DE10120231A1 - Single-channel noise reduction of speech signals whose noise changes more slowly than speech signals, by estimating non-steady noise using power calculation and time-delay stages - Google Patents

Abstract

The method involves a non-steady noise estimation (GSI) which allows effective noise reduction in combination with a selection stage (AS). The selection stage checks if the estimation is acceptable. A stage for calculating the power and power maximum of the input signal and a time delay stage for the non-steady noise component are used in the checking. A time delay stage (TX) extends the analysis range for the GSI stage into the future time domain. A bypass controller (BS) prevents the influence of unwanted speech components in the determination of a noise filter (GF). An Independent claim is included for a single-channel noise reduction apparatus.

Description

The invention relates to the field of noise reduction of extraneous noise in voice communication and is based on a method and an arrangement for single channel noise reduction.

In many applications of voice communication, this can be done by the microphone recorded voice signal by an acoustically superimposed noise signal from the Surrounding the speaker. This is especially the case when a very loud one Noise source is present or if hands-free devices are used. It can cause significant impairments in voice communication occur.

There are a number of proposals for noise reduction techniques in the literature; an overview can be found e.g. B. in "signal processing method for improvement of hands-free voice communication - Part 3: Procedure for Noise reduction "by R. Wehrmann, R. Poltmann, H. Schütze and R. Zelinski, The Telecommunications Engineer, Issue 2, 1995. The procedures can be roughly divided into single-channel Subdivide (only 1 microphone) or multi-channel (2 or more microphones) procedures. Out For reasons of technical complexity, only single-channel processes are described below to be viewed as.

Fig. 1 shows the typical structure of a single-channel noise reduction process. The disturbed input signal x = s + n, which consists of the speech signal s and the superimposed interference signal n, is picked up by the microphone M. Since the amplitude and phase of the current interference signal n are not known and also cannot be estimated directly, the stationary noise component is estimated in the speech pause stage GSS during the speech pauses as a statistical description of the noise signal, e.g. B. in the form of the frequency-dependent short-term power density spectrum of the stationary noise component P _GS (f). This noise component could also be described as equivalent in the form of the short-term autocorrelation function; Therefore, the term "noise component" is intended to encompass both display options in the following. By evaluating the input signal x in the form of the associated short-term power density spectrum P _X (f) or the short-term autocorrelation function and the estimated short-term power density spectrum for the stationary noise component P _GS (f), the adaptive noise filter GF is calculated and continuously Power density spectrum of the speech signal adapted. The noise filter GF can e.g. B. can be realized as a Wiener estimation filter or according to the principle of spectral subtraction. The filtering of the input signal x with the adaptive noise filter GF results in the noise-reduced, reconstructed speech signal, which is transmitted to the remote subscriber.

The time-varying noise filter GF can occasionally lead to additional and conspicuous background noises (artifacts) that are disturbed by the distant participant be perceived. To remedy this, numerous methods have been proposed so contains z. B. DE 198 18 609 C2 a noise filter combined with adaptively controlled Bypass. This bypass function reduces the filter effect, but also reduces it at the same time the strength of such disruptive artifacts. Especially in the time periods with high speech signal energy, the filter effect reduced by the bypass in Purchase because of a partial masking of the noise component the high-energy speech signal occurs.

The single-channel noise reduction method described here is disadvantageous however, that they can usually only be used when the ambient noise is stationary or changes very slowly (e.g. fan noise or road traffic noise). This limitation is caused by the estimate of the amount of noise that is usually activated during the pauses in speech. Because such a language break up to one second can be significant with time-varying noises Deviations between the current noise component and the estimated noise component occur. These deviations can lead to not only the annoying Ambient noise is not attenuated, but also new ones  Noise caused by the incorrectly set noise filter GF in the signal path to the far distance Participants are introduced.

A remedy would be to use noise reduction methods with several Microphones (multi-channel process) possible; however, such procedures are grounded of the significantly increased technical effort often not applicable.

There are hardly any suggestions in the literature for realizing a single-channel noise reduction method that can also be used for less stationary noise signals. If the disturbing noise signal is a second speech signal from the vicinity of the desired speaker (e.g. a background speaker), methods for Solution of the "Co-Channel Speech" problem can be used, such as are described in "Techniques for suppression of an interfering talker in co-channel speech "by J.A. Naylor and S.F. Boll, Proceedings of 1987 International Conf On Acoustics, Speech and Signal Processing, Dallas, pp. 205-208. The disadvantage, however, is that these methods can only be used for this one problem (one Background speaker as a disturbance) and also only under very ideal Boundary conditions work reliably.

The technical task is based on an arrangement and a single-channel method Noise reduction aligned, the speech signal from the disturbed speech signal processed, which contains considerably fewer noise components than that by the known single-channel method for noise reduction processed speech signals.

The solution according to the invention for single-channel noise reduction can be used if two conditions are met:

• A) The spectrum of noise may not change as quickly as that Spectrum of the speech signal. So it should last for at least 60 ms or no longer distinguishable from the spectrum of a stationary noise be.  
• B) The signal-to-noise ratio at the microphone input of the microphone M should amount to at least 6 dB or more, so that the speech signal compared to the Noise signal slightly dominated.

