EP1014340A2 - Method and device for processing noisy audio signals - Google Patents

Abstract

For signal processing of a sound signal y, in which a redundancy is contained, which mainly consists of approximate repetitions of signal profiles, the signal profiles are recorded and correlations between the signal profiles within sections of the sound signal are recorded. Correlated signal components are assigned to a power component and uncorrelated signal components to a noise component of the sound signal. The correlations between the signal profiles are determined using methods of non-linear noise reduction in deterministic systems in reconstructed vector spaces, which are based on the time domain. <IMAGE>

Description

The invention relates to methods for processing noisy Sound signals, in particular for non-linear noise reduction in speech signals, for the non-linear separation of power and noise signals and for the use of non-linear Time series analyzes based on the concept of low-dimensional deterministic chaos. The invention relates also a device for implementing the methods and their Use.

Noise reduction during recording, storage, transmission or rendering human language has a high technical relevance. Noise can be a pure measurement inaccuracy e.g. in the form of the digital error when outputting sound amplitudes, as noise in the transmission channel or as dynamic Noise due to the coupling of the system under consideration with the Outside world occur. Examples of noise reduction of the human language are generally from telecommunications, automatic speech recognition or the use of electronic Known hearing aids. The problem of noise reduction occurs not only in human language, but also in others Types of sound signals, and not just stochastic Noise, but also with all forms of overlay of a relevant sound signal due to extraneous noise. It exists an interest in a signal processing technique, with the strongly aperiodic and non-stationary sound signals in Analyzed, manipulated in relation to power and noise components or can be separated.

weißes Rauschen"). Wenn das Leistungssignal jedoch selbst stark aperiodisch und somit breitbandig ist, wird mit dem Frequenzfilter auch ein Leistungssignalanteil zerstört, woraus sich unzulängliche Ergebnisse ergeben. Soll z.B. bei einer Sprachübertragung die menschliche Sprache durch einen Tiefpaßfilter von hochfrequentem Rauschen befreit werden, so wird das Sprachsignal verzerrt.A typical approach to noise reduction, ie to break down a signal into certain power and noise components, is based on signal filtering in the frequency domain. In the simplest case, filtering is done with bandpass filters, but this creates the following problem. As a rule, stochastic noise is broadband (often so-called

white noise "). If, however, the power signal itself is strongly aperiodic and thus broadband, the frequency filter also destroys a portion of the power signal, which results in inadequate results. For example, if a low-pass filter is to be used to remove high-frequency noise from human speech during speech transmission the speech signal is distorted.

Another well-known approach to noise reduction consists of noise compensation for sound recordings. Here for example with a first microphone from a Noise levels in a room overlaid with human speech and with a second microphone recorded a sound signal that essentially represents the noise level. From the measurement signal of the second A compensation signal is derived from microphones The noise is superimposed with the measurement signal of the first microphone compensated from the surrounding space. This technique is due to the relatively high cost of equipment (use of special Microphones with directional characteristics) and because of the limited Application e.g. disadvantageous in voice recording.

Methods for nonlinear time series analysis based on the concept of low-dimensional deterministic chaos are also known. Since complex dynamic behavior plays an important role in almost all areas of our daily environment, but also in many areas of science and technology, e.g. when processes in medicine, economics, signal technology or meteorology provide aperiodic, difficult to predict and often difficult to classify signals, time series analysis represents a fundamental approach to learn as much as possible about the properties or state of a system from observed data. Known analytical methods for understanding aperiodic signals are described, for example, by H. Kantz et al. in Nonlinear Time Series Analysis ", Cambridge University Press, Cambridge, 1997, or by HDI Abarbanel in Analysis of Observed Chaotic Data ", Springer, New York, 1996. These methods are based on the concept of deterministic chaos. Deterministic chaos means that although a system state at a particular point in time uniquely defines the system state at any later point in time, the However, the system is unpredictable over a long period of time because the current system state is recorded with an inevitable error, the effect of which increases exponentially depending on the equation of motion of the system, so that after a relatively short time a simulated model state with the real state of the system does not exist Resemblance more.

