JP5971646B2 - Multi-channel signal processing apparatus, method, and program - Google Patents

Description

  The present invention relates to a multi-channel signal processing apparatus, method, and program, and more particularly, to a multi-channel signal processing apparatus, method, and program for extracting or suppressing a specific signal included in a multi-channel signal.

  Conventionally, a stereo signal is divided into a plurality of frequency bands for each channel, a similarity between channels for each frequency band is calculated, and an attenuation coefficient for suppressing or enhancing a sound source signal localized near the center from the similarity is calculated. There has been proposed a stereo sound signal processing apparatus that calculates, multiplies each frequency band signal by the attenuation coefficient, re-synthesizes each frequency band signal for each channel, and outputs the result (for example, see Patent Document 1).

  The stereo acoustic signal processing device described in Patent Document 1 is a stereo signal in which an acoustic signal input to a stereo signal input unit is collected so that a target sound source signal to be emphasized or suppressed is localized near the center. It is effective in the case. Specifically, each of the stereo signals (left channel signal sL and right channel signal sR) input to the stereo signal input unit is converted into frequency domain signals (fL (k) and fR (k), N). k = 0,..., N−1), and the similarity a (k) between fL (k) and fR (k) is calculated for each same frequency band. Based on the similarity a (k) calculated for each frequency band, an attenuation coefficient g (k) is calculated for each frequency band. In the same frequency band, the same attenuation coefficient g (k) between the left and right channels is calculated for each frequency band. By multiplying and re-synthesizing the signal fL (k), component sets sL ′ and sR ′ of only components having a high degree of similarity between channels are output, and as a result, only a sound source signal localized near the center remains.

  As described above, the stereo acoustic signal processing device described in Patent Document 1 performs processing on all bands to obtain a sound source signal in which the target sound source signal is localized near the center.

  Further, spectrum data of each of the input sound signals of two channels is generated, and data of each of a plurality of frequency bins belonging to a set frequency band corresponding to a specific sound signal (voice of vocal signal) in the spectrum data is When the predetermined approximate condition is satisfied, the power of the data of the frequency bin is reduced and corrected, and the corrected acoustic signal in the time domain based on the corrected spectrum data and the difference signal with respect to other channels in each of the two channels are obtained. There has been proposed an acoustic signal processing apparatus that generates a two-channel output acoustic signal that constitutes a stereo acoustic signal by combining (for example, see Patent Document 2).

  In the acoustic signal processing device described in Patent Document 2, a difference signal (ΔXL (t) = XL (t) −XR), which is the result of calculating the difference between the input acoustic signals with respect to the other channel, for each of the two channels L and R. (t) and ΔXR (t) = XR (t) −XL (t)) are generated. Then, for each of the L and R channels, the corrected acoustic signals XL ′ (t) and XR ′ (t) in the time domain and the difference signals ΔXL (t) and ΔXR (t) are synthesized by weighted addition, for example. As a result, two-channel output acoustic signals YL (t) and YR (t) that constitute the stereo acoustic signal are generated.

JP 2002-78100 A JP 2008-72600 A

  However, in the technique described in Patent Literature 1, by suppressing a specific frequency component, a portion where the frequency spectrum is isolated occurs, and when it is converted into a time domain signal, it is heard as tone-like musical noise. There is a problem.

  Further, in the technique described in Patent Document 2, the generation of musical noise is prevented by synthesizing the difference signals and complementing the lost frequency band signals. Although the amount of calculation is reduced as compared with the technique described in Patent Document 1, since the same difference signal is synthesized with the left and right signals and corrected, the stereo effect of the generated acoustic signal is reduced, and the sound source signal is more realistic. The feeling is impaired. Further, when the extracted signal is a sound source signal localized near the center, such as a vocal signal, the signal is a monaural signal, and thus a difference signal for correcting the signal cannot be generated.

  As described above, the techniques described in Patent Documents 1 and 2 have a problem that the feeling of stereo disappears and the reproducibility deteriorates.

  The present invention has been made in view of the above problems, and in the case of extracting or suppressing a specific signal included in a multi-channel input signal including stereo and two channels, the reproducibility is good without impairing the sense of stereo. An object of the present invention is to provide a multichannel signal processing apparatus, method, and program capable of outputting a multichannel signal.

In order to achieve the above object, a multi-channel signal processing device according to a first aspect of the present invention provides a time domain of a plurality of channels including a first signal that is commonly included in each channel and a second signal that is different for each channel Each of the observation signals is converted into a spectrum signal in the frequency domain, an estimated first spectrum signal estimated as the spectrum signal of the first signal is extracted based on a ratio of each spectrum signal, and the estimated first signal which is a frequency domain signal is extracted. Extracting means for converting one spectrum signal into a signal in the time domain and extracting an estimated first signal in the time domain estimated as the first signal; and an estimated first signal in the time domain extracted by the extracting means dispersion value, dispersion value of the second signal obtained by subtracting the variance of the estimated first signal from the variance of the observation signal, and using the observation signals of the plurality of channels, wherein the plurality channels A state space model represented by a state equation composed only of the second signal including an element corresponding to, and an observation equation composed of the first signal and the second signal including an element corresponding to the plurality of channels And an estimation means for estimating the first signal or the second signal by applying a Kalman filter with a colored drive source .

According to the multi-channel signal processing device according to the first aspect of the present invention, the extraction means includes a plurality of time domain observation signals including a first signal that is commonly included among the channels and a second signal that is different for each channel. Each is converted into a spectrum signal in the frequency domain, an estimated first spectrum signal estimated as a spectrum signal of the first signal is extracted based on a ratio of each spectrum signal, and an estimated first spectrum signal that is a signal in the frequency domain is extracted. An estimated first signal in the time domain that is converted into a time domain signal and estimated as the first signal is extracted. Then, the variance value of the estimated first signal in the time domain extracted by the extraction means, the variance value of the second signal obtained by subtracting the variance value of the estimated first signal from the variance value of the observation signal, and the observation signal of a plurality of channels Are used to express a state equation composed only of a second signal including elements corresponding to a plurality of channels, and an observation equation composed of a first signal and a second signal including elements corresponding to a plurality of channels. A Kalman filter with a colored drive source is applied to the state space model to estimate the first signal or the second signal.

  The Kalman filter with a colored drive source is a Kalman filter applicable even when the drive source is a colored signal, and is a Kalman filter for estimating a target state quantity (here, the first signal or the second signal) from an observation signal. is there.

  As a result, when extracting or suppressing a specific signal included in a multi-channel signal, a stereo signal with good reproducibility without impairing the stereo effect compared to the case where each channel is complemented with the same difference signal. Can be output. Further, the conversion from the time domain to the frequency domain or the inverse conversion process from the frequency domain to the time domain is reduced once.

  According to a first aspect of the present invention, there is provided a multi-channel signal processing device based on an autocorrelation peak value of each of a plurality of time domain observation signals including the first signal or the second signal estimated by the estimation unit. A second-stage extraction means for extracting an estimated second signal in the time domain estimated as the second signal, a variance value of the estimated second signal in the time domain extracted by the second-stage extraction means, and a variance value of the observation signal The first signal or the second signal is obtained by a Kalman filter with a colored drive source using the variance value of the first signal obtained by subtracting the variance value of the estimated second signal from the observation signal of the plurality of channels. And a post-stage estimating means for estimating.

  The multi-channel signal processing apparatus according to the first aspect of the present invention converts each of the plurality of time domain observation signals including the first signal or the second signal estimated by the estimation means into a frequency domain spectrum signal. Then, an estimated second spectrum signal that is estimated as the spectrum signal of the second signal is extracted based on the spectrum entropy obtained from each spectrum signal, and the estimated second spectrum signal that is a frequency domain signal is converted into a time domain signal. A second-stage extraction means for extracting an estimated second signal in the time domain that is converted and estimated as the second signal; a variance value of the estimated second signal in the time domain extracted by the second-stage extraction means; Colored driving using the variance value of the first signal obtained by subtracting the variance value of the estimated second signal from the variance value and the observation signals of the plurality of channels The urging Kalman filter, and the rear stage estimating means for estimating the first signal or the second signal may be configured further comprising a.

Further, the multi-channel signal processing device according to the second aspect of the invention relates to the autocorrelation of each of the observation signals in the time domain of a plurality of channels including the first signal that is commonly included among the channels and the second signal that is different for each channel. Extraction means for extracting an estimated second signal in the time domain estimated as the second signal based on a peak value of the second signal, a variance value of the estimated second signal in the time domain extracted by the extraction means, and the observation Using the variance value of the first signal obtained by subtracting the variance value of the estimated second signal from the variance value of the signal, and the observation signal of the plurality of channels, the second signal including elements corresponding to the plurality of channels state equation composed only, and the first signal and the state space model represented by composed observation equation and a second signal containing the elements corresponding to the plurality of channels, Karumanfu with colored driving source By applying the filter, it is configured to include a, an estimation unit for estimating the first signal or the second signal.

