JP3431696B2 - Signal separation method - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to various physical quantities such as ambient noise removal and speaker separation in voice input devices such as various communication systems and voice recognition systems, or separation of signal sources when measuring bioelectric signals. The present invention relates to a signal separation method for separating a signal component from noise at the time of detection.

[0002]

2. Description of the Related Art In recent years, there have been many announcements of a technique for dynamically processing a plurality of input signals to achieve, for example, noise reduction or noise reduction, that is, a signal separation method. An example of such a technique will be described below. In the following example, all signals handled are audio signals.

FIG. 8 shows an example of such a signal separation method, and is a block diagram showing a first conventional example of the signal separation method of the present invention. In the figure, 1 is H-J
(Heralt-Jutten) network, 2 is a signal propagation path, 3 is a weighted adder, and 4 is a gain coefficient adjusting unit.

Audio signals x ₁ (t) from I signal sources,
x ₂ (t), ..., X _I (t) are I detection signals e ₁ (t), e ₂ (t), ...
.., detected as e _I (t), processed in the HJ network 1, and separated signals s ₁ (t), s ₂ (t),
.., s _I (t) is output.

Here, it is assumed that the signal propagation path 2 described above has a transfer characteristic given by the following equation.

E (t) = G · X (t) (1) where E (t) = [e ₁ (t), e ₂ (t), ...
e _I (t)] ^T , X (t) = [x ₁ (t), x ₂ (t), ...
.., x _I (t)] ^T , G is I row I where the inverse matrix G ⁻¹ exists
It is a constant matrix of columns. [...] ^T represents transposition, and each element of the matrix G is g _ij .

Each element g _ij of the matrix G representing the static (uniform with no time change) transfer characteristic as in the equation (1) is physically a voice signal x ₁ (t), x _2. (T), ..., x
_I (t) represents the amount of attenuation and mixing in the signal propagation path 2. Therefore, hereinafter, each element g of the matrix G
_ij is called “attenuation amount g _ij ”.

Now, in the signal detecting means (not shown) through such a signal propagation path 2, the detection signal E (t)
(= [E ₁ (t), e ₂ (t), ..., e _I (t)] ^T )
, S ₁ (t), s ₂ (t), ..., S which are the outputs of the HJ network 1.
_I (t) is given by the following equation using a vector expression.

S (t) = E (t) −C · S (t) (2) = [I + C] ⁻¹ G · X (t) where the matrix C is the gain coefficient c _ij (where c _ii = 0) is an I-row I-column matrix, and I is an I-row I-column identity matrix. Note that S (t) = [s ₁ (t), s
₂ (t), ..., s _I (t)] ^T.

From the equation (2), if I · (I−1) elements of the matrix C can be calculated, the speech signal X (t) (= [x
₁ (t), x ₂ (t), ..., x _I (t)] ^T ) can be reconstructed and each separated signal s _i (t) is only 1
Will be proportional to one x _i (t). When this equation (2) is written down, it becomes the following equation (3).

[0011]

[Equation 5]

To simplify the description, a 2-input 2-output HJ network will be illustrated with I = 2. FIG. 9 shows the HJ network in such a case. In the figure, the weighted adder 3 of FIG. 8 is composed of an adder 5 and a multiplier 6.

At this time, the equation (2) is

[0014]

[Equation 6]

The gain coefficient c _ij is expressed by the following equation (5),
If the relationship of (6) is established, the separation signal s _i is the audio signal x _i.
Linearly proportional to.

[0016]

[Equation 7]

Then, s ₁ = g ₁₁ x ₁ and s ₂ = g ₂₂ x ₂ are obtained. Or

[0018]

[Equation 8]

Then, s ₁ = g ₁₂ x ₂ and s ₂ = g ₂₁ x ₁ are obtained.

That is, if the attenuation amount g _ij during which the audio signal x _i reaches the signal detecting means is known in advance, the gain coefficient c _ij is calculated by the equation (5) or the equation (6) and substituted into the equation (3). By doing so, it is possible to obtain the separated signal s _i that is linearly proportional to the audio signal x _i .

When the matrix G is unknown, it is assumed that the speech signals x _i are independent of each other and the separated signals s _i are also independent of each other, and the gain coefficient c _ij is estimated as follows.

Further, the above-described estimated each power split signal s _i (electric power) by adjusting the gain coefficient c _ij to minimize can be accomplished, the gain coefficient c _ij is the following using a gradient method It is adjusted by the equation (7).

[0023]

[Equation 9]

Here, ε is a small positive value that determines the efficiency of learning (the above-mentioned adjustment process), and functions f ₁ () and f ₂ ()
_Are odd functions different from each other introduced to set c _ji ≠ c _ij .

In addition to the above theory, the following two methods have been proposed as noise removal techniques when, for example, a music or voice source is input in a noisy environment. FIG. 10 is a block diagram showing a second conventional example of such a signal separation method. In the figure, 7 is a first microphone for inputting a voice signal x ₁ and 8 is a noise signal x ₂
Is a second microphone for inputting. Reference numeral 9 is an adaptive filter, and 10 is a subtractor.

In the first microphone 7, the audio signal x
_The noise signal x ₂ and the signal x ₂ ^* which propagates to the first microphone 7 are mixed and detected as ₁ and output as the detection signal e ₁ . The second microphone 8 is arranged at such a position that its input is mainly the noise signal x ₂ and obtains the detection signal e ₂ (≈x ₂ ). These detection signals e ₁ and e ₂ are
It is input to the adaptive filter 9. Output y of adaptive filter 9
Is subtracted from the detection signal e ₁ by the subtractor 10,
The signal s ₁ obtained by removing the noise component x ₂ from the detection signal e ₁ is obtained.

