close

Вход

Забыли?

вход по аккаунту

?

JP2002084593

код для вставкиСкачать
Patent Translate
Powered by EPO and Google
Notice
This translation is machine-generated. It cannot be guaranteed that it is intelligible, accurate,
complete, reliable or fit for specific purposes. Critical decisions, such as commercially relevant or
financial decisions, should not be based on machine-translation output.
DESCRIPTION JP2002084593
[0001]
BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a
signal separation system for separating and extracting source signals before mixing using a
mixed signal obtained by mixing a plurality of source signals, and in particular, to a signal
separation process. The present invention relates to a signal separation system and a signal
separation method that can be performed in a short time and with high accuracy.
[0002]
2. Description of the Related Art Conventionally, as a signal separation method for separating
and extracting a source signal before mixing using a mixed signal obtained by mixing a plurality
of source signals (audio signal, music, etc.), blind signal separation is known. The law is known.
The blind signal separation method is a method of separating and extracting a source signal
before mixing from an observation signal (mixed signal) observed by a sensor such as a
microphone, on the condition of only statistical independence between the source signals. It is a
method that does not require a priori knowledge of the special statistical nature of the signal, the
transfer function (the transfer function between the signal source and the sensor), etc. For this
reason, the blind signal separation method is expected to be applied to various fields for source
signals whose position and special statistical properties are not known. Specifically, for example,
(1) measurement of biological signals, (2) control of a machine by the voice of a specific
individual in a state in which voices of plural people and environmental sounds are mixed,
telephone calls with hands free, (3) Machine fault diagnosis by sound in a factory where multiple
mechanical sounds are mixed, (4) An auditory function called cocktail party effect that is
10-05-2019
1
intrinsically provided to humans (while multiple humans are talking simultaneously Therefore, it
is expected to be applied to various fields such as modeling of functions that can distinguish a
specific person's conversation.
[0003]
By the way, a conventional signal separation system using such a signal separation method has a
plurality of microphones for observing a mixed signal of source signals outputted from a
plurality of sound sources, and observation signals observed by each microphone By processing
the (mixed signal) by a predetermined signal separation method, the source signal before mixing
is separated and taken out.
[0004]
However, in the above-described conventional signal separation system, it takes a lot of time for
processing (signal separation processing) to separate and extract the source signal before mixing
from the observation signal (mixing signal). Also, the output signal (separate signal) after
processing is also low in separation accuracy in which other signals etc. are mixed, and it has
been difficult to apply to practical applications.
[0005]
Under these circumstances, in the process of researching a signal separation system, the present
inventors use a directional sensor as a sensor for observing a plurality of source signals (mixed
signals) for signal separation processing. It has been found that it is possible to significantly
reduce such time and to significantly improve the separation accuracy of the signal.
[0006]
The present invention has been made based on such findings, and a signal separation system and
signal separation method capable of significantly reducing the time required for signal separation
processing and significantly improving the signal separation accuracy. Intended to provide.
[0007]
SUMMARY OF THE INVENTION According to the present invention, there are provided a
plurality of directivity sensors for observing a mixed signal of source signals output from a
plurality of signal sources, and an observation signal observed by each of the directivity sensors
A signal processing system for processing the signal according to a predetermined signal
separation method, and using the observation signal to separate and extract a source signal
before mixing.
10-05-2019
2
[0008]
In the present invention, the plurality of signal sources are installed in the same hemispherical
region based on the installation positions of the plurality of directional sensors, and the pointing
direction of each of the directional sensors is the hemispherical region It is preferable to be
directed inward.
Moreover, it is preferable that the said processing apparatus is a parallel computer which
processes the observation signal observed by each said directivity sensor in parallel per sensor
unit.
Furthermore, a blind signal separation method can be used as the predetermined signal
separation method.
[0009]
Further, according to the present invention, there is provided a signal separation method for
separating and extracting a source signal before mixing using a mixed signal obtained by mixing
a plurality of source signals, the mixing of the source signals output from the plurality of signal
sources. A step of observing signals with a plurality of directivity sensors, a calculation step of
taking an observation signal observed by each of the directivity sensors as input and calculating
an output signal corresponding to the observation signal through a predetermined filter, and the
output Evaluating the signal according to a predetermined evaluation function, and correcting
the coefficients of the filter based on the evaluation result, the calculating step and the
calculating step until the output signal output through the filter is sufficiently converged A
method of signal separation characterized in that the correction step is repeated, and when the
output signal is sufficiently converged, the output signal at that point is taken out as a source
signal before mixing. To provide.
[0010]
According to the present invention, since the plurality of source signals (mixed signals) output
from the respective signal sources are observed by the directivity sensor, the source signals
output from the respective signal sources are transmitted to the respective directivity sensors. It
10-05-2019
3
will be observed with a greater intensity difference than non-directional sensors.
