the original sound power estimated value time series satisfying a non-negative constraint, and a parameter normalization unit that updates the indoor impulse response estimated value updated in the indoor impulse response updating step, Normalizing the sum of the element values of the indoor impulse response estimated value to be a constant value, the original sound power estimated value time series update step A parameter normalization step for normalizing the updated original sound power estimated value time series so that the sum of the element values of the original sound power estimated value time series becomes a constant value, and a convergence determining unit in the parameter normalizing step In the convergence determination step for determining whether the normalized indoor impulse response estimated value and the original sound power estimated value time series satisfy a predetermined criterion, and in the convergence determination step, the parameter normalization unit defines a standard When it is determined that the converted indoor impulse response estimated value and the original sound power estimated time series satisfy a predetermined criterion, the parameter output unit outputs the indoor impulse response estimated value and the original sound power estimated time series A parameter output step for outputting the indoor impulse normalized by the parameter normalization unit in the convergence determination step. When it is determined that the response response estimated value and the original sound power estimated value time series do not satisfy a predetermined criterion, the room impulse response estimated value and the original sound power estimated value time series normalized in the parameter normalizing step Based on the above, the reverberant sound power estimated value time series calculating step calculates the reverberant sound power estimated value time series, the indoor impulse response updated step updates the indoor impulse response estimated value, and the original sound power estimated value In the dereverberation method, the original sound power estimated value time series is updated in a time series updating step, and the original sound power estimated value time series is updated in the original sound power estimated value time series updating step.

Further, in the dereverberation method of the present invention, the indoor impulse response update step includes a step in which the observed sound / estimated original sound correlation function calculating unit calculates the observed power time series and the original sound power estimated value time series for each frequency channel. An observed sound / estimated original sound correlation function calculating step for calculating a correlation function between the observed sound / estimated original sound, and an estimated reverberant sound / estimated original sound correlation function calculating section for each frequency channel, value time series and the estimated reverberation Probable original correlation function calculation step of calculating the estimated reverberation Probable original correlation function is a correlation function of the original sound power estimate time series, the room impulse response estimate update coefficient For each frequency channel, the calculation unit divides the time-series element value of the correlation function between the observed sound and the estimated original sound by the time-series element value of the estimated reverberant sound / estimated original sound correlation function. An indoor impulse response estimated value update coefficient calculating step for calculating an indoor impulse response estimated value update coefficient, and an indoor impulse response estimated value update value output unit for each frequency channel. And an indoor impulse response estimated value update value output step of calculating an impulse response estimated value update coefficient for each element and calculating an indoor impulse response estimated value update value.

Further, in the dereverberation method of the present invention, the indoor impulse response update step includes a modeled error ratio sequence / estimated original sound correlation function calculation unit that converts the observed power time series into the reverberant sound power estimated value for each frequency channel. Modeling error ratio sequence / estimated original sound correlation function for calculating a model error ratio sequence / estimated original sound correlation function that is a correlation function between the time series divided for each element in the time series and the original sound power estimated value time series A calculation step and an estimated original sound partial sum series calculation unit calculate an estimated original sound partial sum series that is a series in which the element value is a partial sum of element values of each specific range of the original sound power estimated value time series for each frequency channel. the estimated original partial sum sequence calculation step of, room impulse response estimate update coefficient calculation unit, for each frequency channel, the modeling error ratio sequence-estimation original correlation function, Serial estimation and room impulse response estimate update coefficient calculating a room impulse response estimate update coefficient is a value obtained by dividing the element values of the original partial sum sequence, the room impulse response estimate update value output unit, each frequency channel The indoor impulse response estimated value and the indoor impulse response estimated value update coefficient are integrated element by element, and an indoor impulse response estimated value update value output step of calculating an indoor impulse response estimated value update value is included. Features.

Further, in the dereverberation method of the present invention, the original sound power estimated value time series update step includes: the observed sound time / estimated impulse response correlation function calculation unit, for each frequency channel, the observed power time series and the indoor impulse response estimation. A correlation function calculation step between the observed sound and estimated impulse response that calculates a correlation function between the observed sound and estimated impulse response, which is a correlation function with the value, and a correlation function calculation unit between the estimated reverberant sound and estimated impulse response with a sparse correction term, For each frequency channel, a correlation function between the reverberation sound power estimated time series and the room impulse response estimated value is calculated, and the correlation function and the original sound power estimated time series are multiplied by a constant for each element and further multiplied by a constant. To calculate the correlation function between the estimated reverberant sound with the sparse correction term and the estimated impulse response. Correlation function calculation step between the positive section with estimated reverberation-estimated impulse response, when the original sound power estimate sequence updating coefficient calculating unit, for each frequency channel, the elements of the time series of the observed sound-estimated impulse response correlation function An original sound power estimated value time series update coefficient calculating step for calculating an original sound power estimated value time series update coefficient that is a value divided by a time series element value of the correlation function between estimated reverberant sound / estimated impulse response with sparse correction term; The original sound power estimate time series update value output unit integrates the original sound power estimate time series and the original sound power estimate time series update coefficient element by element for each frequency channel to update the original sound power estimate time series An original sound power estimated value time series update value output step for calculating a value.

  Further, in the dereverberation method of the present invention, in the original sound power estimated value time series update step, the modeled error ratio series / estimated impulse response correlation function calculation unit calculates the reverberation of the observed power time series for each frequency channel. Modeling error ratio sequence / estimation for calculating a correlation function between a time series obtained by dividing the sound power estimation value for each element by a time series and the indoor impulse response estimation value and a correlation function between estimated impulse responses A correlation function calculation step between impulse responses and an estimated impulse response partial sum series calculation unit with a sparse correction term for each frequency channel, wherein the partial sum of element values of each specific range of the indoor impulse response estimation value is an element value And the time series obtained by adding the time series obtained by multiplying the original sound power estimated value time series by a constant power for each element and further multiplying by a constant An estimated impulse response partial sum sequence calculation step with a sparse correction term that calculates an estimated impulse response partial sum sequence with a sparse correction term, and an original sound power estimate time series update coefficient calculation unit, for each frequency channel, the modeling error ratio The original sound power estimated value for calculating the original sound power estimated value time series update coefficient, which is a value obtained by dividing the time series element of the correlation function between the sequence and estimated impulse response by the element value of the estimated impulse response partial sum series with sparse correction term A time series update coefficient calculating step, and an original sound power estimated value time series update value output unit, for each frequency channel, integrates the original sound power estimated value time series and the original sound power estimated value time series update coefficient element by element, An original sound power estimated value time series update value output step for calculating an original sound power estimated value time series update value;

  The present invention is also a computer program for operating a computer as a dereverberation device.

  The present invention is also a computer-readable recording medium that records a computer program for operating a computer as a dereverberation device.

  According to the present invention, it is possible to remove reverberation while flexibly responding to changes in the reverberation environment accompanying movement of a sound source, changes in room temperature, and the like. Moreover, even when there is no change in the reverberation environment, the reverberation can be removed as in the conventional case. Also, the calculation of dereverberation can be performed at high speed.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In the present embodiment, it is assumed that constants λ, p, C _G , and C _S are determined in advance. Also, let the time index be (Equation 1). Further, the frequency index is represented by (Equation 2). Further, the power spectrum time series of the original signal (hereinafter referred to as the original sound power estimated value time series) is S (ω, t). Further, the spectrogram of the impulse response of the indoor transmission system (hereinafter referred to as an estimated value of the indoor impulse response) is assumed to be G (ω, t). (Equation 3) represents a set of whole integers.

