WO2014094709A2

WO2014094709A2 - Method for detecting at least two individual signals from at least two output signals

Info

Publication number: WO2014094709A2
Application number: PCT/DE2013/000788
Authority: WO
Inventors: Daniel Kotulla
Original assignee: Ask Industries Gmbh
Priority date: 2012-12-20
Filing date: 2013-12-10
Publication date: 2014-06-26
Also published as: WO2014094709A3; JP6163211B2; JP2016504622A; DE102012025016B3

Abstract

The invention relates to a method for detecting at least two individual signals from at least two output signals, which are formed of at least two source signals, wherein the mixing ratios of at least two source signals on the output signals are determined. The output signals are multiplied by the factors formed from the mixing ratio, and from the output signals, at least two subtraction signals are determined such that from each subtraction signal, a source signal is eliminated. The subtraction signals are Fourier transformed, whereby transformation signals are produced. From the transformation signals, a residual signal is determined. Based on the transformation signals and the residual signal, at least two individual signals are calculated, and the individual signals are transformed by an inverse Fourier transformation.

Description

Method for determining at least two individual signals from at least two

output signals

DESCRIPTION

The invention relates to a method for determining at least two individual signals according to the preamble of claim 1. It is known to record two or more sound sources so that thereby a stereo signal can be formed. For this, level differences or differences in maturity are used. The stereo signals should create a spatial sound impression.

Known methods of stereophony are the AB method, the intensity stereophony in the form of the XY method or the MS method and mixed methods such as the ORTF method or the OSS method.

In recent years, more and more sound systems have come on the market, which include more than two speakers. In particular, such sound systems are already for many years, for example, in motor vehicles having at least two speakers in the side doors and two speakers in the inner region of the motor vehicle.

Therefore, methods have been developed to provide more than two individual signals, as is the case with stereophony. There are essentially two different approaches:

First, there are methods that can calculate out of a variety of output signals of each source signal. The source signal is understood to mean every instrument and every voice or speaking voice. In choirs, however, a plurality of singers can be considered as a source signal, for example, "the soprano" can be understood as a singing voice, although it includes, for example, twenty singers.

All in all, the problem is to filter out any number of source signals from two output signals as they are in stereo signals or even more output signals.

Such source separation methods are referred to as "blind separation." There are essentially two types of methods statistical methods such as independence analysis, principal component analysis, the maximum-a-postariori method, the maximum likelihood method. In addition, it is also known to use recursive optimization methods. The known methods either allow a determination of any source signals from a smaller number of output signals with high computational complexity or require a real-time calculation, a number of output signals that is greater than or equal to the number of source signals.

It is therefore an object of the present invention to further develop the method with the features of the preamble of claim 1 such that even with a number of output signals which is smaller than the number of source signals, a source separation in real time is made possible. This object is solved by the features of claim 1. Advantageous developments of the invention will become apparent from the dependent claims. As the core of the invention, it is considered that subtraction signals are determined from the output signals, which are Fourier-transformed and individual signals are determined in the transformation signals thus obtained. By eliminating statistical methods and recursive methods, it is possible to determine the individual signals in real time. Real-time means that a resulting data stream with a marginal time offset arises when calculating the first block of signals, the signals to be reproduced can be provided without interruption. It is thus possible, for example, to extract from a recording a certain singing voice and, for example, the piano, to assign certain loudspeakers with a mixing ratio to be selected, whereby the output signals need not be preprocessed. It should thus be possible, for example, to insert a CD in a motor vehicle, to specify a song and possibly the distribution of the signals to be extracted as well as the signals to be extracted, and to start playing the song or the CD without it noticeable delay comes. The mixing ratios of source signals to the output signals can be determined with a direction determination. For this purpose, the output signals are Fourier-transformed and the amplitude values are calculated. These amplitude values are combined into a histogram. For this purpose, pairs of dots are formed from the Fourier-transformed output signals, the pairs of points in the histogram being combined by means of the angle obtained. The angle of the pairs of points assumes values between 0 ° and 90 ° and is calculated as follows:

arctan

Here X _{X org} denotes the Fourier-transformed first output signal X, X _{2 org} the Fourier-transformed second output signal X ₂ and a _x and a _y vectors in which the amounts are stored. The letter I is the counting index of the output signals X _x and X ₂ and the Fourier transform output signals X _{X org} and X _{2 org} , on the basis of which the

Paired points are moored. The first value of the Fourier-transformed output signal X _{X org} and the first value of the Fourier-transformed output signal

X _{2 t0r} g _< which are nothing more than vectors, thus form the first pair of points, with I = 1.

