JP2006323134A - Signal extractor - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a signal extractor which reduces discrimination errors of an objective signal and miscellaneous signals and extracts the objective signal at high SN ratio. <P>SOLUTION: The signal extractor comprises: at least one or more signal input means which receive signals from an objective signal source and a miscellaneous signal source; a signal feature level calculation means for calculating a signal feature level of a signal received by the signal input means; a weight value determination means which determines a weight value α for multiplying thereby the received signal, based on the value of the signal feature level calculated by the signal feature level calculation means; and a weight value multiplying means which multiplies the received signal by the weight value determined by the weight value determination means. As the signal feature levels, a dispersion value about the time base of cepstrum, a dispersion value about the time base of an auto-correlation function, sharpness are used as necessary. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a signal extraction apparatus that suppresses a miscellaneous signal and extracts a target signal with a high S / N ratio in an environment in which signals are emitted from a target signal source in the vicinity and a distant signal source.

The environment where the target signal and the miscellaneous signal are emitted without overlapping in time (for example, when the signal is a sound, the environment in which the speakers change in turn, or the unsteady noise that occasionally sounds suddenly) In the existing environment), a noise gate method that performs threshold processing using power has been proposed as a method for suppressing a distant signal in the distance and emphasizing a target signal in the vicinity (for example, non-existing). Patent Document 1).
Internet <http://www.jiten.com/dicmi/docs/k25/20745s.htm> [Search May 16, 2005]

  However, when only power is used for information, for example, a miscellaneous signal with a large distant power and a target signal with a small power emitted from a nearby position are mistakenly output, and an unnecessary miscellaneous signal is output. There is a problem that a large target signal is excessively suppressed.

  Therefore, an object of the present invention is to provide a signal extraction apparatus that reduces the discrimination error between a target signal and a miscellaneous signal as compared with the conventional method and extracts a target signal with a high S / N ratio.

  If the target signal source is closer to the signal input means than the miscellaneous signal source (for example, a microphone if the signal is sound), the distant miscellaneous signal has a large amount of reflection and reverberation superimposed on the nearby target signal. Has a feature that miscellaneous signals are not superposed. By calculating a physical quantity (signal feature quantity) that can detect this feature, it is determined whether the signal is emitted from a distant position or a close position, and only a target signal that is close is extracted.

It is known that the degree of signal reflection and reverberation can be observed from the magnitude of the fluctuation of the higher-order component of the signal cepstrum (for example, 50 ms or more in quefrency). Similarly, it is known that it can be observed from fluctuations in the autocorrelation function of the signal.
(Reference 1): Alan V. Oppenheim, Ronald W. Schafer, Translated by Gen Date, "DIGITAL SIGNAL PROCESSING", First Edition, Volume 2, Corona, Inc., June 20, 1986

It is also known that the degree of signal reflection and reverberation can be observed from the sharpness of the signal.
(Reference 2): Bradford W. Gillespie, Henrique S. Malvar and Dinei AF Florencio, “Speech Dereverberation Via Maximum-Kurtosis Subband Adaptive Filtering”, International Conference on Acoustics, Speech and Signal Processing, 2001

  Further, the dispersion value on the time axis of higher cepstrum is small for the proximity signal and large for the far signal. The variance value of the autocorrelation function with respect to the time axis is also small for the near signal and large for the far signal. Further, the sharpness is large for the proximity signal and small for the far signal.