  FIG. 1 is a functional block diagram of the dereverberation apparatus according to the present embodiment. The dereverberation apparatus shown in the figure includes an observation power time series generation unit 1, an initial setting unit 2, a reverberation sound power estimation value time series calculation unit 3, an indoor impulse response update unit 4, and an original sound power estimation value time series. The update unit 5, the parameter normalization unit 6, the convergence determination unit 7, and the parameter output unit 8 are provided. The dereverberation apparatus receives an input of an acoustic signal from a microphone or the like (not shown), and calculates an indoor impulse response estimated value G (ω, t) and an original sound power estimated value time series S (ω, t) from the acoustic signal. And output.

  First, the observation power time series generation unit 1 receives an input of an acoustic signal and outputs a time frequency component of a power spectrum of the input acoustic signal. This time frequency component is represented as Y (ω, t). Y (ω, t) is called an observation power time series. Let ω = 1,..., Ω be an index corresponding to the frequency. Further, t = 1,..., T is an index corresponding to the frequency. The observation power time series generation unit 1 calculates a time frequency component Y (ω, t) by time frequency decomposition means using a filter bank output of a plurality of channels such as short-time Fourier transform and wavelet transform.

  Subsequently, the initial setting unit 2 sets and outputs initial values of the original sound power estimated value time series S (ω, t) and the indoor impulse response estimated value G (ω, t). These values may be set by random numbers, but the initial value of the original sound power estimated value time series S (ω, t) is equal to Y (ω, t) output by the observation power time series generation unit 1. It is preferable to set as follows. The initial value of the indoor impulse response estimated value G (ω, t) is preferably set to take a maximum value at t = 1, such as an exponential function, and become smaller as t increases.

  Subsequently, the reverberant sound power estimated value time series calculation unit 3 and the original sound power estimated value time series S (ω, t) output from the initial setting unit 2 or the parameter normalization unit 6 described later and the indoor impulse response estimated value G ( The input of ω, t) is accepted. Subsequently, the reverberant sound power estimated value time series calculation unit 3 calculates and outputs a reverberant sound power estimated value time series X (ω, t) by convolution using (Equation 4).

  Here, the discrete Fourier transform F ′ (ω, k) for the series F (ω, 1), F (ω, 2),... Is expressed as (Equation 5). Also, the discrete inverse Fourier transform F (ω, t) for the series F ′ (ω, 1), F ′ (ω, 2),. Thereby, it can be said that the reverberation sound power estimated value time series X (ω, t) is (Expression 7). Therefore, the reverberant sound power estimated time series X (ω, t) can be calculated at high speed by FFT (Fast Fourier Transform).

In addition, (Equation 8) and (Equation 9) represent those in which S (ω, t) and G (ω, t) are appropriately zeroed for the purpose of reducing the influence of cyclic convolution. Specifically, (Equation 10) and (Equation 11). However, _Mω is the number of time indexes in (Expression 8) and (Expression 9).

  Subsequently, the indoor impulse response updating unit 4 outputs the observation power time series Y (ω, t) output from the observation power time series generation unit 1 and the original sound power output from the initial setting unit 2 or the parameter normalization unit 6 described later. Estimated value time series S (ω, t), indoor impulse response estimated value G (ω, t), and reverberant sound power estimated value time series calculation unit 3 output reverberant sound power estimated value time series X (ω, t) Is accepted. Subsequently, the indoor impulse response update unit 4 updates and outputs the indoor impulse response estimated value G (ω, t). A specific method for updating the indoor impulse response estimated value G (ω, t) will be described with reference to FIGS.

  FIG. 2 and FIG. 3 are diagrams showing configuration examples of functional blocks included in the indoor impulse response update unit 4. First, the configuration example shown in FIG. 2 will be described. In the example shown in FIG. 2, the indoor impulse response updating unit 4 includes an observed sound / estimated original sound correlation function calculating unit 41, an estimated reverberant sound / estimated original sound correlation function calculating unit 42, and an indoor impulse response estimated value update coefficient. A calculation unit 43 and an indoor impulse response estimated value update value output unit 44 are provided.

The observed sound / estimated original sound correlation function calculation unit 41 receives input of the observed power time series Y (ω, t) and the original sound power estimated value time series S (ω, t). Subsequently, the observed sound / estimated original sound correlation function calculation unit 41 calculates (Equation 12) and outputs the observed sound / estimated original sound correlation function R _SY (ω, τ).

Subsequently, the estimated reverberant sound / estimated original sound correlation function calculation unit 42 receives input of the original sound power estimated value time series S (ω, t) and the reverberant power estimated value time series X (ω, t). Subsequently, the estimated reverberant sound / estimated original sound correlation function calculation unit 42 calculates (Equation 13) and outputs the estimated reverberant sound / estimated original sound correlation function R _SX (ω, τ).

Subsequently, the indoor impulse response estimated value update coefficient calculating unit 43 outputs the observed sound / estimated original sound correlation function R _SY (ω, τ) output from the observed sound / estimated original sound correlation function calculating unit 41 and the estimated reverberant sound / The input of the estimated reverberant sound / estimated original sound correlation function R _SX (ω, τ) output from the estimated original sound correlation function calculation unit 42 is received. Subsequently, the indoor impulse response estimated value update coefficient calculation unit 43 performs calculation of (Equation 14) and outputs an indoor impulse response estimated value update coefficient α _G (ω, t).

Subsequently, the indoor impulse response estimated value update value output unit 44 includes the indoor impulse response estimated value G (ω, t) and the indoor impulse response estimated value update coefficient α _G output from the indoor impulse response estimated value update coefficient calculation unit 43. The input of (ω, t) is accepted. Subsequently, the indoor impulse response estimated value update value output unit 44 calculates (Equation 15) and outputs the updated indoor impulse response estimated value G (ω, t). However, “←” means substitution.

  Next, the configuration example shown in FIG. 3 will be described. In the example shown in FIG. 3, the indoor impulse response update unit 4 includes a modeling error ratio sequence / estimated original sound correlation function calculation unit 45, an estimated original sound partial sum sequence calculation unit 46, and an indoor impulse response estimated value update coefficient calculation. Unit 47 and an indoor impulse response estimated value update value output unit 48.

The modeling error ratio sequence / estimated original sound correlation function calculation unit 45 includes an observed power time series Y (ω, t), an original sound power estimated value time series S (ω, t), and a reverberant sound power estimated value time series X. The input of (ω, t) is accepted. Subsequently, the modeling error ratio sequence / estimated original sound correlation function calculation unit 45 calculates (Equation 16) and outputs the modeling error ratio sequence / estimated original sound correlation function L _{SY / X} (ω, τ). .

Subsequently, the estimated original sound partial sum series calculation unit 46 receives an input of the original sound power estimated value time series S (ω, t). Subsequently, the estimated original sound partial sum series calculation unit 46 performs the calculation of (Equation 17) and outputs the estimated original sound partial sum series L _S (ω, τ).