Therefore, the same source signal is included at different intensities in the observation signals
observed by each directivity sensor, and the dependency between the observation signals is
sufficiently reduced.
Therefore, if it is required to have the same degree of signal separation accuracy as that of a
system using a nondirectional sensor, the time required for signal separation processing can be
significantly reduced.
In addition, since the dependency between the observed signals is sufficiently reduced by the
difference in intensity between the source signals, the distance between the directional sensors
can be narrower than when using non-directional sensors, and signal separation can be achieved.
It is also possible to reduce the size of the sensor part of the system.
[0011]
Further, according to the present invention, since a plurality of source signals (mixed signals)
output from each signal source are observed by the directivity sensor, the influence of
reverberation which causes a problem in a real environment is effectively made. Therefore, the
processing time can be shortened, and the separation accuracy of the signal obtained after
processing can be significantly improved.
[0012]
Furthermore, according to the present invention, since processing of observation signals
observed by each directivity sensor is processed in parallel in parallel by a parallel computer, the
time required for signal separation processing can be further shortened. it can.
[0013]
DETAILED DESCRIPTION OF THE INVENTION Embodiments of the present invention will be
described below with reference to the drawings.
1 to 3 are diagrams for explaining an embodiment of a signal separation system according to the
10-05-2019
4
present invention.
[0014]
As shown in FIG. 1, the signal separation system 10 according to the present embodiment
monitors a plurality of mixed signals of source signals output from a plurality of sound sources
(signal sources) 41, 42,. The directional microphones (directional sensors) 21, 22, ..., 2n are
provided.
Here, the sound sources 41, 42, ..., 4n are installed in the hemispheric region located in front of
the installation positions of the directional microphones 21, 22, ..., 2n, and the directional
microphones 21, 22, ..., 2n The pointing direction is directed into the hemispheric region.
[0015]
Further, parallel computers (processing devices) 12 are connected to the directional microphones
21, 22,..., 2 n via the A / D converter 11, and the directional microphones 21, 22,. It is possible to
process the observation signal observed in the above and separate and extract the source signal
before mixing using the observation signal.
Note that speakers 31, 32, ..., 3n for monitoring are connected to the subsequent stage of the
parallel computer 12 via the D / A converter 13, and the source signal before mixing extracted
by the parallel computer 12 is used as sound. It is supposed to be output.
[0016]
The parallel computer 12 processes the observation signals observed by the directional
microphones 21, 22, ..., 2n in parallel in units of microphones (sensors). Here, a blind signal
separation method is used as a processing method (signal separation method) in the parallel
computer.
[0017]
10-05-2019
5
In the blind signal separation method used in this embodiment, the source signal is s (t) = [s1 (t),.
. . , SN (t)] T, the observed signal x (t) = [x1 (t),. . . , XN (t)] T, and s (t) is separated using only x (t),
assuming a model represented by the following equation (1). Here, it is assumed that there are N
sound sources 41, 42, ..., 4n and N directional microphones 21, 22, ..., 2n, and each element of
the source signal s (t) and the observation signal x (t) is a sound source 41. , 42, ..., 4n and the
directional microphones 21, 22, ..., 2n, respectively.
[0018]
In the above equation (1), A (z) = [aij (z)] represents the transfer function between the sound
source and the microphone considering the time delay, and the observation signal xi (t) is It
represents by following Formula (2) (3). In the blind signal separation method, it is assumed that
A (z) is unknown.
[0019]
Here, it is assumed that the source signals s (t) are statistically independent of one another and
the mean value is zero. In this case, the covariance matrix R (t, τ) of the source signal s (t) is
expressed by the following equation (4). It is also assumed that the source signal s (t) is a nonstationary signal (for example, an audio signal, music, etc.). By this assumption, it is possible to
perform signal separation processing using only the correlation of the observation signal x (t) at
the same time.
[0020]
In the above equation (4), diag {... } Has diagonal elements {... Represents a diagonal matrix of}.
Also, E [x] is a set average of x.
[0021]
The source signal s (t) is a non-stationary signal, and in the above equation (4), the elements r1 (t,
10-05-2019
6
τ) (i = 1,...) Of the covariance matrix R (t, τ) , N) independently change with time with respect to
all τ. Also, in the above equation (4), it is assumed that the poles of A (z) and A (z) -1 do not exist
on the unit circle (| z | = 1).
[0022]
Under such premise, the following equation (5) is used to separate and extract the source signal
before mixing from the observation signal x (t). An output signal (separated signal) is output from
y (t) of the following equation (5).
[0023]
Here, when the following equation (6) is substituted into the above equation (5), the following
equation (7) can be obtained as a discrete equation for computer mounting.