According to the multi-channel signal processing device of the second invention, the extraction means includes a plurality of time domain observation signals including a first signal that is included in common among the channels and a second signal that is different for each channel. Based on each autocorrelation peak value, an estimated second signal in the time domain estimated as the second signal is extracted. Unlike the extracting means of the first invention that extracts the estimated first signal, the extracting means of the second invention extracts the estimated second signal. Then, the variance value of the estimated second signal in the time domain extracted by the extraction means, the variance value of the first signal obtained by subtracting the variance value of the estimated second signal from the variance value of the observation signal, and the observation signals of a plurality of channels Are used to express a state equation composed only of a second signal including elements corresponding to a plurality of channels, and an observation equation composed of a first signal and a second signal including elements corresponding to a plurality of channels. A Kalman filter with a colored drive source is applied to the state space model to estimate the first signal or the second signal.

  As a result, when a specific signal included in a multi-channel signal is extracted or suppressed, a stereo signal with good reproducibility can be output without impairing the stereo feeling. Further, since the signal processing is performed only in the time domain, the amount of calculation is reduced as compared with the first invention.

The multi-channel signal processing apparatus according to the third aspect of the present invention provides a multi-channel time domain observation signal including a first signal that is commonly included in each channel and a second signal that is different for each channel. And calculating each power spectral density, extracting an estimated second spectral signal estimated as a spectral signal of the second signal based on a spectral entropy obtained from the power spectral density, and a frequency An extracting means for converting the estimated second spectrum signal, which is a signal in a region, into a time domain signal and extracting an estimated second signal in the time domain estimated as the second signal, and the extraction means extracted by the extracting means A variance value of the estimated second signal in the time domain, a variance value of the first signal obtained by subtracting the variance value of the estimated second signal from the variance value of the observed signal, Using observation signals of the plurality of channels each time, the plurality of channels state equation composed only of the second signal containing the elements corresponding to, and said first and second signals containing the elements corresponding to the plurality of channels And an estimation means for estimating the first signal or the second signal by applying a Kalman filter with a colored drive source to a state space model represented by an observation equation .

According to the multi-channel signal processing device of the third invention, the extraction means includes a plurality of time domain observation signals including a first signal that is commonly included among the channels and a second signal that is different for each channel. Each is converted into a spectrum signal in the frequency domain, each power spectral density is calculated, and an estimated second spectral signal estimated as a spectral signal of the second signal is extracted based on the spectral entropy respectively obtained from the power spectral density. Then, the estimated second spectrum signal, which is a frequency domain signal, is converted into a time domain signal to extract a time domain estimated second signal estimated as the second signal. Unlike the extracting means of the first invention that extracts the estimated first signal, the extracting means of the second invention extracts the estimated second signal. Then, the variance value of the estimated second signal in the time domain extracted by the extraction means, the variance value of the first signal obtained by subtracting the variance value of the estimated second signal from the variance value of the observation signal, and the observation signals of a plurality of channels Are used to express a state equation composed only of a second signal including elements corresponding to a plurality of channels, and an observation equation composed of a first signal and a second signal including elements corresponding to a plurality of channels. A Kalman filter with a colored drive source is applied to the state space model to estimate the first signal or the second signal.

  As a result, when a specific signal included in a multi-channel signal is extracted or suppressed, a stereo signal with good reproducibility can be output without impairing the stereo feeling.

In addition, the multi-channel signal processing method according to the fourth aspect of the present invention is to apply each of the observation signals in the time domain of a plurality of channels including the first signal that is commonly included among the channels and the second signal that is different for each channel to the frequency domain. The estimated first spectrum signal that is estimated as the spectrum signal of the first signal is extracted based on the ratio of each spectrum signal, and the estimated first spectrum signal that is a frequency domain signal is extracted in the time domain The first signal in the time domain estimated as the first signal is extracted, and the estimated first signal in the time domain is extracted from the extracted variance value of the estimated first signal in the time domain and the variance value of the observed signal. variance value of the second signal obtained by subtracting a variance value of the first signal, and by using the observation signals of the plurality of channels, composed only said second signal containing the elements corresponding to the plurality of channels Equation of state is, and the first signal and the state space model represented by composed observation equation and a second signal containing the elements corresponding to the plurality of channels, by applying the Kalman filter with colored driving source, the second This is a method for estimating one signal or the second signal.

In addition, the multi-channel signal processing method according to the fifth aspect of the invention relates to the autocorrelation of each of the observation signals in the time domain of a plurality of channels including the first signal that is commonly included among the channels and the second signal that is different for each channel. Based on the peak value of the second signal, an estimated second signal in the time domain estimated as the second signal is extracted, and the estimated value is extracted from the extracted variance value of the estimated second signal in the time domain and the variance value of the observed signal A state equation composed of only the second signal including elements corresponding to the plurality of channels using the dispersion value of the first signal obtained by subtracting the dispersion value of the second signal and the observation signals of the plurality of channels. and wherein the plurality of channels the first signal comprising the elements corresponding to the state space model represented by composed observation equation and a second signal, by applying a Kalman filter with colored driving source, said first It is a degree or a method of estimating the second signal.

In addition, the multi-channel signal processing method according to the sixth aspect of the present invention provides a frequency domain observation signal in each of a plurality of channels including a first signal that is commonly included among the channels and a second signal that is different for each channel. And calculating each power spectral density, extracting an estimated second spectral signal estimated as a spectral signal of the second signal based on a spectral entropy obtained from the power spectral density, and a frequency The estimated second spectrum signal, which is a signal in a domain, is converted into a signal in the time domain, a second estimated signal in the time domain estimated as the second signal is extracted, and the extracted second estimated signal in the time domain is extracted. , The dispersion value of the first signal obtained by subtracting the dispersion value of the estimated second signal from the dispersion value of the observed signal, and the view of the plurality of channels. Using the signal, observation composed of a plurality channel state equation composed only of the second signal containing the elements corresponding to, and said first and second signals containing the elements corresponding to the plurality of channels This is a method for estimating the first signal or the second signal by applying a Kalman filter with a colored drive source to a state space model represented by an equation .

  A multi-channel signal processing program according to the seventh invention is a program for causing a computer to function as each means constituting the multi-channel signal processing device.

  As described above, according to the multi-channel signal processing apparatus, method, and program of the present invention, when a specific signal included in a multi-channel input signal including stereo and two channels is extracted or suppressed, The effect that a multi-channel signal with good reproducibility can be output without impairing the above is obtained.

It is a block diagram which shows schematic structure of the stereo signal processing apparatus which concerns on 1st Embodiment. It is the schematic for demonstrating the observation condition of an observation signal. It is a figure for demonstrating the process of a frequency domain conversion part. It is a figure for demonstrating the process of a vocal signal extraction part. It is a figure for demonstrating the process of a time domain conversion part. It is a block diagram showing a state space model. It is a flowchart which shows the content of the stereo signal process in 1st Embodiment. It is a flowchart which shows the content of a colored drive moped Kalman algorithm. It is a block diagram which shows schematic structure of the stereo signal processing apparatus which concerns on 2nd Embodiment. It is a flowchart which shows the content of the stereo signal process in 2nd Embodiment. It is a block diagram which shows schematic structure of the stereo signal processing apparatus which concerns on 3rd Embodiment. It is a figure for demonstrating the process of an autocorrelation process part. It is a figure for demonstrating the process of a peak value detection part. It is a flowchart which shows the content of the stereo signal process in 3rd Embodiment. It is a block diagram which shows schematic structure of the stereo signal processing apparatus which concerns on 4th Embodiment. It is a flowchart which shows the content of the stereo signal processing in 4th Embodiment. It is a figure for demonstrating the amount-of-computation reduction type | formula color drive mop Kalman filter. It is a figure for demonstrating the amount-of-computation reduction type | formula color drive mop Kalman filter. It is a figure for demonstrating the amount-of-computation reduction type | formula color drive mop Kalman filter. It is a flowchart which shows the content of the calculation reduction type | mold color drive moped Kalman algorithm.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
<First Embodiment>
In the first embodiment, an example of the first signal of the present invention is a vocal signal having a sound source position near the center of the L channel microphone and the R channel microphone, and an example of the second signal of the present invention is, for example, A case where the music signal is a musical instrument or the like will be described.

  As shown in FIG. 1, the stereo signal processing apparatus 10 according to the first embodiment includes A / D conversion units 12L and 12R, frequency domain conversion units 14L and 14R, a spectrum ratio calculation unit 16, and a vocal signal. The extraction unit 18, the time domain conversion unit 20, the music signal estimation unit 22, and the D / A conversion units 24L and 24R are configured. The stereo signal processing apparatus 10 can be configured by a semiconductor integrated circuit such as an ASIC (Application Specific Integrated Circuit).