For time series of deterministic chaotic systems, methods for noise suppression have been developed which do not separate in the frequency domain, but explicitly use the deterministic structure of the signal. These methods are described, for example, by P. Grassberger et al. in CHAOS ", Vol. 3, 1993, p. 127, by H. Kantz et al. (See above) and by EJ Kostelich et al. In Phys. Rev. E ", vol. 48, 1993, p. 1752. The principle of noise suppression for deterministic systems is described below with reference to FIG. 10a.

Fig. 10 shows schematically the dependence of successive Time series values for noise-free or noisy systems (on Example of a one-dimensional relationship). The noiseless Data from a deterministic system provide this in Fig. 10a shown image. There is an exact (here: one-dimensional) deterministic relationship between a value and the Subsequent value. The time offset vectors, to which details continue explained below are in a low dimensional Diversity in the embedding room. When introducing Noise becomes the deterministic relationship through an approximate Relationship replaced. The data is no longer on the Under manifold, but in their vicinity (Fig. 10b). The Power and noise are differentiated by dimensionality. Everything that leads out of the sub-manifold, is due to the influence of noise.

Accordingly, the noise suppression is deterministic chaotic signals in three steps. First is the Dimension m of the embedding space and the dimension of the manifold, in which the noiseless data were located, estimated. The actual correction is then for each individual point the diversity in its vicinity identified and finally to reduce noise the point under consideration on the manifold projected (Fig. 10c).

The disadvantage of the illustrated noise reduction is their limitation to deterministic systems. In one non-deterministic system, so in the no clear There is a connection between a state and a subsequent state, is the concept of identifying an equation of motion and contemplating a smooth manifold like it is illustrated in Figure 10, not applicable. For example, form the signal amplitudes of speech signals time series, which are unpredictable and the time series are more non-deterministic Systems match.

The applicability of conventional non-linear noise reduction has so far been ruled out for speech signals, in particular for the following reasons. Human speech (but also other sound signals of natural or synthetic origin) is generally highly non-stationary. The language is composed of a concatenation of phonemes. The phonemes alternate constantly, so that the speech dynamics change continuously. For example, hissing sounds contain primarily high frequencies and vowels (e.g. o ") primarily low frequencies. This would require equations of motion that would constantly change over time to describe the language. However, the existence of a uniform equation of motion is a prerequisite for the concept of noise suppression described with reference to FIG. 10.

It is the object of the invention to provide an improved signal processing method for sound signals, especially for noisy ones Speech signals to indicate with which an effective and rapid separation of power and noise components of the considered Sound signal is possible without distortion. The object of the invention is also to implement devices to specify such a method.

These tasks are accomplished by a method or by an apparatus with the features according to claims 1 and 10 respectively solved. Advantageous embodiments and uses of the invention result from the dependent claims.

A first important aspect of the invention is in particular therein, non-stationary sound signals consisting of Power and noise components, with such a high sampling rate to detect that predetermined signal profiles within of the considered sound signal enough redundancy for one Noise reduction included. Phonemes consist of a sequence of periodic or approximately periodic repetitions. On the concepts of periodic or approximately periodic Repetitions are discussed separately below. Hereinafter the concept of the approximately periodic becomes uniform Signal profiles used. The time series of Sound signals provide waveforms that are at least over repeat certain signal sections of the sound signal and a temporary application of the above, per se allow known concept of non-linear noise reduction.

According to another important aspect of the invention almost periodic within a sound signal under consideration Detected signal profiles and correlations between the signal profiles determined to correlated signal components a power component and uncorrelated signal components a noise component assign the sound signal.

Another important aspect of the invention is in the idea of temporal correlations through geometric correlations in the time delay embedding room to replace the be expressed by environments in this room. Points in these environments provide the information needed for nonlinear Noise reduction of the point are necessary for the the environment is constructed.

The invention also provides a device for signal processing for sound signals, in particular a sampling circuit for signal value detection, an arithmetic circuit for signal value processing and an output unit to output noise-free time series.

Finally, it should be emphasized that for the first time the application nonlinear noise reduction method for deterministic Systems for processing non-stationary and non-deterministic Sound signals is described. This is surprising since the prerequisite for the known noise reduction method especially stationarity and determinism of the signals to be processed. Precisely these requirements are considered for non-stationary sound signals of the global signal curve violated. Nevertheless, it delivers certain signal profiles limited application of the nonlinear Noise reduction excellent results.

