EP1103956B1 - Exponential reduction of echo and noise during speech pauses - Google Patents

Abstract

The exponential echo (E) and noise (N) reduction system samples the received signal regularly and has a speech pause detector (SPD) to scale down the cancellation constant (a0(k)) used to weight the stored best estimate of noise power spectral density (g(S/N)) during pauses.

Description

die Abtastwerte durchzählt und T die Periodendauer von einem Abtastwert zum nächsten bedeutet. The invention relates to a method for the reduction of echo and / or noise signals in telecommunications (= TK) systems for the transmission of useful acoustic signals, in particular human speech, in which by means of speech pause detection is determined when in the mixture to be transferred from useful signals and interfering signals, a speech signal is included or when there is a speech break, wherein by means of a multiplier with two inputs the disturbed usually by echo and / or noise signals useful in their amplitude by a time-dependent control signal a ₀ (t) or by a in the rhythm of a sampling rate f _T = 1 / T clocked control signal a ₀ (k) are changed, where k ∈

the samples are counted and T is the period from one sample to the next.

Such a method is known for example from DE 42 29 912 A1.

During a natural communication between people one fits in usually the amplitude of the spoken language automatically to the acoustic environment. In case of voice communication between remote But places are not in the same acoustic Environment and are therefore each not the acoustic situation at the place of other interlocutor aware. Particularly aggravated therefore occurs Problem if one of the partners because of its acoustic environment is forced to speak very loud while the other partner in one quiet acoustic environment produces low-amplitude speech signals.

In addition there is the problem that on a TK channel also an "electronic generated "noise arises and as a background to the useful signal is transferred. Furthermore, it is also advantageous to interference signals such as unwanted background noise (street noise, factory noise, office noise, Canteen noise, aircraft noise, etc.) to reduce or completely suppress. To the To increase comfort when making calls, one is generally anxious to do any kind of Keep noise as low as possible.

Finally, so-called echoes occur in telecommunications connections in two-wire TK networks as line echoes and for example in simple and more uncomfortable TK terminals in the form of acoustic echoes occur.

In general, therefore, it is in the transmission of a mixture of Speech signal and interference signals over TK networks important, the interference signals such Lowering the noise and echo as much as possible in their amplitude.

One known method of noise reduction is the so-called "spectral Subtraction ", for example, in the publication" A new approach to noise reduction based on auditory masking effects "by S. Gustafsson and P. Jax, ITG Symposium, Dresden, 1998. It is about to a spectral noise reduction method, in which an acoustic Masking threshold (for example according to the MPEG standard) is taken into account. A disadvantage of such methods is the complicated Determination of this acoustic masking threshold and the execution of all arithmetic operations associated with this method.

In a spectral subtraction, the noise in the first Speech pauses measured and in the form of a power density spectrum continuously stored in a memory. The power density spectrum is over won a Fourier transformation. When speech occurs then the stored noise spectrum "as the best current estimate" Subtracted from the current disturbed speech spectrum, then in the Time domain transformed back to this way a noise reduction for the disturbed signal.

Another disadvantage of the spectral subtraction is that due to the Process of a principally inaccurate spectral noise estimation and Subsequent subtraction also errors occur in the output signal itself as "musical tones" noticeable. Besides, this is known Method hardly to suppress echo signals in TK connections suitable.

In the extended spectral signal processing, which is also described in the above citation, the power density spectra for the noise and for the speech itself are first estimated by means of a spectral subtraction. From the knowledge of these partial spectra, a spectral acoustic masking threshold R _T (f) for the human ear is then calculated using the rules from the MPEG standard, for example. Using this masking threshold and the estimated spectra for noise and speech, a filter pass curve H (f) is then calculated according to a simple rule, which is designed so that essential spectral parts of the speech are transmitted as unchanged as possible and spectral parts of the noise are lowered as far as possible.

Then the original disturbed speech signal is only through this filter given in order to reduce noise for the disturbed signal to obtain. The advantage of this method is that of disturbed signal "Nothing is added or subtracted" and therefore errors in the Estimates are less to barely perceptible. The disadvantage is again the considerable computational effort for the spectral noise suppression as well as the adaptive filters to be pre-switched for echo cancellation.