FIG. 1 shows that when the signal is a sound, the cepstrum high-order dispersion when the target sound signal from a position about 50 cm away from the microphone and the noise signal from a position about 3 m away from the microphone are observed for about 30 seconds. An example of the value is shown. The horizontal axis represents the number of frames in the frame time when short-time Fourier transform is performed in the cepstrum, and the vertical axis represents the cepstrum higher-order dispersion value. The solid line is the target sound signal (near sound) in the vicinity (about 50 cm from the microphone), and the dotted line is the noise signal (far sound) in the distance (about 3 m from the microphone).
FIG. 2 shows an example of the variance value of the autocorrelation function under the same conditions as described above (in the case of FIG. 1). The horizontal axis represents the number of frames in the frame time when calculating the autocorrelation function, and the vertical axis represents the variance value of the autocorrelation function.
FIG. 3 shows an example of sharpness under the same conditions as above (in the case of FIG. 1). The horizontal axis represents the number of frames in the frame time when calculating the sharpness, and the vertical axis represents the sharpness.

  Therefore, according to the present invention, the cepstrum high-order variance value of the received signal received by the signal input means, the variance value of the autocorrelation function, or the sharpness value is calculated, and the weight when outputting the signal according to the value The value α is calculated.

When the cepstrum high-order variance value of the received signal is used as the signal feature quantity, the cepstrum variance is obtained by utilizing the fact that the cepstrum variance value of the signal from the target signal source is smaller than the cepstrum variance value of the signal from the miscellaneous signal source. A section in which the value is smaller than a predetermined threshold is determined as a target signal, and a weight value α to be multiplied with the received signal is determined to be a predetermined value. The section in which the cepstrum variance value is larger than the threshold is determined as a section of a miscellaneous signal, and the weight value α is determined to be a predetermined value. The value of the weight value α in each determined section is set to an appropriate value so that the signal from the target signal source and the signal from the miscellaneous signal source can be separated. For example, if α is 0 ≦ α ≦ 1.0, the weight value α multiplied by the received signal in the section determined as the target signal is determined to be 1.0 or a value close to 1.0, and determined as a miscellaneous signal. The weight value α to be multiplied with the received signal in the section may be determined to be 0 or a value close to 0.
Similar processing is performed when the variance value of the autocorrelation function of the received signal is used as the signal feature amount.

  In addition, when using the sharpness of a received signal as a signal feature quantity, the sharpness of a signal from a nearby target signal source is large, and the sharpness of a signal from a distant signal source is small. A section where the sharpness is greater than or equal to a threshold value is determined as a target signal, and the weight value α to be multiplied with the received signal is determined to be a predetermined value. If the sharpness is less than or equal to the threshold value, it is determined as a miscellaneous signal component, and the weight value α is determined to be a predetermined value. The value of the weight value α in each determined section is set to an appropriate value so that the signal from the target signal source and the signal from the miscellaneous signal source can be separated. For example, if α is 0 ≦ α ≦ 1.0, the weight value α multiplied by the received signal in the section determined as the target signal is determined to be 1.0 or a value close to 1.0, and determined as a miscellaneous signal. The weight value α to be multiplied with the received signal in the section may be determined to be 0 or a value close to 0.

  The weight value multiplying means multiplies each band of the received signal x (t) (t: sampling time) by the determined weight value α. That is, the reception signal in the section determined as the target signal is multiplied by a predetermined weight value α (in the above example, 1.0 or a value close to 1.0), and the reception signal in the section determined as the miscellaneous signal is received. The signal is multiplied by a predetermined weight value α (in the above example, 0 or a value close to 0). The weighted reception signal is output as an output signal.

  In addition, the signal feature values (the cepstrum high-order variance, the autocorrelation function variance, and the sharpness) may be used alone or in combination. When there is a margin in the amount of calculation, an improvement in discrimination accuracy can be expected by combining a plurality of calculations. The signal feature amount can be combined with power.

  Furthermore, the above-mentioned plurality of signal feature amounts are calculated from a known target signal and a miscellaneous signal, a regression equation based on multiple regression analysis is obtained from these values, and the signal feature amount of the received signal is applied to the regression equation. The weight value can also be determined from the obtained objective variable.

  According to the signal extraction apparatus of the present invention, it is possible to reduce the discrimination error between the target signal and the miscellaneous signal and extract the target signal with a high S / N ratio as compared with the conventional method in which discrimination is performed using only power. it can.