Of course, the count index I can also start at 0, it then runs to L - 1. Here L denotes the number of Fourier transform points of the output signals X and X ₂ , for example 4096. Θ is an angle between 0 ° and 90 °. and ψ _ι the amount of the pair of points with index I.

The vector ψ contains the sum of the sums of the vectors a _x and a _y :

Ψι = ^a n + <*

The Fourier transformation is performed block by block. In each case a power of two, ie a power to the base 2, is used as the input signal. It has been found that powers of 10, 11 or 12, ie 1024, 2048 or 4096 data points, are particularly efficient. Particularly preferred are 4096 data points, since in these the computing time is optimized in terms of computational effort.

The division of the histogram is preferably carried out in 1 "increments, ie the histogram contains 90 elements, the values of the angle Θ being rounded to integers in order to narrow the range of values of the histogram.

Then the histogram is smoothed to improve the detection of the maxima. For example, the following function is used as the smoothing function:

The parameter T is an integer and specifies how many neighbor points are included in the smoothing. So there is an averaging over several data points. At the boundaries of the histogram, where there are no neighboring points, the value 0 is used at the corresponding points to be supplemented. When calculating the first numerical value of the histogram, the value 0 must be set eight times on one side, while the existing numerical values are used on the other side. In the smoothed histogram, all local extreme values are determined and sorted according to the height of the numerical value, ie according to frequency. Each position in the histogram corresponds to an angle as described above, so each angle also corresponds to an angle. The corresponding angle is determined for the maxima determined and either a predefined number of maxima are used or all maxima whose frequency is above a predetermined threshold value.

The smoothed histogram vector thus results in

Μ = Σ ( ^Γ ) ^χ Μ ^' The problem may arise that a source signal does not appear in the output signals over the entire duration. For example, certain instruments or voices have pauses. In order for these pauses not to cause errors in the angle determination, it is possible to multiply the following histograms by a weighting function, such as a low pass, after determining two or more histograms and corresponding angle determination. The low-pass can have the following appearance: y (n) = axx (n) + bxy {n-1), where a = 0.1 and b = 0.9. The index n is the index of the histogram, ie the first one

Histogram or the first 4096 data points is n = 1. The application of the low pass results in the histogram h, _TP .

This weighting stabilizes the angle detection. Starting from the determined angles, the mixing ratios of two source signals are determined by the angles γ _Ν = index (max ( _hgl- _TP )). be used in the following equation:

If a maximum of the histogram is, for example, 18 °, this results in a mixing ratio of V = 0.325.

Based on the mixing ratios, subtraction signals are calculated, which result for output signals as follows:

X =

X ₂ - x X _t v _N > \

* N with N = 1, 2, .... N is the index of the source signal to be filtered out. Depending on whether the mixing ratio is greater or less than 1, the corresponding subtraction signal is calculated. It is possible to use more than two output signals, but this only leads to an increased computational effort. If there are not two but more output signals as in the case of a stereo signal, then those two output signals in which the source signal is most strongly represented are preferably selected for the extraction of a source signal.

In the case of the subtraction signals thus determined, a source signal is hidden in each case.

These subtraction signals are then Fourier transformed. This happens block by block, whereby subsequent blocks always start with a distance that corresponds to half the size of the previously transforming data points. That is, the first half of a block has already been Fourier transformed as the second part of the previous block.

This allows continuous signal processing and is known under the name overlap-add method. To minimize the leak effect, the input blocks are multiplied by a window, for example a Hanning window. The window function f (n) for N data points is:

/ («) = Iu il - cos with O <n <N - l

From the Fourier-transformed subtraction signals X, referred to below as transformation signals, a residual signal is determined. For two transformation signals X, the residual signal results as a simple subtraction of the second transformation signal from the first:

X ₃ = X _x - X ₂ .

In the following, two transformation signals X and X ₂ and a residual signal X ^ are assumed, resulting in a total of three signals. For the extraction of a

Single signal are then compared these three signals.