Such noise elimination can be achieved by assuming that the voice signal x ₁ and the noise signal x ₂ are independent and adjusting the transfer characteristic of the adaptive filter 9 so as to minimize the power of the signal s _1. .

Such an adaptive filter 9 is provided by a finite impulse response (hereinafter abbreviated as "FIR") digital filter composed of a recurrence formula represented by the formula (8).

[0029]

[Equation 10]

Here, K is the number of filter stages. In addition,
The filter coefficient a _k (n) is adjusted by the least square method algorithm represented by the recurrence formula of the formula (9).

A _k (n + 1) = a _k (n) + εs ₁ (n) e ₂ (n−k) (9) where ε is a small positive value that determines the efficiency of learning by the least squares algorithm. (For example, JP-A-5-22
392 or JP-A-6-118967).

Now, FIG. 11 is a block schematic diagram of a speaker selection system of a voice conference system using the signal separation method in the third conventional example. In the figure, 11 and 12 are microphones, 13 and 14 are variable delay units, 15 is a delay time setting unit, 16 is an adder, and two microphones 11 and 12 are installed at position A,
This is to detect the voices of the speakers in B and C.

To emphasize the voice of the speaker located at A,
Sound wave propagation time t ₁ from the position A to the microphone 11
Then, the sound wave propagation time t ₂ from the position A to the microphone 12 is measured in advance, and the time difference between them is corrected by the variable delay devices 13 and 14, respectively, and the delay time is changed to the same phase as the voice waveform of the speaker at the position A. Output from the devices 13 and 14. In this example, the delay time setting unit 15 sets (t ₂ −t ₁ ) in the variable delay unit 13 and sets the delay time 0 in the variable delay unit 14. By adding these two outputs by the adder 16, the voice of the speaker at the position A is emphasized (for example, Japanese Patent Laid-Open No. 5-158492).

[0034]

However, the above-mentioned three conventional signal separation methods have the following problems, respectively.

That is, H-J shown in the first conventional example
In the network, as is clear from the equation (1), the matrix G
Is a constant matrix, so the audio signal x
_{Since the} phase shift due to the time difference until _i (t) is detected by the signal detecting means is not taken into consideration, there is a problem that it cannot be practically used in most cases.

For example, in the case where sound waves are emitted from two sound sources as shown in FIG. 8, the intensity of the sound transmitted from the two sound sources in the signal propagation path 2 is simply the propagation distance if there is no barrier. Inversely proportional to, so the audio signals x ₁ and x ₂
In order to eliminate the phase shift, the distance from when the detection signals e ₁ (t) and e ₂ (t) are obtained must be made equal. However, if they are equidistant, when sound waves emitted from sound sources are detected by the detection means, the sound waves from the two sound sources are mixed at the same rate. With this, the inverse matrix does not exist in the matrix G of the equation (1), and the signal separation cannot be performed. In an actual environment, if the phase shift is so small that it can be ignored and the detection signals e ₁ and e ₂ are obtained with the mixing ratios of the audio signals x ₁ and x ₂ sufficiently different, then they hardly exist. .

Furthermore, in the first prior art example, as is clear from the fact that the matrix G is a constant matrix in the equation (1), the audio signal x _i (t) is transmitted in all its signal propagation paths 2. It is assumed that the frequency components have uniform attenuation, and if the attenuation frequency characteristics are not uniform, the effect of signal separation deteriorates.

In the second conventional example, although the frequency characteristics of the phase and the attenuation are taken into consideration in the adaptive filter 5, the second microphone 4 which detects the noise signal x ₂ detects the voice signal x _1. There is an assumption that it is not done or is sufficiently small compared to the noise signal x ₂ .
As a result, the audio signal x _{1 is} also transmitted to the second microphone 4.
Is detected, there is a problem that an echo occurs in the output signal s ₁ .

Since such an echo occurs, the signal s ₁ used in the adaptive process of the adaptive filter 5 is
There is also a problem that the minimization of the power of 1 does not necessarily coincide with the minimization of the power of the noise component.

Furthermore, in the third conventional example, in the audio signals detected by the two microphones 11 and 12,
Since the voice signals of the selected speaker are added in phase, the voices of the selected speaker are not echoed, but the voices of other speakers cannot be sufficiently removed. Further, since there is no means for adaptively adjusting the delay time given to the variable delay devices 13 and 14, there is a problem that the position of the speaker is fixed.

The present invention has been made in view of the above problems, and it is an object of the present invention to provide a signal separation method which is capable of accurately separating signals even when there is a phase shift and which does not cause an echo.

[0042]

In order to achieve the above object, the signal separating method of the present invention is such that at time t, original signals x _i (t) (where i = 1,
2, ..., when I) is propagated are mixed in a plurality of paths having a predetermined transfer characteristic, is detected as a detection signal e _i (t) in a plurality of signal detection means, these detection signals e _i ( estimating the original signal x _i (t) before mixing from t),
In a signal separation method for separating as a separated signal s _i (t), using a predetermined gain coefficient c _ij (t) and a predetermined delay time d _ij (t) determined according to the transfer characteristic, The separated signal s _i (t) is

[0043]

[Equation 11]

It is derived by the following equation .

Further , the delay time d _ij (t) is set so that the correlation between the detected signal e _i (t) and the separated signal s _i (t) becomes high by using the cross-correlation function. The gain coefficients c _ij (t) are sequentially adjusted so that the powers of the separated signals s _i (t) corresponding to the respective gain coefficients are sequentially adjusted.