In the known Kompander method, as for example in the the initially cited DE 42 29 912 A1 is described, the degree of Noise and echo reduction according to a fixed preset Transfer function, among other things, a level reduction also at makes very small input signals.

The compander initially has the property of having speech signals certain (pre-set) "normal speech signal level" (possibly normal Called volume) virtually unchanged from its entrance to the exit transferred to.

But if the input signal becomes too loud, e.g. because a speaker too comes close to his microphone, so limited a dynamic compressor the Output level to almost the same value as in the normal case by the current gain in compander with increasing input volume linear is lowered. This property leaves the language at the output of the Kompandersystems about the same volume - regardless of how strong the Input volume fluctuates.

Now, on the other hand, a signal with a level that is smaller than the Normal level is given to the input of the compander, so will that Signal is additionally attenuated by the gain being back-regulated to If possible, transmit background noise only attenuated.

The compander thus consists of a compressor for speech signal levels, which are greater than or equal to a normal level and an expander for Signal levels that are less than the normal level. The gain reduction in the expander becomes stronger with increasingly smaller input levels.

A disadvantage of the compander solution is the considerable computational effort, the is required to carry out the known method. By the Compression of the speech signal level on the one hand and through the expansion on the other hand, there is also a modulation in the speech volume which changes the speech signal in a manner that the The subjective result is often perceived as unsatisfactory, i. one unsatisfactory hearing impression leaves behind.

Another method of noise reduction is in document US-5,533,133. A. This document discusses Speechless Segments linear unit superimposes speech segments.

Object of the present invention is in contrast, a method with to present the features described above, in which possible inexpensive and inexpensive way without big Computing and with low demand for computer memory and Data storage is an echo and noise reduction is effected with simple means as pleasant as possible for the human ear Acoustic overall impression produced, depending on the taste in addition to individual needs can be adjusted.

According to the invention, this object is achieved in a manner that is simple and effective in that the control signal a ₀ (t) or a ₀ (k) is varied such that the amplitude of the control signal a ₀ (t ₁ ) during the presence of speech signals in the useful signal ) or a ₀ (k) is set to a predetermined constant gain value c ₀ and at the beginning of a speech break in the payload the amplitude of the control signal a ₀ (t) or a ₀ (k) from one sample to the next according to the recursion formula a 0 (k + 1) = a 0 (k) · β with β <1 is steadily lowered,
and that after the end of a speech pause again a ₀ (k) = c _{0 is} set.

This is a very simple and very cost-effective method ready, the also gives a surprisingly good quality with respect to a noise reduction, by preferably in speech pauses the disturbing echo and Noise signals reduced. During the language phases themselves the Noise at least partially masked and therefore from the human ear much less clearly perceived. By omitting the compression according to the known compander method, the original speech signal significantly less changed, so in the end a generally better sounding voice signal arrives at the other end of the line. Furthermore the method according to the invention requires less computing power than The compander method, since at least the compression is omitted. Accordingly, lower capacities of data memories and Compensators required, which is the inventive method in Contrary to the known methods simpler and cheaper designed.

In order to achieve an effective noise reduction, the to be transmitted Signal during the pauses in its power value according to a temporal exponential function, as opposed to one of the input level dependent lowering, as in the compander method, lowered. In order to Already a significant noise reduction is achieved. In addition, that a reduction of the noise during a speech break the hearing significantly less burdened by making the deafness effect louder Sound effect significantly reduces. The ear can be reinserted when the Speech more responsive and listen more carefully.

Advantageously, the factor β is chosen so that the steady time reduction in about a time constant τ ₁ corresponds to the perceptibility of the human ear. This means that after a strong sound event, the human ear does not perceive new sound events in time and in their amplitude below a curve which decays with the time constant τ ₁ after the end of the strong sound event. Therefore, a variant of the method according to the invention is preferred in which the factor β is determined from the sampling rate f _T , from a time constant τ ₁ and from a predetermined constant prefactor c ₁ according to the relationship β = c ₁ · exp (-1 / τ ₁ f _T )

The value of the time constant τ ₁ in humans is usually between 50 ms to 150 ms and is preferably about 65 ms.