Subsequently, the indoor impulse response estimated value update coefficient calculation unit 47 outputs the modeling error ratio sequence / estimated original sound correlation function L _{SY / X} (ω, τ) and the estimated original sound partial sum sequence L _S (ω, τ) output from the estimated original sound partial sum sequence calculation unit 46 are received. Subsequently, the indoor impulse response estimated value update coefficient calculation unit 47 calculates (Equation 18) and outputs an indoor impulse response estimated value update coefficient β _G (ω, t).

Subsequently, the indoor impulse response estimated value update value output unit 48 outputs the indoor impulse response estimated value G (ω, t) and the indoor impulse response estimated value update coefficient β _G output from the indoor impulse response estimated value update coefficient calculation unit 47. The input of (ω, t) is accepted. Subsequently, the indoor impulse response estimated value update value output unit 48 calculates (Equation 19), and outputs the updated indoor impulse response estimated value G (ω, t). However, “←” means substitution.

  As described with reference to FIGS. 2 and 3, the indoor impulse response update unit 4 updates and outputs the indoor impulse response estimated value G (ω, t).

  Returning to the description of FIG. Subsequently, the original sound power estimated value time series update unit 5 outputs the observation power time series Y (ω, t) output from the observation power time series generation unit 1 and the initial setting unit 2 or the parameter normalization unit 6 described later. Original sound power estimated value time series S (ω, t), indoor impulse response estimated value G (ω, t), and reverberant sound power estimated value time series calculation unit 3 output reverberant sound power estimated value time series X (ω , T). Subsequently, the original sound power estimated value time series update unit 5 updates and outputs the original sound power estimated value time series S (ω, t). A specific method of updating the original sound power estimated value time series S (ω, t) will be described with reference to FIGS. 4 and 5.

  FIG. 4 and FIG. 5 are diagrams showing configuration examples of functional blocks included in the original sound power estimated value time-series updating unit 5. First, the configuration example shown in FIG. 4 will be described. In the example shown in FIG. 4, the original sound power estimated value time-series updating unit 5 includes an observed sound / estimated impulse response correlation function calculating unit 51 and an estimated reverberant sound / estimated impulse response correlation function calculating unit 52 with a sparse correction term. And an original sound power estimated value time series update coefficient calculating unit 53 and an original sound power estimated value time series update value output unit 54.

The observed sound / estimated impulse response correlation function calculation unit 51 accepts inputs of the observed power time series Y (ω, t) and the indoor impulse response estimated value G (ω, t). Subsequently, the observed sound / estimated impulse response correlation function calculation unit 51 calculates (Equation 20), and outputs the observed sound / estimated impulse response correlation function R _GY (ω, τ).

Subsequently, the correlation function calculation unit 52 between the estimated reverberation sound / estimated impulse response with sparse correction term, the original sound power estimated value time series S (ω, t), the indoor impulse response estimated value G (ω, t), and the reverberation. The input of the sound power estimated value time series X (ω, t) is received. Subsequently, the estimated reverberant sound / estimated impulse response correlation function calculation unit 52 with the sparse correction term performs the calculation of (Equation 21), and the estimated reverberant sound / estimated impulse response correlation function R _GX (ω, τ) with the sparse correction term. ) Is output.

Subsequently, the original sound power estimated value time series update coefficient calculation unit 53 calculates the correlation function R _GY (ω, τ) between the observed sound and the estimated impulse response output from the correlation function calculation unit 51 between the observed sound and the estimated impulse response, and the sparse. An input of the estimated reverberant sound / estimated impulse response correlation function R _GX (ω, τ) with a sparse correction term output from the estimated reverberant sound / corrected impulse response correlation function calculation unit 52 with a correction term is received. Subsequently, the original sound power estimated value time series update coefficient calculation unit 53 calculates (Equation 22) and outputs the original sound power estimated value time series update coefficient α _S (ω, t).

Subsequently, the original sound power estimated value time series update value output unit 54 outputs the original sound power estimated value time series S (ω, t) and the original sound power estimated value time series update coefficient calculation unit 53. An input with the update coefficient α _S (ω, t) is received. Subsequently, the original sound power estimated value time series update value output unit 54 calculates (Equation 23) and outputs the updated original sound power estimated value time series S (ω, t). However, “←” means substitution.

  Next, the configuration example shown in FIG. 5 will be described. In the example shown in FIG. 5, the original sound power estimated value time series update unit 5 includes a modeling error ratio sequence / estimated impulse response correlation function calculation unit 55, an estimated impulse response partial sum series calculation unit 56 with a sparse correction term, The original sound power estimated value time series update coefficient calculating unit 57 and the original sound power estimated value time series update value output unit 58 are provided.

The modeling error ratio sequence / estimated impulse response correlation function calculation unit 55 includes an observation power time series Y (ω, t), an indoor impulse response estimated value G (ω, t), and a reverberant power estimated value time series X. The input of (ω, t) is accepted. Subsequently, the modeling error ratio sequence / estimated impulse response correlation function calculation unit 55 calculates (Equation 24), and calculates the modeling error ratio sequence / estimated impulse response correlation function L _{GY / X} (ω, τ). Output.

Subsequently, the estimated impulse response partial sum series calculation unit 56 with a sparse correction term receives inputs of the original sound power estimated value time series S (ω, t) and the indoor impulse response estimated value G (ω, t). Subsequently, the estimated impulse response partial sum series calculation unit 56 with a sparse correction term performs the calculation of (Equation 25), and outputs an estimated impulse response partial sum series L _G (ω, τ) with a sparse correction term.

Subsequently, the original sound power estimated value time-series update coefficient calculation unit 57 performs the modeling error ratio sequence / estimated impulse response correlation function calculation unit 55 output by the modeling error ratio sequence / estimated impulse response correlation function L _{GY / X.} The input of (ω, τ) and the estimated impulse response partial sum series L _G (ω, τ) with sparse correction term output from the estimated impulse response partial sum series calculation unit 56 with sparse correction term is received. Subsequently, the original sound power estimated value time series update coefficient calculation unit 57 calculates (Equation 26) and outputs the original sound power estimated value time series update coefficient β _S (ω, t).

Subsequently, the original sound power estimated value time series update value output unit 58 outputs the original sound power estimated value time series S (ω, t) and the original sound power estimated value time series update coefficient calculating unit 57. The input of the update coefficient β _S (ω, t) is received. Subsequently, the original sound power estimated value time series update value output unit 58 performs calculation of (Equation 27), and outputs the updated original sound power estimated value time series S (ω, t). However, “←” means substitution.

  As described with reference to FIGS. 4 and 5, the original sound power estimated value time series update unit 5 updates and outputs the original sound power estimated value time series S (ω, t).

  Returning to the description of FIG. Subsequently, the parameter normalization unit 6 outputs the room impulse response estimated value G (ω, t) output from the room impulse response update unit 4 and the original sound power estimated value time series output from the original sound power estimated value time series update unit 5. An input with S (ω, t) is accepted. Subsequently, the parameter normalization unit 6 calculates (Equation 28), corrects the indoor impulse response estimated value G (ω, t) that has received the input, and corrects the corrected indoor impulse response estimated value G (ω, t). ) Is output. Further, the parameter normalization unit 6 calculates (Equation 29), corrects the original sound power estimated value time series S (ω, t) that has received the input, and corrects the original sound power estimated value time series S (ω, t). t) is output.