  FIG. 4 is a functional block diagram of the signal extraction device (A) according to the first embodiment and the second embodiment. FIG. 5 shows a flowchart of signal extraction processing in the signal extraction apparatus (A) according to the first embodiment and the second embodiment. In these embodiments, the signal is described as an acoustic signal such as voice or musical sound. The acoustic signal input unit (1) which is a signal input means is a microphone, for example. Let s (t) be the acoustic signal (target sound signal) of the target sound source that is the target signal source, and n (t) be the acoustic signal (noise signal) of the noise source that is the miscellaneous signal source. In order to simplify the description, a single noise source is described here, but in general, a plurality of noise sources may be used.

<First Embodiment>
First, a first embodiment of the signal extraction device of the present invention will be described.

The acoustic feature quantity calculation unit (2), which is a signal feature quantity calculation means, calculates an acoustic feature quantity τ that is a signal feature quantity of the received signal x (t) received by the acoustic signal input unit (1) (S100). ). The acoustic feature amount τ is, for example, one of a cepstrum high-order dispersion value τ ₁ , an autocorrelation function dispersion value τ ₂ , and a sharpness τ ₃ of the received signal.

  In the first embodiment, any one of these acoustic feature quantities is used alone, but an improvement in accuracy can be expected by using a plurality of them in combination. This case will be described in the second embodiment.

  Below, the definition of cepstrum, autocorrelation function, and sharpness will be explained.

《Cepstrum》
The cepstrum of the received signal x (t) is defined by equation (1).

Where fft (・) is the Fourier transform of the input ・ abs (・) is the absolute value of the input ・ log (・) is the common logarithm of the input ・ ifft (・) is the inverse Fourier transform of the input ・ and real (•) represents the real part of input. The acoustic feature quantity τ calculated by the acoustic feature quantity calculation unit (2) is the variance value of the high-order component of the cepstrum defined by the expression (1) in the case of the cepstrum high-order variance value τ ₁ of the received signal. .

《Autocorrelation function》
The autocorrelation function of the received signal x (t) is defined by equation (2).

Here, N represents the length of the signal for calculating the correlation, and m represents the correlation shifted by m samples. The acoustic features tau calculated in acoustic feature amount calculation unit (2), if the variance value tau ₂ of the autocorrelation function is a variance value of the autocorrelation function defined by equation (2).

《Sharpness》
Let y (t) be the linear prediction residual signal of the received signal x (t). The sharpness of the signal y (t) is defined by the following formula (3).

Here, E is the expected value of the signal. The acoustic features tau calculated in acoustic feature amount calculation unit (2), if the sharpness tau _3, a sharpness defined by equation (3).

  In the weight value determining unit (3), which is a weight value determining means, a weight value for multiplying each band of the received signal x (t) based on the value of the acoustic feature amount τ calculated by the acoustic feature amount calculating unit (2). α is determined (S101). For example, when a cepstrum high-order variance value is used as the acoustic feature amount, the cepstrum variance value is obtained by utilizing the fact that the cepstrum variance value of the acoustic signal from the target sound source is smaller than the cepstrum variance value of the acoustic signal from the noise source. A section where is smaller than a predetermined threshold is determined as a target sound signal, and a weight value α to be multiplied with the received signal is determined to be α = 1.0 (or a value close to 1.0). A section in which the cepstrum variance value is larger than the threshold is determined as a section of a noise signal, and a weight value α (0 ≦ α <1) close to zero is determined. Similar processing is performed when the variance value of the autocorrelation function of the received signal is used as the acoustic feature quantity.

  In addition, when the sharpness of the received signal is used as the acoustic feature amount, the sharpness of the acoustic signal from the nearby target sound source is large, and the sharpness of the acoustic signal from the distant noise source is small. The section where the sharpness is greater than or equal to a certain threshold is determined as the target sound signal, and the weight value α is determined to be α = 1.0 (or a value close to 1.0). When the sharpness is less than or equal to the threshold value, it is determined as a noise signal component, and a weight value α (0 ≦ α <1) close to zero is determined.