In order to dimension the factor β exactly according to the time constant τ ₁ , it is favorable if c ₀ = 1 is selected.

If the steady exponential decay of the interfering signals is not limited according to the recursion formula described above, the value of a ₀ (k) grows very rapidly with increasing k and approaches 0. However, this is not always desirable since in many cases it is better want to hear a small residual noise to avoid during a speech break the impression that the telecommunications line is suddenly "dead" or interrupted. A variant of the method according to the invention is therefore preferred in which a ₀ (k + 1) assumes a predetermined constant value c ₂ during a speech pause and / or a value c ₃ <c ₂ if an echo signal is present if the predecessor value a ₀ (k ) ≤ c ₂ has become.

Furthermore, it is desirable to reduce the degree of signal level reduction in the Language pauses of the current situation in the TK channel adapt.

For example, one sounds preferably depending on instantaneous noise level N or as a function of a function g (S / N) Lower the signal-to-noise ratio S / N, but short-term Lower echoes more and lower the echo to the end of the echoes reduce the value of the noise reduction.

Particularly preferred is therefore a method variant, which is characterized in that in the presence of a noise signal and / or echo signal and for a ₀ (k) ≤ c ₂ , where c _{2 is} a predetermined constant, the power value of the noise level N in the currently used TK- Channel continuously measured and / or estimated, and that depending on the current noise level N, the control signal a ₀ (k + 1) is continuously set according to a ₀ (k + 1) = f (N), where f (N) a predetermined Function of N is.

Thus, the degree of noise reduction from the currently occurring Power value N of the noise automatically controlled and the current Noise value in the telephone channel adapted and defined in a predetermined Tracked manner. By choosing the function f (N), the subjective Impression of the generated overall signal to be adjusted. Another advantage This variant of the method consists in that in a bundle of Telephone channels, for example between international exchanges, the noise situation in each individual channel, which is very good from channel to channel can be different, adjusted automatically and optimized individually can.

Particularly preferred is a variant of the method according to the invention is characterized in that the given function f (N) is a function g (S / N), which is derived from the quotient S / N from the power value of the signal level S the useful signals to be transmitted and the power value of the noise level N or that the given function f (N) is a function g '(N / S), which depends on the reciprocal N / S of this quotient. For the sake of one simpler practical realization can also be a function of (S + N) / N or from (S + N) / S.

The advantage of the above process variant is that when strong varying signal level S in the telephone channels of a bundle always the correct setting for the noise reduction is found. At a Controlling the noise reduction proportional to the reciprocal N / S can be the Function g '(N / S) easily on a digital signal processor (= DSP) with fixed Computer word lengths of, for example, 16 bits using implement a particularly simple software, as for N / S preferably a Number range 0 <N / S <1 relevant for controlling the noise reduction or interesting is.

Acoustic hearing tests have shown that with S / N = 0 db the speech is already so strongly disturbed that it is only possible to reduce the noise by a value f ₀ or g ₀ between 5 and 10 dB, preferably between 6 and 8 dB in order not to impair the overall acoustic impression with regard to the naturalness of the language. With even less favorable values of the signal-to-noise ratio S / N <0 dB, the value f ₀ or g ₀ can then be maintained, since any further noise reduction only worsens the overall impression.

At medium S / N, a greater noise reduction can be made according to these studies. A maximum results in the range of 10 to 15 dB. The value of the noise reduction f _max or g _max should be at a maximum between 20 and 30, preferably about 25 dB.

With very good noise S / N> 40 dB should only a minimum Lowering between 0 and 3 dB can be set to the naturalness of the receive as much as possible translated language.

The sound of the language and the intelligibility are especially good when the Function f (N) or g (S / N) over the three areas discussed above in FIG constantly related to each other, with rapid changes in N or in S (N) can be advantageously smoothed by filtering.