Further, according to the method of claim 2 , at time t, the original signals x _i (t) (where i = 1, 2, ..., I) emitted from a plurality of signal sources have predetermined transfer characteristics. Of the original signal x _i (t) before being mixed from the detection signals e _i (t) when the signals are mixed and propagated in a plurality of paths having a plurality of paths and are detected as the detection signals e _i (t) by the plurality of signal detecting means. )
Estimating a separation signal s _i in the signal separation method for separating a (t), the filter coefficients are determined in accordance with the impulse response of the transfer characteristics a _ijk (t) (where k =
0, 1, ..., K−1) and a predetermined discrete time T, the separated signal s _i (t) is

[0047]

[Equation 12]

It is derived by the following equation .

Further , the filter coefficient a _ijk (t)
Is to minimize the power of the separated signal s _i (t) when ε _ij is an arbitrary small positive value,

[0050]

[Equation 13]

It is characterized in that it is sequentially adjusted according to the following equation.

[0052]

[0053]

[0054]

[0055]

According to the method of claim 1, when I = 2,
The original signals x ₁ (t) and x ₂ (t) emitted from two sound sources placed in a free space without obstacles are g i _ij from the j-th sound source in the two signal detecting means. , I = 1 or 2, j = 1 or 2, where τ _ij is the propagation delay time, and e ₁ (t) = g ₁₁ x ₁ (t −τ ₁₁ ) + g ₁₂ x ₂ (t−
τ ₁₂ ) e ₂ (t) = g ₂₁ x ₁ (t−τ ₂₁ ) + g ₂₂ x ₂ (t−
τ ₂₂ ) are detected as detection signals e ₁ (t) and e ₂ (t), where d _ij is the delay time and c _ij is the gain coefficient, d ₁₂ = τ ₁₂ −τ ₂₂ d ₂₁ = τ ₂₁ − When given by τ ₁₁ c ₁₁ = g ₁₂ / g ₂₂ c ₂₁ = g ₂₁ / g ₁₁ , the relationship between the separated signal s ₁ (t) and the original signal x ₁ (t) is s ₁ (t) -g ₁₁ x ₁ (t-τ ₁₁ ) = c ₁₂ c ₂₁ (s ₁ (t
−d ₁₂ −d ₂₁ ) −g ₁₁ x ₁ (t−τ ₁₁ −d ₁₂ −d ₂₁ )), and if c ₁₂ c ₂₁ <1, s ₁ (t) → g ₁₁ x as time t passes. _The original signal x ₁ (t) and x ₂ (t) are separated as asymptotic as ₁ (t−τ ₁₁ ) s ₂ (t) → g ₂₂ x ₂ (t−τ ₂₂ ).

Furthermore, assuming that each original signal x _i (t) is independent, the delay time d _ij for each separated signal s _i (t) is
When determining (t), the detection signal e _i (t) and the separation signal s _j
The cross-correlation function R _ij (d) of (t) is calculated according to the following equation (E
[] Represents the calculation of the expected value. ), R _ij (d) = E [e _i (t) s _j (t−d)] The delay time d (d ≧ 0) having the highest correlation is set to the delay time d.
It will be the estimated value of _ij (t). Further, the separation signal s _i (t) corresponding to each of the gain coefficients c _ij is sequentially adjusted so as to be minimized, so that the separation signal s _{i is obtained.}
(T) will be independent of each other and signal separation will be achieved.

According to the method of claim 2 , a FIR digital filter whose tap coefficient is the filter coefficient a _ijk (t) is formed, and each frequency component of the separated seasonal issue s _i (t) is defined by claim 1. Since it can be considered that the gain coefficient c _ij (t) and the delay time d _ij (t) in are similar to those defined, the original signals x ₁ (t) and x ₂ (for each frequency component). t) will be separated.

Further, when each original signal x _i (t) is independent of each other,

[0059]

[Equation 15]

The following equation is defined and s _i ² (t)
_Is partially differentiated with the filter coefficient a _ijk (t) to obtain the gradient,

[0061]

[Equation 16]

Each filter coefficient a _ijk (t) is gradually adjusted in the direction opposite to the gradient direction, and each separated signal s
_i Power s _i ² minimized and thus the performed in (t) of (t).

[0063]

[0064]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of an apparatus to which the signal separation method according to the first embodiment of the present invention is applied. In the figure, 17 and 18 are sound sources, 19 and 20 are microphones, 21 and 22 are low-pass filters, 23 and 24 are analog / digital converters (hereinafter abbreviated as "AD converters"), and 25 and 26 are subtractions. , 27 and 28 are multipliers, 29
Reference numerals 30 and 30 are delay devices, 31 and 32 are correlators, 33 and 34 are delay time adjusting units, and 35 and 36 are gain coefficient adjusting units.

The subtractors 25 and 26 and the multipliers 27 and 2
8, delay units 29 and 30, correlators 31 and 32, delay time adjusting units 33 and 34, and gain coefficient adjusting units 35 and 36.
Constitute a signal separation filter 100 having a MIMO (multi-input other output) structure. The operation of the embodiment of the present invention configured as described above will be described below.

Sound wave x emitted from sound sources 17 and 18
_The microphones 19 and 20 detect ₁ (t) and x ₂ (t). The output of the microphone 19 is band-limited to prevent aliasing in the low-pass filter 21, sampled at a predetermined sampling period T in the AD converter 23, and output as a detection signal e ₁ (n). Similarly, the output of the microphone 20 is band-limited by the low-pass filter 22 to prevent aliasing, sampled at a predetermined sampling period T by the AD converter 24, and output as a detection signal e ₂ (n). . Here, t = nT.

These detection signals e ₁ (n) and e ₂ (n) are
In the signal separation filter 100,

[0068]

[Equation 17]

The above calculation is performed for i = 1 and 2, and the separated signals s ₁ (n) and s ₂ (n) are output. Here, d _ij (t) = dn _ij (n) T.