  In the weight value multiplication unit (4) which is a weight value multiplication means, the received signal x (t) is multiplied by the weight value α determined by the weight value determination unit (3) (S102). That is, the received signal in the section determined as the target sound signal is multiplied by the predetermined weight value α (1.0 or a value close to 1.0), and the received signal in the section determined as the noise signal is set to the predetermined weight. The value α (0 or a value close to 0) is multiplied, and this α × x (t) is output as an output signal. This output signal is extracted as the target sound signal.

<Second Embodiment>
Next, a second embodiment of the signal extraction device of the present invention will be described.
The signal extraction device according to the second embodiment has the same configuration as the signal extraction device (A) described in the first embodiment. Hereinafter, a different part from 1st Embodiment is demonstrated.

  In the second embodiment, a plurality of acoustic feature quantities are used for determining the weight value. That is, the weight value determining unit (3), which is a weight value determining unit, multiplies each band of the received signal x (t) based on the values of the plurality of acoustic feature values calculated by the acoustic feature value calculating unit (2). The weight value α to be determined is determined (S101p).

An example of processing when combining a plurality of acoustic feature values described above is shown in the following program format <a>. Here, ₁ a variance value of the cepstrum higher _tau, th1 threshold of tau _{1, 2} the variance of the autocorrelation function _tau, the threshold tau ₂ th2, sharpness and tau _3, and th3 the threshold of tau ₃ To do. The symbol ∪ in the program format <a> represents “or”.

The program format <a> determines whether or not at least one of τ ₁ is smaller than th1, τ ₂ is smaller than th2, and τ ₃ is larger than th3 [program format <a> first line], and if at least one of them is satisfied, the weight value α is determined to be 1.0 (second line of the program format <a>); otherwise, the weight value is determined. It represents that α is determined to be 0.0 [the third line of the program format <a>].

  Of course, in the if statement on the first line of the program format <a>, it may be determined using ∩ (and) instead of ∪ (or), or a combination of these may be used. is there.

Moreover, it is not essential to make a determination using all of τ ₁ , τ ₂ , and τ ₃ . For example, the determination may be made by a combination of τ ₁ <th1 and τ ₂ <th2, a combination of τ ₁ <th1 and τ ₃ > th3, or a combination of τ ₂ <th2 and τ ₃ > th3. Further, not only the above τ ₁ , τ ₂ , τ ₃ but also the power value of the received signal x (t) is τ ₄ , and the threshold value of this τ ₄ is th4, and it is possible to make a determination using τ ₄ > th4. It is. That is, for example, it is possible to make a determination based on a combination of τ ₁ <th1 and τ ₄ > th4, and it is also possible to make a determination based on a combination of τ ₁ <th1, τ ₃ > th3, and τ ₄ > th4. A more specific example is shown in the program format <b>.

The program format <b> determines whether or not at least _one of τ ₁ is smaller than th1, or τ ₂ is smaller than th2, and τ ₄ is larger than th4 [ The first line of the program format <b>], if at least one of them is satisfied, the weight value α is determined to be 1.0 [the second line of the program format <b>], otherwise the weight The value α is determined to be 0.0 [the third line of the program format <b>].

In the above-described program format, for example, for τ ₁ , τ ₁ <th1 is determined, but conversely, it may be changed to determine th1 ≦ τ ₁ . In order to explain this, program formats <c> and <d> are illustrated.