For each data point of the signals,, X ₂ and X _} , usually each signal has 4096 data points, the extreme values are calculated from the amplitudes of the three signals.

The starting point is an array with 3 x 4096 data points. The number 3 indicates the number of transform signals and the residual signal, and the 4096 indicates the number of Fourier transformed data points in a block. Looking at the first data point and the first frequency bin of the vectors X _x, X ₂ and X, so there are three numerical values for comparison. In place of the minimum value of these three values, the maximum value of the other two values is set and the other values are set to zero.

A numerical example for clarification: Let the first value of X _x 5, the first value of X ₂ 10 and the first value of X ₃ 15 then set 15 to the position of 5 and the first value of X ₂ and J ₃ 0. The values are therefore considered column by column. One obtains an array of 4096 columns and three rows in which two-thirds of the values are zero. The values unequal to zero are distributed irregularly over the individual vectors. The individual vectors to X _l and X ₂ are called S, and S ₂ and are the Fourier transform of the calculated source signals S! and S ₂ . The resulting vector of the residual signal X ₃ has no further meaning. The determination of S _x and S ₂ is given by the following formula:

E _m M if X _m (k) = E _rBm (k)

S "{k)

otherwise

m = 1, 2, 3, 4, 5, ....

K is the counting index of the data points or frequency bins and passes through the values 1 to 2048 for 4096 data points. Only half of the frequency bins need to be traversed, since the data points occur twice for reasons of symmetry in the course of the Fourier transformation. In the example above, k = 1.

If one takes the lines thus converted, which have belonged to the signals X and X ₂ , namely the individual signals S _x and S ₂ , one can use them to calculate the calculated source signals ST and

Determine S ₂ by an inverse Fourier transform. As a phase, the phase of the signals X _x and X _{2 can} also be used directly. S _x and S ₂ will be the respective

Phase assigned from the vectors X _x and X ₂ . The assignment is of course dependent on the counting index k.

The individual signals S _x and S ₂ or the Fourier-transformed individual signals S _! and S ₂ may be slightly different from the source signals due to computational inaccuracies and rounding errors. This means that a complete reproduction of a source signal is not possible, but the differences are so small that they are usually not noticeable. To improve the separation of the individual signals, the following steps are possible:

In order to avoid jumps in the minima detection it can be provided that the definition of a minimum in the signals X _X , X ₂ and X ₃ depends on the detection of preceding minima. For example, a conditional low pass filter may be used:

P _ho i _d ik) ^{is the} value of a frequency bin k ^{to be} memorized, k again being the count index. The parameter b is freely adjustable between 0 and 1, whereby the note of the frequency is switched off at b = 0.

The determination of S _i and S ₂ is then given by the following formula: ^ (k), X _m {k) = E _mm (k)

The parameter η (0 <η <\) specifies how strongly the low-pass filtered signal enters the individual signals S _m , m = 1, 2, 3,.

Furthermore, the minima of the transformation signals X _m , m = 1, 2, 4, 5,... Can be compared separately with the residual value signal X, ie the minimum determination Em, _n (k),

Assignment with 0 or the maximum value of the column E _max (k), etc. each between a transformation signal X _m and the residual value signal X _{2 is} performed. Thus, the vectors Emmi, E _min2 , E _max i and Em _a x2 are determined. The minima are multiplied by a factor ß (0 <ß <2). Depending on the choice of the factor, different effects result.

For ß <1 unwanted frequencies are suppressed and for ß> 1 results in a more harmonious sound. The individual signals arise too

In addition, the signals can be separated based on the phase angle. If the amplitude of the transformation signals X _m {k) is substantially identical and has in the residual value signal

X ₃ (k) a minimum, the phase is taken into account. If this is the same, the maximum becomes

S ₃ (k), otherwise S (k) and S ₂ (k). Of course, this consideration is to be performed separately for each frequency bin k.

It applies accordingly: ₌ jmin { _m ( ₃ (*)}

0

m = 1, 2, 4, 5, ...

