JP2003271168A - Method, device and program for extracting signal, and recording medium recorded with the program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a signal extracting method, a signal extracting device and a signal extraction program for estimating a separation filter with a length sufficient for separation while holding the assumption of the independence a signal in each band by adopting a configuration for applying subband analysis combination to a plurality of mixed signals to calculate an output signal corresponding to the original signal, and to provide a recording medium recorded with the program. <P>SOLUTION: This signal extracting method is for separatingly extracting the original signal from the plurality of mixed signals observed through a path having a long impulse response on the basis of its independence. By using the signal extracting method, the signal extracting device, the signal extraction program and the recording medium recorded with the program, the plurality of mixed signals are inputted to a subband analyzing part 51, the mixed signals are respectively subjected to a subband analysis into N (N is positive integer) bands, the subband-analyzed signal of each band is inputted to a time domain BSS part 53 of a corresponding band to be subjected to sound source separation in each band, and the sound source-separated signal is inputted to a subband combining part 55 to calculate an output signal corresponding to the original signal. <P>COPYRIGHT: (C)2003,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a signal extracting method, a signal extracting device, a signal extracting program and a recording medium having the program recorded therein. In particular, an original signal to be observed cannot be directly observed and noise and other Under the condition that signals are observed in a superposed state, it is a technology that estimates the original signal to be observed. A signal extracting method and a signal extracting apparatus capable of extracting a target speaker's voice and performing voice recognition with a high recognition rate even in a situation where sound is picked up,
The present invention relates to a signal extraction program and a recording medium recording the program.

[0002]

The problem of estimating multiple linearly mixed signals without using any knowledge of the original signal and the mixing process is the Blind Source Separation:
BSS), but the invention of this application also belongs to this blind sound source separation technology. The technique of separating linearly mixed signals based on the statistical independence between the signals is called Independent Component Analysis (ICA). A signal that is linearly mixed with the impulse response of the recording device convoluted into the signal, like a sound pickup in a real sound field,

[0003]

[Equation 1]

It is expressed as follows. Here, x _j : signal s _i observed by sensor _j : signal h _{ji of} signal source i: impulse response (linear system) of P tap from signal source i to microphone j. When the blind sound source separation is practically used, the impulse response length P is often large. For example, in the case of mixed speech observed in a real environment, reverberation will be convoluted with the speech. Even in the reverberation time of a general conference room of about 150 to 300 ms, the impulse response length P is as long as several thousand.

In independent component analysis, there are N signal sources.
The signals emitted from them are assumed to be statistically independent of each other
However, the observation signal obtained in the form of equation (1) and the length are Q taps.
Separation filter group w_ijUsing a separation system consisting of
Put out. This separation filter group w _ijSeparation extraction using
Signal y obtained by_i(N) is

[0006]

[Equation 2]

It is expressed as FIG. 1 is a diagram for explaining the above mixing and separating process when N = M = 2. The separating system estimates the outputs y _i to be independent of each other using some learning rule w _ij ^{k + 1} = w _ij ^k + Δw _ij ^k (3). k indicates the number of learning updates. Since this problem is a complicated problem of convolutional mixing, it is difficult to directly obtain the separation filter group. Therefore, the discrete Fourier transform (DF
The method of converting to the frequency domain by T) is often adopted. This is called frequency domain blind source separation (frequency domain BBS). First, the equation (1) is transformed into the frequency domain by DFT.

X (ω, m) = H (ω) S (ω, m) (4) Thus, the convolutional mixing problem can be expressed as an instantaneous mixing problem at each frequency, and the problem can be simplified. As described above, the estimation of the separation process is performed so that the output signals Y ₁ (ω, m) and Y ₂ (ω, m) are independent of each other at each frequency. It suffices to estimate the separation matrix W (ω) of 2). Y (ω, m) = W (ω) X (ω, m) (5) The reverberation time is 1 with a large impulse response length P as in a real environment.
In the case of about 50 to 300 ms or more, it is necessary to find a separation filter having a length comparable to this. Therefore, when obtaining the separation matrix W (ω) in the frequency domain BSS, it is necessary to perform the DFT analysis using the frame T longer than the impulse response length P of the room to increase the number of frequency bins. However, when the learning data of a fixed length is analyzed using a long frame, the number of data at each frequency is reduced, and the statistical property of the data is deteriorated at each frequency.

