 |
Description  |
|
|
BACKGROUND OF THE INVENTION
1. Field of the Invention
The invention relates to a source characterization system which, starting
from first signals E(t) formed by linear convolutive mixing of primary
signals X(t) from respective primary signal sources, supplies at least one
second signal characterizing at least one respective primary signal X(t),
the system comprising source separation means.
The invention also relates to the use of such a system for the control of
apparatuses for transmitting and/or receiving electric, acoustic or
electromagnetic signals. This may be, for example, a car radio or a
hands-free telephone.
2. Description of the Related Art
The technique of separating primary signals from independent sources by
processing mixtures of primary signals so as to separately extract each
primary signal from these mixtures is known. This technique applies to
primary signals which are only available in the form of said mixtures.
This may concern general linear convolutive mixtures. The mixtures may
have various origins. They may be produced by mechanisms of propagation of
primary signals and/or by mechanisms of superposition of signals from a
plurality of sources or by other causes.
In general, the separation technique is blind, i.e., the sources are
assumed to be unknown and independent with unknown mixtures. For this
purpose, a plurality of samples of these mixtures are detected, from these
samples, one or a plurality of the original primary signals can be
restored by means of separation algorithms.
Such a technique is known, for example, from the document "Blind separation
of sources", C. JUTTEN, J. HERAULT, Signal processing 24 (1991), pages
1-10.
This document describes source separation means comprising a neuron network
which receives at its input a plurality of components of signal mixtures
E(t) and which restores the separate primary signals at its output. The
neuron network operates recursively to calculate the synaptic coefficients
by means of an adaptive algorithm. In this way, it is possible to process
instantaneous linear convolutive mixtures, i.e., for which at any instant,
each signal E(t) is a linear combination, with fixed or slowly varying
real coefficients, of values of the primary signals X(t) at the same
instant. The system adapts itself continuously to variations of the
mixtures. Also known is the document "New algorithms for separation of
sources" by C. JUTTEN and H. L. NGUYEN THI, Congres Europeen de
Mathematiques, 2-3 Jul. 1991, PARIS (France), which relates to general
linear convolutive mixtures. However, such a system has the drawback that
the blind separation of all the primary sources of the mixture requires a
substantial computing power. This may be a handicap for large-scale
applications. Moreover, problems arise with the accuracy or stability of
the separated sources at the output.
SUMMARY OF THE INVENTION
It is the object of the invention to reduce this computing power by taking
into account the specific character of applications utilizing such a
source separation.
This object is achieved with a source characterization system which
comprises pre-processing means arranged before the source separation
means, the preprocessing means pre-processing the signals E(t) by
determining at least one characteristic variable of the signals E(t), this
variable being supplied in the form of third signals I(t, p) formed by
linear combinations, by fixed or slowly varying coefficients, of
characteristic variables of the same nature related to the primary signals
X(t).
By thus pre-processing the signals E(t), the source separation means are
simplified considerably and only have to process linear combinations of
characteristic variables of the primary signals, i.e., instantaneous
convolutive linear mixtures. At the output, this does not yield the
separated sources X(t) but one or more characteristic variables
characterizing these sources. For many uses, it is in fact merely required
to know these characteristic variables. If necessary, they may
subsequently be processed to obtain the separated sources at the output.
This characteristic variable may be: an average energy, a spectral density,
an autocorrelation function or other variables. This is very interesting,
for example, if the output power of an apparatus, such as a car radio,
should detect the voice of a user of a hands-free telephone to validate
the processing of voice information.
Obviously, it is also possible to extend the system so as to supply either
a plurality of signals F(t) for a plurality of characteristic variables
for one and the same primary signal X(t), or a plurality of signals F(t)
for an identical or non-identical characteristic variable relating to a
plurality of primary signals X(t).