The invention has the following advantages. It will be the first time created a noise reduction method for sound signals, which works essentially without distortion and with a low one equipment expenditure can be implemented. The invention can be implemented in real time or almost in real time. Certain parts of the signal processing according to the invention are compatible with conventional noise reduction methods, so that additional correction methods known per se or fast data processing algorithms easily on the Invention are transferable. The invention allows the effective Separation of power and noise components regardless of the frequency spectrum of noise. So is so-called in particular colored noise or isospectral noise separable. The Invention is not only with stationary noise, but also applicable to non-stationary noise if the time scale, on which the intoxication process changes its properties, longer than typically 100 ms (this is an example value that relates in particular to the processing of speech signals and can also be shorter in other applications).

The invention is not limited to human language, but also with other sound sources natural or synthetic Applicable origin. When processing speech signals it is possible for human speech signals from background noise to separate. However, it is not possible separate individual speech signals. This would assume that e.g. one vote as a share of performance and one other voice is considered a noise component. The the However, a voice representing noise would become an untreatable Show non-stationary noise on the same time scale.

Fig. 1

Kurvendarstellungen zur Illustration eines Sprachsignals;

Fig. 2

eine Kurvendarstellung eines Zeitausschnitts des in Fig. 1 illustrierten Schallsignals;

Fig. 3

ein Flußdiagramm zur Illustration des erfindungsgemäßen Verfahrens;

Fig. 4

Kurvendarstellungen zur Illustration einer erfindungsgemäßen Rauschreduzierung an einem Pfeifsignal;

Fig. 5

Kurvendarstellungen zur Illustration des erfindungsgemäßen Verfahrens an Sprachschallsignalen;

Fig. 6

eine Darstellung der Rauschreduzierung in Abhängigkeit vom Rauschpegel;

Fig. 7

eine Kurvendarstellung zur Illustration von Korrelationen zwischen Signalprofilen in einem Sprachsignal;

Fig. 8

eine Kurvendarstellung zur Illustration eines rauschbereinigten Sprachsignals;

Fig. 9

eine schematische Blockdarstellung einer erfindungsgemäßen Vorrichtung; und

Fig. 10

Kurvendarstellungen zur Illustration der nichtlinearen Rauschreduzierung in deterministischen Systemen (Stand der Technik).

Further details and advantages of the invention are described below with reference to the accompanying drawings. Show it:

Fig. 1

Curve representations to illustrate a speech signal;

Fig. 2

a graph of a time segment of the sound signal illustrated in Fig. 1;

Fig. 3

a flowchart to illustrate the inventive method;

Fig. 4

Curve representations to illustrate a noise reduction according to the invention on a whistle signal;

Fig. 5

Curve representations to illustrate the method according to the invention on speech sound signals;

Fig. 6

a representation of the noise reduction depending on the noise level;

Fig. 7

a graph to illustrate correlations between signal profiles in a speech signal;

Fig. 8

a graph to illustrate a noise-cleared speech signal;

Fig. 9

a schematic block diagram of a device according to the invention; and

Fig. 10

Curve representations to illustrate the non-linear noise reduction in deterministic systems (state of the art).

The invention is described below using the example of noise reduction on speech signals by utilizing intra-phonem redundancy explained. The power component of the sound signal is formed by a speech component x, which by a noise component r is superimposed. The sound signal is in signal sections divided, in the language example by spoken syllables or phonemes are formed. However, the invention is not limited to speech processing. With other sound signals the assignment of the signal sections becomes application-dependent chosen differently. The signal processing according to the invention is every sound signal is accessible, which in itself is non-stationary is, but approximately within predetermined signal sections periodically repeating signal profiles.

Nonlinear noise reduction in deterministic systems

The following are first details of the nonlinear Noise Reduction explains how they are per se from those cited above Publications by E. J. Kostelich et al. and P. Grassberger et al. are known. These explanations are for your understanding conventional technology. In terms of details non-linear noise reduction are the ones mentioned here Publications by E. J. Kostelich et al. and P. Grassberger et al. fully incorporated into the present description. The Explanation relates to deterministic systems. The invention Transfer of conventional technology to non-deterministic Systems is described below.