In the program format <c>, it is determined whether τ ₁ is smaller than th1 and τ ₂ is smaller than th2 (the first line of the program format <c>). The weight value α is determined to be 1.0 (second line of the program format <c>). Otherwise, the weight value α is determined to be 0.0 (3 lines of the program format <c>). Eyes]. On the other hand, the program format <d> is either tau ₁ is th1 or more or, it is determined whether or at least one is tau ₂ is th2 or more is satisfied [1 line of source form <d> First, if at least one of them is satisfied, the weight value α is determined to be 0.0 (second line of the program format <d>). Otherwise, the weight value α is determined to be 1.0. This indicates that [the third line of the program format <d>]. Eventually, the program formats <c><d> represent equivalent processing contents.

  In this way, as exemplified in the program format <c> <d>, although the processing contents are equivalent, the processing may be performed according to different judgments. In the present invention, the processing according to a certain judgment is performed. It is not limited to. In addition, the determination of the weight value α using a plurality of acoustic feature quantities (signal feature quantities) can be changed as appropriate without departing from the spirit of the present invention.

<Third Embodiment>
Next, a third embodiment of the signal extraction device of the present invention will be described.
FIG. 6 shows a functional block diagram of the signal extraction device (B) in the third embodiment. FIG. 7 shows a flowchart of signal extraction processing in the signal extraction apparatus (B) according to the third embodiment.
The signal extraction device (B) according to the third embodiment is provided with a statistical analysis unit (5) described later in addition to the signal extraction device (A) described in the first embodiment and the second embodiment. is there. Hereinafter, parts different from the first embodiment or the second embodiment will be described.

  The feature of the signal extraction device (B) according to the third embodiment is that it is not necessary to set the threshold value every time by using a technique called multiple regression analysis when determining the threshold value of the acoustic feature value. .5 or more is the target sound signal, and if it is less than 0.5, it can be determined as a noise signal (though the threshold value is 0.5 in the third embodiment, the threshold value is not limited to 0.5). It can be changed as appropriate.) Multiple regression analysis is a technique that can examine the correlation between a plurality of acoustic feature quantities and signal features in a multidimensional manner.

In multiple regression analysis, unknown data can generally be discriminated using data whose correct answer is known. For example, for a nearby target sound signal and a distant noise signal recorded in advance, a numerical value “1” is assigned to the close target sound signal and a numerical value “0” is assigned to the far noise signal as a correct answer. Furthermore, a plurality of acoustic feature quantities are calculated as the acoustic feature quantities of the known target sound signal and noise signal (this is the above-described cepstrum high-order variance value, variance value of autocorrelation function, sharpness, power, etc. Is). By performing multiple regression analysis on these acoustic feature amounts, a regression equation for determining whether the unknown signal is a close sound or a far sound is derived. The regression equation is expressed by the following equation (4) using the regression coefficients b _{1 to} b _k for the explanatory variables p _{1 to} p _k (k = 1, 2,...).

Here, the explanatory variables p _{1 to} p _k represent the acoustic feature quantities τ _{1 to} τ _k , and y is an objective variable obtained from the regression equation (4). Further, a ₀ is a y-intercept (a technique for successfully obtaining the y-intercept a ₀ and the regression coefficients b _{1 to} b _k is a multiple regression analysis).

In the statistical analysis unit (5), which is a statistical analysis means, the objective variable y ₁ is calculated as a result of substituting the acoustic feature amount calculated by the acoustic feature amount calculation unit (2) into the explanatory variable of the equation (4). (S200).

Using the objective variable y ₁ calculated in this way, the weight value calculation unit (3) determines the weight value α as in the program format <e> (S201).

Program Format <e> is, determines whether or not the value of y ₁ is greater than 0.5, when it is determined to be greater determines the weight value α to 1.0, otherwise Represents that the weight value α is determined to be 0.0.

  By determining in this way, the weight value α can be determined without having to visually check the threshold distribution in advance.