For some characteristic variables, such as a spectral density of a signal,
pre-processing may be followed by post-processing to transform a computed
characteristic variable into another characteristic variable. According to
the invention, post-processing means are arranged at the output of the
source separation means. By combining the pre-processing effected by the
pre-processing means and the post-processing effected by the
post-processing means, it is possible to determine spectral densities for
one or a plurality of primary signals or other characteristic variables.
These and other aspects of the invention will be apparent from and
elucidated with reference to the embodiments described hereinafter.
BRIEF DESCRIPTION OF THE DRAWINGS
The invention will be more fully understood with the aid of the following
Figures, given by way of non-limitative examples, in which:
FIG. 1 is a block diagram of a source characterization system comprising
pre-processing means.
FIG. 2 shows a source characterization system comprising control means for
retroacting on primary signal sources.
FIG. 3 is an example of a diagram of a part of the pre-processing means for
determining the average energy of a signal.
FIG. 4 is an example of a diagram of a part of the pre-processing means for
determining autocorrelation functions.
FIG. 5 is a block diagram of a source characterization system comprising
pre-processing means and post-processing means.
FIG. 6 is a diagram similar to that shown in FIG. 3 with filter means for
processing general convolutive mixtures.
DESCRIPTION OF EMBODIMENTS
A distinction is made between several types of signals E(t) in accordance
with the nature of the mixtures from which they originate. General
convolutive linear mixtures produce generic signals E.sub.i (t) such that
##EQU1##
where c.sub.ij (t) are the impulse responses of the mixture filters, and
where the symbol (*) represents a convolution product.
These general mixtures include a family of mixtures for which an arbitrary
signal X(t) propagates with a fixed propagation delay .theta. and with a
constant attenuation 1/.alpha..sub.ij. This corresponds, in particular, to
the propagation of sound waves in free air.
The generic signals E.sub.i (t) are then of the type:
##EQU2##
where .theta..sub.ij are constants defining a propagation delay and where
the symbol (.multidot.) denotes a classical multiplication. These are
linear convolutive mixtures with a fixed non-zero delay.
This family includes a sub-family of mixtures for which the propagation
delay .theta..sub.ij is zero. The generic signals E.sub.i (t) are then of
the type:
##EQU3##
where a.sub.ij are fixed or slowly varying coefficients (which may differ
from the coefficients .alpha..sub.ij). These are linear instantaneous or
linear convolutive mixtures with zero delay.
The cited document by C. JUTTEN and J. HERAULT relates to signals in
accordance with equation (3). The invention also relates to signals of two
other types.
By way of example, the case is considered that the volume of sound of a car
radio mounted in a vehicle is to be controlled. The car radio comprises
means (not shown) which can be influenced to control the volume of sound
produced by the car radio depending on ambient sound sources. Thus, if the
ambient noise increases (open windows, higher speed, driving noise . . . )
it may de desirable to increase the sound level produced by the car radio.
This is not the case if the ambient sound sources are formed by the voices
of passengers. The problem is then to reduce rather than increase the
sound volume when the passengers talk. This requires identification of
passenger voices. FIG. 1 shows primary sources 5, S1 to Sn, formed by, for
example, the passengers voices, by various noise sources (engine, car
body, air circulation through the windows etc.) and by the car radio
itself. To identify the voices, transducers C1-Cn, for example,
microphones, are placed inside the passenger space. A transducer may
directly pick up the sound emitted by the loudspeaker. The microphones
detect first signals E.sub.1 (t) to E.sub.n (t) from mixtures of primary
signals X.sub.1 (t) to X.sub.n (t) supplied by the sources S1 to Sn.
The mixtures formed inside the passenger space may be regarded as being
related directly to the propagation of sound signals in air. In a first
approximation, these mixtures may be characterized by a non-zero
attenuation coefficient and a non-zero propagation time constant
characteristic of each source. The signals E.sub.i (t) detected by the
microphones can be defined by the expression:
##EQU4##
in which: i is a current index of a microphone
j is a current index of a source
1/.alpha..sub.ij, .theta..sub.ij are attenuation coefficients and
propagation time constants, respectively, characteristic of the
propagation from the source S.sub.j to the microphone E.sub.i.