This calculation process will be described in accordance with the signal flow in the configuration of the block diagram of FIG. 1. The separated signal s ₂ (n)
Is input to the delay unit 30 and delayed by the delay time dn ₁₂ (n) set by the delay time adjusting unit 34. Delay device 3
0 outputs a signal s ₂ to the multiplier _{28 (n-dn 12 (n} )), the signal _{s 2 (n-dn 12 (} n)) is the multiplier 2
8 is multiplied by the gain coefficient c ₁₂ (n), and the adder 2
In 5, the signal is subtracted from the detection signal e ₁ (n) and output as a separation signal s ₁ (n).

On the other hand, the separated signal s ₁ (n) is input to the delay unit 29 and delayed by the delay time dn ₂₁ (n) set by the delay time adjusting unit 33. Delayer 29 outputs the multiplier 27 to the signal _{s 1 (n-dn 21 (} n)) , the signal _{s 1 (n-dn 21 (} n)) , the gain coefficient c ₂₁ at multiplier 27 (n ), Is subtracted from the detection signal e ₂ (n) in the adder 26, and is output as a separation signal s ₂ (n).

The above-mentioned delay time dn ₁₂ (n) is calculated in the following process. The detection signal e ₁ (n) and the separated signal s ₂ (n) are input to the correlator 32, and in the following equation (11), i = 1 and j = 2, and the cross-correlation function R ₁₂ (d ₀ ).
Seeking

[0073]

[Equation 18]

Output to the delay time setting unit 34. here,
d ₀ is d ₀ ≧ 0, and is an estimated value of the delay time dn ₁₂ ().

In the delay time setting unit 34, the cross correlation function R
Select the delay time d ₀ that maximizes the absolute value of ₁₂ (d ₀ ),
This is sent to the delay unit 30 as the delay time dn ₁₂ (n + 1) and used for the calculation at the next time (n + 1).

Similarly, the above-mentioned delay time dn ₂₁ (n) is
It is calculated in the following process. The detection signal e ₂ (n) and the separated signal s ₁ (n) are input to the correlator 31, and the above equation (11)
The cross-correlation function R ₂₁ (d ₀ ) is obtained in accordance with the above, and is output to the delay time setting unit 33. The delay time setting unit 33 selects the delay time d ₀ that maximizes the absolute value of the cross-correlation function R ₂₁ (d ₀ ), sends it to the delay device 29 as the delay time dn ₂₁ (n + 1), and the next time (n + 1 ) Is used for calculation.

The above-mentioned gain coefficient c ₁₂ (n) is calculated in the following process. The separated signal s ₁ (n) and the output signal s ₂ (n−dn ₁₂ (n)) of the delay device 30 are input to the gain coefficient adjustment unit 36, and the gain coefficient adjustment unit 36 outputs the separated signal s ₁ (n). The gain coefficient c ₁₂ (n) is updated with i = 1 and j = 2 in the recurrence formula represented by the following formula (12) so as to minimize the power.

C _ij (n + 1) = c _ij (n) + ε _ij s _i (n) s _j (n-dn _ij (n)) (12) At this time, multiplication is performed as a gain coefficient c ₁₂ (n + 1). Bowl 28
Output to. This gain coefficient c ₁₂ (n + 1) will be used for calculation in the next (n + 1). Where ε
_ij is a small positive value that determines the learning efficiency.

Similarly, the above-mentioned gain coefficient c ₂₁ (n) is
It is calculated in the following process. The separated signal s ₂ (n) and the output signal s ₁ (n−dn ₂₁ (n)) of the delay unit 29 are input to the gain coefficient adjustment unit 35, and the gain coefficient adjustment unit 35
In order to minimize the power of the separated signal s ₂ (n), the gain coefficient c ₂₁ (n) is updated as i = 2 and j = 1 in the recurrence formula expressed by the above formula (12), and the gain coefficient c ₂₁ (n) is updated. ₂₁ (n +
1) to the multiplier 27.

This gain coefficient c ₂₁ (n + 1) is used for calculation in the next (n + 1). In this way, the signal separation filter 100 can be realized discretely, so that it can be easily realized by software on a digital signal processor, a microprocessor or the like, not to mention dedicated hardware.

Now, the signal separation filter 10 as described above is used.
From 0, the detection signals e ₁ (n) and e ₂ (n), which are detected by mixing the original signals x ₁ (t) and x ₂ (t),
It will be described below whether the separation signal s ₁ (n) proportional to the original signal x ₁ (t) and the separation signal s ₂ (n) proportional to the original signal x ₂ (t) are separated. In the following, the discrete time nT
Is approximated to the continuous time t, but the motion is considered to be essentially the same by this treatment.

Original signals x ₁ (t) and x emitted from two sound sources 17 and 18 placed in a free space without obstacles
₂ (t) are mixed in the microphones 19 and 20 and detected as detection signals e ₁ (t) and e ₂ (t). At this time, the detection signals e ₁ (t) and e ₂ (t)
Is e ₁ (t) = g ₁₁ x ₁ (t−τ ₁₁ ) + g ₁₂ x ₂ (t−τ ₁₂ ) (13) e ₂ (t) = g ₂₁ x ₁ (t−τ ₂₁ ) + g ₂₂ x ₂ (t−τ ₂₂ ) is represented as (14). Here, g _ij is the amount of attenuation and τ _ij is the propagation delay time. Also, i = 1 or 2, j = 1 or 2
Is.

Now, these detection signals e ₁ (t) and e
₂ (t) is passed through the signal separation filter 100, and the separated signals s ₁ (t) and s ₂ (t) that are outputs are

[0084]

[Formula 19]

It is assumed that they are related by the following expression. Here, c _ij (t) is a gain coefficient, and d _ij (t) is a delay time.