  The reverberant sound power estimated time series calculating unit 3, the indoor impulse response updating unit 4, the original sound power estimated value time series updating unit 5, and the parameter normalizing unit 6 described above repeatedly execute each process in this order. Thus, the accuracy of the indoor impulse response estimated value G (ω, t) output from the parameter normalization unit 6 and the original sound power estimated value time series S (ω, t) is improved.

  Subsequently, the convergence determination unit 7 has sufficient accuracy of the indoor impulse response estimated value G (ω, t) output from the parameter normalization unit 6 and the original sound power estimated value time series S (ω, t). Determine whether or not. When the convergence determination unit 7 determines that the accuracy of the indoor impulse response estimated value G (ω, t) output by the parameter normalization unit 6 and the original sound power estimated value time series S (ω, t) is sufficient The parameter output unit 8 to be described later executes processing. When the convergence determination unit 7 determines that the accuracy of the indoor impulse response estimated value G (ω, t) output from the parameter normalization unit 6 and the original sound power estimated value time series S (ω, t) is not sufficient Again, the reverberant sound power estimated value time series calculating unit 3, the indoor impulse response updating unit 4, the original sound power estimated value time series updating unit 5, and the parameter normalizing unit 6 execute the respective processes in this order. .

  A method for determining whether or not the accuracy of the indoor impulse response estimated value G (ω, t) output by the parameter normalization unit 6 and the original sound power estimated value time series S (ω, t) is sufficient is as follows. There is a way.

  For example, the convergence determination unit 7 includes a reverberation sound power estimated value time-series calculating unit 3, an indoor impulse response updating unit 4, an original sound power estimated value time-series updating unit 5, and a parameter normalizing unit 6, a predetermined number of times It is determined whether the above processing has been performed. Subsequently, when the convergence determination unit 7 determines that the processing has been performed a predetermined number of times or more, the accuracy of the indoor impulse response estimated value G (ω, t) and the original sound power estimated value time series S (ω, t) is high. It is determined that it is sufficient, and otherwise it is determined that the accuracy is not sufficient.

  Further, for example, the convergence determining unit 7 has a rate of change of the indoor impulse response estimated value G (ω, t) updated by the parameter normalizing unit 6 and the original sound power estimated value time series S (ω, t) below a predetermined value. It is determined whether or not. Subsequently, when the rate of change of the objective function is equal to or less than a predetermined value, the convergence determination unit 7 has the accuracy of the indoor impulse response estimated value G (ω, t) and the original sound power estimated value time series S (ω, t). It is determined that it is sufficient, and otherwise it is determined that the accuracy is not sufficient.

  For example, the convergence determination unit 7 calculates an objective function and determines whether or not the calculated change rate of the objective function is equal to or less than a predetermined value. Subsequently, when the rate of change of the objective function is equal to or less than a predetermined value, the convergence determination unit 7 has the accuracy of the indoor impulse response estimated value G (ω, t) and the original sound power estimated value time series S (ω, t). It is determined that it is sufficient, and otherwise it is determined that the accuracy is not sufficient. The convergence determination unit 7 calculates (Expression 30) or (Expression 31), and calculates an objective function J (S, G) or K (S, G). However, Λ = {1,..., Ω} and Γ = {1,. A multiplicative update algorithm for minimizing the objective functions J (S, G) and K (S, G) will be described later.

  The parameter output unit 8 has sufficient accuracy between the indoor impulse response estimated value G (ω, t) output from the parameter normalization unit 6 and the original sound power estimated value time series S (ω, t). If it is determined that, the indoor impulse response estimated value G (ω, t) and the original sound power estimated value time series S (ω, t) are output.

  Next, a multiplicative update algorithm for minimizing the objective functions J (S, G) and K (S, G) will be described. As described above, in the present embodiment, the time index is (Expression 32). The frequency index is represented by (Expression 33). The power spectrum time series of the original signal is S (ω, t). Hereinafter, the power spectrum time series is referred to as a spectrogram. Further, (Expression 34) represents a set of whole integers. Further, the spectrogram of reverberant speech is assumed to be X (ω, t). Further, the spectrogram of the impulse response of the indoor transmission system (hereinafter referred to as the indoor impulse response) is G (ω, t). Further, it is assumed that the spectrogram X (ω, t) of the reverberant voice is approximately expressed by (Expression 35) using the room impulse response G (ω, t).

  When the speech observation spectrogram is Y (ω, t), X (ω, t) ≈Y (ω, t), and the non-negative impulse is such that the spectrogram S (ω, t) of the original signal is as sparse as possible. The purpose of this embodiment is to obtain the response G (ω, t). Therefore, considering the problem of approximating the speech observation spectrogram Y (ω, t) in the range shown in (Expression 36) and (Expression 37), the spectrogram S (ω, t) of the original signal is expressed as (Expression 38). The impulse response G (ω, t) is represented by (Equation 39). Here, Γ is a set of time indexes. Λ is a set of frequency indexes. T is an element of a time index set. Ω is an element of a set of frequency indexes.

Further, Υ _ω = {1,..., T _ω }. T _ω is the time length (number of time indexes) of the indoor impulse response that can be different for each _ω , and is hereinafter referred to as a filter length. When the closeness between the speech observation spectrogram Y (ω, t) and the spectrogram X (ω, t) of the reverberant speech is measured by a square error, the spectrogram S (ω, t) of the original signal and the indoor impulse response G (ω , T), the constrained optimization problem is as shown in (Equation 40) to (Equation 43).

  The second term of (Equation 40) is sparsity cost. The smaller the sparseness cost, the sparser the spectrogram S (ω, t) of the original signal. λ is a constant for adjusting the balance between the cost for modeling error and the sparsity cost of the speech spectrogram. Further, p is a real constant that may be arbitrarily determined within the range of 0 <p ≦ 2.

  In the above description, the square error is used as the reference for the closeness between the speech observation spectrogram Y (ω, t) and the spectrogram X (ω, t) of the reverberant speech. On the other hand, the spectrogram S (ω, t) of the original signal and the room in the case where the I divergence is used as a reference for the closeness between the spectrogram Y (ω, t) of the sound observation spectrogram Y (ω, t) Constrained optimization problems related to the impulse response G (ω, t) are as shown in (Expression 44) to (Expression 43). As for I divergence, for example, the document “I. Csiszar,“ I-Diverence Geometry of Probability Distributions and Minimization Problems, ”The Annals of Probability 3, p. It is described in.

  λ is a constant for adjusting the balance between the cost for modeling error and the sparsity cost of the speech spectrogram. Further, p is a real constant that may be arbitrarily determined within the range of 0 <p ≦ 2.

  Next, a multiplicative update algorithm for minimizing the objective function J (S, G) will be described. The objective function J (S, G) can be efficiently reduced by an iterative algorithm similar to the multiplicative update algorithm. The multiplicative update algorithm is described in, for example, the document “DD Lee and HS Seung,“ Learning the Part of Objects by Non-Negative Matrix Fac-torization, ”Nature Vol. 401, pp. 788-79.79.79. ."It is described in.