The A / D converters 12L and 12R are observation signals x _L (n) and x _R (n) which are analog signals inputted from the outside (indicated as observation signal L and observation signal R in FIG. 9, 11, and 15) are converted into digital signals, and the observation signals x _L (n) and x _R (n) converted into digital signals are output to the frequency domain converters 14 L and 14 R, respectively.

Here, the observation signals x _L (n), x _R (n) are composed of a music signal (L channel signal i _L (n), R channel signal i _R (n)) and a vocal signal, as shown in FIG. This is a signal obtained by observing d (n). At time n, the L channel observation signal observed by the L channel microphone is x _L (n), and the R channel observation signal observed by the R channel microphone is x _R (n). The observation signals x _L (n) and x _R (n) are expressed by the following formulas (1) and (2).

x _L (n) = d (n) + i _L (n) (1)
_{x R (n) = d (} n) + i R (n) (2)
The frequency domain transform units 14L and 14R respectively convert the observation signals x _L (n) and x _R (n), which are time domain signals input from the A / D conversion units 12L and 12R, to the frequency domain observation signals X _L ( l, k), X _R (l, k) and output to the spectrum ratio calculation unit 16 and the vocal signal extraction unit 18. Specifically, as shown in FIG. 3, the frequency domain transform units 14L and 14R perform the following observation signal x _L (l, n) and x _R (l, n) in a frame having a predetermined frame length ( 3) Fourier transform is performed according to equations (4) and (4) to convert the spectrum into each frequency bin. Here, 2M is the number of samples per frame, l is a frame number, and k is a frequency bin number. Hereinafter, an observation signal converted into a frequency domain signal is also referred to as an “observation spectrum”.

The spectrum ratio calculation unit 16 calculates a spectrum ratio between the observed spectrums | X _L (l, k) | and | X _R (l, k) | input from the frequency domain conversion units 14L and 14R, and extracts a vocal signal. To the unit 18. The vocal signal is equally included in the L channel observation signal x _L (n) and the R channel observation signal x _R (n) as shown in the equations (1) and (2). Therefore, also in the observed spectrum, as shown in the following equations (5) and (6), the L-channel observed spectrum | X _L (l, k) | and the R-channel observed spectrum | X _R (l, k) And | include the spectrum | D (l, k) | of the vocal signal equally. In addition, | I _L (l, k) | and | I _R (l, k) | are the spectrum of the L channel music signal and the spectrum of the R channel music signal.

| X _L (l, k) | = | D (l, k) | + | I _L (l, k) | (5)
| X _R (l, k) | = | D (l, k) | + | I _R (l, k) | (6)
From this, when the spectrum ratio between the L channel observation spectrum and the R channel observation spectrum is small, the signal is determined as a vocal signal, and when the spectrum ratio is large, the signal is determined as a music signal. it can. Therefore, the spectrum ratio calculation unit 16 calculates the spectrum ratio between the L channel observation spectrum and the R channel observation spectrum. In Patent Documents 1 and 2, the similarity between the left channel signal and the right channel signal converted into a frequency domain signal is calculated for each same frequency band. In the present embodiment, L is calculated by the following equation (7). A spectrum ratio A _e (l, k) between the channel observation spectrum and the R channel observation spectrum is calculated.

Based on the spectrum ratio A _e (l, k) input from the spectrum ratio calculation unit 16, the vocal signal extraction unit 18 observes the spectrum | X _L (l, k) input from the frequency domain conversion units 14L and 14R. A spectrum (hereinafter referred to as “estimated vocal spectrum”) of a signal estimated as a vocal signal (hereinafter referred to as “estimated vocal signal”) is extracted from ||| _R (l, k) | Output to. Specifically, based on the spectral ratio A _e (l, k), determines whether the vocal signal or music signal for each frequency bin of the observed spectrum of each frame. Then, as shown in the following equation (8), when the vocal signal is determined, the observed spectrum is extracted as it is, and when it is determined as the music signal, the observed spectrum is suppressed, so that the estimated vocal spectrum | D ^ (l, k) | is extracted. In Patent Document 2, a music signal that is a target sound source signal is extracted, but here, an estimated vocal spectrum is extracted instead of a music signal that is a final extraction target.

Here, α is a threshold value that determines how much a sound source signal (here, a music signal) other than a sound source signal (here, a vocal signal) localized near the center of the L channel microphone and the R channel microphone is allowed. Yes, 0 ≦ α ≦ 1. K ₀ is a coefficient for adjusting the degree of suppression of the music signal, and 0 ≦ k ₀ ≦ 1. As shown in FIG. 4, when k ₀ = 0, the music signal is completely suppressed. Note that the expression (8) shows the case where the estimated vocal spectrum | D ^ (l, k) | is extracted from the observed spectrum | X _L (l, k) |, but the observed spectrum | X _R (l, k) ) | May be used.

Note that the above processing may be performed only for the vocal band W ₀ as shown in FIG. W ₀ is a coefficient that specifies the band of the vocal signal included in the observed signal. In the case of male vocals, the vocal signal is concentrated in a low band, and in the case of female vocals, the vocal signal is concentrated in a high band. Therefore, by providing a processing band such as W ₀ , it is possible to reduce the amount of calculation compared to the case where processing is performed over the entire band of the observation signal as in the method of Patent Document 1. it can. In this embodiment, since the first signal is a vocal signal, the processing band W ₀ corresponding to the characteristics of the vocal signal is set. However, depending on what kind of signal the first signal is to be used. The processing band W ₀ corresponding to the characteristics of the signal may be set.

  The time domain transform unit 20 performs an inverse Fourier transform on the estimated vocal spectrum | D ^ (l, k) | input from the vocal signal extraction unit 18 according to the following equation (9) to obtain an estimated vocal signal d ^ in the time domain. Convert to (l, n) (see also FIG. 5). In addition, compared with the method of patent document 1 and 2, the frequency | count of an inverse Fourier transform may be 1 time.

Next, the time domain estimated vocal signal d ^ (l-1, n + M) using the latter half M samples one frame before by the overlap add method and the time domain estimated vocal signal d ^ (l using the first half M samples of the current frame are used. , N) are added together to obtain an M sample time domain estimated vocal signal d ^ (n) (1 ≦ n ≦ M) of the current frame. When the overlap add method is expressed by a mathematical expression, it can be expressed as follows.

The music signal estimation unit 22 is estimated as a music signal based on the estimated vocal signal d ^ (n) input from the time domain conversion unit 20 and the observation signals x _L (n), x _R (n). A signal (hereinafter referred to as “estimated music signal”) is extracted. In the present embodiment, a specific signal (music signal in this case) included in the observation signal is extracted by a Kalman filter with a colored drive source that does not use estimation of the AR coefficient.

  Specifically, the observation signal is replaced with a state equation composed of only a music signal represented by the following equation (10) and a state space model represented by an observation equation composed of a vocal signal and a music signal.

However, the vectors x _p2 , δ _p2 , y _p2 , ε _p2 , Φ _p2, and M _p2 in the equation (10) are respectively defined by the following equation (11). The vector x _p2 is a 2L _p2 × first-order state vector composed of a desired music signal, the vector δ _p2 is a 2L _p2 × primary drive source vector, the vector y _p2 is a 2 × first-order observation signal vector, and the vector ε _p2 is This is a 2 × first-order vocal signal vector. The matrix Φ _p2 is a state transition matrix composed of only 0 and 1, and the matrix M _p2 is a 2 × 2L _p- order observation transition matrix. FIG. 6 is a block diagram showing this state space model. 2L _p2 is the size of the state transition matrix. Further, p2 is a subscript indicating that it is a variable of the state equation and the observation equation to which the colored drive moped Kalman filter is applied.

The state equation in equation (10) describes the system to be estimated (here, the music signal) with a state space model, and represents the internal state, that is, the time change of the state variable (here, the state vector x _p2 ). Yes. The observation equation in equation (10) describes the process of observation through some kind of observation device, and the observation result (here, the observation signal vector y _p2 ) is the observed quantity, that is, the input (here, the state vector). The state of time evolution depending on x _p2 ) is shown. “State vector x _p2 (n) at time n” means a state vector composed of music signals up to time n.

  The L / R channel coupled Kalman algorithm shown below is derived by the state equation and the observation equation shown in the equation (10).

The above algorithm is roughly divided into an initialization process [Initialization] and an iterative process [Iteration]. In the iterative process, steps 1 to 5 are sequentially repeated. A detailed processing flow of each process and procedure will be described later, and an outline of each process and procedure will be described here.