The states x of a dynamic system are determined by an equation of motion x n + 1 = F (x n ) described in a state space. The equation of motion is usually a complicated differential equation. If the function F is not known, it can, however, from long time series {x _k }, k = 1, ..., N , can be approximated linearly by considering all points in an environment (or: neighborhood) U _{n of} a point x _n and minimizing the function (1). s n 2nd = k: x k εU n (A n x k + b n -x k + 1 ) 2nd ,

The quantity s _n ² represents a prediction error in relation to the factors A _n and b _n . The implicit expression A n x k + b n x k + 1 = 0 illustrates that the values which correspond to the above-mentioned equation of motion are limited to a hyperplane within the state space under consideration.

If the state x _{k is} due to a statistical noise r _k to a real state y k = x k + r k is superimposed, the points belonging to the environment U _{n are} no longer limited to the hyperplane formed by A _n and b _n , but are scattered in an area around the hyperplane. The nonlinear noise reduction now means to project the noisy vectors y _n onto this hyperplane. The projection of the vectors onto the hyperplane is carried out using known methods of linear algebra.

With time series, like with speech signals, only a sequence of Scalar values recorded. These become those to be reconstructed Phase space vectors with the concept of time offset vectors ascertained in detail by F. Takens under the Title "Detecting Strange Attractors in Turbulence" in "Lecture Notes in Math ", Vol. 898, Springer, New York, 1981, or from T. Sauer et al. in "J. Stat. Phys.", Vol. 65, 1991, p. 579, and is illustratively described below. These publications too are hereby fully incorporated into the present description involved.

Starting from a scalar time series s _k , time offset vectors in an m-dimensional space are corresponding s n = (see n , s n-τ , ... p n- (m-1) τ ) educated. The parameter m is the embedding dimension of the time offset vectors. The embedding dimension is chosen depending on the application and is greater than twice the value of the fractal dimension of the attractor of the dynamic system under consideration. The parameter τ is a sampling interval (or: "time lag"), which represents the time interval between the successive elements of the time series. The time offset processor is thus an m-dimensional vector, the components of which comprise a specific time series value and the (m-1) previous time series values. It describes the temporal development of the system during a time range or embedding window with the duration m · τ. With each new sample, the embedding window shifts by one sample distance within the entire development over time. The scanning distance τ is in turn a variable selected depending on the application. If the system changes little, the scanning distance can be chosen larger to avoid processing redundant data. If the system changes rapidly, the sampling distance must be chosen smaller, since otherwise the correlations that occur between neighboring values would introduce errors into the further processing. The choice of the sampling distance τ is therefore a compromise between the redundancy and the correlation between successive states.

The above-mentioned projection of the states onto the hyperplane takes place using the time offset vectors in accordance with a calculation which has been carried out in detail by H. Kantz et al. in "Phys. Rev. E", vol. 48, 1993, p. 1529. This publication is also fully included in the present description. All neighbors in the time delay space are considered for each time offset vector s ∧ _n , ie the environment U _n is formed. The covariance matrix is then calculated in accordance with equation (2), the symbol ^ meaning that the mean value on the environment U _{n has been} subtracted. C. ij = U n ( s k ) i ( s k ) j

The singular or eigenvalues are determined for the covariance matrix C _ij . The vectors corresponding to the largest singular values represent the directions spanning the hyperplane defined by the A _n and b _n above.

In order to reduce the noise from the values s ∧ _n , the associated time offset vectors are projected onto the dominant directions that span the hyperplane. For each element of the scalar time series, this results in m different corrections, which are combined in a suitable manner. The described process can be repeated for the new projection with the noise-reduced values.

Identification of the neighbors, calculation of the covariance matrix and determining dominant vectors that match a predetermined one Number Q of largest singular values the search for correlations between successive System states. This search is for the deterministic Systems based on the known or assumed equation of motion of the system. Like the search according to the invention for correlations between system states in non-deterministic Systems is described below.

Non-linear noise reduction in non-deterministic Systems

To determine the correlation between the states, in the deterministic system, the assumed temporal invalidity the equation of motion is used as additional information. In contrast to this, the correlation is determined between neighboring states in the invention Signal processing in the non-deterministic system based on the following additional information.