In addition, the Fourier-transformed output signals X _{m org} can furthermore be taken into account. If there is a minimum at X _{l org} (k), this is also assigned to S _{ (k): (k) v X _mfirg (k) = (k)

otherwise

The invention is explained in more detail with reference to embodiments in the drawing figures. These show:

a block diagram of the method according to the invention for a stereo signal, a block diagram of the method according to the invention in a second embodiment, a flow chart for determining the mixing ratios, a flow scheme for source separation, an arrangement for recording stereophonic signals, a scatter plot of data pairs, a scatter plot of FT data pair , a histogram, the overlap-add method, a 3D amplitude spectrum of two transformation and a residual signal, a 3D amplitude spectrum of the individual signals and 12 transformation signals and individual signals in vector form

Fig. 5 shows an arrangement for receiving a stereo signal. By way of example only, an arrangement for recording the source signals by means of XY stereophony is shown; in principle, however, the method according to the invention can be used not only in all other stereophonic methods, but also in methods in which more than two output signals are generated. Shown are four signal sources, each generating a source signal 1, 2, 3 and 4. Source signal 1 is, for example, a singing voice, ie a singer, with source signal 2 around the background choir consisting of a plurality of singers, who, however, reproduce the same text and the same notes, in the case of the source signal 3 Instrument, for example a piano, and in the case of source signal 4 a group of instruments which, however, like the choir, reproduce the same notes. An example would be a set of violins playing the same melody. This is to show that a source signal can be formed not only by a single singer or a single singer or by a single instrument, but also by a plurality of singers or instruments. Due to the diverging directivity of the microphones 5 and 6, the source signals 1, 2, 3 and 4 are recorded at different levels, which is why the mixing ratio of the source signals 1, 2, 3 and 4 is always different.

From the source signals 1, 2, 3 and 4, the two output signals of the stereo signal are thus obtained. The output signals are those signals which form the starting point of the method according to the invention, the original source signals are no longer available.

Of course, more source signals can be used to generate the output signals X _m , but in order to carry out the method, it is necessary to have at least two source signals.

In the following, reference is frequently made to two output signals X,, X ₂ . The inventive method is not limited to stereo signals, in principle, any number of output signals X _m , m = 1, 2, 3, ... are available.

FIG. 1 shows a block diagram of the method according to the invention in a first embodiment. In this case, two output signals X and X ₂ are used and the Mixing ratio for two source signals, for example, the source signals 1 and 2, determined. A possible method for determining the mixing ratios is explained in detail below. For the source signal 1, for example, a mixing ratio V results and for the source signal 2 a mixing ratio V ₂ . From the output signals X _! and X ₂ are then Subtraktionsignale X _x and X ₂ calculated by the output signals Xi and X ₂ are subtracted as follows, using a mixing ratio:

X _x - V _N x X ₂ _V _N <\

X =

X ₂ ~ x X ^ V _N > \

N stands for the respective index of the source signal.

The thus calculated subtraction signals X _x and X ₂ are Fourier-transformed according to the overlap-add method.

The output signals Xi and X ₂ are of course digital signals, which accordingly correspond to pure sequences of numerical values or data points. One could also represent an output signal as a vector with a very large number of numerical values, usually many tens of thousands. This also applies to the subtraction signals. In order to keep the computational requirements small and in particular to keep the time until the provision of the first individual signals low, the output signals Xi and X ₂ and the subtraction signals X _x and X ₂ are further processed block by block. The

Fourier transformation is applied, for example, to the first 4,096 data points of the subtraction signal X _x and X _2, respectively. Preferably, powers of two of data points are Fourier-transformed, since the fast Fourier transformation (FFT) can be used for these. The number 4,096 represents an optimum value with regard to the used computing time and the used computing resources. It is also possible to take smaller or larger data blocks, for example data blocks of 1024 or 2048 data points. Of course, these are always consecutive data points. By the Fourier transform of _x Substraktionssignaie X and X ₂ to obtain the transformation of signals X and _{2. FIG.} By subtracting the

Transformation signal X ₂ from the transformation signal X, one obtains the residual signal X _3rd

The residual signal X ₃ contains exactly as many data points as the transformation signals X and ₂ , for example 4096 data points. Since these data points are in the frequency domain, they are also typically Called Bins. A frequency bin is therefore a data point of a transformation or residual signal. In the example, each transformation residual signal therefore has 4096 frequency bins. In the signals Χ, X ₂ and X ₃ then minima are searched, since these are synonymous with the hidden signals. For this purpose, the minima are calculated for each frequency bin from the amplitudes of the signals X, X ₂ and X ₃ and sent to them

Minimum digits set the maximum values of the respective frequency bins set the other values for this frequency bin to zero. The phases to the non-zero values are obtained from the In this way one holds the individual signals S _x and S ₂ . These must be transformed back into the time domain, which gives the calculated source signals Si and S ₂ .