A relatively simple realization in hardware and / or software results, by passing the functions f (N) or g (S / N) or g '(N / S) through even Characteristic sections between the three operating points described above approximates (sectionally linear approximation).

In a somewhat more complicated variant of the method according to the invention, but the result in the result to a better sound, is one Polynomial function for implementing the continuous functions f (N) and g (S / N) or g '(N / S) in the three areas discussed, which resulted leads to a kind of asymmetrical bell function.

Particularly preferred is a variant of the method according to the invention the functions f (N) or g (S / N) or g '(N / S) are chosen so that the Reduction of the noise level N is correct according to the psychoacoustic Means of the human auditory spectrum is done. The value for S and / or N not only from the instantaneous power value alone, but also determined from a weighted spectral curve of S and N and in total, a hearing aid, i. a psychoacoustically pleasant sounding noise reduction achieved. Since there is no easily representable measure for an acoustically pleasant sounding Noise reduction is there, all quality assessments are extensive Hearing tests dependent, which then optimized by means of statistical Methods are evaluated to a rating scale, (similar to Speech codecs).

A good noise level estimation requires a good speech pause detector, because one can only be sure that in the speech pause sections just annoying noise and not some mix between noise and speech fragments, as is common in practice occurs.

Therefore, a method variant which is characterized by this is particularly preferred characterized in that in the speech pause detector from the input signal x a short-term output sam (x) by means of a short-term level estimator, by means of a middle-time level estimator, a middle-time output signal mam (x) and by means of a long-term level estimator, a long-term output signal lam (x) is formed, that the three output signals sam (x), mam (x) and lam (x) via suitable gain coefficients are set to be approximately equal large, if the input signal x is a pure noise signal, where sam (x) < mam (x) <lam (x), that the three output signals sam (x), mam (x) and lam (x) be monitored by comparators, and that the presence of a Speech signal is assumed as the input signal x, if sam (x) and mam (x) initially each become larger than lam (x), and the presence of a Speech pause, after which sam (x) and / or mam (x) becomes smaller again than lam (x).

With the help of these relatively simple ways of forming different Averages of the time signal can already be a surprisingly good Speech pause detection are performed, which is only a very small Computing effort required.

A development of this variant of the method provides that the three output signals sam (x), mam (x) and lam (x) for speech pause estimation be supplied to a neural network, which with a variety was trained by scenarios with different input signals x. One Neural network may advantageously have linear and non-linear relationships between a large amount of input parameters and the desired ones Map output values. A prerequisite for this is that the neural Net once with a sufficient amount of input values and training associated output values. Therefore, neural are suitable Networks especially for the task of speech pause detection Presence of different disturbing noises.

Preferably, in addition to the detection and reduction of noise signals, the presence of echo signals is also detected and / or predicted, and the corresponding echo signals are suppressed or reduced. If, in addition to noise, echoes also occur in a telephone channel, they can generally be predicted on the basis of a previously determined signal propagation time τ _{E of} an echo and the previously determined echo coupling ERL in the channel and the signal strength ES which triggers the echo in the return channel. This estimation can be carried out in such a way that the magnitude of the delayed incoming echoes is estimated as a function of the transmitted speech signal and its instantaneous power. If the respectively estimated echo signal exceeds a predefined threshold value thrs in certain short time segments, then this echo-loaded signal is preferably additionally briefly damped, for example by the abovementioned exponential depression, to a value which is necessary for a substantial reduction of the echo signal. In the same sense, a compander characteristic curve can also be briefly moved in the direction of greater input volume for echoes and returned to its original position after the echoes have faded away.

Particularly preferred is a development of this method variant, in which the control signal a ₀ (k + 1) is set continuously according to a ₀ (k + 1) = h (N, S, ES, τ _E , ERL), where h (N, S, ES, τ _E , ERL) is a predetermined function of N, S, the useful signal ES in the opposite direction of a speaking TK partner, τ _{E is} a constant delay time of the echo signal and ERL is an attenuation constant of the amplitude of the echo signal.

It can be advantageous to aurally sound reduction with a regardless of working echo reduction. That's special then important if in the phone channel as well as no background noise exists, since then no noise reduction is effective, and thus occurring echoes unhindered can reach the speaker.