At this time, the gain coefficient c _ij (t) and the delay time d _ij (t) are constants, and d ₁₂ = τ ₁₂ −τ ₂₂ (16) d ₂₁ = τ ₂₁ −τ ₁₁ (17) c ₁₁ = If g ₁₂ / g ₂₂ (18) c ₂₁ = g ₂₁ / g ₁₁ (19), the relationship between the separated signal s ₁ (t) and the original signal x ₁ (t) is given by the following equation (20).

S ₁ (t) -g ₁₁ x ₁ (t−τ ₁₁ ) = c ₁₂ c ₂₁ (s ₁ (t−d ₁₂ −d ₂₁ ) −g ₁₁ x ₁ (t−τ ₁₁ −d ₁₂ −) d ₂₁ )) (20) Here, if c ₁₂ c ₂₁ <1, s ₁ (t) → g ₁₁ x ₁ (t−τ ₁₁ ) (21) as time t elapses. Asymptotic.

Similarly, the relationship between the separated signal s ₂ (t) and the original signal x ₂ (t) is asymptotically s ₂ (t) → g ₂₂ x ₂ (t-τ ₂₂ ) (22). Then, the original signals x ₁ (t) and x ₂ (t) are
The separated signals s ₁ (t) and s ₂ (t) are separated and output. In addition, in the separated signals s ₁ (t) and s ₂ (t),
It can be seen from equations (21) and (22) that no echo is included.

Also, assuming that each original signal x _i (t) is independent, the delay time d for each separated signal s _i (t) is
_When determining _ij (t), the cross-correlation function R _ij (d) of the detection signal e _i (t) and the separated signal s _j (t) is calculated as R _ij (d) = E [e _i (t) s _j ( t-d)] (23). Here, E [] represents the calculation of the expected value.

At this time, the delay time d (d
≧ 0) is an estimated value of the delay time d _ij (t). Thus, if the delay times d _ij (t) are adjusted sequentially, it is always optimal even when the delay times τ of the original signals x ₁ (t) and x ₂ (t) are unknown or have time fluctuations. It is adjusted. In addition, when there are more than two signal sources, the optimum delay time d _ij (t) may not always be obtained in the initial operation, but in the process of successive adjustment, the optimum value gradually increases. Approaching.

Further, by sequentially adjusting the powers of the separated signals s _i (t) corresponding to the gain coefficients c _ij so as to be minimized, the separated signals s _i (t) become independent from each other, and the signal separation is performed. Will be achieved.

[0092] Next, the minimization of the power of the separated signal s _i (t) is described to lead to independently each separated signal s _i (t) of each other. Here, "each separated signal s _i (t) is independent of each other" means that the original signals x _i (t) are independent and are mixed in the propagation path to become the detection signal e _i (t). , This process is the original signal x
_This is equivalent to taking the linear sum of _i (t). Reverting to independent again means obtaining a separation signal s _i (t) that is linearly proportional to the original signal x _i (t), that is, separation of the signals.

Now, in order to simplify the explanation, the original signal x _i
(T) is a random process, and its expected value E [x
It is assumed that _i (t)] is 0. At this time, the detection signals e ₁ (t) and e ₂ (t) have a cross correlation only with respect to the delay times d ₁₂ and d ₂₁ . FIG. 2 shows the separation signals s ₁ (t) and s ₂ (t) and the detection signals e ₁ (t) and e ₂ (t) as two-dimensional vectors.

In the figure, the separated signal s ₁ (t ₁ ) at a certain time t ₁ is separated from the detected signal e ₁ (t ₁ ) by the separated signal s ₂ (t ₁ -d ₁ ) as shown in equation (15). ₁₂ ) is subtracted. At this time, when the gain coefficient c ₁₂ is changed in FIG. 2, the vector s ₁ (t ₁ ) of the separated signal is obtained.
Is on a line L that is in contact with the end point B of the detection signal vector e ₁ (t ₁ ) and is parallel to the separation signal vector s ₂ (t ₁ −d ₁₂ ).
Can be moved.

As is clear from FIG. 2, the separation signal vector s ₁ (t ₁ ) is minimized when the separation signal vectors s ₁ (t ₁ ) and s ₂ (t ₁ -d ₁₂ ) are orthogonal. It is time to do.
At this time, E [s ₁ (t ₁ ) s ₂ (t ₁ −d ₁₂ )] = 0 (24), and the equation (24) expresses that the separated signals s ₁ (t ₁ ) and s ₂ are related to the delay time d _12. It means that (t ₁ −d ₁₂ ) becomes uncorrelated. That is, the separation signals s ₁ (t) and s ₂ (t) are independent, and the separation signal s ₁ (t) is linearly proportional to the original signal x _i (t).
Obtaining _i (t), in other words signal separation, is achieved.

As described above, according to the present embodiment, in the equation (15) for realizing the signal separation filter 100, the delay time d _ij (t) is considered in consideration of the time variation and the gain is adjusted. Since the coefficient is c _ij (t) and the time variation is taken into consideration, even when the original signal x _i (t) is transmitted to the detection signal e _i (t), a dynamic delay with time variation is included. It is possible to separate signals with high accuracy, and it is possible to sufficiently suppress the echo as is clear from the expressions (21) and (22).

FIG. 3 is a block diagram of an apparatus to which the signal separation method according to the second embodiment of the present invention is applied. In the figure, 25 and 26 are subtractors, 37 and 38 are adaptive filters, and 39 and 40 are filter coefficient adjusting units. Also, the subtracters 25 and 26, the adaptive filters 37 and 3
8. The filter coefficient adjusting units 39 and 40 constitute the signal separation filter 200. The operation of the present embodiment configured as described above will be described below.

Although not shown, the detection signal e ₁ (t)
And e ₂ (t), sound waves x ₁ (t) and x ₂ (t) emitted from two sound sources are detected by two microphones as in the first embodiment, and are respectively passed through a low pass filter and an AD. It is detected via the converter.