FIG. 2 shows an embodiment for determining more than two calculated source signals ST us S ₂ . For this purpose, from the output signals X _{1 (X} _2, X ₃ and X ₄ are determined more than two mixing ratios, thereby more subtraction signals and thus more transformation signals are obtained. Also, in the thus-obtained transform signals X _x, X _2, j ₄ and X _5, and the residual signal X ₃ determined therefrom is followed by a minimization determination and, moreover, by the determination of the individual signals S _{ , S ₂ , S ₃ and S ₄ and from this the calculation of the calculated source signals Si, S ₂ , S ₃ and S ₄ .

In the following, a possibility for determining the mixing ratios is disclosed. In principle, all methods for determining the mixing ratios can be used.

FIG. 3 shows a flow chart for determining the mixing ratios on the basis of a direction determination. In this case, the output signals Xi and X _{2 are} Fourier-transformed in step S1, whereby the transformed output signals X _{X org} and X _{2 org} are obtained. In step S2, for each data point or frequency bin of the Fourier-transformed output _signals X _{X org} and X _2t0rg, the amplitude _values are determined on the basis of _FIG

formulas

determined. In the next step S3, the calculation of the angle Θ and the absolute value ψ takes place from the previously determined values. In this case, the magnitude values of the Fourier-transformed output _signals X _{l org} and X _firg can be calculated on a _data point-by- _pixel basis or as a vector, at least value pairs of angles Θ and magnitude values ψ are to be determined therefrom.

In step S4, a histogram is calculated from the value pairs Θ and ψ, which is smoothed in step S5.

In step S6, the determination of the maximums of the histogram follows, whereby either a predetermined number or all maxima above a threshold value are used.

Each of these maxima is assigned an angle, which is calculated in each case from the maximum values from the formula given above. In step S7, the determination of the desired mixing ratios then takes place on the basis of the determined angles.

FIG. 4 shows a flow chart for source separation. Based on the transformation signals X _x, X _2, and optionally X ₄ and X _5, as well as the residual signal X _3, the minima and maxima for each frequency bin is calculated (step SB). In a further embodiment, the Fourier transform of the output _signals X _{l org} and X _2jorg and optionally X _{3 org} and X _{4 org} may additionally be taken into account.

In step S9, the maximum value of the frequency bin or the respective column is set in each frequency bin at the location of the minimum value, all other values are assigned either a zero value or the hold value P _h ' _old , as described above.

According to step 10, the respective phase of the transformation signals X, X ₂ , X ₄ and X _{5 is} assigned to each non-zero value of the individual signals S, S ₂ , S ₄ and S ₅ .

The assignment is based on the numbering of the frequency bins. If the frequency bin 28 of the individual signal 5 has a value other than zero, this receives the phase of the frequency bin 28 of the transformation signal Χ. Due to the

Shifting the maximum values to the minimum values in each frequency bin and zeroing the remaining values is not a single value of a single signal equal to the value of the associated transformation signal. Nevertheless, the phases can be taken over. Instead of zeroing the non-minimum points, the hold value _ηΡ Μά can also be used. This prevents jumps in the individual signals. FIG. 6 shows a scatter plot of the amplitudes of two mixed output signals X 1 and X ₂ , wherein the output signals X ₁ and X _{2 are} mutually shifted sinusoidal signals. A scatter plot represents pairs of values, whereby the value pairs are formed on the basis of the number of the data point, so to speak of the data bin. The first point of the data point of the output signal X ^ has a first amplitude and the first data point of the output signal X ₂ an equal or a different amplitude. These two amplitudes are used to set a point in the scatter plot 9. By using a multiplicity of data points in each case of the output signal X 1 and of the output signal X ₂ , the same number of data-point pairs is produced. In this case, on the axis 10, the amplitude of the output signal X, and indicated on the axis 11, the amplitude of the output signal X ₂ . By using, for example, 4096 data points on the output signal X ^ and the output signal X ₂ , one obtains 4096 pairs of data points. These form in the scatter diagram a point cloud 12, which is formed from a plurality of the individual pairs of data points 13 determined points. FIG. 7 shows a correspondingly formed scattergram 14, wherein the output signals X _! and X _2, however, were Fourier transformed and the amplitudes of the signals according to

