EP1091349B1 - Method and apparatus for noise reduction during speech transmission - Google Patents

Abstract

The method involves using a multi-layer self-organising neural network with feedback. A minima detection layer, a reaction layer, a diffusion layer and an integration layer define a filter function (F(f,T)) for noise filtering. The filter function is used to convert a spectrum B(f,T) free of noise, into a noise-free speech signal (y(t)) by inverse Fourier transformation. The signal delay caused by processing the signal is so short that the filter can operate in real-time for telecommunication. All neurons are supplied with an externally set parameter K, the size of which defines the degree of noise suppression of the whole filter. An Independent claim is included for an apparatus for noise suppression during speech transmission.

Description

The invention relates to a method for noise suppression in speech transmission by multiplication of spectra of a speech signal with a filter, which by arithmetic operations on spectrums of the input signal with the help of a multi-layered, self-organizing, to determine the feedback neural network is.

In telecommunications as well as in recording This happens with speech in portable storage devices Problem on that the speech intelligibility by noise is severely impaired. Especially when Making calls in the car using a hands-free device this problem is evident. To suppress the Noise filters are built into the signal path. Classic band-pass filters offer little benefit since noise is generally in the same frequency ranges lie like the speech signal. Therefore, be requires adaptive filters that independently the existing noise and the characteristics of the adjust the transmitted speech signal. There are several Known concepts.

Derived from the optimal filter theory is the Wiener-Komolgorov filter. (S.V. Vaseghi, Advanced Signal Processing and Digital Noise Reduction ", John Wiley and Teubner-Verlag, 1996). This procedure is based on the Minimizing the mean square error between the actual and the expected speech signal. This filter concept requires a considerable Computational effort. Moreover, as with most known Method a stationary interference signal theoretical Requirement.

A similar filter principle is based on the Kalman filter (E. Wan and A. Nelson, Removal of noise from Speech using the Dual Extended Kalman Filter algorithm, Proceedings of the IEEE International Conference on Acoustics and Signal Processing (ICASSP'98), Seattle 1998). The disadvantage of this filter concept has an effect the long workout time needed to get that To determine filter parameters.

Another filter concept is H. Hermansky and N. Morgan, RASTA processing of speech, IEEE Transactions on Speech and Audio Processing, Vol. 2, no. 4, p. 587, 1994, known. Also with this procedure is a training procedure required, as well as require different Noise different parameter settings.

A method known as LPC requires the elaborate Calculation of correlation matrices to use with the help of a linear prediction method filter coefficients too calculate, as from T. Arai, H. Hermansky, M. Paveland, C. Avendano, Intelligence of Speech with Filtered Time Trajectories of LPC Cepstrum, The Journal of the Acoustical Society of Maerica, Vol. 4, Pt. 2, p. 2756, 1996, is known.

Other known methods use neural networks of the type of multilayer perceptron for speech amplification as in H. Hermansky, E. Wan, C. Avendano, Speech Enhancement Based on Temporal Processing. Proceedings of the IEEE International Conference on acoustics and signal processing (ICASSP'95), Detroit, 1995, described.

Furthermore, US Pat. No. 5,878,389 A discloses a method for noise suppression in speech signals, which is based on a smoothing of short-term spectra over time and frequency. The smoothing can be carried out by a neural network. Here, however, the signal spectrum itself is subjected to the smoothing, which always has a negative effect on the voice quality.
Object of the present invention is to provide a method that recognizes a speech signal in terms of its temporal and spectral characteristics with little computational effort and can be freed from noise.

This task is solved by a minimadetection layer, which minima over past signal spectra detected, a reaction layer, which a non-linear response to that of the minimadetection layer detects detected minima, one Diffusion layer, which with only local couplings adjacent compute node in the diffusion layer a spectral smoothing on the outputs of the reaction layer performs, an integration layer, which the Output of the diffusion layer without weighting in one Compute nodes added up, with a filter function F (f, T) for noise filtering is created by coupling of compute nodes of successive layers Minimadetektionsschicht, reaction layer, diffusion layer and integration layer, where f is the frequency a spectral component, which for Time T is to analyze.

Such a method recognizes a speech signal on its temporal and spectral properties and frees this from noise. Compared to known methods, the required computational effort is low. The procedure is characterized by a particularly short adaptation time, within which the system on the type of noise. The signal delay when processing the signal very short, so that the filter in real time for telecommunications is operational.

The invention also relates to a device For noise suppression according to claim 9. Further advantageous measures are in the Subclaims described. The invention is in the enclosed.