[0024]
In the above equation (7), B (k) = [bij (k)] represents the coefficient of the adaptive control type
filter.
Also, L represents a fixed time delay, and M represents the number of taps of the filter.
[0025]
Here, the coefficient bij (k) of the filter of the above equation (7) is updated using the steepest
descent method based on the evaluation function of the following equation (8), and the signal
corresponding to the source signal before mixing ( The separated signal is output from y (t).
[0026]
In the above equation (8), diagE [y (t−L) y (t−L) T] is estimated by the following equation (9).
[0027]
Using φi (t) of the above equation (9), the correction amount ΔB (k) of the filter coefficient B (k)
= [bij (k)] is given by the following equation (10) 11).
10-05-2019
7
[0028]
Note that, in the case of N = 2, the following equation (12) is obtained as a discrete equation for
computer implementation by the above equations (10) and (11).
In the following equation (12), γ is a small positive constant.
[0029]
[Equation 9]
[0030]
Next, the operation of the present embodiment having such a configuration will be described.
[0031]
First, mixed signals of source signals output from the sound sources 41, 42,..., 4n are observed by
the directional microphones 21, 22,.
[0032]
The observation signals observed by the directional microphones 21, 22,..., 2n are input to the
parallel computer 12 through the A / D converter 11.
In the parallel computer 12, the observation signals sent from the directional microphones 21,
22, ..., 2n are processed by the above-mentioned blind signal separation method, and the source
signals before mixing are separated and taken out using the observation signals.
[0033]
Specifically, as shown in FIG. 2, the parallel computer 12 first sets initial values of parameters N,
L, M, β, γ (step 101).
10-05-2019
8
[0034]
Next, observed signals x (t) = [x1 (t),. . . , XN (t)] T (step 102), the output signal (separated signal)
yi (t) (i = 1,..., N) is calculated according to the above equation (7) (step 103) .
[0035]
Thereafter, after the evaluation formula φi (t) is calculated according to the above equation (9)
(step 104), the filter coefficient bij (k) is corrected according to the above equation (12) (step
105).
[0036]
Then, the convergence of the output signal (separated signal) yi (t) is judged, and the processing
of steps 103 to 105 is repeated until it is judged that the convergence is sufficient (step 106).
[0037]
Note that the processing of step 103 and step 105 is carried out according to the observation
signal x (t) = [x1 (t),. . . , XN (t)] T are processed in parallel in units of the directional microphones
21, 22,.
[0038]
As described above, according to the present embodiment, a plurality of source signals (mixed
signals) output from the respective sound sources 41, 42,... Are observed by the directional
microphones 21, 22,. The source signals output from the respective sound sources 41, 42,... Are
observed by the directional microphones 21, 22,.
Specifically, as shown in FIG. 3, the source signals output from the sound source 41 are observed
at different intensities by the directional microphones 21 and 22 (see symbols A1 and A2), and
the sources output from the sound source 42 The signals are also observed at different
intensities by the directional microphones 21 and 22 (see symbols B1 and B2).
In FIG. 3, dotted lines attached to the front of each of the directional microphones 21 and 22
indicate contours of the sound collection intensity of the microphones, and solid lines indicate
directional characteristics of the microphones.
10-05-2019
9
[0039]
Therefore, the same source signal is included at different intensities in the observation signals
observed by the directional microphones 21, 22, ..., 2n, and the dependency between the
observation signals is sufficiently reduced. It becomes.
For this reason, if the same degree of signal separation accuracy as that of a system using a
nondirectional sensor is required, the above-described steps 103 to 105 can be performed with a
smaller number of repetitions. The time required for signal separation processing can be
significantly reduced.
In addition, since the dependency between the observed signals is sufficiently reduced by the
difference in intensity between the source signals, the distance between the directional
microphones 21, 22, ..., 2n may be narrower than when using a nondirectional sensor. It becomes
possible to reduce the size of the sensor part of the signal separation system.
[0040]
Further, according to the present embodiment, a plurality of source signals (mixed signals)
outputted from the respective sound sources 41, 42,..., 4n are observed by the directional
microphones 21, 22,. Therefore, the influence of the reverberation which is a problem in the real
environment can be effectively reduced, so that the processing time can be shortened and the
separation accuracy of the signal obtained after the processing can be remarkably improved.
[0041]
Further, according to the present embodiment, the processing of the observation signals
observed by the directional microphones 21, 22, ..., 2n is processed in parallel by the parallel
computer 12 so that signal separation can be achieved. Processing time can be further shortened.
[0042]
In the embodiment described above, the blind signal separation method using non-stationarity of
10-05-2019
10
signals is used as the signal separation method, but the present invention is not limited to this,
and non-Gaussianity of signals is used. Other blind signal separation techniques are also possible.