A separation of the control of a noise reduction from that of an echo reduction is expedient, since noises and echoes occur independently of each other and also generally have completely different physical causes. Mathematically, however, one can specify a general reduction function R which describes a reduction of signal levels for both noises and echoes: R (S, N, ES, τ e , ERL, thrs) ~ g (S / N) d (ES, τ e , ERL, thrs), where g (S / N) is the above-described noise reduction and d (...) is the independently additionally occurring echo reduction when the estimated echo signal exceeds the predetermined threshold value thrs.

Particularly advantageous is a process variant in which during the Duration of an echo reduction to the useful signal in addition an artificial Noise signal is added.

A noise reduction is also at a constant noise level constant. A sudden onset of echo reduction in the rhythm of Language also means a noise reduction (at least in the short Period) in the speech rhythm. This leads to a pulsed Background noise, which does not sound natural. Therefore, it is advantageous, in the moments of an additional echo reduction Synthetic noise of a suitable noise generator in the Magnitude of the normal background noise to the processed signal to add. This should be as consistent as possible background noise be conveyed to the listener.

The noise generator can be designed so that the artificial Sound signal a psychoacoustically perceived as pleasant acoustic Signal sequence (= comfort noise) includes.

But instead of a synthetic background noise can also Section of a previously recorded true background noise in appropriate strength in the echo time slots. The added noise then differs as well as not at all from the previous noise and therefore no annoying acoustic Cause change in the listener.

The addition of sounds for the acoustic masking of effects as well as the measures of separate treatment of noise and Echoes, if properly matched, will be extra special understandable and pleasant language impression even with "difficult" Environment (echoes plus noise).

Particularly preferred is also a variant of the invention Method in which the useful signal to be transmitted to a spectral Subtraction is subjected. The advantage of a spectral subtraction with Downstream level reduction in the speech pauses is that First, by means of spectral subtraction, a part of the noise from the Speech signal itself is eliminated and only then the speech pauses in the described type of noise and echoes are released. Total results This combination in subjective tests better hearing impressions than just one simple spectral subtraction.

Another particularly advantageous variant of the invention Finally, the method provides that the useful signal to be transmitted is a subjected to human hearing adapted spectral filtering. Again, the means of a spectral subtraction first a Estimation of noises, speech and echoes performed, then determines an auricular masking threshold and then the entire Signal via a suitably set transmission filter so edited that the Language components as unadulterated as possible and the echo and noise components be suppressed as much as possible.

A combination with the downstream level reduction in the Speech pauses further improve the listening experience.

In the context of the present invention also falls to a server Support of the method according to the invention described above as well a computer program for carrying out the method. The procedure can both as a hardware circuit, as well as in the form of a computer program will be realized. Nowadays, software programming for powerful DSP's preferred because new knowledge and additional features easier by changing the software on an existing hardware basis can be implemented. However, methods can also be used as hardware components For example, be implemented in telecommunications terminals or telephone systems.

Further advantages of the invention will become apparent from the description and the Drawing. Likewise, those mentioned above and even further listed features according to the invention in each case individually or for themselves several in any combination use find. The shown and described embodiments are not to be exhaustive enumeration but rather have an exemplary character for the description the invention.

Fig. 1

das Steuersignal ao bei Vorliegen von Sprachsignalen, während einer Sprachpause und bei erneutem Einsetzen der Sprachsignale;

Fig. 2

ein Schema einer Anordnung zur gesteuerten Signalabsenkung;

Fig. 3a

die Funktion g(S/N) in der linearen Näherung;

Fig. 3b

die entsprechende Funktion g'(N/S);

Fig. 4a

die Funktion g(S/N) als unsymmetrische Glockenkurve; und

Fig. 4b

die entsprechende Funktion g'(N/S).

The invention is illustrated in the drawing and will be explained in more detail with reference to embodiments. Show it:

Fig. 1

the control signal ao in the presence of speech signals, during a speech pause and on re-insertion of the speech signals;

Fig. 2

a schematic of a controlled signal reduction arrangement;

Fig. 3a

the function g (S / N) in the linear approximation;

Fig. 3b

the corresponding function g '(N / S);

Fig. 4a

the function g (S / N) as unbalanced bell curve; and

Fig. 4b

the corresponding function g '(N / S).