In the initial setting process, the initial value x ^ _p2 (0 | 0) of the optimal estimated value of the state vector (hereinafter referred to as “optimum estimated value vector”) indicating the music signal to be estimated, and the error when the state vector is estimated the initial value _{P p2} of a covariance matrix (0 | 0), the dispersion value of the vocal signals _{R εp2 (n) [i,} j], and variance of the music signal _{R δp2 (n) [i,} j] values of , Set as above. The music signal variance is obtained by subtracting the vocal signal variance from the observed signal variance. * [I, j] is an element of i row and j column of the variable name *, and I is a unit matrix.

In the iteration process, in step 1, a covariance matrix P _p2 (n + 1 | n), which is an error when the state vector at time n + 1 is estimated from information up to time n, is calculated. Next, in step 2, the estimated error of the observed signal vector is multiplied by the Kalman gain matrix and the optimum estimated value vector x ^ _p2 (n + 1 | n) at time n + 1 based on the information up to time n is added. A Kalman gain matrix K _p2 (n + 1) is calculated so as to be the optimum estimated value vector x ^ _p2 (n + 1 | n + 1) at that time based on the information up to n + 1.

Next, in procedure 3, the optimum estimated value vector x ^ _p2 (n + 1 | n) at time n + 1 is calculated from information up to time n. Next, in step 4, the optimum estimated value vector x ^ _p2 (n + 1 | n + 1) at that time based on the information up to time n + 1 is calculated. In steps 3 and 4, the state quantity is updated. Next, in step 5, the covariance matrix at that time is updated with information up to time n + 1.

The music signal estimation unit 22 repeats the above iterative process a predetermined number of times, and the L channel estimated music signal i ^ in the first row and the first column of the optimum estimated value vector x ^ _p2 (n + 1 | n + 1) obtained in the procedure 4 is obtained. _{As L} (n), the (L _p2 +1) row and first column are output to the D / A converters 24L and 24R as the estimated music signal i ^ _R (n) of the R channel.

The D / A converters 24L and 24R convert the estimated music signals i ^ _L (n) and i ^ _R (n), which are digital signals input from the music signal estimation section 22, into analog signals, respectively, and finally Output signals L and R.

  Next, the operation of the stereo signal processing apparatus 10 according to the first embodiment will be described with reference to FIG.

In step 100, the A / D converters 12L and 12R convert the observation signals x _L (n) and x _R (n), which are analog signals input from the outside, into digital signals, respectively. Next, in step 102, the frequency domain converters 14L and 14R convert the observation signals x _L (n) and x _R (n), which are time domain signals converted into digital signals in step 100, into frequency domains. It is converted into an observation spectrum | X _L (l, k) |, | X _R (l, k) |

Next, in step 104, the spectrum ratio calculation unit 16 calculates the spectral ratio A _e (l between the observed spectrum | X _L (l, k) | converted in step 102 and | X _R (l, k) | , K). Next, in step 106, the vocal signal extraction unit 18 determines whether or not the spectrum ratio A _e (l, k) calculated in step 104 is larger than a predetermined threshold value α. If A _e (l, k)> α, the process proceeds to step 108, where the signal is regarded as a music signal and a coefficient k ₀ (0 ≦ k ₀ ≦ 1) for adjusting the degree of suppression of the music signal. For example, k ₀ = 0, and the observed spectrum | X _L (l, k) | or | X _R (l, k) | is multiplied to suppress the music signal. On the other hand, if A _e (l, k) ≦ α, the process proceeds to step 110, where the signal is regarded as a vocal signal, for example, k ₀ = 1, and the observed spectrum | X _L (l, k) | | X _R (l, k) | is extracted as an estimated vocal spectrum | D ^ (l, k) |.

  Next, in step 112, the time domain transforming unit 20 performs an inverse Fourier transform on the estimated vocal spectrum | D ^ (l, k) | extracted through the processing of steps 108 and 110 according to the equation (9). And converted into an estimated vocal signal d ^ (l, n) in the time domain. Next, the time domain estimated vocal signal d ^ (l-1, n + M) using the latter half M samples one frame before by the overlap add method and the time domain estimated vocal signal d ^ (l using the first half M samples of the current frame are used. , N) are added together to obtain an M sample time domain estimated vocal signal d ^ (n) (1 ≦ n ≦ M) of the current frame.

Next, in step 114, the music signal estimation unit 22 performs music signal estimation processing based on the estimated vocal signal d ^ (n) and the observation signals x _L (n), x _R (n). Thus, the estimated music signal is extracted. The music signal estimation process corresponds to the colored drive moped Kalman algorithm shown in FIG. Here, with reference to FIG. 8, the flow of the colored drive moped Kalman algorithm will be described.

In step 1140, the state space model is defined by the state equation and the observation equation shown in the equation (10), the initial value x ^ _p2 (0 | 0) of the optimum estimated value vector, and the error when the state vector is estimated. the initial value _{P p2} of the dispersion matrix (0 | 0), the dispersion value of the vocal signals _{R εp2 (n) [i,} j], and variance of the music signal _{R δp2 (n) [i,} j] , and the above-mentioned initial Setting process Set to the initial state shown in [Initialization]. Also, a variable n indicating time is set to 0.

Next, in step 1142, the state transition matrix Φ _p2 in the state space model defined in step 1140 above, the initial value P _p2 (0 | 0) of the covariance matrix of the set state vector (when n = 0), or Using the covariance matrix P _p2 (n | n) (when n ≧ 1) updated in step 1150 described later one time ago and the value of the variance value R _δp2 (n + 1) [i, j] of the music signal Then, the covariance matrix P _p2 (n + 1 | n), which is an error when the state vector at time n + 1 is estimated from the information up to time n, is calculated (procedure 1 of the above iteration process [Iteration]).

Next, in step 1144, the covariance matrix P _p2 (n + 1 | n) calculated in step 1142, the observed transition matrix M _p2 in the state space model defined in step 1140, and the variance value R _εp2 (n of the vocal signal) ) The Kalman gain matrix K _p2 (n + 1) is calculated using [i, j] (same procedure 2).

Next, in step 1146, the state transition matrix Φ _p2 and the initial value x ^ _p2 (0 | 0) of the optimum estimated value vector set in step 1140 (if n = 0), or this step one time before _Is used to obtain the optimum estimated value vector x ^ _p2 (n + 1 | n) at time n + 1 from the information up to time n using the optimum estimated value vector x ^ _p2 (n | n) (when n ≧ 1) Calculate (same procedure 3). Then, using the calculated optimum estimated value vector x ^ _p2 (n + 1 | n), the Kalman gain matrix _Kp2 (n + 1) calculated in step 1144, the observed vector _yp2 (n + 1), and the observed transition matrix _Mp2 , The optimum estimated value vector x ^ _p2 (n + 1 | n + 1) at that time based on the information up to time n + 1 is calculated (same procedure 4).

  Next, in step 1148, it is determined whether or not to end the process. In this determination, when the time n reaches a predetermined number N of samples, it may be determined that the process is ended, or may be determined when the sample is exhausted. If the process is not terminated, the process proceeds to step 1150. If the process is terminated, the process proceeds to step 1154.

In step 1150, using the unit matrix I, the Kalman gain matrix K _p2 (n + 1), the observed transition matrix M _p2 , and the covariance matrix P _p2 (n + 1 | n) calculated in step 1142, information up to time n + 1 is obtained. To update the covariance matrix P _p2 (n + 1 | n + 1) at that time. Next, in step 1152, n is incremented by 1, and the process returns to step 1142.

On the other hand, in step 1154, the first channel and the first column of the optimum estimated value vector x ^ _p2 (n + 1 | n + 1) calculated in step 1146 are set as the L channel estimated music signal i ^ _L (n), and ( _Lp2 + 1). ) The first row and the first column are output as the R channel estimated music signal i ^ _R (n), and the process returns to the process of FIG.

Next, in step 116, the D / A converters 24L and 24R respectively convert the estimated music signals i ^ _L (n) and i ^ _R (n), which are digital signals output by the processing of step 114, into analog signals. And output as final output signals L and R, and the process is terminated.

  As described above, according to the stereo signal processing apparatus of the first embodiment, a Kalman filter with a color driving source is used for the estimated vocal signal and the observation signal extracted based on the ratio of the observation spectrum of the L channel and the R channel. Is applied to extract the estimated music signal, so that a stereo signal with good reproducibility can be output without impairing the stereo feeling.

Further, only one inverse Fourier transform is required to convert the extracted vocal spectrum into a time domain signal.
<Second Embodiment>
In the second embodiment, an example of the first signal of the present invention is a vocal signal having a sound source position near the center of the L channel microphone and the R channel microphone, and an example of the second signal of the present invention is, for example, A case where the music signal is a musical instrument or the like will be described.