  The signal extraction device of the present invention includes the signal input means (for example, a microphone), a storage device (for example, a RAM, a ROM, and a hard disk), an arithmetic processing device (for example, a CPU), an input / output device (for example, a keyboard, a display), It can be realized by a computer provided with a bus or the like that allows data exchange between these devices (see FIG. 8). In this case, a program (acoustic feature amount calculation program, statistical analysis) required to calculate the variance value of the higher-order components of the cepstrum, the variance value of the autocorrelation function, the sharpness, the statistical analysis, the weight value, the output signal, etc. Program, weight value determination program, weight value multiplication program, control program for controlling the processing of these programs, etc. However, in the first embodiment and the second embodiment, a statistical analysis program is unnecessary.) Other reception Data such as the signal x (t) is stored in the storage device, and the processing unit reads the program and interprets and executes it as necessary, thereby realizing the functions of the above-described units (acoustic feature amount calculation unit, A statistical analysis unit, a weight value determination unit, a weight value multiplication unit, and a control unit for controlling the processing of these units). The output signal output by the weight value multiplication unit may be stored in the storage device. Each program can also be recorded on a computer-readable recording medium.

  The signal extraction device of the present invention is useful for acoustic signal analysis such as speech recognition and noise signal suppression when the target signal is speech, for example. This is particularly useful when the target signal source is closer to the signal extraction device than the miscellaneous signal source in an environment where the target signal from the target signal source and the miscellaneous signal from the miscellaneous signal source do not overlap in time. is there.

An example of a higher cepstrum dispersion value when a target sound signal from a position about 50 cm away from a microphone and a noise signal from a position about 3 m away from the microphone are observed for about 30 seconds. An example of the dispersion value of the autocorrelation function under the same conditions as in FIG. An example of sharpness under the same conditions as in FIG. The functional block diagram of the signal extraction apparatus (A) concerning 1st Embodiment and 2nd Embodiment. The flowchart of the signal extraction process in the signal extraction apparatus (A) concerning 1st Embodiment and 2nd Embodiment. The functional block diagram of the signal extraction apparatus (B) concerning 3rd Embodiment. The flowchart of the signal extraction process in the signal extraction apparatus (B) concerning 3rd Embodiment. The hardware structural example of a signal extraction apparatus.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Microphone 2 Acoustic feature-value calculation part 3 Weight value determination part 4 Weight value multiplication part 5 Statistical analysis part

Claims (6)

A signal extraction apparatus that suppresses a miscellaneous signal and extracts a target signal from a received signal received in an environment where a target signal source and a miscellaneous signal source exist,
At least one signal input means for receiving signals from a target signal source and a miscellaneous signal source;
Signal feature amount calculating means for calculating the signal feature amount of the received signal received by the signal input means;
Weight value determining means for determining a weight value α for multiplying the received signal received by the signal input means based on the value of the signal feature value calculated by the signal feature value calculating means;
A signal extraction apparatus comprising weight value multiplying means for multiplying the received signal received by the signal input means by the weight value determined by the weight value determining means. 2. The signal extraction apparatus according to claim 1, wherein the signal feature amount calculated by the signal feature amount calculating means is a variance value with respect to a time axis of the cepstrum. 2. The signal extraction device according to claim 1, wherein the signal feature amount calculated by the signal feature amount calculating means is a variance value with respect to a time axis of the autocorrelation function. The signal extraction apparatus according to claim 1, wherein the signal feature amount calculated by the signal feature amount calculation unit is sharpness. The signal feature amount calculation means
Calculates any two or more of power, variance value with respect to time axis of cepstrum, variance value with respect to time axis of autocorrelation function, and sharpness as signal feature amount,
The weight value determining means is
Weight to multiply the received signal based on multiple combinations of power, signal cepstrum time axis variance, autocorrelation function time axis variance, and sharpness 2. The signal extraction device according to claim 1, wherein the value α is determined. Statistical analysis means for calculating a target variable by multiple regression analysis from the value of the signal feature quantity calculated by the signal feature quantity calculation means,
The weight value determining means is
6. The signal extraction apparatus according to claim 1, wherein the weight value α is determined based on the objective variable calculated by the statistical analysis means.

Images