The invention is based on the use of redundancy in the Signal. Because of the non-stationarity is between a real one Redundancy and random similarities of signal parts, which, however, are uncorrelated. This is through the use of a higher embedding dimension and one larger embedding window than would be necessary to to dissolve the current dynamics. A voice signal is a concatenation of phonemes. Every single phoneme is characterized by a characteristic waveform that repeated several times almost unchanged. A time offset embedding vector that completely covers such a wave, can thus be clearly assigned to a given phoneme without misinterpretation of another phoneme occurs with a different characteristic waveform. Within a phoneme, these waveforms change in one certain way so that no absolutely exact repetitions occur. Because of the latter property is almost periodic repetitions spoken.

Human language is a series of phonemes or syllables related to the amplitudes and Frequencies have characteristic patterns. These patterns can for example by observing electrical signals Sound transducer (e.g. microphone). On medium Language is not a time scale (e.g. in the context of a word) stationary and on long time scales (e.g. in the context of a sentence) highly complex, with many active degrees of freedom and possibly long-range correlations occur. On short time scales (Time ranges that are essentially the length of a phoneme or correspond to a syllable), occur repetitively in the signal curve Patterns or repeating signal profiles based on the following are explained. Details of the concrete calculations are implemented in the same way as conventional noise reduction and can do the above Publications are taken.

Fig. 1 shows an example of the Italian greeting "Buon giorno" as a wave train. This is the signal amplitude recorded with a sampling frequency of 10 kHz with the (arbitrarily standardized) time series values y _n as a function of the dimensionless time counter scale. This signal amplitude was derived from an extremely low-noise, digital voice recording. The total time course from n = 0 to n = 20,000 corresponds to a time range of approx. 2 seconds.

Representing a time segment of that shown in FIG. 1 Amplitude curve with extremely stretched time scale results the picture in Fig. 2. It shows that the amplitude curve within certain signal sections (e.g. phonemes) has periodic repetitions. A signal profile repeats in the illustrated example at time intervals a width of around 7 ms. A special advantage the invention is that the effectiveness of the invention Noise reduction not of absolute accuracy depends on the periodicity shown. It is possible, that there are no exact repetitions, but a systematic one Modification of the typical waveform of a signal profile takes place within a phoneme. However, this variation will taken into account in the procedure explained in detail below, since they have freedom in those remaining after projection Q. Directions. To take into account the Variation (deviation from exact repetitions) becomes the Concept of approximately periodic signal profiles used differ from exactly periodic signal profiles by only one distinguish systematic variability.

With time offset embedding (with suitably chosen parameters m and τ, see above) form the repetitions shown neighboring Points in the state space (or vectors that refer to this Points are directed). Now is the variability in these Points due to noise overlay greater than that natural variability due to non-stationarity, see above becomes an approximate identification of the manifold and the projection on it will reduce the noise more than it does affects the actual signal. This is the basic approach of the inventive method, the following with reference to the flowchart of FIG. 3 is explained.

Fig. 3 is an overview diagram that schematically basic Shows steps of the method according to the invention. The However, the invention is not limited to this process. Depending on the application can be a modification in terms of data collection, the parameter determination, the actual calculation for noise reduction, the separation of power and Noise components and the output of the result can be provided.

3, data acquisition 101 takes place after start 100 and the parameter determination 102. The data acquisition 101 comprises recording a sound signal by converting the sound in an electrical size. Data acquisition can be analog or digital sound recording. Depending on the application is the sound signal in a data memory or with real-time processing in a buffer memory (see FIG. 9) saved. The parameter determination 102 includes the selection of Parameters that are used for the later search for correlations between neighboring states in the sound signal are suitable. These parameters include in particular the embedding dimension m, the scanning distance τ, the ε diameter of the surroundings U im Time offset embedding room to identify neighbors, and the number Q of the time offset vectors to which the state projection should be done.