In accordance with the invention, the signals E.sub.i (t) enter the source
characterization system 8, which comprises pre-processing means 20
followed by source separation means 10.
Source separation means capable of separating instantaneous linear signal
mixtures, i.e., mixtures for which the terms .theta..sub.ij are all zero
(equation 3), are much simpler and therefore easier to realize than source
separation means capable of separating non-zero linear convolutive
mixtures in accordance with equation 2. The means for processing linear
instantaneous mixtures can be those described in the cited document by C.
JUTTEN and J. HERAULT.
In accordance with the invention, source separation means (10), capable of
separating linear instantaneous signal mixtures is selected, which is
preceded by pre-processing means 20, which converts the non-zero delay
linear mixtures formed by the first signals E.sub.i (t) so as to obtain
said linear combinations formed by the third signals I.sub.i (t, p). Here,
t is the time and p is a significant parameter for the relevant
characteristic variable (p is, for example, a frequency). The parameter p
is not to be used if the characteristic variable is an average energy.
The structure of the pre-processing means 20 depends on the characteristic
variable of the primary signals X(t) to be identified. This variable may
be:
the average energy of a primary signal X(t) of one or more sources. In this
case, it is an overall characteristic variable determined for a given
duration, for successive time intervals,
an autocorrelation function of a primary signal X(t) or a plurality of
primary signals X(t), determined for a given duration, for one or more
successive time intervals,
a spectral density giving the spectral distribution of a primary signal
X(t) or a plurality of primary signals X(t), determined for a given
duration, for one or more successive time intervals.
One skilled in the art will be able to carry the invention into effect for
other characteristic variables without departing from the scope of the
invention.
To carry out the invention, the d.c. components of the signals E(t) are
first removed, for example by low-pass filtering. This yields signals E(t)
whose average value is zero, as in the document by C. JUTTEN and J.
HERAULT.
There is a particular case for which there are two sources S1 and S2 and
two mixed signals E1(t) and E2(t) among which one of them E1(t) is a
mixture of sources signal and the other is a pure signal (or considered as
such) coming directly from one of the sources. In that case, sources
separation means 10 can be simplified into an adaptive filtering device
realizing a substraction of the pure signal from the other signal after
weighting adaptively the signals.
The case can be generalized to the one having a pure signal and a plurality
of mixed signals containing the pure signal.
Practically, a pure signal can be measured in a car radio directly at the
output of the car radio just before its transmission by the loudspeaker.
By way of example, the case will be considered in which the average energy
of a primary signal X(t) is determined. For this purpose, the
pre-processing means 20 computes the average signal energy E.sub.i (t).
For the signal E.sub.1 (t), this average energy is determined for a
duration .delta., starting from an instant t=T.sub.0, such that:
##EQU5##
In the case of two sources S1 and S2 with mixtures in accordance with
equation (2), the signal measured by the transducer C1 may be written as:
E1(t)=.alpha..sub.11 .multidot.X.sub.1 (t-.theta.)+.alpha..sub.12 X.sub.2
(t-.theta..sub.12)
The average energy is:
.epsilon..sub.E1 (T.sub.0,.delta.).congruent..alpha..sub.11.sup.2
.multidot..epsilon..sub.X1 (T.sub.0,.delta.)+.alpha..sub.12.sup.2
.multidot..epsilon..sub.X2 (T.sub.0,.delta.),
in the case of non-correlated sources X.sub.1 and X.sub.2, and by selecting
.delta. in such a manner that:
.theta..sub.11,.theta..sub.12 <<.delta.. .
It is found that the delay terms .theta..sub.11, .theta..sub.12 have
disappeared and that the term .epsilon..sub.E1 is a linear mixture of the
average energies of the sources S1 and S2, with fixed or slowly varying
coefficients and zero delay.