Figur 1

das Gesamtsystem zur Sprachfilterung;

Figur 2

ein eine Minimadetektions-Schicht, eine Reaktions-Schicht, eine Diffusions-Schicht und eine Integrations-Schicht enthaltendes neuronales Netzwerk;

Figur 3

ein Neuron der Minima-Detektions-Schicht, welche M(f,T) ermittelt;

Figur 4

ein Neuron der Reaktions-Schicht, welches mit Hilfe einer Reaktionsfunktion r[S(T-1)] aus dem Integralsignal S(T-1) und einem frei wählbaren Parameter K, welcher den Grad der Geräuschunterdrückung bestimmt, aus A(f,T) und M(f,T) das Relativspektrum R(f,T) ermittelt;

Figur 5

Neuronen der Diffusionsschicht, in welcher eine der Diffusion entsprechende, lokale Kopplung zwischen den Moden hergestellt wird;

Figur 6

ein Neuron der gezeigte Ausführung der Integrationsschicht;

Figur 7

ein Beispiel für Filtereigenschaften der Erfindung bei verschiedenen Einstellungen des Kontrollparameters K.

Drawing shown and will be described in more detail below; it shows:

FIG. 1

the overall system for language filtering;

FIG. 2

a neural network including a minimum detection layer, a reaction layer, a diffusion layer, and an integration layer;

FIG. 3

a neuron of the minima detection layer which detects M (f, T);

FIG. 4

a neuron of the reaction layer, which by means of a reaction function r [S (T-1)] from the integral signal S (T-1) and a freely selectable parameter K, which determines the degree of noise suppression, from A (f, T ) and M (f, T) determines the relative spectrum R (f, T);

FIG. 5

Neurons of the diffusion layer in which a diffusion-dependent local coupling between the modes is established;

FIG. 6

a neuron of the illustrated embodiment of the integration layer;

FIG. 7

an example of filter characteristics of the invention at various settings of the control parameter K.

In the figure 1 is schematically and exemplarily an overall system presented for language filtering. This consists from a sampling unit 10, which is the noisy one Speech signal in the time t samples and discretizes and thus generates samples x (t) that are in time T to frames from n samples are summarized.

From each frame, Fourier transforms the Spectrum A (f, T) determined at time T and a filter unit 11 supplied with the help of a neural Network, as shown in Figure 2, a Filter function F (f, T) is calculated with which the spectrum A (f, T) of the signal is multiplied by the noise-removed one To produce spectrum B (f, T). Subsequently becomes the thus filtered signal of a synthesis unit (12) passed by inverse Fourier transform from the filtered spectrum B (f, T) the noise-free Speech signal y (t) synthesized.

FIG. 2 shows a minimadetection layer, a reaction layer, a diffusion layer and a Integration layer containing neural network, which is in particular the subject of the invention and to which the spectrum A (f, T) of the input signal is supplied from which the filter function F (f, T) is calculated becomes. Each of the modes of the spectrum that goes through distinguish the frequency f, corresponds to one single neuron per layer of the network except the integration layer. The individual layers become specified in the following figures.

Thus, FIG. 3 shows a neuron of the minima detection layer, which determines M (f, T). M (f, T) is in fashion with frequency f the minimum of over m frames averaged Amplitude A (f, T) within an interval of Time T, which corresponds to the length of 1 frame.

FIG. 4 shows a neuron of the reaction layer which by means of a reaction function r [S (T-1)] from the integral signal S (T-1), as shown in detail in FIG is shown, and a freely selectable parameter K, which determines the degree of noise suppression, from A (f, T) and M (f, T) determines the relative spectrum R (f, T). R (f, T) has a value between zero and one. The Reaction layer distinguishes speech from noise based on the temporal behavior of the signal.

FIG. 5 shows a neuron of the diffusion layer in which a diffusion corresponding local coupling between the fashions is made. The diffusion constant D determines the strength of the resulting Smoothing over the frequencies f at fixed time T. The diffusion layer determined from the relative signal R (f, T) the actual filter function F (f, T), with the the spectrum A (f, T) is multiplied to noise to eliminate. In the diffusion layer is Language of sounds based on spectral properties distinguished.

Figure 6 shows that in the chosen embodiment of the invention only neuron of the integration layer that the Filter function F (f, T) at fixed time T over the frequencies f integrated and the integral signal thus obtained S (T) is fed back into the reaction layer, as Figure 2 shows. This global coupling ensures that is heavily filtered at high noise while noise-free speech is transmitted unadulterated.

FIG. 7 shows exemplary details of the filter properties the invention for various settings of the control parameter K. The remaining parameters of Invention have the values n = 256 samples / frame, m = 2.5 Frames, 1 = 15 frames, D = 0.25. The picture shows the Damping of amplitude modulated white noise in Dependence of the modulation frequency. At modulation frequencies between 0.6 Hz and 6 Hz is the attenuation less than 3 dB. This interval corresponds to the typical modulation of human speech.