In speech signal processing, the embedding dimension m for example in the range of 10-50, preferably 20-30, and the scanning distance τ is in the range from 0.1 to 0.3 ms, so that the embedding window m · τ preferably approx. 3 to 8 ms covers. These data refer to a phoneme duration of approx. 50 to 200 ms and the complexity of the human voice. Typical Signal profiles are due to the pitch of the human Voice of approximately 100 Hz between 3 and 15 ms. Fig. 2 shows for example repetitions of the signal profile after each 7 ms. The parameter determination 102 (FIG. 3) can interact with the data acquisition 101 or within the framework of a Preliminary analysis has been carried out. With a preliminary analysis becomes the embedding dimension m and the dimension of diversity (corresponding to the parameter Q) in which the noiseless Data was lying, estimated. It can also be provided be that the parameter determination 102 during the process is repeated. This can be used as a correction in Response to the result of power / noise separation 109 (see below).

The signal sample 103 follows on the basis of the recorded measured values and the specified parameters. The signal sample 103 is provided to determine the values of the time series y _{n in} accordance with the previously determined sample parameters from the data. The following steps 104 to 109 represent the actual calculation of the projections of the real sound signals onto noiseless sound signals or states.

Step 104 comprises the formation of the first time offset vector at the beginning of the time series (for example according to FIG. 2). This first time offset vector does not necessarily have to refer to the first signal profile that appears first in time. However, this is particularly preferred for real-time or quasi-real-time processing. The first time offset vector comprises m signal values y _n as m components which follow one another with the time offset τ. Then, in step 105, adjacent time offset vectors (neighboring vectors) are formed and recorded. The neighboring vectors refer to signal profiles that are very similar to the signal profile represented by the first vector. They form the first environment U. If the first vector represents a profile that is part of a phoneme, the neighboring vectors essentially correspond to the approximately repeating signal profiles within the same phoneme. In speech processing, around 15 signal profiles are repeated within a phoneme. The number of neighboring vectors determined is less than or equal to the number of repeating signal profiles and is, for example, around 5 to 15.

The covariance matrix 106 is then calculated accordingly of equation (2) given above. The one in this matrix inserted vectors are the vectors from the base environment U as determined in step 105. step 106 includes then the determination of the Q largest singular values of the covariance matrix and the associated singular vectors in the m-dimensional Room.

In the following projection 107 all parts of the first time offset vector that is not in the Q determined dominant vectors spanned subspace are eliminated. The value Q is in the range of around 2 to 10, preferably 4 to 6. In a modified procedure, the value can be Q be zero (see below).

The relatively small number Q, which is the dimension of the subspace represents, onto which the states or signals are projected represents a particular advantage of the invention It was found that the dynamic range of the waves only a few degrees of freedom within a given phoneme owns once inside a high dimensional Space has been identified. Therefore, they are also proportional few neighboring states for the projection calculation required. To capture the correlation between the Signal profiles are only the largest singular values and corresponding ones Singular vectors of the covariance matrix are relevant. This The result is surprising since the non-linear noise reduction in itself for deterministic systems with extensive Time series was developed. It also emerges as special Advantage of a relatively small amount of time for the Calculation.

Then the next time offset vector becomes step 108 selected and the sequence 105-107 repeated, with new Environments and new covariance matrices are formed. This Repeat until all time offset vectors that result from the Time series can be constructed, have been processed.

The formation or acquisition of the neighboring vectors (step 105) Incidentally, it takes place at a higher dimension than the projection 107. The high dimension in the search for neighbors guarantees that Choosing the right neighbors to represent the profiles are derived from the same phonemes. The invention thus chooses implicitly without any language model phonemes. As above has been explained represents the dynamics within a phoneme however, significantly fewer degrees of freedom, so that within of the subspace spanned by the singular vectors low-dimensional and can be worked quickly. For real-time applications the sound signal processing takes place essentially for the phonemes in succession, so that phoneme for Phonem completely processed and so a noise-free output signal is produced. This output signal is compared to that recorded sound signal (input signal) delayed by around 100-200 ms (Real time or quasi real time application).

Steps 109 and 110 relate to the formation of the actual output signal. Step 109 is directed to the separation of power and noise signals. A noise-free time series element s _k is formed by averaging over the corresponding elements from all time offset vectors which contain this element. A weighted averaging can be introduced instead of a simple averaging. After step 109, a jump back can be provided before step 104. The noise-free time series elements then form the input variables for the renewed formation of time offset vectors and their projection onto the subspace in accordance with the singular vectors. This process repetition is not necessary, but can be provided, for example, 2 or 3 times to improve the noise reduction. After step 109, however, a return to parameter determination 102 can also be provided if the power component present after step 109 differs less than expected (for example by less than a predetermined threshold value) from the unprocessed sound signals. For this purpose, decision mechanisms, not shown, can be built in. At step 110, data output follows. With noise reduction, the noise-reduced voice signal is output as a power component. Alternatively, the output or storage of the noise component can also be provided depending on the application.