These detection signals e ₁ (n) and e ₂ (n) are
In the signal separation filter 200,

[0100]

[Equation 20]

The above calculation is performed for i = 1 and 2, and the separated signals s ₁ (n) and s ₂ (n) are output.

The separated signal s ₂ (n) is supplied to the adaptive filter 38.
Is input to the input terminal, and its output y ₁₂ (n) is calculated by the following equation (26).

[0103]

[Equation 21]

Where a _ijk (n) (k = 0, 1, ...
., K-1) are filter coefficients.

The output y ₁₂ (n) of the adaptive filter 38 is subtracted from the detection signal e ₁ (n) in the subtractor 25, and the separated signal s ₁ (n) is output. Similarly, the separated signal s
₁ (n) is input to the adaptive filter 37 and its output y
₂₁ (n) is calculated by the equation (26), subtracted from the detection signal e ₂ (n) in the subtractor 26, and output as the separation signal s ₂ (n).

Furthermore, the separated signals s ₁ (n) and s ₂ (n) are
It is input to the filter coefficient adjusting units 39 and 40, and the filter coefficient adjusting unit 40 _{calculates the} filter coefficient a _12k (n) by the following equation (27) a _ijk (n + 1) so as to minimize the power of the separated signal s ₁ (n). = A _ijk (n) + ε _ij s _i (n) s _j (n−k) (27). Here, ε _ij is a small positive value that determines the learning efficiency.

Similarly, the filter coefficient adjusting unit 39 adjusts the filter coefficient a _21k (n) by the above equation (27) so as to minimize the power of the separated signal s ₂ (n).

The signal separation filter 200 according to the present embodiment.
Is also calculated by the signal separation filter 10 in the first embodiment.
Similar to the operation of 0, it can be realized by software such as a digital signal processor or a microprocessor. Let's apply the respective values to the equation (25) to calculate the separated signals s ₁ (n) and s ₂ (n). Then, s on the right side
Since it includes _j (n), it cannot be calculated as it is.

Therefore, from the right side of equation (25), time step n
Of the separated signal s _j (n) of (formulas (28) to (3
It transforms into 1). In order to simplify the description, the signal separation filter 200 has two inputs and two outputs as shown in FIG.

[0110]

[Equation 22]

[0111]

[Equation 23]

Here,

[0113]

[Equation 24]

[0114]

[Equation 25]

It is Note that equations (29) to (31) are
It is completely equivalent because it is simply a modification of equation (25).

The above equation (27) is given by the following equation (32)

[0117]

[Equation 26]

It is a discretized version of. This formula (32)
Is derived as follows.

Power s _i ² (t) of each separated signal s _i (t)
In order to minimize

[0120]

[Equation 27]

The filter coefficient a is expressed by the following equation (34).
Partial differentiation may be performed with _ijk (t).

[0122]

[Equation 28]

By gradually adjusting the filter coefficient a _ijk (t) in the direction opposite to the gradient direction given by the equation (34), the power s _i ² (t) of the separated signal s _i (t) is minimized. Can be converted.

Equation (32) is such a filter coefficient a
_ijk (t) is represented by adjusting the filter coefficient a _ijk (t) according to this equation, and the separation signal s _i (t) (here, s
₁ (t) and s ₂ (t) are independent of each other and signal separation is achieved.

As described above, according to this embodiment, in the equation (25) for realizing the signal separation filter 200, the delay time is taken into consideration with the time step k including the time variation, and the filter coefficient is considered. Since a _ijk (t) is a format that takes time variation into consideration, the accuracy of the transmission from the original signal x _i (t) to the detection signal e _i (t) is high even when a dynamic delay with time variation is included. There is an effect that it is possible to separate signals with good quality.

FIG. 4 is a circuit block diagram of an apparatus to which the signal separating method according to the third embodiment of the present invention is applied. In the figure, 41-1 to 41- (K-1) are transconductance amplifiers and 42-1 to 42- (K
-1) is a condenser, 43-0 to 43- (K-1) and 4
4-0 to 44- (K-1) is a multiplier, 45-0 to 45-
(K-1) is an integrator. The operation of the present embodiment configured as described above will be described below.

The circuit shown in the figure corresponds to the adaptive filter 38 and the filter coefficient adjusting unit 40 in the second embodiment, and generates the output y ₁₂ (t) from the separated signals s ₁ (t) and s ₂ (t). It represents only the process of doing.

The separated signal s ₂ (t) is supplied to the K-threshold of the first-order low pass filter (follower integration circuit) composed of the transconductance amplifier 41-1 and the capacitor 42-1.
It is input to the delay network 400 in which one cascade connection is made. At this time, the capacitance of the capacitor is designed such that the delay time for one stage of the low-pass filter is equivalent to the sampling time T in the second embodiment.

The output signal at each node k of the delay network 400 is input to both of the multipliers 43-k and 44-k, and the respective output signals of the multiplier 44-k and the separated signal s ₁ (t). The product is taken, and the product is input to the integrator 45-k and integrated. At this time, in the integrator 45-k, the filter coefficient a _12k (t) is adjusted by the equation (28). Further, the learning efficiency ε ₁₂ in the equation (28) is set by the time constant given to each integrator 45-k.

The filter coefficient a _12k (t) which is the output of the integrator 45-k is input to each multiplier 43-k, and the output signal s ₂ (t) of each node k of the delay network 400 is input.
-(K-1) T). Then each multiplier 4
The respective products obtained in 3-k are all added by the adder 46-k, and the output signal y ₁₂ (t) is output. This signal y
₁₂ (t) corresponds to the output of the adaptive filter 38 in FIG. Note that the adaptive filter 37 and the filter coefficient adjusting unit 39 in FIG. 3 can be configured by using exactly the same circuits.