  In the second embodiment, a case where a vocal signal or a music signal is selectively extracted will be described. Note that, in the stereo signal processing device of the second embodiment, the same parts as those of the stereo signal processing device 10 of the first embodiment are denoted by the same reference numerals, and detailed description thereof is omitted.

  As shown in FIG. 9, the stereo signal processing apparatus 210 according to the second embodiment includes A / D conversion units 12L and 12R, frequency domain conversion units 14L and 14R, a spectrum ratio calculation unit 16, and a vocal signal. The extraction unit 18, the time domain conversion unit 20, the specific signal estimation unit 222, and the D / A conversion units 24L and 24R are configured.

The specific signal estimation unit 222 and the estimated vocal signal d ^ (n) input from the time domain conversion unit 20 and the observation signal x _L according to a selection signal for selecting whether to extract a music signal or a vocal signal Based on (n) and x _R (n), an estimated music signal or an estimated vocal signal is extracted. When the selection signal indicates that the music signal is to be extracted, the specific signal estimation unit 222 extracts the estimated music signal by the same processing as that of the music signal estimation unit 22 of the first embodiment.

  On the other hand, if the selection signal indicates that a vocal signal is to be extracted, an estimated vocal signal is extracted by the following L / R channel combined color drive moped Kalman algorithm. Note that the initialization process [Initialization] is the same as that in the first embodiment, and thus the description thereof is omitted.

When extracting the estimated vocal signal, the variance value R _δp2 (n + 1) [i, j] of the music signal in the procedure 1 of the iteration process [Iteration] in the first embodiment and the vocal signal of the procedure 2 are _extracted . The variance R _εp2 (n + 1) [i, j] is replaced. As a result, the first row and the first column or the (L _p2 +1) row and the first column of the optimum estimated value vector x ^ _p2 (n + 1 | n + 1) calculated in the procedure 4 are used as the estimated vocal signal d ′ ^ _p2 (n). Can be obtained. The estimated vocal signal obtained here is a signal without musical noise.

  Next, the operation of the stereo signal processing apparatus 10 according to the second embodiment will be described with reference to FIG. In addition, about the process same as the process in 1st Embodiment, the same code | symbol is attached | subjected and detailed description is abbreviate | omitted.

  Through steps 100 to 112, the estimated vocal spectrum | D ^ (l, k) | extracted based on the spectrum ratio is converted into an estimated vocal signal d ^ (n) in the time domain.

Next, in step 200, the specific signal estimation unit 222 determines whether to extract a music signal or a vocal signal based on the selection signal. If it is determined that the music signal is to be extracted, the process proceeds to step 114 where the music signal estimation unit 22 determines the estimated vocal signal d ^ (n) and the observation signals x _L (n), x _R (n). Based on the above, an estimated music signal is extracted by executing a music signal estimation process.

On the other hand, if it is determined that the vocal signal is to be extracted, the process proceeds to step 202, where the music signal estimation unit 22 determines the estimated vocal signal d ^ (n) and the observed signals x _L (n), x _R (n). Based on the above, the vocal signal estimation process is executed to extract the estimated vocal signal.

  The vocal signal estimation process corresponds to the colored drive moped Kalman algorithm shown in FIG. 8 as in the first embodiment. Here, processing that is different from the flow of the Kalman algorithm with colored driving executed as music signal estimation processing will be described.

In the Kalman algorithm executed as vocal signal estimation process, replaced in step 1142, the variance value of the music signal _{R δp2 (n + 1) [} i, j] , and variance _{R εp2 (n) [i,} j] of the vocal signals Then, the covariance matrix P _p2 (n + 1 | n), which is an error when the state vector at time n + 1 is estimated from the information up to time n, is calculated (procedure 1 of the above iteration process [Iteration]).

In step 1144, the variance value R _εp2 (n) [i, j] of the vocal signal is replaced with the variance value R _δp2 (n + 1) of the music signal to calculate the Kalman gain matrix K _p2 (n + 1) (same as above). Procedure 2).

In step 1154, the estimated vocal signal d ′ ^ _p2 (1st row, first column or (L _p2 + 1) th row, first column of the optimum estimated value vector x ^ _p2 (n + 1 | n + 1) calculated in step 1146 is used. n) and return to the process of FIG. The estimated vocal signal obtained here is a signal without musical noise.

As described above, according to the stereo signal processing apparatus of the second embodiment, in addition to the effects of the first embodiment, it is possible to selectively extract a desired signal (vocal signal or music signal). it can.
<Third Embodiment>
In the third embodiment, an example of the first signal of the present invention is a vocal signal (voice signal) having a sound source position near the center of the L channel microphone and the R channel microphone, for example. For example, a case where a noise signal close to white noise is used will be described.

  In the third embodiment, a case will be described in which a noise signal is suppressed from an observation signal observed in the situation shown in FIG. Note that in the stereo signal processing device of the third embodiment, the same parts as those of the stereo signal processing device 10 of the first embodiment are denoted by the same reference numerals, and detailed description thereof is omitted.

  As shown in FIG. 11, the stereo signal processing apparatus 310 according to the third embodiment includes A / D conversion units 12L and 12R, autocorrelation processing units 26L and 26R, peak value detection units 28L and 28R, The noise determination units 30L and 30R, the noise suppression unit 322, and the D / A conversion units 24L and 24R are configured.

The autocorrelation processing units 26L and 26R calculate the autocorrelation functions of the observation signals x _L (n) and x _R (n), which are time domain signals input from the A / D conversion units 12L and 12R, Output to the value detectors 28L and 28R, respectively. Specifically, the autocorrelation processing units 26L and 26R divide the observation signal into frames by L samples as shown in FIG. Observation signals x _L (l, i), x _R (l, i) relating to the i-th sample in the l frame are expressed by the following equations (12) and (13).

x _L (l, i) = d (l, i) + i _L (l, i) (12)
_{x R (l, i) =} d (l, i) + i R (l, i) (13)
The autocorrelation processing units 26L and 26R set the delay time as τ, and the autocorrelation functions R _xL (l, τ), R _xR (of the L channel and R channel observation signals according to the following equations (14) and (15): l, τ) (τ = 0,..., L−1).

The peak value detectors 28L and 28R detect the peak values in the autocorrelation functions R _xL (l, τ) and R _xR (l, τ) respectively input from the autocorrelation processing units 26L and 26R, and each noise determination unit Output to 30L and 30R. Specifically, in autocorrelation functions R _xL (l, τ), R _xR (l, τ), peak values other than τ = 0 are detected by the following equations (16) and (17) (FIG. 13). See also). Note that max {*} is processing for finding the maximum value of the function *.

The noise determination units 30L and 30R are signals (hereinafter, referred to as noise signals) estimated for each frame based on the peak values p _L (l) and p _R (l) input from the peak value detection units 28L and 28R, respectively. Each of which is referred to as “estimated noise signal” and output to the noise suppression unit 322. Specifically, the peak values p _L (l) and p _R (l) are compared with the threshold σ ₁ according to the following formulas (18) and (19), and the peak value is greater than the threshold σ _1. Determines that frame l is a vocal signal (speech signal), copies the estimated noise signal one frame before, and estimates estimated noise signals i ^ _L (l, i), i ^ _R (l, i) of frame l. And On the other hand, when the peak value is smaller than the threshold σ ₁ , the frame l is determined as a noise signal, and is directly used as the estimated noise signals i ^ _L (l, i) and i ^ _R (l, i). The threshold σ ₁ is a value determined by the SNR of the observation signal.

The noise suppression unit 322, the noise determination unit 30L, estimated noise input from the 30R signal _{i ^ L (l, i)} , i ^ R (l, i) and observed signal _{x L (n), x R} (n ) To suppress the noise signal. Specifically, the noise signal is suppressed by the following L / R channel coupling type color driving moped Kalman algorithm. Since the iteration process [Iteration] is the same as that in the first embodiment, the description thereof is omitted.

In the course of initialization, the initial value x ^ _p2 optimum estimate vector (0 | 0), the initial value P _p2 of the covariance matrix is the error in estimating the state vector (0 | 0), the variance of the noise signal The values of the element R _δp2 (n) [i, j] in the i row and j column of the value matrix R _δp2 (n) and the variance value matrix R _εp2 (n) of the vocal signal are set as described above. The variance value of the vocal signal is obtained by subtracting the variance value of the noise signal from the variance value of the observation signal. Thereafter, as in the first embodiment, the iterative process is executed, and the first row or the first column of the optimum estimated value vector x ^ _p2 (n + 1 | n + 1) calculated in step 4 of the iterative process or (L _p2 + 1) row 1st column can be obtained as the estimated vocal signal d ^ (n). That is, a signal in which the noise signal is suppressed in the observation signal is obtained.