were calculated. Correspondingly, the magnitude of the amplitude of X _{l org} , ie the Fourier transform of the output signal X _{1 (} and the magnitude of the amplitude of X _{lfir are} plotted on axis _15. By calculating the slope of the straight lines 17 and 18, the mixing ratios are obtained However, instead of forming a scattergram and determining the slope of the resulting straight line, the output signals Xi and X ₂ of the _{A scattering} diagram is used to determine a histogram from the pairs of angles Θ and vector ψ in which the angles are shortened to an integer and corresponding frequencies are derived therefrom Angles from each pair of points this results in a single frequency for the histogram.

This histogram is smoothed with a smoothing function, reducing the number of local minima and maxima. As a smoothing function, for example, the function

be used, which has already been described above. Such a smoothed histogram is shown in FIG. On the axis 19 are given in degrees, from 0 ° to 90 °, while on the axis 20, the respective frequencies are shown. The histogram 21 has an absolute maximum 22 and several local maximum 23, 24 and 25. The number of frequencies is sorted by the maximum 22 has the largest value, followed by the maxima 25, 24 and 23. From these maxima is starting from the Maximum with the highest frequencies either a predetermined number selected, for example, for two maxima the maximum 22 and 25. However, so long maxima can be selected as long as the frequency is above a threshold. If one uses this, for example, at 10, then further calculations for the maxima 22, 25 and 24, but not for the maximum 23, which is below this threshold.

Every maximum stands for an angle. The maximum 22 for example for 18 ° and the maximum 25 for 72 °.

The indication of the angle can take place according to the formula γ _Ν = index (max (h _{gl TP} )). From the thus determined angles results for each angle in each case a mixing ratio according to the formula

V _N = ton ( _N ). N is the index of the source signal to be extracted. Thus, the determination of the mixing ratios is completed on the basis of a direction information. The determination of the subtraction signals X _l and X ₂ is carried out according to the formula given above, in particular the subtraction signals X _y and X _{2 can} also be determined in blocks. Thus, in each case a predefined set of data points, for example 4096 data points each of the output signals and X ₂ to subtraction signals X and X ₂ can be calculated using the mixing ratios. The subtraction signals Χ and X ₂ are Fourier-transformed according to the overlap-add method shown in FIG. Shown are data blocks 26, 27, 28 and 29 each consisting of 2048 data points of the subtraction signal Χ and the subtraction signal X ₂ . The numerical value 2048 results from the fact that it is half of the data points used for Fourier transformation. The number of data points of the data blocks 26 to 29 thus results from the number of data points of a data block 30 to 33.

The data block 30 is composed of the data blocks 26 and 27 and accordingly has twice the number of data points of a data block 26 or 27. This data block 30 is the data block which is Fourier-transformed. The data block 31 consists of the data blocks 27 and 28, the data block 32 of the data blocks 28 and the following data block. The data block 27 is thus contained in the data block 30 and 31, the data block 18 in the data block 31 and 32nd

Before the Fourier transformation, the data blocks 30, 31, 32 and 33 are multiplied by a Hanning window. As a result, leakage effects in the Fourier transformation can be minimized.

FIG. 10 shows a 3D amplitude spectrum of the transformation signals,, X ₂ and of the residual value signal X 1. On the axis 35, the respective number is a frequency bin, that is a numbered point in the vector of a signal shown, while on axis 36 of the amount value is indicated. Axis 37 indicates the respective amount values. In place 38 is the signal X-. in place 39 the signal and in place 40 the signal

X,

The signals shown in FIG. 10 were determined from two pure sinusoidal signals. For real signals, more or less all frequency bins are filled, but for purposes of illustration the sinusoidal signals are particularly suitable. From the illustration in FIG. 10 it follows directly that the signals X, X, and X each have only three peaks which are of different heights but otherwise have only zeros. 11 shows a 3D amplitude spectrum of the individual signals S _lt S ₂ and the residual signal mixture S ₃ . The axis 35 again indicates the frequency bin and the axis 36 the amount. On the other hand, the obtained individual signals are plotted on the axis 42. In place 43 is the single signal IsJ, at position 44, the residual signal \ s ₃ \ and 45 instead of the single signal \ S ₂ \. To the respective non-zero digits of the individual signals and are from the transformation signals Χ and X ₂ still each use the phases in order to be able to Fourier transform the individual signals and backward.