The delay circuit network 400 can also be realized by the circuit shown in FIG. In the figure, 52-1 to 5-5
2- (K-1) and 53-1 to 53- (K-1) are analog switches, 54-1 to 54- (K-1) and 55-1.
~ 55- (K-1) is a transconductance amplifier,
56-1 to 56- (K-1) and 57-1 to 57- (K-
1) is a capacitor. These form sample and hold circuits 500-1 to 500- (K-1).

Opening / closing of the analog switch 52-k is controlled by the sample / hold signal S / H1, and opening / closing of the analog switch 53-k is controlled by the sample / hold signal S / H2. These sample and hold signals S / H1 and S / H2 have a phase difference of .pi.
This is a phase signal, and the signal s ₂ (t−kT) of each node k is held by opening / closing the analog switch 52-k with the sample hold signal S / H1, and after a predetermined time T, the analog switch 53-k is turned on. By opening and closing, the signal s ₂ (t− (k + 1) T) is transmitted to the next stage (k + 1).

Note that nodes 1 to (K-1) in FIG.
4 correspond to points 1 to (K-1) in FIG. 4, and the continuous operation functions as a delay circuit network.

Note that such a circuit configuration has an analog C
It can be easily designed by MOS (Complementary Metal Oxide Semiconductor) circuits.

As described above, according to this embodiment, since the signal separation method in the second embodiment is realized by the analog circuit, the AD converter can be omitted and the structure can be simplified. effective.

Now, FIG. 6 is a schematic block diagram showing an apparatus to which the signal separation method according to the fourth embodiment of the present invention is applied. In the figure, the sound sources 17 and 18, the microphones 19 and 20, and the subtractors 25 and 26 are the same as those in the first embodiment, and detailed description thereof will be omitted. 6
0 and 61 are analog filters. Note that the filter 6
0 and 61 and subtractors 25 and 26 are signal separation filters 6
00 is configured. The operation of the present embodiment configured as described above will be described below.

The sound wave x ₁ (t) emitted from the sound source 17 and the sound wave x ₂ (t) emitted from the sound source 18 are transfer functions H ₁₁ (ω), H ₁₂ (ω), and H _{21 as shown in FIG.} (Ω), H ₂₂
Sounds are collected by the microphones 19 and 20 via the transmission path having (ω) and become the detection signals e ₁ (t) and e ₂ (t). Here, H _ij (ω) is a transfer function from the sound wave x _j (t) to the detection signal e _i (t), ω is an angular frequency, and i = 1, 2, j = 1, 2, Is.

These transfer functions H ₁₁ (ω), H ₁₂ (ω),
If H ₂₁ (ω) and H ₂₂ (ω) have already been measured and are known, the transfer functions G of the analog filters 60 and 61 are
₁₂ (ω) and G ₂₁ (ω) are obtained by the following equations (35) and (36).

G ₁₂ (ω) = H ₁₂ (ω) / H ₂₂ (ω) (35) G ₂₁ (ω) = H ₂₁ (ω) / H ₁₁ (ω) (36) These equations (35) and The transfer function G ₁₂ represented by (36)
Let g ₁₂ (t) and g ₂₁ (t) be the impulse responses of (ω) and G ₂₁ (ω), in other words, the result of inverse Fourier transform of the transfer function described in the frequency domain into the time domain. , The output separation signal s of the signal separation filter 600
₁ (t) and s ₂ (t) are given by the following equation (37).

[0140]

[Equation 29]

Here, i = 1, 2 and j = 1, 2.

In the equation (37), the time integral term is the impulse response g ₁₂ (t) (or g ₂₁ (t)) and the separated signal s ₂ (t) (or s ₁ (t)). Represents the convolution between, and is an analog filter 60 (or 61)
Through the separated signal s ₂ (t) (or s ₁ (t)).

The impulse response g _ij (t) represented by the equation (37) is obtained by the equation (25) in the second embodiment.
In the equation of discrete time expression represented by, the filter coefficient a _ijk (n) is merely fixed value in continuous time. Therefore, by the same discussion as in the second embodiment, the separation signals s ₁ (t) and s ₂ (t) are made independent, and the detection signal e ₁
It is self-evident that (t) and e ₂ (t) can be separated.

As described above, according to this embodiment, the transfer functions of the transfer paths from the sound sources 17 and 18 are measured in advance to determine the analog filters 60 and 61.
There is an effect that a complicated adjustment process is not required unlike the first to third embodiments.

For reference, as an application example of the apparatus using the above-mentioned signal separation method, a handset of a telephone for removing ambient noise components from transmitted voice is shown in FIG. In the figure, in addition to the first microphone 48 for detecting the transmitted voice, the handset 47 of the mobile phone has a second background noise detection second unit on the back surface of the handset 47.
The microphone 49 of is installed. The second microphone 49 is installed on the back surface of the receiver 50 in order to avoid mixing the voice of the sender as much as possible.

Detection signal e ₁ of the first microphone 48
(T) and the detection signal e of the second microphone 49
₂ (t) is input to the signal separation circuit 51. The signal separation method used in the signal separation circuit 51 may be any of the above-described first to third embodiments.

At this time, as the separation signal s ₁ (t), a signal obtained by removing the ambient noise from the detection signal e ₁ (t) is output, and this is transmitted as a transmission signal. This allows S
The transmission signal having a good / N ratio (signal to noise ratio) can be transmitted. Further, the signal separation circuit 51 is configured by an analog circuit as shown in the third embodiment, and this is a CMO.
If the S circuit is used for design, a device with extremely low power consumption can be configured.