This will be described with reference to FIG. In the case of FIG. 2A where the frame length is short, the number of data at each frequency is sufficient, so that the statistical properties are sufficiently guaranteed at each frequency. However, the length of the separation filter that can be estimated is short, which is insufficient. on the other hand,
In the case of FIG. 2B where the frame length is long, a long separation filter can be prepared, but the statistical property of the data deteriorates because the number of data at each frequency is small.
The problem that the statistical properties deteriorate will be described with reference to FIG.

In FIG. 3, the correlation coefficient γω (ω is a subscript) is calculated at all frequencies and averaged.

[0011]

[Equation 3]

[0012] is shown. It can be seen that when the frame length T is large, the correlation between signals becomes high and the assumption of independence necessary for independent component analysis is broken. Thus, in the frequency domain BSS, it is necessary to perform the DFT analysis using the frame T longer than the impulse response length P of the room and increase the number of frequency bins. Separation becomes difficult because it deteriorates.

[0013]

Conventional frequency domain BS
At S, a large frame length T corresponding to long reverberation
When using, the correlation between signals became high, and the assumption of independence necessary for independent component analysis collapsed, and high performance could not be obtained. Therefore, the present invention adopts a configuration in which a mixed signal of a plurality of mixed signals is subjected to subband analysis and synthesis to obtain an output signal corresponding to an original signal in a BSS having a large impulse response length P of several thousand, and a signal in each band is adopted. The present invention provides a signal extraction method and a signal extraction device for estimating a separation filter having a sufficient length for separation while maintaining the independence assumption of (1), a signal extraction program, and a recording medium recording the program.

[0014]

In a signal extraction method for separating and extracting an original signal based on its independence from a plurality of mixed signals observed through a path having a long impulse response, the plurality of mixed signals are subbanded. The signals are input to the analysis unit and subband-analyzed into N (N: integer) bands, and the signals of each band subjected to the subband analysis are input to the time-domain BSS unit of the corresponding band to separate the sound sources for each band. Then, a signal extraction method for obtaining the output signal corresponding to the original signal by inputting the source-separated signal to the subband synthesis unit was constructed.

Then, in a signal extracting device for separating and extracting an original signal from a plurality of mixed signals observed through a path having a long impulse response on the basis of their independence, a plurality of mixed signals are input to each of N signals. A time domain BSS that includes a subband analysis unit that analyzes (N: integer) bands and separates sound sources for each signal in each band subjected to subband analysis.
A signal extraction device having a subband synthesizing unit for inputting a signal separated from a sound source and obtaining an output signal corresponding to the original signal is configured. Here, each of the plurality of mixed signals is subband-analyzed into N (N: integer) bands, the signals of each band subjected to the subband analysis are separated into sound sources for each band, and the original signals are extracted from the signals separated in each band. A signal extraction program was constructed to execute the calculation of the output signal corresponding to the signal.

Then, the plurality of mixed signals are subband-analyzed into N (N: integer) bands, the subband-analyzed signals of each band are separated into sound sources for each band, and the sound source is separated into each band. A recording medium having a signal extraction program for executing the operation of obtaining an output signal corresponding to the original signal is constructed.

[0017]

BEST MODE FOR CARRYING OUT THE INVENTION The present invention utilizes subband analysis to perform signal separation in each band. This is a sub band BS
It is called S. Since the subband BSS can freely select the number of subbands to be divided, as shown in FIG. 2C, it is possible to select the number of band divisions that sufficiently satisfies the statistical properties of each subband. And frequency domain BS
S was able to estimate only a 1-tap filter at each frequency, but subband BSS can have a long filter for each band as shown in FIG. It is possible to estimate a sufficiently long filter when viewed in the full band even if there are few.

From the above two points, the subband BSS is used to estimate a long separation filter capable of coping with reverberation while maintaining the statistical property of data in each band even when the impulse response length P is long. can do.

[0019]

Embodiments of the present invention will be described with reference to the embodiment of FIG. FIG. 4 is a diagram showing the entire subband BSS. (1) Subband analysis process First, the input observed signals x ₁ (n) and x ₂ (n) are input to the subband analysis units 51 ₁ and 51 ₂ , respectively, and subband analyzed. (2) Sound source separation process Next, the component of the signal subjected to the subband analysis in each band of the observed signals x ₁ (n) and x ₂ (n) has the number of band divisions as N,
Time domain BSS units 53 ₁ , ...
Input to 53 _N to separate sound sources.