Calculation of the average energies of the signals E.sub.i (t) thus yields
signals I(t, p)=.epsilon..sub.Ei (T.sub.0,.delta.) from which the energies
of the primary signals can be extracted.
FIG. 3 shows an element 22 forming part of the pre-processing means 20.
This element 22 comprises means 24 for multiplying the signal E.sub.i (t)
by itself, followed by means for storing the multiplication results for
the duration .delta.. At the output, this yields the average energy
.epsilon..sub.E1 corresponding to the signal E.sub.i (t).
The diagram in FIG. 3 represents a variant for digital signal processing.
The same result as with an analog version can be obtained, for example, by
integrating the rectified signal by means of capacitances, without thereby
departing from the scope of the invention.
The pre-processing means 20 then comprises a plurality of elements 22, each
being associated with a signal E.sub.i (t). It is possible to apply time
multiplexing At the output, a plurality of average energies
.epsilon..sub.E1 is obtained, corresponding to each signal E.sub.i (t).
All these average energies .epsilon..sub.E1 are input into the source
separation means 10 in a manner as shown in FIG. 1. This means effect
source separation and supplies output signals F.sub.i (t) which, in the
present case, are the average energies of each of the primary signals
X.sub.i (t) in the successive time intervals considered.
Preferably, the source separation means 10 comprises a neuron network. This
is, for example, a network as described in the cited document by C. JUTTEN
and J. HERAULT. In accordance with the technique known to those skilled in
the art, the operation of a neuron network comprises two phases:
a learning phase in which it learns to perform a task, and
a resolution phase in which it uses the data thus learned to determine the
results corresponding to current values.
These learning techniques are known and will not be explained here. It is
adequate to know that they consist in applying samples to the input and
modifying the characteristics of the neuron network (basically its
synaptic coefficients) to produce, at its output, the envisaged results
corresponding to the samples. Use is made of "unsupervised" learning,
i.e., on the basis of the samples, an algorithm is applied to adapt the
synaptic coefficients in such a manner that the outputs of the neuron
network supply second signals F(t) which are independent of one another.
Such a learning process is known to those skilled in the art.
With the configuration of synaptic coefficients thus determined, the neuron
network is now used with new input data. It is thus capable of determining
results depending on the new input data for the function it has learned to
perform. In order to allow the system to effect a continuous
characterization of sources by adapting itself to the continual variations
of the mixtures, learning is effected iteratively so as to update the
synaptic coefficients for the characterization of new mixtures.
Measured signals E.sub.i (t), which correspond to primary sources X.sub.i
(t) with unknown characteristics, are applied to the input and the means
10 supplies measures of these unknown characteristics (in the present case
average energies).
The characteristic properties of the primary signals X(t) to be measured
are determined in successive time intervals of a duration .delta.. In each
time interval this may concern:
for the average energy, a single value,
for determining the autocorrelation function, a series of values obtained
by comparing a signal with itself at instants which are shifted by a time
.tau.,
for determining spectral densities, a plurality of values determined at
different frequencies.
By way of example, the case will be considered in which autocorrelation
functions .GAMMA.(.tau.) are to be determined for primary signals X.sub.i
(t) with .GAMMA..sub.Xi (.tau.)=.SIGMA.X.sub.i (t).multidot.X.sub.i
(t+.tau.). For this purpose, the pre-processing means 20 determines the
autocorrelation functions of the signals E.sub.i (t) such that
.GAMMA..sub.E.sbsb.1 (.tau.)=.SIGMA.E.sub.i (t).multidot.E.sub.i
(t+.tau.). The means 20 comprises pre-processing blocks 25 as shown in
FIG. 4. The number of blocks 25 is equal to the number of signals E(t) to
be processed. It is also possible to apply time multiplexing. To effect
the correlation of a signal E(t) with itself, the signal E(t) is inputted
into acquisition means 32, which takes successive samples of the signal
E(t) at a given rate (duration D). This may be a memory comprising, inter
alia, storage elements 32a, 32b in which a series of samples is stored.