In the above description, the signal separation filter has two inputs and two outputs, but the MIM having more inputs and outputs.
The same effect can be obtained with an O filter.
Besides, the present invention is not limited to the above-mentioned embodiment, and various modifications can be made.

[0149]

As described above, according to the present invention, claim 1
In the above method, the separation factor s at the time t is calculated by using the gain coefficient c _ij (t) and the delay time d _ij (t) which are determined in relation to the transfer characteristic between each signal source and each signal detecting means.
By derived using equations described _i (t) to the claim 1, the original signal x _i (t) is propagated to each signal detecting means,
Even if there is a time lag between the original signals x _i (t) detected by the respective signal detecting means, it is possible to separate the signals with high accuracy and sufficiently suppress the echo.

Further, in determining the delay time d _ij (t) for each separated signal s _i (t), the cross-correlation function between the detection signal e _i (t) and the separated signal s _j (t) is obtained, and the correlation is high. By adopting the delay time, the delay time d _ij (t)
By sequentially adjusting the gain coefficients c _ij (t) so as to minimize the power of the corresponding separated signals s _i (t), and when the transfer characteristic is unknown or the signal source There is an effect that signals can be separated even when there is movement.

Further, according to the method of claim 2 , the filter coefficient a _ijk (t) (k = 0, 1, ..., Determined in relation to the transfer characteristic between each signal source and each signal detecting means). K-1)
To obtain the separated signal s _i (t) at time t

The original signal x _i is calculated by the formula described in
Even if the delay characteristic and the attenuation characteristic in each detection process of (t) are not uniform on the frequency axis, there is an effect that the signals can be accurately separated.

Further, the filter coefficient a _ijk (t) is

By adjusting with the formula described in (Equation 3) , there is an effect that the signal can be separated even when the transfer characteristic is unknown or the signal source moves.

[0153]

[Brief description of drawings]

FIG. 1 is a block diagram of an apparatus to which a signal separation method according to a first embodiment of the present invention is applied.

FIG. 2 is a vector diagram illustrating an adaptive operation of a gain coefficient in the embodiment.

FIG. 3 is a block diagram of an apparatus to which a signal separation method according to a second embodiment of the present invention is applied.

FIG. 4 is a block diagram of an apparatus to which a signal separation method according to a third embodiment of the present invention is applied.

FIG. 5 is a circuit diagram showing another configuration example of the delay circuit network in the embodiment.

FIG. 6 is a block diagram of an apparatus to which a signal separation method according to a fourth embodiment of the present invention is applied.

FIG. 7 is a configuration diagram in the case where the device to which the signal separation method of the present invention is applied is applied to a handset of a telephone.

FIG. 8 is a block diagram of an apparatus to which the signal separation method according to the first conventional example of the present invention is applied.

FIG. 9 is a block diagram when the device of the conventional example is simplified to have two inputs and two outputs.

FIG. 10 is a block diagram of an apparatus to which a signal separation method according to a second conventional example of the present invention is applied.

FIG. 11 is a block diagram of an apparatus to which a signal separation method according to a third conventional example of the present invention is applied.

[Explanation of symbols]

17, 18 sound source 19, 20 microphone 21, 22 Low pass filter 23, 24 AD converter 25,26 Subtractor 27, 28 Multiplier 29, 30 delay device 31, 32 correlator 33, 34 Delay time adjustment unit 35, 36 Gain coefficient adjusting unit 37, 38 Adaptive filter

─────────────────────────────────────────────────── ─── Continuation of front page (72) Hiroaki Niwamoto 22-22 Nagaike-cho, Abeno-ku, Osaka-shi, Osaka Prefecture Sharp Corporation (72) Tatsuya Nomura 22-22 Nagaike-cho, Abeno-ku, Osaka City, Osaka Sharp shares In-house (56) References JP-A-6-75591 (JP, A) Jikkyo 2-10559 (JP, Y2) (58) Fields investigated (Int.Cl. ⁷ , DB name) G10L 21/02 G10L 15/02

Claims (2)

(57) [Claims] 1. At time t, original signals x _i (t) emitted from a plurality of signal sources (where i = 1, 2, ...,
I) are mixed and propagated in a plurality of paths having a predetermined transfer characteristic, and the detection signals e are detected in a plurality of signal detecting means.
These signal detections e when detected as _i (t)
_In a signal separation method for estimating the original signal x _i (t) before mixing from _i (t) and separating it as a separated signal s _i (t), a predetermined gain determined according to the transfer characteristic. Coefficient c
_{Using ij} (t) and a predetermined delay time d _ij (t), the separated signal s _i (t) is calculated as follows. The delay time d _ij (t) is calculated by the following equation and the detected signal e _i (t)
Using the correlation function between the separated signals s _j (t),
It is adjusted sequentially so that the correlation between the two becomes high.
The in-coefficients c _ij (t) correspond to the corresponding separated signals s _i.
A signal separation method, which is sequentially adjusted so as to minimize the power of (t) . 2.Emit from multiple signal sources at time t
Original signal x _i (T) (However, i = 1, 2, ...,
I) are mixed in multiple paths with predetermined transfer characteristics
The signal e propagates and is detected by a plurality of signal detecting means.
_i When detected as (t), these signal detections e
_i The original signal x before mixing from (t) _i Estimate (t) and
Separation signal s _i A signal separation method for separating as (t)
Be careful A filter that is determined according to the impulse response of the above transfer characteristics.
Coefficient a _ijk (T) (however, k = 0, 1, ..., K-
1) and a predetermined discrete time T, the separated signal s
_i (T) [Equation 2] Is derived by The filter coefficient a _ijk (T) is ε _ij Any positive small
Assuming a small value, the separated signal s _i The power of (t)
To minimize [Equation 3] Is adjusted sequentially according to Characterized byBelief
No. separation method.