The invention will be described below with reference to an embodiment explained in more detail. First, a Speech signal that interferes with random noise is sampled in a sampling unit 10 and digitized, as Figure 1 shows. To this We obtain the samples x (t) in the time t. From each of these samples is combined into a frame, from that at time T by means of Fourier transformation a spectrum A (f, T) is calculated.

The modes of the spectrum differ by their Frequency f. In a filter unit 11 is from the Spectrum A (f, T) produces a filter function F (f, T) and multiplied by the spectrum. This gives you that filtered spectrum B (f, T), from which in a synthesis unit by inverse Fourier transform the noise-freed Speech signal y (t) is generated. This can after digital-to-analog conversion in a speaker be made audible.

The filter function F (f, T) is a neural Network, which is a minimadetection layer, a reaction layer, a diffusion layer and a Contains integration layer, as Figure 2 shows. The spectrum A (f, T) generated by the sampling unit 10 is first fed to the Minimadetektions layer, as it shows the figure 3.

A single neuron of this layer works independently from the other neurons of the minimadetection layer a single mode, represented by the frequency f is marked. The neuron averages for this fashion the amplitudes A (f, T) in the time T over m frames. From the neuron then determines these average amplitudes over a period of time in T, which is 1 frame in length corresponds to the minimum for his fashion. To this The neurons create the minimadetection layer the signal M (f, T), which is then fed to the reaction layer becomes.

Also, every neuron of the reaction layer, as shown in FIG. 4 shows, processes a single mode of frequency f, independent of the other neurons in this layer. In addition, all neurons will be externally adjustable Paramter K, whose size is the degree of Noise suppression of the entire filter is determined additionally the integral signal S (T-1) stands for these neurons from the previous frame (time T-1), the in the integration layer, as shown in FIG. 6 has been.

This signal is the argument of a nonlinear reaction function r, with whose help the neurons of the reaction layer the relative spectrum R (f, T) at the time T calculate.

The value range of the reaction function is on an interval [r1, r2] restricted. The value range of the in this way resulting relative spectrum R (f, T) is limited to the interval [0, 1].

In the reaction layer is the temporal behavior of the speech signal for distinguishing useful and Interference signal evaluated.

Spectral properties of the speech signal are in the Diffusion layer, as shown in FIG 5, evaluated, whose neurons have a local mode coupling according to Art perform a diffusion in the frequency space.

In the generated by the neurons of the diffusion layer Filter function F (f, T) leads to an approximation neighboring modes, their strength by the diffusion constant D is determined. Similar mechanisms, as in the reaction and diffusion layers on the Works are, lead in so-called dissipative media to structure-forming phenomena, which is a research subject of nonlinear physics.

All modes of the filter function F (f, T) are at the time T multiplied by the respective amplitudes A (f, T). This results in noise freed spectrum B (f, T) by means of inverse Fourier transformation into the noise-free speech signal y (t) is transformed. About the modes of the filter function F (f, T) is integrated in the integration layer, so that the integral signal S (T) results, such as it shows Figure 6.

This integral signal is fed back into the reaction layer. This global coupling causes the strength of signal manipulation in the filter from the noise level is dependent. Speech signals with low noise level pass the filter virtually unaffected, while at high noise level a strong filter effect takes effect. This makes the difference Invention of classical bandpass filters whose influence on the signal only from the chosen, fixed predetermined Depends on parameters.

Unlike a classic filter, the item has the invention no frequency response in the conventional Sense. When measuring with a tunable sinusoidal test signal would already be the modulation speed of the test signal, the filter characteristics influence.

A suitable method for analyzing the properties the filter uses an amplitude modulated noise signal, in dependence on the modulation frequency the To determine damping of the filter, as the figure 7 shows. To do this, set the input and output side mean integral performance in relation to each other and carries this value against the modulation frequency of the Test signal on. In Figure 7, this "modulation path" for different values of the control parameter K shown.

For modulation frequencies between 0.6 Hz and 6 Hz the attenuation for all shown values of the control parameter K less than 3 dB. This interval corresponds to the modulation of human speech, which is the Filter can therefore happen optimally. Signals outside of the said modulation frequency interval, on the other hand identified as noise and in dependence the setting of the parameter K is strongly damped.