Here, a multiplicative update formula for the spectrogram S (ω, t) of the original signal is derived. The spectrogram of the original signal output from the parameter normalization unit 6 is S (ω, t). Further, before outputting the spectrogram S (ω, t) of the original signal, the spectrogram of the original signal output when the parameter normalization unit 6 performs processing is assumed to be S ′ (ω, t). That is, the updated value of the spectrogram S (ω, t) of the original signal one step before is S ′ (ω, t). The indoor impulse response output by the parameter normalization unit 6 is assumed to be G (ω, t). In addition, before outputting the indoor impulse response G (ω, t), the indoor impulse response output when the parameter normalization unit 6 performs processing is defined as G ′ (ω, t). That is, the updated value of the indoor impulse response G (ω, t) one step before is G ′ (ω, t). Here, when m _τ (ω, t) is (Expression 48), (Expression 49) is established.

  The right side of (Expression 49) is defined as (Expression 50). If the spectrogram S (ω, t) of the original signal is updated so as to minimize (Equation 50), the non-increasing property of J (S, G) is guaranteed. Therefore, solving (Equation 51) yields an update equation (Equation 52) that guarantees the non-increasing property of J (S, G). However, X ′ (ω, t) is (Formula 53).

  As shown in (Formula 52), the updated value of the spectrogram S (ω, t) of the original signal is the product of the updated value S ′ (ω, t) one step before and the update coefficient. Such an update formula is called a multiplicative update formula. Further, from (Equation 52), if all the elements of S ′ (ω, t) and G ′ (ω, t) are non-negative values, all the elements of S (ω, t) are updated to non-negative values. As can be seen, the condition of (Equation 41) is satisfied. In addition, if (Formula 54), it will be (Formula 55). Therefore, if (Formula 55) is set in the initial setting, this update does not deviate from the condition of (Formula 43).

  Next, a multiplicative update formula for the indoor impulse response G (ω, t) is derived. Similar to the case of deriving the multiplicative update formula of the spectrogram S (ω, t) of the original signal, the update value one step before the spectrogram S (ω, t) of the original signal is S ′ (ω, t), The updated value of the impulse response G (ω, t) one step before is G ′ (ω, t). If (Formula 48) is used, (Formula 56) is established.

  The right side of (Expression 56) is defined as (Expression 57). When (Formula 58) is solved in the same manner as when the multiplicative update formula of the spectrogram S (ω, t) of the original signal is derived, the multiplicative update formula (Formula 59) is obtained. However, in the derivation of the multiplicative update formula (Formula 59), the constraint of (Formula 60) is not taken into consideration, and thus it is necessary to standardize after the update of (Formula 59).

  Next, a multiplicative update algorithm for minimizing the objective function K (S, G) will be described. Similar to the objective function J (S, G), the objective function K (S, G) can be efficiently reduced by an iterative algorithm similar to the multiplicative update algorithm.

Here, a multiplicative update formula for the spectrogram S (ω, t) of the original signal is derived. The spectrogram of the original signal output from the parameter normalization unit 6 is S (ω, t). Further, before outputting the spectrogram S (ω, t) of the original signal, the spectrogram of the original signal output when the parameter normalization unit 6 performs processing is assumed to be S ′ (ω, t). That is, the updated value of the spectrogram S (ω, t) of the original signal one step before is S ′ (ω, t). The indoor impulse response output by the parameter normalization unit 6 is assumed to be G (ω, t). In addition, before outputting the indoor impulse response G (ω, t), the indoor impulse response output when the parameter normalization unit 6 performs processing is defined as G ′ (ω, t). That is, the updated value of the indoor impulse response G (ω, t) one step before is G ′ (ω, t). Here, when m _τ (ω, t) is (Expression 61), (Expression 62) is established.

  The right side of (Expression 62) is defined as (Expression 63). If the spectrogram S (ω, t) of the original signal is updated so as to minimize (Equation 63), the non-increasing property of K (S, G) is guaranteed. Therefore, solving (Equation 64) yields a multiplicative update equation (Equation 65). However, X ′ (ω, t) is (Expression 66).

  (Expression 65) indicates that if all the elements of S ′ (ω, t) and G ′ (ω, t) are non-negative values, all the elements of S (ω, t) are updated to non-negative values. .

  Next, a multiplicative update formula for the indoor impulse response G (ω, t) is derived. Similar to the case of deriving the multiplicative update formula of the spectrogram S (ω, t) of the original signal, the update value one step before the spectrogram S (ω, t) of the original signal is S ′ (ω, t), The updated value of the impulse response G (ω, t) one step before is G ′ (ω, t). When (Expression 61) is used, (Expression 67) is established.

  The right side of (Expression 67) is defined as (Expression 68). When (Equation 69) is solved in the same manner as when the multiplicative update equation of the spectrogram S (ω, t) of the original signal is derived, the multiplicative update equation (Equation 70) is obtained. However, in the derivation of the multiplicative update equation (Equation 63), the constraint of (Equation 71) is not taken into consideration, and thus it is necessary to standardize after the update of (Equation 70).

  Next, a convolution calculation unit 9 that is an example of a configuration for calculating a convolution between two sequences in the reverberation sound power estimated value time series calculation unit 3 will be described. The convolution calculator 9 can calculate the convolution at high speed. The convolution calculation unit 9 receives an input of a value necessary for calculation of convolution from the reverberation sound power estimated value time series calculation unit 3, performs convolution calculation based on this value, and calculates the calculation result as a reverberation sound power estimated value time series. Input to the calculation unit 3.

  FIG. 6 is a diagram illustrating a configuration example of functional blocks included in the convolution calculation unit 9. The convolution calculation unit 9 includes a zero padding unit 91, a fast Fourier transform unit 92, a Fourier transform product calculation unit 93, and a fast inverse Fourier transform unit 94.

Hereinafter, for W (1), W (2),..., W (T _W ) and Z (1), Z (2),..., Z (T _Z ), (Equation 72) Each part of the convolution calculation unit 9 will be described using an example of outputting the result V (t) of the convolution calculation given in the form of However, _TW and _TZ are the number of elements of each series.

  The zero padding portion 91 outputs W ′ (t) according to (Expression 72). Further, the zero padding portion 91 outputs Z ′ (t) according to (Expression 73). Here, U is the number of elements of W ′ (t) and Z ′ (t).

  Subsequently, the fast Fourier transform unit 92 receives input of W ′ (t) and Z ′ (t) output from the zero padding unit 91. Subsequently, the fast Fourier transform unit 92 performs calculation of (Expression 75) and (Expression 76) by FFT (Fast Fourier Transform) and outputs w (k) and z (k).

  Subsequently, the Fourier transform product calculation unit 93 receives the input of w (k) and z (k) output from the fast Fourier transform unit 92. Subsequently, the Fourier transform product calculation unit 93 calculates (Expression 77) and outputs v (k).

  Subsequently, the fast inverse Fourier transform unit 94 receives the input of v (k) output from the Fourier transform product calculation unit 93. Subsequently, the fast inverse Fourier transform unit 94 performs calculation of (Equation 78) by IFFT (Inverse Fast Fourier Transform) and outputs V (t).