(3) Sub-band synthesis process Finally, the signal components separated into each band are divided into each time domain B
From SS section 53 ₁ , ..., 53 _N to subband synthesis section 55
_The signals y ₁ and y ₂ corresponding to the signals s ₁ (n) and s ₂ (n) are synthesized and output by inputting to ₁ and 55 ₂ . As described above, the signal extraction device according to the present invention, which separates and extracts the original signal based on its independence from a plurality of mixed signals observed via a path having a long impulse response, inputs a plurality of mixed signals and A sub-band analysis unit 51 that analyzes N (N: integer) bands, a time-domain BSS unit 53 that separates sound sources for each signal of each band subjected to sub-band analysis, and a signal that has been subjected to sound source separation as input And a subband synthesis unit 55 for obtaining an output signal corresponding to the signal.

Next, referring to FIG. 5, in detail,
First, the subband analysis process will be described. (1)
The subband analysis process is performed by the subband analysis units 51 ₁ and 5
1 ₂ and SSB modulators 52 ₁ and 52 ₂ . When the number of band divisions is N and the thinning rate is M, x
The signal after thinning in the k-th band of _j (n) is

[0022]

[Equation 4]

Is calculated as Where W _N = exp (j
2π / N). Further, h (n) is a low-pass filter in the band [−π / N, π / N] used for analysis, and h (n) = {sin (n / N)} / (n / N) (8) Is used. At this time, X (k, m) is obtained as a complex number, but the time domain BSS unit 53 in the sound source separation process in the latter stage is
_1, ..., if 53 _N is a real algorithm, in order to handle the signal in real each band, for example, it can be used SSB (single side band) modulation unit 52. The subband using the SSB modulator 52 has a thinning rate M of M when the number of band divisions is N in order to avoid aliasing in the frequency domain.
= N / 4.

The real number signal by SSB modulation is X
_When expressed as _j ^ssb (k, m), X _j ^ssb (k, m) = Re [X _j (k, m)] cos (mπ / 2) + Im [X _j (k, m)] sin (mπ / 2) Obtained by (9). Next, the sound source separation process will be described. (2) The sound source separation process includes N time-domain BSS units 53.
_1, ..., constituted by 53 _N. Here, as an example of the time domain BSS algorithm used in each band, the one derived from the evaluation function based on the non-stationarity of the signal is shown. For simplicity, the output signal Y _i ^ssb (k, n) is y
_It is abbreviated as _i (n).

Non-negative evaluation function having a minimum value of 0 when the cross-correlation of output signals becomes 0 in all time blocks

[0026]

[Equation 5]

Consider Where y (n) = [y
₁ (n), y ₂ (n)] ^T is an output signal, and R _y ^b (τ)
Covariance matrix of the output signal (= <y (n) y T (n-
τ)> _b ) and <x> _b is a block b (b = 1, ...
..., B) is a time average. Separation filter w
_In the natural gradient calculated by differentiating this evaluation function Q with w (k), the update formula of _ij is not only R _y ^b (0) but also the time-shift correlation R _y ^b (for increasing the performance. By taking τ) into consideration, we can obtain

[0028]

[Equation 6]

Equation (11) is derived by T. Nishikawa, H. Saruw
atari, K. Shikano, “Comparison ofblind source separ
ation methods based on time-domain ICA using nonst
ationarity and multistage ICA, ”IEICE Tech.Rep., Ja
n.2002. Using w _ij (element of w) obtained by this updating formula, the separated signal is obtained by using formula (2). This separated signal is Y in FIG.
_i ^ssb (k, n). Although the above method uses the second-order statistics, it is also possible to adopt the method using the higher-order statistics.

Finally, the subband synthesis process will be described. (3) The SSB demodulation unit 5 performs the subband synthesis process.
4 ₁ and 54 ₂ and sub-band synthesis units 55 ₁ and 55 ₂ . First, the signals Y _i ^ssb (k,
m) is SSB demodulated by the SSB demodulators 54 ₁ and 54 ₂ . Re [Y _i (k, m)] = Y _i ^ssb (k, m) cos (m
π / 2) Im [Y _i (k, m)] = Y _i ^ssb (k, m) sin (m
π / 2) This demodulated signal Y _i (k, m) is output to the subband synthesis unit 5
Entered in 5,

[0031]

[Equation 7]

A signal y _i (n) synthesized by is obtained. Here, f (n) is a band [−π /
M, π / M] low-pass filter, and f (n) = {sin (n / M)} / (n / M) (13) is used. Here, the operation flow chart of the embodiment is shown in FIG. Then, the above signal extraction device may be configured by using an electronic computer as a main constituent member.
Further, the present invention can be implemented by installing a program downloaded from a storage medium such as a CD or downloaded via a communication line into this electronic computer.