Claims (13)

1. Method for noise reduction during speech transmission by multiplication of spectra of a speech signal, using a filter that is to be determined by arithmetic operations on spectra of the input signal with the aid of a multiple-film, self-organising, closed-loop neural network, characterised by

   a minima detection film which detects minima via rear signal spectra,

   a reaction film which performs a non-linear reaction function on the minima being detected by the minima detection film,

   a diffusion film which performs spectral smoothing on the outputs from the reaction film, using only local couplings of adjacent arithmetic nodes in the diffusion film,

   an integration film which totalizes the output from the reaction film without weighting in an arithmetic node,

   a filtering function F (f, T) for noise filtering being produced by coupling arithmetic nodes from the successive films, viz. the minima detection film, reaction film, diffusion film and integration film, where f designates the frequency of a spectral component that is to be analysed at instant T.
2. Method according to claim 1, characterised in that an adjustable parameter K is to be multiplied by the reaction function in the reaction film, in order to fix the magnitude of the noise reduction by the filter in the entirety thereof.
3. Method according to claims 1 and 2, characterised in that an arithmetic node from the integration film cumulatively integrates the filtering function F (f, T) at a fixed instant T over the frequencies f, to give a value S (T) which is to be fed back into the reaction film.
4. Method according to claims 1 to 3, characterised in that a signal scanned by a sampling unit and the signal spectrum generated therefrom by means of Fourier transformation is to be sent to the minima detection film, which determines a minimum of the smoothed amplitudes of the spectral components A (f, T), the smoothing corresponding to a temporal averaging over m frames and the minima detection extending over I frames, where a frame corresponds to the time interval on which a Fourier transformation is to be performed.
5. Method according to claims 1 to 4, characterised in that a multiple-film neural network generates a filtering function F (f, T) from a signal spectrum A (f, T) which is to be generated by Fourier transformation on a frame of the input signal x (t), and the spectrum A (f, T) is to be multiplied by the filtering function F (f, T) in order to generate a noise-reduced spectrum B (f, T) from which, by applying an inverse Fourier transformation in a synthesis unit, a noise-reduced speech signal y (t) is to be generated, where t designates the time taken to treat a sample of the signals x and/or y.
6. Method according to claims 1 to 5, characterised in that the spectral characteristics of the speech signal are evaluated in the diffusion film, the arithmetic nodes of which perform a coupling of adjacent spectral components in the spectrum, in the manner of a diffusion in the frequency area, with a diffusion constant D>0.
7. Method according to claims 1 to 6, characterised in that all the spectral components of the filtering function F (f, T) at instant T are multiplied by the corresponding amplitudes A (f, T).
8. Method according to claim 7, characterised in that the attenuation of the filter for signal components having modulation frequencies between 0.6 and 6 Hz is less than 3 dB in respect of all values of the control parameter K, with the result that these signal components pass through the filter, whereas frequency components having modulation frequencies outside the 0.6 to 6 Hz interval are identified as noise and attenuated more severely, in dependence on the setting of the control parameter K.
9. Apparatus for noise reduction during speech transmission, more particularly in conjunction with a method according to claims 1 to 8, characterised by a neural network having

   a minima detection film which detects minima via rear signal spectra,

   a reaction film which performs a non-linear reaction function on the minima being detected by the minima detection film,

   a diffusion film which performs spectral smoothing on the outputs from the reaction film using only local couplings of adjacent arithmetic nodes in the diffusion film,

   an integration film which totalises the output from the diffusion film without weighting in an arithmetic node,

   a filtering function F (f, T) for noise filtering being produced by coupling arithmetic nodes of the successive films, viz. the minima detection film, reaction film, diffusion film and integration film, the spectral components differing from one another by the frequency f and corresponding to individual arithmetic nodes of any one film of the neural network, with the exception of the integration film, and each arithmetic node of the minima detection film determining a value M (f, T) for the frequency component f at instant T, where M (f, T) is produced by temporally averaging the amplitudes A (f, T) over a time interval corresponding to m frames, and the minima detection extending over a time interval corresponding to m frames, where I > m.
10. Apparatus according to claim 9, characterised in that with the aid of a reaction function r [S (T-1)] from the integral signal S (T-1) and a freely selectable parameter K which determines the degree of noise reduction, any one arithmetic node of the reaction film determines a component of the relative spectrum R (f, T) to give R (f, T) = 1 - M (f, T) r [S (T - 1)] K/A (f, T) with the reaction function r [S (T-1)], from A (f, T) and M (f, T).
11. Apparatus according to claims 9 and 10, characterised in that the range of values of the reaction function is confined to an interval [r1, r2], where R (f, T) = 1 applies if R (f, T) > 1; and R (f, T) = 0 applies if R (f, T) < 0.
12. Apparatus according to claims 9 to 11, characterised in that an integral signal S (T - 1), computed in the integration film and fed back into the reaction film, is available from the previous frame (instant T - 1) to the arithmetic nodes of the reaction film.
13. Apparatus according to claims 9 to 12, characterised in that for modulation frequencies between 0.6 and 6 Hz the attenuation is less than 3 dB in respect of all the indicated values of the control parameter K.

Images