  The convolution calculation unit 9 can calculate the convolution at high speed by the operations of the zero padding unit 91, the fast Fourier transform unit 92, the Fourier transform product calculation unit 93, and the fast inverse Fourier transform unit 94.

  Next, the observed sound / estimated original sound correlation function calculating unit 41, the estimated reverberant sound / estimated original sound correlation function calculating unit 42, the modeling error ratio sequence / estimated original sound correlation function calculating unit 45, and the observed sound / estimated sound The correlation function calculation unit 51 between impulse responses, the estimated reverberation sound with sparse correction term / estimated impulse response open correlation function calculation unit 52, and the modeling error ratio sequence / correlation function between estimated impulse responses 55 The correlation function calculation unit 10, which is an example of a configuration for calculating the correlation function, is described. The correlation function calculation unit 10 can calculate the correlation function at high speed. The correlation function calculation unit 10 receives an input of a value necessary for calculating the correlation function from each unit that requires the calculation result of the correlation function, calculates a correlation function based on this value, and inputs the calculation result to each unit. .

  FIG. 7 is a diagram illustrating a configuration example of functional blocks included in the correlation function calculation unit 10. The correlation function calculation unit 10 includes a zero padding unit 101, a fast Fourier transform unit 102, a complex conjugate unit 103, a Fourier transform product calculation unit 104, and a fast inverse Fourier transform unit 105.

Hereinafter, W (1), W ( 2), ···, and _{W (T W), Z (} 1), Z (2), ···, against _{Z (T} Z), (Formula 79) Each part of the correlation function calculation unit 10 will be described using an example of outputting a correlation function calculation result V (t) given in the form of However, _TW and _TZ are the number of elements of each series.

  The zero padding unit 101 outputs W ′ (t) according to (Expression 80). Further, the zero padding portion 101 outputs Z ′ (t) according to (Expression 81). Here, U is the number of elements of W ′ (t) and Z ′ (t).

  Subsequently, the fast Fourier transform unit 102 receives input of W ′ (t) and Z ′ (t) output from the zero padding unit 101. Subsequently, the fast Fourier transform unit 102 performs calculation of (Expression 82) and (Expression 83) by FFT, and outputs w (k) and z (k).

Subsequently, the complex conjugate unit 103 receives the input of w (k) and z (k) output from the fast Fourier transform unit 102. Subsequently, the complex conjugate unit 103 performs the operations of (Expression 84) and (Expression 85), and outputs w (k) and z (k). However, (·) ^* represents a complex conjugate. “←” represents substitution.

  Subsequently, the Fourier transform product calculation unit 104 receives the input of w (k) and z (k) output from the complex conjugate unit 103. Subsequently, the Fourier transform product calculation unit 93 calculates (Equation 86) and outputs v (k).

  Subsequently, the fast inverse Fourier transform unit 105 receives the input of v (k) output from the Fourier transform product calculation unit 104. Subsequently, the fast inverse Fourier transform unit 105 performs calculation of (Expression 87) by IFFT and outputs V (t).

  The correlation function calculation unit 10 calculates the correlation function at high speed by the operations of the zero padding unit 101, the fast Fourier transform unit 102, the complex conjugate unit 103, the Fourier transform product calculation unit 104, and the fast inverse Fourier transform unit 105. can do.

  The dereverberation apparatus of this embodiment estimates the power envelope of the reverberation component using the algorithm described above with reference to the sparsity of the audio signal. Therefore, the dereverberation apparatus of the present embodiment operates robustly against reverberation changes accompanying movement of the sound source, room temperature changes, and the like. Further, the dereverberation apparatus of the present embodiment can perform dereverberation with the same level of performance as a conventionally known dereverberation apparatus even when there is no change in the reverberation environment. In addition, the dereverberation apparatus of the present embodiment can perform dereverberation calculation at high speed.

  The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration, program, and system are not limited to this embodiment, and the design and the like without departing from the gist of the present invention. Is also included.

  In addition, a program for realizing the function of the dereverberation apparatus is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read into the computer system and executed, thereby executing dereverberation of the audio signal. May be performed. Here, the “computer system” may include an OS and hardware such as peripheral devices.

  Further, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used. The “computer-readable recording medium” means a flexible disk, a magneto-optical disk, a ROM, a writable nonvolatile memory such as a flash memory, a portable medium such as a CD-ROM, a hard disk built in a computer system, etc. This is a storage device.

  Further, the “computer-readable recording medium” means a volatile memory (for example, DRAM (Dynamic DRAM) in a computer system that becomes a server or a client when a program is transmitted through a network such as the Internet or a communication line such as a telephone line. Random Access Memory)), etc., which hold programs for a certain period of time.

  The program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the “transmission medium” for transmitting the program refers to a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, and what is called a difference file (difference program) may be sufficient.

It is a functional block diagram of the dereverberation apparatus in one Embodiment of this invention. It is the figure which showed the structure of the functional block with which the indoor impulse response update part in this embodiment is provided. It is the figure which showed the structure of the functional block with which the indoor impulse response update part in this embodiment is provided. It is the figure which showed the structure of the functional block with which the original sound power estimated value time series update part in this embodiment is provided. It is the figure which showed the structure of the functional block with which the original sound power estimated value time series update part in this embodiment is provided. It is the figure which showed the structure of the functional block with which the convolution calculation part in this embodiment is provided. It is the figure which showed the structure of the functional block with which the correlation function calculation part in this embodiment is provided.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Observation power time series production | generation part, 2 ... Initial setting part, 3 ... Reverberation sound power estimated value time series calculation part, 4 ... Indoor impulse response update part, 5 ... Original sound power estimation Value time series update unit, 6 ... parameter normalization unit, 7 ... convergence determination unit, 8 ... parameter output unit, 9 ... convolution calculation unit, 10 ... correlation function calculation unit, 41. .. Correlation function calculation unit between observed sound / estimated original sound, 42... Estimated reverberation sound / estimated original sound correlation function calculation unit, 43, 47... Indoor impulse response estimated value update coefficient calculation unit, 44, 48. Indoor impulse response estimated value update value output unit, 45 ... modeling error ratio sequence / estimated original sound correlation function calculating unit, 46 ... estimated original sound partial sum sequence calculating unit, 51 ... observed sound / estimated impulse Response correlation function calculation unit, 52... Reverberation sound / estimated impulse response correlation function calculation unit, 53, 57... Original sound power estimated value time series update coefficient calculation unit, 54, 58... Original sound power estimated value time series update value output unit, 55. Modeling error ratio sequence / estimated impulse response correlation function calculation unit, 56... Estimated impulse response partial sum sequence calculation unit with sparse correction term, 91, 101. Transformer, 93, 104 ... Fourier transform product calculator, 94, 105 ... Fast inverse Fourier transform, 103 ... Complex conjugate unit

Claims (12)