[0033]

As described above, according to the present invention, it is possible to estimate a long filter in each band while maintaining the statistical properties of each band. Therefore, high separation performance of the original signal is expected. To be done. The effect of this subband BSS is shown in FIG. Here, the number of band divisions N =
64, thinning rate M = 16, separation filter length Q in each band
_{su b} = 64. This corresponds to the decimation rate M of "32" in the frequency domain BSS, and the filter length of the full band to 1024 taps. The experiment was conducted for two cases where the room reverberation was 150 ms and 300 ms.

In FIG. 6, the numbers on the horizontal axis indicate the frequency domain B.
It is the frame length (= filter length) in SS, and is the result of the frequency domain BSS using the frame length. Further, “SUB” is the result of the subband BSS. In the frequency domain BSS, the performance deteriorates when a long separation filter with a length of 1024 is obtained, but with the subband BSS, a separation filter with a length of 1024 can be obtained, and high separation performance is obtained. The value of the correlation coefficient in equation (6) obtained by dividing the original signal into subbands is
The male and female voices were 0.028, the male and female voices were 0.018, and the female and female voices were 0.020. Therefore, it can be considered that the assumption of independence is sufficiently maintained.

[Brief description of drawings]

FIG. 1 is a diagram showing a signal separation model.

FIG. 2 is a diagram for explaining the statistical properties and filter length of each frequency band.

FIG. 3 is a diagram showing that the assumption of independence is broken at a large frame size in the frequency domain BSS.

FIG. 4 is a diagram illustrating an example of a subband BSS.

FIG. 5 is a diagram illustrating details of an embodiment of a subband BSS.

FIG. 6 is a diagram for explaining the effect of the embodiment.

FIG. 7 is a flowchart of an embodiment.

[Explanation of symbols]

51 Sub-band analysis section 53 hours BSS department 55 Subband synthesizer

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Ryo Mukai             2-3-1, Otemachi, Chiyoda-ku, Tokyo             Inside Telegraph and Telephone Corporation (72) Inventor Hiroshi Sawada             2-3-1, Otemachi, Chiyoda-ku, Tokyo             Inside Telegraph and Telephone Corporation (72) Inventor Hiroshi Saruwatari             8916-5 Takayama, Ikoma City, Nara D-307 F-term (reference) 5D015 EE05

Claims (4)

[Claims] 1. A signal extraction method for separating and extracting an original signal from a plurality of mixed signals observed through a path having a long impulse response based on their independence, and inputting the plurality of mixed signals to a subband analysis unit. Then, subband analysis is performed on each of N (N: integer) bands, and the signals of each band subjected to the subband analysis are time-domain B of the corresponding band.
A signal extraction method characterized by inputting to a SS section, separating sound sources for each band, and inputting the separated sound source signals to a subband combining section to obtain an output signal corresponding to an original signal. 2. A signal extraction device for separating and extracting an original signal from a plurality of mixed signals observed via a path having a long impulse response based on their independence, and inputting a plurality of mixed signals, each of which is N in number. A sub-band analysis unit for analyzing the (N: integer) band is provided, and a time-domain BSS unit for separating the sound source for each signal of each band subjected to the sub-band analysis is provided. A signal extraction device comprising a subband synthesis unit for obtaining an output signal corresponding to a signal. 3. A signal obtained by subband-analyzing a plurality of mixed signals into N (N: integer) bands, separating the subband-analyzed signals in each band into sound sources, and separating the sound sources into each band. A signal extraction program that executes the process of obtaining the output signal corresponding to the original signal from. 4. A signal obtained by subband-analyzing a plurality of mixed signals into N (N: integer) bands, separating the subband-analyzed signals of each band into sound sources, and separating the sound sources into respective bands. A recording medium that records a signal extraction program that causes the output signal corresponding to the original signal to be obtained.