The shift .tau. is successively given the values D, 2D, 3D etc. The
samples which appear, spaced by a time D, for example, on the output 2a,
2b of the storage elements 32a, 32b, are applied to the processing means
22 already described, in order to determine the values
E(t).multidot.E(t+D). By sliding addressing (arrow 29) of the memory, the
value of the autocorrelation function .GAMMA..sub.E is determined for the
totality of the relevant signal E(t) corresponding to .tau.=D. Other
values of the autocorrelation function are computed likewise with sliding
addressing for a shift 2D and so on for 3D, 4D . . . All of these
autocorrelation values form the autocorrelation function.
It will be obvious to those skilled in the art that the acquisition means
32 may be given another structure than that shown in FIG. 4 without
departing from the scope of the invention. At the output of the element
22, the autocorrelation function .GAMMA..sub.E corresponding to the signal
E(t) is obtained:
.GAMMA..sub.E.sbsb.1 (.tau.)=.SIGMA.E.sub.i (t).multidot.E.sub.i (t+.tau.)
The different signals E(t) in FIG. 1 are thus processed similarly to
provide autocorrelation functions .GAMMA..sub.E for each signal E(t). In
that case, the signals I(t, p) in FIG. 1 are the autocorrelation functions
.GAMMA..sub.E. By applying the same principles as already described for
the average energies, the source separation means 10 learns, on the basis
of examples, to compute autocorrelation functions .GAMMA..sub.X
corresponding to each of the sources.
When the example is limited to two sources S1 and S2, the following is
valid:
E1(t)=.alpha..sub.11 .multidot.X.sub.1 (t-.theta..sub.11)+.alpha..sub.12
X.sub.2 (t-.theta..sub.12)
E2(t)=.alpha..sub.21 .multidot.X.sub.1 (t-.theta..sub.21)+.alpha..sub.22
X.sub.2 (t-.theta..sub.22)
yielding autocorrelation functions .GAMMA.hd E such that:
.GAMMA..sub.E1 (.tau.)=.alpha..sub.11.sup.2 .multidot..GAMMA..sub.X1
(.tau.)+.alpha..sub.12.sup.2 .multidot..GAMMA..sub.X2 (.tau.)
.GAMMA..sub.E2 (.tau.)=.alpha..sub.21.sup.2 .multidot..GAMMA..sub.X1
(.tau.)+.alpha..sub.22.sup.2 .multidot..GAMMA..sub.X2 (.tau.)
The source separation means 10 derives the autocorrelation functions
.GAMMA..sub.X of the source signals such that:
.GAMMA..sub.X (.tau.)=.SIGMA.X(t).multidot.X(t+.tau.)
When the system 8 determines the autocorrelation functions of the primary
signals X(t), it is also possible to derive the energies of these signals.
For this purpose it suffices to select, in FIG. 4, a delay .tau.=0,
yielding at output the values of .GAMMA..sub.X (0) which are equal to the
energies.
This mode of operation is faster than the calculation of energies as set
forth above, because it allows the source characterization system to
effect learning in a single time interval and not in a series of time
intervals. However, this requires more hardware.
In another example, the signals I(t, p) may be signals representing
spectral densities. In that case the pre-processing means 20
computes the autocorrelation functions, subsequently effects a Fourier
transform for each signal E(t) and supplies signals I(t, f), where f is
the frequency, or
it effects a Fourier transform (for example, a fast Fourier transform) and
subsequently calculates the square of the modulus of the Fourier transform
thus determined.
All of the signals I(t, f) derived from the signals E(t) and determined for
the same frequency f are processes in the source separation means 10 as
described hereinbefore. The signals F(t) at the output then represent the
spectral densities D.sub.X corresponding to the sources S for the
frequency f.
Successive operation at d | | |