An observation power time series generation unit that receives an input of an acoustic signal and generates an observation power time series that is a time series of amplitude or power of a subband signal for each frequency channel by short-time frequency analysis;
An initial setting unit for setting an indoor impulse response estimated value having a non-negative constraint for each frequency channel and an original sound power estimated value time series that is a power estimated value time series for each frequency channel of the original sound;
Reverberation sound power estimation time series calculation that convolves the indoor impulse response estimation value and the original sound power estimation time series to calculate a reverberation power estimation time series that is a power time series of a reverberation sound model for each frequency channel. And
Based on the observation power time series, the reverberation sound power estimate time series, the room impulse response estimate, and the original sound power estimate time series, the room impulse response estimate is updated by satisfying a non-negative constraint. An indoor impulse response updating unit
Based on the observation power time series, the reverberation sound power estimation time series, the room impulse response estimation value, and the original sound power estimation time series, the original sound power estimation time series satisfying non-negative constraints An original sound power estimated value time series update unit to be updated;
The room impulse response estimated value updated by the room impulse response updating unit is normalized so that the sum of the element values of the room impulse response estimated value becomes a constant value, and the original sound power estimated value time series updating unit is updated. A parameter normalization unit that normalizes the original sound power estimated value time series so that a sum of element values of the original sound power estimated value time series becomes a constant value;
A convergence determination unit that determines whether the indoor impulse response estimated value and the original sound power estimated value time series normalized by the parameter normalization unit satisfy a predetermined criterion;
When the convergence determining unit determines that the room impulse response estimated value normalized by the parameter normalizing unit and the original sound power estimated value time series satisfy a predetermined criterion, the room impulse response estimated value And a parameter output unit for outputting the original sound power estimated value time series,
With
When the convergence determining unit determines that the room impulse response estimated value and the original sound power estimated value time series normalized by the parameter normalizing unit do not satisfy a predetermined criterion, the parameter normalizing unit Based on the normalized room impulse response estimated value and the original sound power estimated value time series, the reverberant power estimated value time series calculating unit calculates the reverberant power estimated value time series, and updates the room impulse response Unit updates the indoor impulse response estimated value, the original sound power estimated value time series update unit updates the original sound power estimated value time series, and the original sound power estimated value time series update unit performs the original sound power estimated value time series The dereverberation apparatus characterized by renewing. The indoor impulse response update unit,
For each frequency channel, an observed sound / estimated original sound correlation function calculating unit that calculates a correlation function between the observed power / estimated original sound, which is a correlation function between the observed power time series and the original sound power estimated value time series,
For each frequency channel, calculating estimated reverberation Probable original correlation function for calculating the estimated reverberation Probable original correlation function is a correlation function of the reverberation power estimate time series and the original sound power estimate time series And
Update of the room impulse response estimated value, which is a value obtained by dividing the time series element value of the correlation function between the observed sound and the estimated original sound by the time series element value of the estimated reverberation sound and the estimated original sound for each frequency channel. An indoor impulse response estimated value update coefficient calculation unit for calculating a coefficient;
For each frequency channel, the indoor impulse response estimated value and the indoor impulse response estimated value update coefficient are integrated element by element, and an indoor impulse response estimated value update value output unit that calculates an indoor impulse response estimated value update value;
The dereverberation apparatus according to claim 1, further comprising: The indoor impulse response update unit,
For each frequency channel, a modeled error ratio sequence / estimated original sound, which is a correlation function of the time series obtained by dividing the observed power time series by the reverberant power estimate time series for each element and the original sound power estimate time series A modeled error ratio sequence / estimated original sound correlation function calculation unit for calculating an inter-correlation function;
An estimated original sound partial sum sequence calculating unit that calculates an estimated original sound partial sum sequence that is a sequence having element values of partial sums of element values of each specific range of the original sound power estimated value time series for each frequency channel;
An indoor impulse response estimated value update coefficient for calculating an indoor impulse response estimated value update coefficient, which is a value obtained by dividing the correlation function between the modeled error ratio sequence and the estimated original sound by the element value of the estimated original sound partial sum series for each frequency channel A calculation unit;
For each frequency channel, the indoor impulse response estimated value and the indoor impulse response estimated value update coefficient are integrated element by element, and an indoor impulse response estimated value update value output unit that calculates an indoor impulse response estimated value update value;
The dereverberation apparatus according to claim 1, further comprising: The original sound power estimated value time series update unit,
For each frequency channel, an observed sound / estimated impulse response correlation function calculating unit that calculates a correlation function between the observed sound / estimated impulse response that is a correlation function between the observed power time series and the indoor impulse response estimated value;
For each frequency channel, a correlation function between the reverberation sound power estimated time series and the room impulse response estimated value is calculated, and the correlation function and the original sound power estimated time series are multiplied by a constant for each element and further multiplied by a constant. An estimated reverberation sound with a sparse correction term and an estimated impulse response correlation function that calculates a correlation function between the estimated reverberation sound and the estimated impulse response with a sparse correction term, which is a time series obtained by adding the calculated time series,
For each frequency channel, the original sound is a value obtained by dividing the time series element of the correlation function between the observed sound and the estimated impulse response by the time series element value of the estimated reverberant sound with the sparse correction term and the correlation function between the estimated impulse responses. An original sound power estimated time series update coefficient calculating unit for calculating a power estimated time series update coefficient;
For each frequency channel, the original sound power estimated value time series and the original sound power estimated value time series update coefficient are integrated element by element, and the original sound power estimated value time series updated value is calculated. And
The dereverberation apparatus according to any one of claims 1 to 3, further comprising: The original sound power estimated value time series update unit,
For each frequency channel, a modeled error ratio sequence / estimated impulse response that is a correlation function between the time series obtained by dividing the observed power time series by the reverberant sound power estimated value time series for each element and the indoor impulse response estimated value A modeled error ratio sequence / estimated impulse response correlation function calculation unit for calculating an inter-correlation function;
For each frequency channel, calculate a series having element values that are partial sums of the element values of each specific range of the indoor impulse response estimation values, multiply the series and the original sound power estimation value time series by a constant power for each element, and An estimated impulse response partial sum series calculation unit with a sparse correction term that calculates an estimated impulse response partial sum sequence with a sparse correction term that is a time series obtained by adding a time series multiplied by a constant,
For each frequency channel, when the original sound power estimated value is a value obtained by dividing the time series element of the correlation function between the modeled error ratio series and the estimated impulse response by the element value of the estimated impulse response partial sum series with the sparse correction term. An original sound power estimated value time series update coefficient calculation unit for calculating a series update coefficient;
For each frequency channel, the original sound power estimated value time series and the original sound power estimated value time series update coefficient are integrated element by element, and the original sound power estimated value time series updated value is calculated. And
The dereverberation apparatus according to any one of claims 1 to 3, further comprising: An observation power time series generation unit that receives an input of an acoustic signal and generates an observation power time series that is a time series of amplitude or power of a subband signal for each frequency channel by short-time frequency analysis; and ,
An initial setting step in which an initial setting unit sets an indoor impulse response estimated value having a non-negative constraint for each frequency channel and an original sound power estimated value time series that is a power estimated value time series for each frequency channel of the original sound;
A reverberant sound power estimated time series calculation unit convolves the room impulse response estimated value with the original sound power estimated time series, and a reverberant sound power estimated time series that is a power time series of a reverberant sound model for each frequency channel. Reverberation sound power estimation time series calculation step for calculating