FIG. 12 schematically shows the extraction of the individual signals from the transformation signals X, X ₂ and the residual value signal X ₃ . The vector 46 is there

Section of the transformation signal X again, according to vector 47 the

Numerical values of the transformation signal X ₂ , vector 48 the values of the residual signal X ₃ and

Vector 49 the corresponding frequency bins.

From these, the single signal S i shown in vector 50, the single signal S ₂ shown in vector 51 and the residual signal S ₃ shown in vector 52 result as follows: For frequency bin 1, the minimum lies in vector 48, since the numbers 5 and 7 are greater than 1. The following values are therefore used in the frequency bin 1 of the corresponding vectors 50, 51 and 52: instead of the minimum value 1, the maximum value 7 is set. This is therefore to be set to the frequency bin 1 of the vector 52. The remaining values in frequency bin 1, viz. In vectors 50 and 51, are set to zero. The minimum consideration thus takes place column wise, data pointwise or frequency binwise. These expressions are identical.

Accordingly, the values of the columns for frequency bin 2, 3, 4, 5, etc. are calculated so that the values shown in Fig. 12 result respectively.

In vector 50, the values or minimum values for vector 46 are collected, in vector 51 those for vector 47 and in vector 52 those for vector 48.

To the vectors 50 and 51 of the individual signals 5, and S ₂ , the phases of the transformation signals X, and X ₂ are subsequently assigned. Purely by way of example, this is described for the frequency bin 4 of the vector 50. For vector 50 frequency bin 4 becomes the phase of the frequency bin 4 is taken from vector 46, since the vector 50 maps the minimum locations of the vector 46, if numerical values are present there. Accordingly, the phase of the identical frequency bins is transmitted. In all embodiments in the figures, instead of explicitly mentioned two output signals, two transformation signals or individual signals, more than two signals can be used, furthermore all embodiments mentioned in the introduction to the description can also be applied with regard to the description of the figures, unless this is expressly excluded.

B EZU G SZEI C H A LIST

1 source signal 27 data block

2 source signal 28 data block

3 source signal 29 data block

4 source signal 30 data block

5 Microphone 35 31 Data block

6 microphone 32 data block

7 Determination of mixing ratios 33 Data block

8 Signal separation 34 3D amplitude spectrum

9 scatter plot 35 axis

10 axis 40 36 axis

11 axis 37 axis

12 point cloud 38 point

13 point 39 point

14 scatter plot 40 digit

15 Axis 45 41 3D amplitude spectrum

16 axis 42 axis

17 Just 43 jobs

18 Straight 44 place

19 axis 45 point

20 axis 50 46 vector

21 histogram 47 vector

22 maximum 48 vector

23 maximum 49 frequency bin

24 maximum 50 vector

25 maximum 55 51 vector

26 data block 52 vector

Claims

1. A method for determining at least two individual signals (S _lt S ₂ , S ₃ , S ₄ ) from at least two output signals (X ^ X ₂ , X3, X ₄ ), the at least two source signals (1, 2, 3, 4) are formed,

characterized in that a) the mixing ratios (Vi, V ₂ , V ₃ , V ₄ ) of at least two source signals (1, 2, 3, 4) to the output signals (X ^ X ₂ , X ₃ , X ₄ ) are determined .

b) the output _signals (X _1t X ₂ , X ₃ , X ₄ ) are multiplied by factors formed from the mixing ratio (\, V ₂ , V ₃ , V ₄ ) and from the output signals (X ^ X ₂ , X ₃ , X ₄ ) at least two subtraction signals (Χ, X ₂ , X, X ₄ ) are determined so that in each subtraction signal (X _x , X ₂ , X ₃ , X ₄ ) a source signal (1, 2, 3, 4) is eliminated,

c) the subtraction signals {X, X ₂ , X ₃ , X ₄ ) are Fourier-transformed, resulting in transformation signals {Χ, Χ ₂ , X ₄ , X ₅ ),

d) a residual signal (X ₃ ) is determined from the transformation signals (X, X ₂ , X ₄ , X ₅ ),

e) at least two individual signals (S ^ S ₂ , S ₃ , S ₄ ) are calculated on the basis of the transformation signals (X _X , X ₂ , X ₄ , X ₅ ) and the residual signal (^ ₃ ), and

f) the individual signals (5, S ₂ , S ₃ , S ₄ ) are transformed by an inverse Fourier transformation into calculated source signals (S ₂ , S ₃ and S ₄ ).