  Next, the operation of the stereo signal processing apparatus 310 according to the third embodiment will be described with reference to FIG. In addition, about the process same as the process in 1st Embodiment, the same code | symbol is attached | subjected and detailed description is abbreviate | omitted.

In step 100, the A / D converters 12L and 12R convert the observation signals x _L (n) and x _R (n) into digital signals, respectively. Next, in step 300, the autocorrelation processing units 26L and 26R set the delay time as τ, and the autocorrelation functions R _xL (l, τ) of the L channel and the R channel according to the equations (14) and (15), respectively. , R _xR (1, τ) (τ = 0,..., L−1).

Next, in step 302, the peak value detectors 28L and 28R determine the peak values of the autocorrelation functions R _xL (l, τ) and R _xR (l, τ) calculated in step 300 other than τ = 0. p _L (l) and p _R (l) are detected.

Next, in step 304, the noise determination unit 30L determines whether or not the peak value p _L (l) is greater than the threshold σ ₁ . If p _L (l)> σ ₁ , the process proceeds to step 306, where the frame l is determined to be a vocal signal, the estimated noise signal of the previous frame is copied, and the estimated noise signal i ^ _L (1) of the frame l is copied. l, i). On the other hand, if p _L (l) ≦ σ ₁ , the process proceeds to step 308, where the frame l is determined as a noise signal and is directly used as the estimated noise signal i ^ _L (l, i).

Similarly, for the R channel, the noise determination unit 30R executes steps 304 to 308 to determine the estimated noise signal i ^ _R (l, i) of the frame l.

Next, at step 310, the noise suppressor 322, the estimated noise signal _{i ^ L (l, i)} , and _{i ^ R (l, i)} , the observed signal _{x L} (n), and _{x R} (n) Based on this, a noise signal is suppressed by executing a noise suppression process.

  The noise suppression processing corresponds to the colored drive moped Kalman algorithm shown in FIG. 8 as in the first embodiment. Here, a process different from the flow of the colored drive moped Kalman algorithm executed as the music signal estimation process of the first embodiment will be described.

In the colored drive moped Kalman algorithm executed as noise suppression processing, in step 1140, the initial value x ^ _p2 (0 | 0) of the optimum estimated value vector and the initial value of the covariance matrix that is an error when the state vector is estimated P _p2 (0 | 0), noise signal variance value matrix R _δp2 (n), i row j column element R _δp2 (n) [i, j], and vocal signal variance value matrix R _εp2 (n) Set the values as described above.

In step 1154, the estimated vocal signal d ^ (n) in the first row and first column or the (L _p2 +1) row and first column of the optimum estimated value vector x ^ _p2 (n + 1 | n + 1) calculated in step 1146 is used. That is, an estimated speech signal in which the noise signal is suppressed in the observation signal is output, and the process returns to the process of FIG.

As described above, according to the stereo signal processing apparatus of the third embodiment, in addition to the effects of the first embodiment, the color drive source type is used by using the estimated noise signal estimated using the autocorrelation. By applying the Kalman filter, it is possible to enhance the suppression effect against white noise. Moreover, since only the signal processing in the time domain is performed, the amount of calculation can be reduced.
<Fourth embodiment>
In the fourth embodiment, an example of the first signal of the present invention is a vocal signal (voice signal) having a sound source position near the center of the L channel microphone and the R channel microphone, for example. For example, a case where a noise signal is used will be described.

  A fourth embodiment will be described. In the fourth embodiment, a case will be described in which a noise signal is suppressed from an observation signal observed in the situation shown in FIG. For the stereo signal processing apparatus of the fourth embodiment, the same reference numerals are used for the same parts as the stereo signal processing apparatus 10 of the first embodiment and the stereo signal processing apparatus 310 of the third embodiment. Detailed description will be omitted.

  As shown in FIG. 15, the stereo signal processing apparatus 410 according to the fourth embodiment includes A / D conversion units 12L and 12R, frequency domain conversion units 14L and 14R, spectral density calculation units 32L and 32R, Spectral entropy calculation units 34L and 34R, noise determination units 430L and 430R, time domain conversion units 20L and 20R, a noise suppression unit 322, and D / A conversion units 24L and 24R are configured.

Based on the observed spectrums | X _L (l, k) | and | X _R (l, k) | inputted from the frequency domain conversion units 14L and 14R, the spectral density calculation units 32L and 32R are expressed by the following equation (20). and by (21), L-channel and R-channel observation signal of each of the power spectral density _{P L (l, k),} P R (l, k) is calculated, and spectral entropy calculation unit 34L, and inputs to 34R. l is a frame number, and k is a frequency bin number.

Here, the spectrum of the vocal signal (audio signal), considering that exist within the frequency band of 250～4000Hz, in the case of k ≦ 250 Hz or k ≧ 4000 Hz _{is, | X L (l, k} ) | = Let | X _R (l, k) | = 0.

Spectral entropy calculation unit 34L, 34R, the spectral density calculation unit 32L, the spectrum input from the 32R density _{P L} (l, k), based on _{P R} (l, k), the following expression (22) and (23) The spectral entropies H _L (l) and H _R (l) of the L channel and R channel observation signals are calculated according to the equations, and input to the noise determination units 430L and 430R.

The noise determination units 430L and 430R are based on the spectrum entropy H _L (l) and H _R (l) input from the spectrum entropy calculation units 34L and 34R, respectively, and the spectrum of the estimated noise signal (hereinafter referred to as “estimation”). Each of which is referred to as “noise spectrum” and output to the time domain conversion units 20L and 20R. Specifically, according to the following formulas (24) and (25), the spectral entropy H _L (l), H _R (l) is compared with the threshold σ ₂ and the spectral entropy is smaller than the threshold σ _2. Determines that the frame l is a vocal signal (voice signal), copies the estimated noise spectrum of the previous frame, and estimates the estimated noise spectrum | I ^ _L (l, k) |, | I ^ _R (l , K) | On the other hand, when the spectral entropy is larger than the threshold σ ₂ , the frame l is determined as a noise signal, and the estimated noise spectrum | I ^ _L (l, k) |, | I ^ _R (l, k) | .

Here, the threshold value σ ₂ is determined as follows. First, the threshold value σ ′ ₂ (l) is derived as follows using the average value of the spectral entropy for N frames.

Then 'compared ₂ (l) and a spectral entropy of the current frame, the threshold value sigma' threshold sigma alpha times the threshold sigma _'2 (l) If even smaller for spectral entropy of the current frame than ₂ (l) To do.

Then, after determining whether or not the past three frames are audio signals in succession, a final threshold σ ₂ (l) is obtained.

If the audio signal is not continuous, the final threshold σ ₂ (l) is obtained after determining whether or not the past three frame noise signals have been continuous.

The time domain transforming units 20L and 20R receive the estimated noise spectrums | I ^ _L (l, k) |, | I ^ _R (l, k) |, which are frequency domain signals input from the noise determination units 430L and 430R. Inverse Fourier transform is performed to convert the estimated noise signals i ^ _L (l, n) and i ^ _R (l, n), which are time domain signals. Next, using the overlap add method, the time domain estimated music signal i ^ _L (l-1, n + M), i ^ _R (l-1, n + M) using the latter half M samples of the previous frame and the first half of the current frame are used. The time-domain estimated music signal i ^ _L (l, n), i ^ _R (l, n) using M samples is added together to obtain the M-sample time-domain estimated music signal i ^ _L (n), i ^ _R (n) (1≤n≤M) is obtained.

  Next, the operation of the stereo signal processing device 410 according to the fourth embodiment will be described with reference to FIG. In addition, about the process same as the process in 1st Embodiment, the same code | symbol is attached | subjected and detailed description is abbreviate | omitted.

In step 100, the A / D converters 12L and 12R convert the observation signals x _L (n) and x _R (n) into digital signals, respectively, and then in step 102, the frequency domain converters 14L and 14R , The observed spectrum | X _L (l, k) |, | X _R (l, k) |

Next, in step 400, the power of each of the L channel and R channel observation signals based on the spectral density calculation units 32L and 32R and the observation spectrums | X _L (l, k) |, | X _R (l, k) | spectral density _{P L} (l, k), calculates the _{P R} (l, k).

Next, at step 402, spectral entropy calculation unit 34L, 34R is, the power spectrum calculated in the step 400 density _{P L} (l, k), based on _{P R (l, k),} L -channel and R-channel Spectral entropy H _L (l), H _R (l) of each observation signal is calculated.

Next, in step 404, the noise determination unit 430L is, the threshold value sigma ₂ was determined as described above, using the determined threshold value sigma _2, spectral entropy H _L (l) is whether a threshold value sigma ₂ is smaller than judge. If H _L (l) <σ ₂ , the process proceeds to step 406, where frame l is determined as a vocal signal, the estimated noise spectrum of the previous frame is copied, and the estimated noise spectrum | I ^ _{L of} frame l Let (l, k) | On the other hand, if H _L (l) ≧ σ ₂ , the process proceeds to step 408, where the frame l is determined to be a noise signal and is directly used as the estimated noise spectrum | I ^ _L (l, k) |.