The control signal a ₀ shown in FIG. 1 as a function of the time t or the sampling number k is maintained at a value c ₀ = 1 during a first phase T1 in which speech signals are detected. During a speech pause in the time period T2, the control signal a _{0 is} lowered exponentially to a constant value c ₂ , which is just above 0, and then abruptly returns to the value c ₀ = 1 (or another, arbitrary) when the speech signals are reinserted during a phase T3 selectable constant). As a result, during the speech phases T1, T3, no (or in other examples only a small) suppression of interference signals in the overall signal is carried out, so that the speech signal is forwarded as unadulterated as possible and unhindered. During the speech pause in the phase T2 as effective as possible suppression of echoes and noise signals is effected as quickly as possible (in the present example, not to the value 0, but to a small residual value c _{2 is} lowered, so as not at the other end of the Impression of a "dead" line to awaken. When echoes occur, a reduction to a residual value c ₃ <c _{2 is} made.

FIG. 2 schematically shows the mode of operation of an arrangement for noise and echo reduction in accordance with the abovementioned reduction function R (S, N, ES, τ _E , ERL, thrs) with a speech pause detector SPD.

For all the curves illustrated in FIGS. 3a to 4b, the function value g or g 'for the case S / N <0 dB, ie for an extremely high background noise, changes into a constant value g _{0 of} the noise reduction of approximately 6 dB. Starting from S / N = 0 dB, an increased noise reduction is achieved with increasing improvement of the signal-to-noise ratio S / N, reaching a maximum g _max ≈ 25 dB at approximately S / N≈12 dB. With increasing S / N, the degree of noise reduction finally drops to zero in order to perform as little manipulation in the transmitted useful signal with low background noise.

Claims (23)