2. The method according to claim 1,

characterized in that the determination of the mixing ratios (V ^ V ₂ , V ₃ , V) takes place via a direction determination.

3. The method according to claim 1 or 2,

characterized in that the determination of the mixing _ratios (V _1t V ₂ , V ₃ , V ₄ ) _{comprises the} following steps: Fourier transformation of the output signals (Χι, X2, X3, X ₄ ),

Calculation of the amplitude values of the Fourier-transformed output signals

Determining a histogram (21) from the amplitude values or determining the slope of the straight lines (17, 18) of the amplitude values in a scattergram (14).

Method according to claim 3,

characterized in that the histogram (21) is multiplied by a smoothing function.

Method according to claim 4,

characterized in that as a smoothing function a Hanning window, in particular

, is used.

6. The method according to any one of claims 3 to 5,

characterized in that the angular distance in the histogram (21) is.

7. The method according to any one of claims 3 to 6,

characterized in that the local extreme values, in particular the local maxima (22, 23, 24, 25), of the histogram (21) are determined. Method according to claim 7,

characterized in that beginning with the absolute maximum (22) each peak (22, 23, 24, 25) is assigned angle.

9. The method according to claim 8,

characterized in that the assignment of the angle is terminated at a predetermined number or falls below a threshold value.

10. The method according to any one of claims 3 to 9, characterized in that a histogram (21) is weighted with a low pass.

1 1. A method according to claim 10,

characterized in that the low pass is determined from temporally earlier determined histograms.

12. The method according to any one of the preceding claims,

characterized in that a subtraction signal {X _x, X _2, X _3, X ₄₎ (for two output signals X ^ X ₂₎ is determined by the multiplied (with the mixing ratio V _N) second output signal (X ₂₎ of the first ( X ^ is subtracted when the mixing ratio is less than 1 and otherwise the first output signal (Xi) is multiplied by the reciprocal of the mixing ratio (1 / V _N ) and subtracted from the second output signal (X ₂ ).

13. The method according to any one of the preceding claims,

characterized in that a power of two at data points, in particular 1024, 2048 or 4096 data points are used for the Fourier transformation.

Method according to one of the preceding claims,

characterized in that the data points for Fourier transformation are used quasi-continuously and according to the overlap-add method.

Method according to one of the preceding claims,

characterized in that a residual value signal (X _} ) is determined as a subtraction signal of the transformation signals

Method according to one of the preceding claims,

characterized in that a single signal (5,, S ₂ , S _} , S ₄ ) by means of an extreme value determination of at least one transformation signal {Χ, Χ ₂ , X ₄ , X ₅ ) and / or the residual value signal (X ₃ ) is determined.

Method according to one of the preceding claims,

characterized in that the frequencies of a single signal (£,, S ₂ , S ₃ , S _A ) are detected by minimization.

18. The method according to claim 16 or 17,

characterized in that in the transformation signals (X ₁ , X ₂ , X ₄ , X ₅ ) and the residual value signal (X ₃ ) for generating the individual signals (S, S ₂ , S ₃ , S ₄ ) for each frequency bin (49)

Minimum is determined, in place of which the maximum of the respective frequency bins (49) is set and the remaining values are set to zero.

19. The method according to claim 18,

characterized in that the respective phase values of the transformation signals (Χ, X ₂ , X ₅ , X ₄ ) are determined for the amplitude values of the individual signals (? ₁ , _{2 2} , S ₃ , S ₄ ).

Method according to one of the preceding claims,

characterized in that the number of the source signals to be calculated (S ^ S ₂ , S ₃ and S ₄ ) and / or the one or more source signals (1, 2, 3, 4) are themselves predefined.