Similarly, for the R channel, the noise determination unit 430R executes steps 404 to 408 to determine the estimated noise spectrum | I ^ _R (l, k) |

Next, in step 112, the time domain transforming units 20L and 20R convert the estimated noise spectrum | I ^ _L (l, k) |, | I ^ _R (l, k) | into the time domain using Fourier inverse transform. Are converted into estimated noise signals i ^ _L (l, n), i ^ _R (l, n). Next, using the overlap add method, the time domain estimated music signal i ^ _L (l-1, n + M), i ^ _R (l-1, n + M) using the latter half M samples of the previous frame and the first half of the current frame are used. The time-domain estimated music signal i ^ _L (l, n), i ^ _R (l, n) using M samples is added together to obtain the M-sample time-domain estimated music signal i ^ _L (n), i ^ _R (n) (1≤n≤M) is obtained.

Next, in step 410, the noise suppressor 322, based on the estimated noise signal i ^ _L (n), and i ^ _R (n), the observed signal _{x L} (n), _{x R} (n), and the noise The noise signal is suppressed by executing the suppression process. The noise suppression process is the same as that in the third embodiment.

  As described above, according to the stereo signal processing device of the fourth embodiment, in addition to the effects of the first embodiment, the color drive source type is used by using the estimated noise signal estimated using the spectral entropy. By applying this Kalman filter, it is possible to enhance the suppression effect against various white and colored noises.

  In addition, you may use the Kalman filter with a calculation amount reduction type color drive source which reduced the calculation amount of the Kalman filter with a color drive source used in the said 1st-4th embodiment. In the Kalman filter with a colored drive source that reduces the amount of computation, only processing necessary for estimating a desired signal is extracted.

  Specifically, as shown in FIG. 17, when only the portions indicating the estimated music signal of the L channel and the R channel are extracted in the state quantity update in the procedure 4, four elements of the Kalman gain matrix in the procedure 2 are necessary. I understand that. Therefore, as shown in FIG. 18, when only the necessary four element portions are taken out, it is understood that the four elements of the covariance matrix in the procedure 1 are necessary. Therefore, as shown in FIG. 19, when only the necessary four element portions are taken out, it can be seen that the dispersion value of the music signal is necessary.

  Summarizing the above, the Kalman algorithm with a colored drive source with reduced calculation amount is as shown below, and the amount of calculation can be reduced by reducing the number of steps. Note that p3 is a subscript indicating that it is a variable of the state equation and the observation equation to which the color reduction moped Kalman filter with a reduced calculation amount is applied.

Here, the amount-of-computation-reduced color driving original Kalman algorithm that is executed when the above-described amount-reduction-type colored driving mop Kalman filter is applied to the music signal estimation process (step 114 in FIG. 7) in the first embodiment. This flow will be described with reference to FIG.

In step 2140, the state space model is defined by the state equation and the observation equation shown in the equation (10), and the initial value x ^ _p3 (0 | 0) of the optimum estimated value vector and the error when the state vector is estimated are shared. the initial value _{P p3} of the dispersion matrix (0 | 0), the dispersion value of the vocal signals _{R εp3 (n) [i,} j], and variance of the music signal _{R δp3 (n) [i,} j] , and the above-mentioned initial Setting process Set to the initial state shown in [Initialization]. Also, a variable n indicating time is set to 0.

Next, in step 2142, a covariance matrix that is an error when the state vector at time n + 1 is estimated from information up to time n using the value of the variance value R _δp3 (n + 1) [i, j] of the music signal. P _p3 (n + 1 | n) is calculated (step 1 of the above iteration process [Iteration]).

Next, in step 2144, using the covariance matrix P _p3 (n + 1 | n) calculated in step 2142 and the variance value R _εp3 (n) [i, j] of the vocal signal, the Kalman gain matrix K _p3 ( n + 1) is calculated (same procedure 2).

Next, in step 2146, using the Kalman gain matrix K _p3 (n + 1) calculated in step 2144 and the observation vector y _p3 (n + 1), the optimum estimated value vector x ^ at that time by the information up to time n + 1. _p3 (n + 1 | n + 1) is calculated (same procedure 3).

Next, in step 2148, it is determined whether or not to end the process. In this determination, when the time n reaches a predetermined number N of samples, it may be determined that the process is ended, or may be determined when the sample is exhausted. If the process is not terminated, the process proceeds to step 2152, n is incremented by 1, and the process returns to step 1142. When the processing is ended, the process proceeds to step 2154, and the L-channel estimated music signal i ^ _L (in the first row and the first column of the optimum estimated value vector x ^ _p3 (n + 1 | n + 1) calculated in step 2146 above. n), (L _p3 +1) row 1st column is output as an R channel estimated music signal i ^ _R (n), and the process returns to the process of FIG.

Further, in the third and fourth embodiments, the case where the noise signal is suppressed has been described. As in the second embodiment, in the colored drive motivated Kalman algorithm, the variance value of the noise signal and the vocal signal By exchanging the variance value, a signal in which the vocal signal is suppressed, that is, an estimated noise signal may be extracted. Specifically, in step 1 of the iterative process [Iteration] of the colored drive motivated Kalman algorithm, the noise signal variance R _δp2 (n + 1) [i, j] and the vocal signal variance R _εp2 (n) [ i, j] are replaced, and a covariance matrix P _p2 (n + 1 | n), which is an error when the state vector at time n + 1 is estimated from information up to time n, is calculated. In the same procedure 2, the Kalman gain matrix K _p2 (n + 1) is replaced by replacing the variance value R _εp2 (n) [i, j] of the vocal signal with the variance value R _δp2 (n + 1) [i, j] of the noise signal. ). Then, the estimated noise signal i ′ ^ _L (l, i) is calculated from the first row and first column and the (L _p2 +1) row and first column of the optimum estimated value vector x ^ _p2 (n + 1 | n + 1) calculated in the procedure 4. ), I ′ ^ _R (l, i).

In addition, when the above-described computation amount-reducing Kalman algorithm is applied to the first embodiment (or when applied to the third and fourth embodiments), as in the second embodiment, By exchanging the variance value of the music signal (or noise signal) and the variance value of the vocal signal, a signal in which the vocal signal is suppressed, that is, an estimated music signal (or estimated noise signal) can be extracted. Specifically, in step 1 of the iteration process [Iteration] of the colored drive motivated Kalman algorithm with reduced amount of computation, the variance value R _δp3 (n + 1) [i, j] of the music signal (or noise signal) is expressed as a vocal signal. The covariance matrix P _p3 (n + 1 | n), which is an error when the state vector at time n + 1 is estimated from the information up to time n, is replaced with the variance value R _εp3 (n) [i, j]. Also, in the same procedure 2, the vocal signal variance R _εp3 (n) [i, j] is replaced with the variance value R _δp3 (n + 1) [i, j] of the music signal (or noise signal) to obtain the Kalman gain. The matrix K _p3 (n + 1) is calculated. Then, the estimated noise signal i ′ ^ _L (l, i) is calculated from the first row and first column and the (L _p3 +1) row and first column of the optimum estimated value vector x ^ _p3 (n + 1 | n + 1) calculated in the same procedure 3. ), I ′ ^ _R (l, i).

  The above embodiments can be applied in appropriate combination. For example, after extracting a desired signal according to the first or second embodiment, noise can be suppressed according to the third or fourth embodiment.

  In the first and second embodiments, the first signal is a vocal signal and the second signal is a music signal. In the third and fourth embodiments, the first signal is a vocal signal (audio). Signal), the case where the second signal is a noise signal has been described, but the present invention is not limited to this. In the input signals of a plurality of channels, each of the first signals is a signal included in common between the channels, and the second signal may be a signal that is different for each channel.

  Further, although cases have been described with the above embodiment where each unit is configured by hardware, a program for causing a computer to execute the processing of each unit may be employed. The program may be installed in the apparatus in advance, may be provided by being stored in a computer-readable recording medium, or may be provided via a network.