1. Method for reduction of echo and/or noise signals in telecommunications (= TK) systems for the transmission of acoustic useful signals, in particular human speech in which by means of speech-pause detection it is determined when a speech signal is contained in the mixture of useful signals and interference signals to be transmitted or when a speech pause is present, where by means of a multiplier with two inputs the useful signals usually disrupted by echo and/or noise signals are modified in their amplitude by a time-dependent control signal a₀(t) or by a control signal a₀(k) pulsed at the rhythm of a sampling rate f_T=1/T, where k∈N counts through the sampled values and T indicates the period from one sampled value to the next, characterised in that the control signal a₀(t) or a₀(k) is varied so that during the presence of speech signals in the useful signal the amplitude of the control signal a₀(t) or a₀(k) is set to a prespecified constant value c₀, and on start of a speech pause in the useful signal the amplitude of the control signal a₀(t) or a₀(k) is lowered constantly from one sampled value to the next according to the recursion formula a0(k+1) = a0(k) . β   with β < 1 and that after the end of a speech pause a₀(k)=c₀ is set.
2. Method according to claim 1, characterised in that the factor β is determined from the sampling rate f_T, a time constant τ₁ and a prespecified constant prefactor c₁ according to the relation β = c₁.exp(-1/τ₁f_T).
3. Method according to claim 2, characterised in that the time constant τ₁ is selected between 50 ms and 150 ms, preferably τ₁ ≈ 65 ms.
4. Method according to any of the previous claims, characterised in that the constant value c₀ = 1.
5. Method according to any of the previous claims, characterised in that a₀(k+1) during a speech pause and/or the presence of an echo signal assumes a prespecified constant value c₂ if the preceding value a₀(k) ≤ c₂.
6. Method according to any of claims 1 to 4, characterised in that during a speech pause and/or the presence of an echo signal and for a₀(k) ≤ c₂, where c₂ is a prespecified constant, the power value of the noise level N in the TK channel currently used is measured and/or estimated continuously, and that depending on the current noise level N the control signal a₀(k+1) is adjusted constantly according to a₀(k+1) = f(N), where f(N) is a prespecified function of N.
7. Method according to claim 6, characterised in that the prespecified function f(N) is a function g(S/N) which depends on the quotient S/N of the power value of the signal level S of the useful signal to be transmitted and the power value of the noise level N, or that the prespecified function f(N) is a function g'(N/S) which depends on the reciprocal value N/S of this quotient.
8. Method according to claim 7, characterised in that the function f(N) or g(S/N) at 1/N<<1 or S/N = 0dB begins at a constant value f₀>0 or g₀>0, rises to a maximum f_max or g_max in the range between N or S/N = 10dB to 15dB preferably at N or S/N = 12dB, and then falls to a minimum value f_min or g_min preferably at 0dB, where 5dB ≤ f₀, g₀ ≤ 10dB, preferably 6dB ≤ f₀, g₀ ≤ 8dB, and where 20dB ≤ f_max, g_max ≤ 30dB, preferably f_max, g_max ≈ 25dB.
9. Method according to any of claims 6 to 8, characterised in that the function f(N) or g(S/N) runs at least in sections, preferably in all part sections, linear with N or S/N.
10. Process according to any of claims 6 to 8, characterised in that the function f(N) or g(S/N) is constructed from polynomials and runs as an asymmetric bell curve over N or S/N.
11. Method according to any of claims 6 to 10, characterised in that the functions f(N) and g(S/N) and g' (N/5) are selected so that the reduction in noise level N is acoustically correct according to the psycho-acoustic mean value of the human hearing spectrum.
12. Method according to any of the previous claims, characterised in that in addition to the detection and reduction of noise signals, the presence of echo signals is detected and/or predicted and the echo signals are suppressed or reduced.
13. Method according to claim 12 and one of claims 6 to 11, characterised in that the control signal a₀(k+1) is adjusted continuously according to a₀(k+1) = h(N,S,ES,τ_E,ERL), where h(N,S,ES,τ_E,ERL) is a prespecified function of N, S, in the opposite direction to the useful signal ES of a speaking TK partner, τ_E a constant delay time of the echo signal and ERL a damping constant of the amplitude of the echo signal.
14. Method according to claim 12, characterised in that the reduction of noise signals and the reduction of echo signals are controlled separately.
15. Method according to any of claims 12 to 14,
    characterised in that for the duration of an echo reduction, a synthetic noise signal is also added to the useful signal.
16. Method according to claim 15, characterised in that the synthetic noise signal comprises an acoustic signal sequence perceived as pleasant psycho-acoustically (= comfort noise).
17. Method according to claim 15, characterised in that the synthetic noise signal comprises a noise signal previously recorded during the current TK connection.
18. Method according to any of the previous claims, characterised in that in a speech-pause detector (SPD) from an input signal x by means of a short-term level estimator a short-term output signal sam(x) is formed, by means of a medium-term level estimator the medium-term output signal mam(x) and by means of a long-term level estimator a long-term output signal lam(x), that the three output signals sam(x), mam(x) and lam(x) are set via suitable amplification co-efficients so that they are approximately equally large when the input x is a pure noise signal, where sam(x) < mam(x) < lam(x), that the three output signals sam(x), mam(x) and lam(x) arc monitored by comparators, and that the presence of a speech signal is taken as input signal x if sam(x) and mam(x) at first are each greater than lam(x), and the presence of the speech pause if thereafter sam(x) and/or mam(x) become smaller than lam(x).
19. Method according to claim 18, characterised in that for speech-pause estimation, the three output signals sam(x), mam(x) and lam(x) are supplied to a neuronal network which has been trained with a multiplicity of scenarios with different input signals x.
20. Method according to any of the previous claims, characterised in that the useful signal to be transmitted is subjected to spectral subtraction.
21. Method according to any of the previous claims, characterised in that the useful signal to be transmitted is subjected to spectral filtering adapted to human hearing.
22. Server unit with means for performance of each individual step of the method according to any of claims 1 to 21.
23. Computer program which has instructions for performance of each individual step of the method according to any of claims 1 to 21 on a computer when the computer program is loaded on the computer.

Images