The procedure explained above can be used in relation to the parameter determination considering the following points be modified. First, the dimension of Manifold (according to the parameter Q) in which the noise-free data would lie in the course of a signal vary. The dimension Q can vary from phoneme to phoneme. The dimension can, for example, also during a break between two spoken words or any other resting phase Be zero. Second is a selection of relevant ones inherent time offset vectors onto which the state is projected should be excluded if the noise is relatively high is (about 50%). In this case, all eigenvalues of the Correlation matrix to be approximately the same.

Accordingly, the following can be done in the process flow Variation of the parameter Q may be provided. Instead of an unchangeable one Projection dimension Q becomes the dimension for each covariance matrix is adjusted or individually determined. At step 102, a constant f = 1 is determined. This constant f is determined empirically. It depends on the Signal type down and is, for example, with language f = 0.1. The maximum singular value of a given covariance matrix, multiplied with the constant f, represents a threshold. The number of singular values that are greater than the threshold is then used as the value for Q for the projection, provided that this value does not exceed a certain one Maximum value. This maximum value is e.g. 8. In the latter The case is all singular values of a given covariance matrix so similar that no pronounced linear subspace is selected can be selected and therefore Q = 0 must be selected. Instead of the current time offset vector then becomes a projection replaced by the mean of its surroundings.

This modification increases the efficiency of the process drastically increased especially at high noise levels.

Examples

The signal processing according to the invention is described below illustrated two examples. In the first example, this is processed Sound signals a human whistle (see Fig. 4). The second example concerns the above words "Buon giorno" (see Figs. 5 to 8).

Fig. 4 shows the power spectrum for a human whistle lasting 3 s. A whistle is an essentially periodic signal with characteristic harmonics and only minor non-stationarities. 4a shows the amplitude profile of the original recording. After the numerical addition of a 10% noise, the spectrum shown in FIG. 4b results. This provides the input data for step 101 of the process sequence (FIG. 3). After the noise reduction according to the invention, the image shown in FIG. 4c results. This shows the complete restoration of the original, noiseless signal. Figures 4a to 4c show a particular advantage of the invention over a conventional filter in the frequency domain. A filter in the frequency domain would cut off all power components with amplitudes below 10 ^-6 , so that the noisy spectrum would only contain the peak at 0 and the peak around the fundamental frequency. Accordingly, the time series obtained from the back transformation would be completely harmonic, which would sound very synthetic. These disadvantages are avoided with the noise reduction according to the invention.

5 shows corresponding results using the example of curve representations for processing voice signals. In Fig. 5a is a section of the noiseless wave train of the words "Buon giorno" based on the signal curve according to FIG. 1 shown analogously to FIG. 2. It is the time-limited repetition of signal profiles recognizable, which are used to reduce the Noise contains the necessary redundancy. 5b shows the wave train after adding a synthetic noise. After Noise reduction according to the invention results in the image Fig. 5c. It turns out that the original signal for the most part could be reconstructed.

The functionality of the noise reduction according to the invention was tested in different noise types and amplitudes. The attenuation D (in dB) according to equation (3) can be considered as a measure of the performance of the noise reduction. D = 10 log ((Σ ( y k -x x ) 2nd ) / (Σ (y k -x k ) 2nd ))

In equation (3), X _k stands for the noiseless signal (power component), y _k for the noiseless signal (input sound signal) and y ∧ _k for the signal after the noise reduction according to the invention.

Fig. 6 illustrates the dependence of the damping D of the non-linear Noise reduction depending on the relative Noise amplitude (variance of the noise component: variance of the Performance share). It turns out that the damping itself at relatively high noise amplitudes (in the range of more than 100%) is reinforced.