10, 210, 310, 410 Stereo signal processor 12L, 12R A / D converter 14L, 14R Frequency domain converter 16 Spectrum ratio calculator 18 Vocal signal extractor 20, 20L, 20R Time domain converter 22 Music signal estimator 24L, 24R D / A conversion units 26L, 26R Autocorrelation processing units 28L, 28R Peak value detection units 30L, 30R, 430L, 430R Noise determination units 32L, 32R Spectral density calculation units 34L, 34R Spectral entropy calculation units 222 Specific signal estimation 322 Noise suppression unit

Claims (11)

Each of the time domain observation signals of a plurality of channels including a first signal commonly included in each channel and a second signal different for each channel is converted into a spectrum signal in the frequency domain, and based on the ratio of each spectrum signal The estimated first spectrum signal estimated as the spectrum signal of the first signal is extracted, and the estimated first spectrum signal, which is a frequency domain signal, is converted into a time domain signal to be estimated as the first signal. Extracting means for extracting an estimated first signal of the region;
The variance value of the estimated first signal in the time domain extracted by the extraction means, the variance value of the second signal obtained by subtracting the variance value of the estimated first signal from the variance value of the observation signal, and the plurality A state equation composed only of the second signal including an element corresponding to the plurality of channels using an observation signal of the channel, and composed of the first signal and the second signal including an element corresponding to the plurality of channels An estimation means for estimating the first signal or the second signal by applying a Kalman filter with a colored drive source to the state space model represented by the observed equation ;
A multi-channel signal processing apparatus. Based on the autocorrelation peak value of each of the time domain observation signals of the plurality of channels including the first signal or the second signal estimated by the estimation means, the time domain estimation A subsequent extraction means for extracting two signals;
A variance value of the estimated second signal in the time domain extracted by the subsequent extraction means, a variance value of the first signal obtained by subtracting the variance value of the estimated second signal from the variance value of the observation signal, and the Post-stage estimation means for estimating the first signal or the second signal by a Kalman filter with a colored drive source using observation signals of a plurality of channels;
The multi-channel signal processing apparatus according to claim 1, comprising: Each of the time domain observation signals of a plurality of channels including the first signal or the second signal estimated by the estimation means is converted into a spectrum signal in the frequency domain, and the first signal is obtained based on the spectrum entropy obtained from each spectrum signal. An estimated second spectrum signal that is estimated to be a spectrum signal of two signals is extracted, the estimated second spectrum signal that is a frequency domain signal is converted into a time domain signal, and a time domain signal that is estimated to be the second signal is converted. A subsequent extraction means for extracting the estimated second signal;
A variance value of the estimated second signal in the time domain extracted by the subsequent extraction means, a variance value of the first signal obtained by subtracting the variance value of the estimated second signal from the variance value of the observation signal, and the Post-stage estimation means for estimating the first signal or the second signal by a Kalman filter with a colored drive source using observation signals of a plurality of channels;
The multi-channel signal processing apparatus according to claim 1, comprising: The second signal is estimated based on the autocorrelation peak value of each of the observation signals in the time domain of a plurality of channels including the first signal that is commonly included among the channels and the second signal that is different for each channel. Extracting means for extracting an estimated second signal in the time domain;
The variance value of the estimated second signal in the time domain extracted by the extraction means, the variance value of the first signal obtained by subtracting the variance value of the estimated second signal from the variance value of the observation signal, and the plurality A state equation composed only of the second signal including an element corresponding to the plurality of channels using an observation signal of the channel, and composed of the first signal and the second signal including an element corresponding to the plurality of channels An estimation means for estimating the first signal or the second signal by applying a Kalman filter with a colored drive source to the state space model represented by the observed equation ;
A multi-channel signal processing apparatus. Each of the time domain observation signals of a plurality of channels including a first signal commonly included in each channel and a second signal different for each channel is converted into a spectrum signal in the frequency domain, and each power spectral density is calculated. The second spectrum signal estimated as the spectrum signal of the second signal is extracted based on the spectrum entropy respectively obtained from the power spectrum density, and the estimated second spectrum signal, which is a frequency domain signal, is extracted in the time domain. Extracting means for converting to a signal and extracting an estimated second signal in the time domain estimated as the second signal;
The variance value of the estimated second signal in the time domain extracted by the extraction means, the variance value of the first signal obtained by subtracting the variance value of the estimated second signal from the variance value of the observation signal, and the plurality A state equation composed only of the second signal including an element corresponding to the plurality of channels using an observation signal of the channel, and composed of the first signal and the second signal including an element corresponding to the plurality of channels An estimation means for estimating the first signal or the second signal by applying a Kalman filter with a colored drive source to the state space model represented by the observed equation ;
A multi-channel signal processing apparatus. The extraction means converts the observation signal into a spectrum signal in the frequency domain for each frame of a predetermined frame length, obtains the spectrum entropy for each frame, and an average σ ′ of spectrum entropy for a first predetermined number of frames is When the spectral entropy of the current frame is smaller than σ ′, and when larger, σ ′ is multiplied by a predetermined coefficient α to obtain a value σ ″ that is ασ ′, and the first signal continues for the second predetermined number of frames in the past. Σ ″ if the first signal is not continuous for the second predetermined number of frames in the past, and σ ′ if the second signal is continuous for the second predetermined number of frames in the past. not continuous past second number given frame, and when said second signal is not continuous past number second predetermined frame is a threshold sigma "sigma, less than the spectral entropy threshold sigma of the current frame If stomach, the current frame is determined as the first signal, when the spectral entropy is above a threshold value σ of the current frame, the multi-channel signal processor of the present frame is judged according to claim 5, wherein the second signal. Each of the time domain observation signals of a plurality of channels including a first signal commonly included in each channel and a second signal different for each channel is converted into a spectrum signal in the frequency domain, and based on the ratio of each spectrum signal The estimated first spectrum signal estimated as the spectrum signal of the first signal is extracted, and the estimated first spectrum signal, which is a frequency domain signal, is converted into a time domain signal to be estimated as the first signal. Extracting means for extracting an estimated first signal of the region;
The variance value of the estimated first signal in the time domain extracted by the extraction means, the variance value of the second signal obtained by subtracting the variance value of the estimated first signal from the variance value of the observation signal, and the plurality An estimation means for estimating the first signal or the second signal by a Kalman filter with a colored driving source with a reduced amount of calculation using an observation signal of the channel,
The Kalman filter with a colored drive source that reduces the amount of computation is an element related to the time n + 1 among the elements of the Kalman filter with a colored drive source that estimates the first signal or the second signal at the time n + 1 using the observation signal up to the time n. This is a Kalman filter that extracts only
Multi- channel signal processing device. Each of the time domain observation signals of a plurality of channels including a first signal commonly included in each channel and a second signal different for each channel is converted into a spectrum signal in the frequency domain, and based on the ratio of each spectrum signal The estimated first spectrum signal estimated as the spectrum signal of the first signal is extracted, and the estimated first spectrum signal, which is a frequency domain signal, is converted into a time domain signal to be estimated as the first signal. Extracting an estimated first signal of the region;
The extracted variance value of the estimated first signal in the time domain, the variance value of the second signal obtained by subtracting the variance value of the estimated first signal from the variance value of the observed signal, and the observed signals of the plurality of channels And a state equation composed only of the second signal including elements corresponding to the plurality of channels, and an observation equation composed of the first signal and the second signal including elements corresponding to the plurality of channels. A multichannel signal processing method for estimating the first signal or the second signal by applying a Kalman filter with a colored drive source to a state space model represented by : The second signal is estimated based on the autocorrelation peak value of each of the observation signals in the time domain of a plurality of channels including the first signal that is commonly included among the channels and the second signal that is different for each channel. Extract an estimated second signal in the time domain;
The extracted variance value of the estimated second signal in the time domain, the variance value of the first signal obtained by subtracting the variance value of the estimated second signal from the variance value of the observed signal, and the observed signals of the plurality of channels And a state equation composed only of the second signal including elements corresponding to the plurality of channels, and an observation equation composed of the first signal and the second signal including elements corresponding to the plurality of channels. A multi-channel signal processing method including: estimating the first signal or the second signal by applying a Kalman filter with a colored drive source to a state space model represented by : Each of the time domain observation signals of a plurality of channels including a first signal commonly included in each channel and a second signal different for each channel is converted into a spectrum signal in the frequency domain, and each power spectral density is calculated. The second spectrum signal estimated as the spectrum signal of the second signal is extracted based on the spectrum entropy respectively obtained from the power spectrum density, and the estimated second spectrum signal, which is a frequency domain signal, is extracted in the time domain. Extracting a second estimated signal in the time domain that is converted into a signal and estimated as the second signal;
The extracted variance value of the estimated second signal in the time domain, the variance value of the first signal obtained by subtracting the variance value of the estimated second signal from the variance value of the observed signal, and the observed signals of the plurality of channels And a state equation composed only of the second signal including elements corresponding to the plurality of channels, and an observation equation composed of the first signal and the second signal including elements corresponding to the plurality of channels. A multichannel signal processing method for estimating the first signal or the second signal by applying a Kalman filter with a colored drive source to a state space model represented by : The multi-channel signal processing program for functioning a computer as each means which comprises the multi-channel signal processing apparatus of any one of Claims 1-7.