Further, not only the blind signal separation method but also any other signal separation method
in which the statistical properties of the signal, the transfer function and the like are considered
to some extent can be used.
[0043]
In the embodiment described above, although the case where the signal source is a sound source
has been described as an example, the present invention is not limited to this. The same applies
to the case of signal sources).
In this case, a directional electrode (an electrode used with an electroencephalograph or an
electrocardiograph and attached to a living body) is used as a sensor for brain waves or an
electrocardiogram, and a sensor for brain images is used as a sensor for brain images. It is
preferable to use an imaging device.
[0044]
Here, in the case where the present invention is applied, the signal source needs to be “pointlike”, that is, spatially localized. It can be localized as (dipolar).
Moreover, in the case of, for example, memory or thinking among brain activities, only a part of
the brain is active, and it is possible to localize the signal source also with respect to brain
images.
In the case of a living body, there are echoes inside the living body due to reflections from the
skull and other tissues, etc., and a situation similar to that of sound separation in a room is
created.
10-05-2019
11
[0045]
EXAMPLE Next, a specific example of the above-described embodiment will be described.
[0046]
The present embodiment In this embodiment, two sound sources (speakers) 41 and 42 and two
directional microphones 21 and 22 are echoed in a configuration as shown in FIGS. 4 (a) and 4
(b). It was placed in a normal room.
4 (a) is a plan view showing the arrangement relationship of the directional microphones 21, 22,
and FIG. 4 (b) is a side view when the arrangement relationship shown in FIG. 4 (a) is viewed
from the IVb direction.
As directional microphones 21 and 22, a Sony electret condenser microphone ECM-670 having
directional characteristics as shown in FIG. 5 was used.
Further, as the parallel computer 12, a UNIX (registered trademark) workstation model J7000
manufactured by Hewlett-Packard having two CPUs (PA-8500) was used.
Further, N = 2, L = 100, M = 800, β = 0.9, γ = 0.000005, and the filter coefficients b12 (k) and
b21 (k And 0), and the initial value of the evaluation formula φi (t) is 1.
[0047]
Comparative Example As a sensor for observation, a conventional nondirectional microphone
(manufactured by Rion) was used instead of the directional microphone. The other points are the
same as the above-described embodiment.
[0048]
Results FIGS. 6 (a) (b) (c) (d) (e) (f) are diagrams showing the results of signal separation
10-05-2019
12
processing in the present embodiment and the comparative example. 6 (a) and 6 (b) show the
source signals s1 (t) and s2 (t) output from the sound sources 41 and 42, respectively, and FIGS.
6 (c) and 6 (d) show the outputs obtained in this embodiment. 6 (e) and 6 (f) show the output
signals (separated signals) y1 (t) and y2 (t) obtained in the comparative example. ing.
[0049]
As shown in FIGS. 6 (a), (b), (c) and (d), in this example, a sufficient separation accuracy was
secured as compared with the comparative example. Further, in the present embodiment, the
output signal (separation signal) is output in about 45 seconds of calculation time. On the other
hand, in the comparative example, the number of loops (three times in the present embodiment)
is three times more than in the present embodiment until the output signal (separating signal)
shown in FIGS. On the other hand, it takes 30 times in the comparative example. From this result,
it can be seen that the present embodiment has realized a significant reduction in processing
time and an improvement in separation accuracy.
[0050]
As described above, according to the present invention, it is possible to greatly shorten the time
required for signal separation processing and to significantly improve the signal separation
accuracy.
[0051]
Brief description of the drawings
[0052]
1 is a diagram showing an embodiment of a signal separation system according to the present
invention.
[0053]
Flow chart for explaining the processing content in the parallel computer (processing apparatus)
of the signal separation system shown in FIG.
[0054]
3 is a diagram for explaining the observation state of each directional microphone of the signal
10-05-2019
13
separation system shown in FIG.
[0055]
4 is a diagram showing the relationship between the sound source (signal source) and the
directional microphone (directivity sensor) in the specific embodiment of the signal separation
system shown in FIG.
[0056]
5 is a diagram showing the characteristics of the directional microphone of the specific
embodiment shown in FIG.
[0057]
6 is a diagram showing the result of the signal separation process of the specific embodiment
shown in FIG.
[0058]
Explanation of sign
[0059]
DESCRIPTION OF SYMBOLS 10 Signal separation system 11 A / D converter 12 Parallel
computer (processing apparatus) 13 D / A converter 21, 22, ..., 2 n Directional microphone
(directivity sensor) 31, 32, ..., 3 n Speaker 41 for monitor , 42, ..., 4n sound source
10-05-2019
14
Документ
Категория
Без категории
Просмотров
0
Размер файла
24 Кб
Теги
jp2002084593
1/--страниц
Пожаловаться на содержимое документа