Figures 7 and 8 show further details of the speech noise reduction. Fig. 7 illustrates the occurrence of repetitive signal profiles within the phoneme train shown in the upper part of the figure. Depending on a (arbitrary) time index i, a graph is printed in the lower part of the figure, which consists of points formed under the following conditions. The associated time offset vector s ∧ _i and the set of all time offset vectors s ∧ _{j, i are} considered for each time i. If the amount of the difference vector between the s ∧ _i and each s ∧ _{j is} less than a predetermined limit, a dot is printed. The points form more or less extended lines. The line structures show that the periodicities of the signal profiles explained above occur within the phonemes. The gaps in these line segments show that the environments are suitable for differentiating between different phonographs. For line structures that are particularly extended in the ordinate direction, the number of intra-phoneme neighbors is particularly large. However, it also turns out that for | ij | > 2000 no repetitions occur.

Fig. 8 again shows the example of the words "Buon giorno" in upper part of the figure the noiseless signal, in the middle Part of the synthetically added noise and in the lower part the noise remaining after the noise reduction. The ordinate scaling is identical in all three cases. The rest Noise (bottom part of the figure) shows a systematic Variation indicating that the success of the invention Noise reduction even from the sound signal, i.e. depends on the specific phoneme.

The invention also relates to a device for implementation of the method according to the invention. 9 includes a noise reduction arrangement a transducer 91, a data memory 92 and / or a buffer memory 93, a sampling circuit 94, an arithmetic circuit 95 and an output unit 96.

The components of the device according to the invention presented here are preferably used as a permanently connected circuit arrangement or manufactured as an integrated chip.

Preferred applications of the invention are mentioned below. In addition to the already mentioned noise reduction on speech signals the invention is also in noise reduction Hearing aids and to improve the computerized automatic Speech recognition applicable. Regarding speech recognition can be provided, in particular, the noise Compare time series values or sectors with table values. The table values represent corresponding values or Vectors of predetermined phonemes. An automatic speech recognition can be integrated with the noise reduction process become.

Other applications are in the field of telecommunications and when processing signals from sources other than that human language. These include animal voices, for example or music.

Claims (12)

Method for signal processing of a sound signal y, at which within predetermined signal sections of the Sound signal detects redundant signal profiles and correlations are determined between the signal profiles, where correlated signal components are a power component and uncorrelated signal components a noise component of the Sound signal can be assigned. The method of claim 1, wherein the correlations between the signal profiles using non-linear methods Noise reduction in deterministic systems be determined. Method according to Claim 1, in which the sound signal y, which is composed of a speech component x and a noise component r, is processed in each signal section according to the following steps: a) Acquisition of a large number of sound signal values y k = x k + r k with a scanning distance τ, b) Formation of time offset vectors, each of which consists of components y _k , the number m of which is an embedding dimension and the indices k of which are derived from the embedding window of width m · τ, whereby for each of these time offset vectors an environment U is formed from all the time offset vectors and their spacing at a given time offset vector is less than a predetermined value ε; c) determining correlations between the time offset vectors and projecting the time offset vectors onto predetermined singular value vectors, and d) Determination of useful signal values which form a speech signal which essentially corresponds to the speech component x _k and / or a noise signal which essentially corresponds to the noise component r _k . A method according to claim 3, wherein the number k of those formed Time offset vectors that form the environment of The redundancy depends on the approximate repetitions the signal profiles is saved. The method of claim 3, wherein the correlations between the time offset vectors by identification the environment and by calculating a covariance matrix are extracted on the vectors, too belong to the environment. The method of claim 3, wherein steps b) to c) repeated at least once with the useful signal values be used to improve the outcome can be repeated if the entire time series is noisy. A method according to claim 3, wherein the sound signal is on Speech signal is. Method according to one of claims 3 to 7, wherein the Embedding window m · τ is in the range of 1 to 20 ms. A method according to claim 3, wherein in step c) the Time offset vectors on a Q-dimensional manifold projected with an adaptively set Q value. Device for carrying out a method according to a of the preceding claims, which are a sensor (91), a data memory (92) and / or a buffer memory (93), a sampling circuit (94), an arithmetic circuit (95) and an output unit (96). Use of a method according to one of claims 1 up to 9 for noise reduction on voice signals in telecommunications, in hearing aids or in automatic Voice recognition. Use of non-linear noise reduction methods for deterministic systems for noise reduction of Voice signals.

Images