The indoor impulse response update unit satisfies the non-negative constraint based on the observation power time series, the reverberation power estimation value time series, the indoor impulse response estimation value, and the original sound power estimation value time series, and An indoor impulse response update step for updating the indoor impulse response estimate;
An original sound power estimated value time series update unit performs non-negative constraints based on the observed power time series, the reverberant sound power estimated time series, the room impulse response estimated value, and the original sound power estimated value time series. An original sound power estimate time series update step that satisfies and updates the original sound power estimate time series; and
The parameter normalization unit normalizes the room impulse response estimated value updated in the room impulse response update step so that the sum of the element values of the room impulse response estimated value becomes a constant value, and the original sound power estimated value A parameter normalizing step for normalizing the original sound power estimated value time series updated in the series updating step so that the sum of the element values of the original sound power estimated value time series becomes a constant value;
A convergence determining step for determining whether the room impulse response estimated value and the original sound power estimated value time series normalized by the parameter normalizing step satisfy a predetermined criterion; and
When it is determined in the convergence determination step that the room impulse response estimated value normalized by the parameter normalization unit and the original sound power estimated value time series satisfy a predetermined criterion, the parameter output unit A parameter output step for outputting the impulse response estimated value and the original sound power estimated value time series;
Have
In the convergence determination step, when it is determined that the room impulse response estimated value and the original sound power estimated value time series normalized by the parameter normalization unit do not satisfy a predetermined criterion, the parameter normalizing step Based on the normalized room impulse response estimated value and the original sound power estimated value time series, the reverberant sound power estimated value time series is calculated in the reverberant power estimated value time series calculating step, and the room impulse response update is performed. Updating the room impulse response estimated value in the step, updating the original sound power estimated value time series in the original sound power estimated value time series updating step, and updating the original sound power estimated value time series in the original sound power estimated time series The dereverberation method characterized by updating the above. The indoor impulse response update step includes:
The observed sound / estimated original sound correlation function calculator calculates an observed sound / estimated original sound correlation function that is a correlation function between the observed power time series and the original sound power estimated value time series for each frequency channel. A step of calculating a correlation function between estimated original sounds;
Estimated reverberation Probable original correlation function calculating unit, for each frequency channel, it estimates the reverberation Probable original correlation function is a correlation function of the reverberation power estimate time series and the original sound power estimate time series An estimated reverberation sound / estimated original sound correlation function calculating step,
The indoor impulse response estimated value update coefficient calculation unit calculates, for each frequency channel, the time series element value of the correlation function between the observed sound and the estimated original sound as the time series element value of the estimated reverberant sound and the estimated original sound correlation function. An indoor impulse response estimated value update coefficient calculating step for calculating an indoor impulse response estimated value update coefficient that is a divided value;
An indoor impulse response estimated value update value output unit integrates the indoor impulse response estimated value and the indoor impulse response estimated value update coefficient element by element for each frequency channel, and calculates an indoor impulse response estimated value update value An impulse response estimated value update value output step;
The dereverberation method according to claim 6, further comprising: The indoor impulse response update step includes:
A modeling error ratio sequence / estimated original sound correlation function calculation unit, for each frequency channel, divides the observed power time series by the reverberant sound power estimated value time series for each element, and the original sound power estimated time A modeling error ratio sequence / estimated original sound correlation function calculating step for calculating a modeling error ratio sequence / estimated original sound correlation function that is a correlation function with the sequence;
An estimated original sound partial sum sequence calculation unit calculates an estimated original sound partial sum sequence that is a sequence having element values of partial sums of element values of each specific range of the original sound power estimated value time series for each frequency channel Sum series calculation step;
The indoor impulse response estimated value update coefficient calculation unit updates the indoor impulse response estimated value that is a value obtained by dividing the modeled error ratio sequence / estimated original sound correlation function by the element value of the estimated original sound partial sum sequence for each frequency channel. An indoor impulse response estimated value update coefficient calculation step for calculating a coefficient;
An indoor impulse response estimated value update value output unit integrates the indoor impulse response estimated value and the indoor impulse response estimated value update coefficient element by element for each frequency channel, and calculates an indoor impulse response estimated value update value An impulse response estimated value update value output step;
The dereverberation method according to claim 6, further comprising: The original sound power estimated value time series update step includes:
Observation sound / estimated impulse response correlation function calculation unit calculates, for each frequency channel, an observation sound / estimated impulse response correlation function that is a correlation function between the observed power time series and the indoor impulse response estimated value A step of calculating a correlation function between estimated impulse responses;
An estimated reverberant sound / estimated impulse response correlation function calculation unit with a sparse correction term calculates a correlation function between the reverberant power estimated value time series and the indoor impulse response estimated value for each frequency channel, and the correlation function With a sparse correction term for calculating a correlation function between estimated reverberant sound and estimated impulse response, which is a time series obtained by adding the time series obtained by multiplying the original sound power estimated value time series by a constant power for each element and further multiplying by a constant A step of calculating a correlation function between the estimated reverberant sound and the estimated impulse response;
The original sound power estimated value time series update coefficient calculation unit calculates, for each frequency channel, the time series elements of the correlation function between the observed sound and the estimated impulse response, and the correlation function between the estimated reverberant sound with the sparse correction term and the estimated impulse response. An original sound power estimated value time series update coefficient calculating step for calculating an original sound power estimated value time series update coefficient that is a value divided by a time series element value;
An original sound power estimated value time series update value output unit integrates the original sound power estimated value time series and the original sound power estimated value time series update coefficient element by element for each frequency channel, and the original sound power estimated value time series updated value An original sound power estimated value time series update value output step for calculating
The dereverberation method according to any one of claims 6 to 8, further comprising: The original sound power estimated value time series update step includes:
The modeling error ratio sequence / estimated impulse response correlation function calculation unit, for each frequency channel, divides the observed power time series into elements by the reverberant power estimate time series, and the indoor impulse response estimation A modeling error ratio sequence / estimated impulse response correlation function calculating step for calculating a modeling error ratio sequence / estimated impulse response correlation function, which is a correlation function with a value;
An estimated impulse response partial sum series calculation unit with a sparse correction term calculates, for each frequency channel, a series having element values as partial sums of element values of each specific range of the indoor impulse response estimated values, and the series, Estimated impulse response partial sum sequence with sparse correction term to calculate the estimated impulse response partial sum sequence with sparse correction term, which is a time series obtained by multiplying the original sound power estimated value time series by a constant power for each element and adding a time series multiplied by a constant A calculation step;
The original sound power estimated value time-series update coefficient calculation unit calculates, for each frequency channel, the time-series element of the modeling error ratio series / estimated impulse response correlation function as the element of the estimated impulse response partial sum series with the sparse correction term. An original sound power estimated value time series update coefficient calculating step for calculating an original sound power estimated value time series update coefficient that is a value divided by the value;
An original sound power estimated value time series update value output unit integrates the original sound power estimated value time series and the original sound power estimated value time series update coefficient element by element for each frequency channel, and the original sound power estimated value time series updated value An original sound power estimated value time series update value output step for calculating
The dereverberation method according to any one of claims 6 to 8, further comprising:   A computer program for causing a computer to operate as the dereverberation apparatus according to any one of claims 1 to 5.   A computer-readable recording medium having recorded thereon a computer program for operating the computer as the dereverberation apparatus according to any one of claims 1 to 5.

Images