The basic idea of the new process is to stop the noise component in the (Possibly long ago) language breaks, but in the immediate time Proximity of the signal section to be filtered from the disturbed input signal x estimate. This makes the deviation compared to conventional methods a lot between the noise estimate and the current noise spectrum less if the noise spectrum changes over time. Because the noise estimate is no longer carried out during the language breaks, however, it also contains these unwanted parts of the speech signal. However, this disadvantageous influence can be caused by complementary measures are further weakened. So through a Selection level AS controlled only accepts those noise estimates at which the Signal level is significantly below the maximum speech level. An additional Time delay TX before the entrance of the noise filter GF allows, too Use "future" signal sections to estimate the noise component. This significantly increases the chance of a signal section of the input signal x to be found with only slight speech influence to estimate the noise component. Finally, the influence of remaining language parts in the Noise estimation through the use of an adaptive bypass function, bypass control BS, weakened for the noise filter GF. The bypass function reduces the Filter effect for high-energy sections of the input signal x. The control of this However, function does not take place as described in the known solutions.

A block diagram of the method arrangement is shown in FIG. 2. The basic core is formed by the elements known from FIG. 1, microphone M, noise filter GF and noise estimation GSS for the stationary noise component. Furthermore, a bypass control BS known per se, stages known per se for time delay and a stage known per se for calculating the maximum power of the input signal LM are used. The time delay stage for the input signal TX is arranged between the input of the nearby subscriber E and the noise filter GF. The output of the time delay stage for the input signal TX is connected to an input of the bypass control BS, an input of the noise filter GF and an input of the filter adaptation of the noise filter GF. The input of the near participant E is on the one hand via a known stage for calculating the maximum power of the input signal LM with the bypass control BS and a selection stage AS and on the other hand with the known stage for stationary noise estimation GSS and an additional stage non-stationary noise estimation connected to GSI. The stage for stationary noise estimation GSS is connected to the bypass control BS and the selection stage AS. The stage for the transient noise estimation GSI is connected on the one hand directly and on the other hand via an assembly for the time delay TG to the selection stage AS. The selection stage AS is connected to the filter adaptation module of the noise filter GF via a multiplier MU. The multiplier MU has a connection to the output of the bypass control BS, the output of the stage for stationary noise estimation GSS, the output of the stage for transient noise estimation GSI and the output of the time delay stage for the transient noise component TG.

In the stage for calculating the maximum power LM, the power L _{x of} the input signal x is continuously measured as a short-term mean square over about 30 ms according to the relationship

L _x = x² (1)

calculated and from this the sliding power _maximum L _XM according to the relationship

L _XM.new = max {L _x ; µL _{XM, old} }; µ <1. (2)

certainly. The decay factor µ means that a high power value does not persist forever, so that the power maximum L _XM can adapt to gradually decreasing signal volume. The power maximum of the input signal L _XM calculated according to the relationship (2) serves as a control variable for the selection stage AS and the bypass control BS.

The input signal x is evaluated in the already known stage for the stationary noise estimation GSS. The estimated short-term power density spectrum for the stationary noise component is then available at the output of the stage for stationary noise estimation GSS. The additional stage for the transient noise estimation GSI supplies the current value of the estimated short-term power density spectrum for the transient noise component P _GI (f).

FIG. 3 illustrates the mode of operation of the stage for stationary noise estimation GSS in connection with the stage for transient noise estimation GSI. The signal evaluation and noise filtering takes place block by block with a block length of about 30 ms. The signal block to be currently filtered is a section of the input signal x and has the index n. In the stage for stationary noise estimation GSS, after a speech pause, SBS = {n, n - 1 in a search area. . ., n-K}, which has a length of time of typically about 1 s and comprises the current block and blocks in time ( FIG. 3a). The signal block with the smallest block power within the search area SBS indicates a speech pause and the short-term power density spectrum P _GS (f) assigned to this signal block is therefore used as an estimate for the stationary noise component.

In the additional new stage GSI, only the search area SBI = {n-1, n, n + 1} is evaluated, which comprises the current signal block, the previous signal block and also the future signal block with the index n + 1 ( FIG. 3b) . The evaluation of the future signal block is made possible by an additional time delay stage for the input signal TX, which is arranged between the input of the nearby subscriber E and the adaptive noise filter GF. The concrete time delay of the time delay stage TX corresponds exactly to the length of a signal block. Assuming that the signal block from the search area SBI, which contains the smallest block power, presumably also contains the lowest speech component, the short-term power density spectrum assigned to this signal block is used further as an estimate for the unsteady noise component P _GI (f).

Since the estimate for the transient noise component P _GI (f) can still contain considerable speech components, particularly in the vocal language areas, the selection level AS is used to select values of the estimated short-term power density spectrum for the transient noise component P _GI (f) with a high signal power, which correspond to a indicate possible language content, excluded from further processing. Various performance comparisons are carried out in the AS selection level. First, the power maximum of the input signal L _XM according to (2) becomes a power threshold

L _SW = cL _XM ; c <1 (3)

determined, which is significantly below the performance of the current vocal language areas. The estimated power of the transient noise component L _{GI is checked} , which is calculated from the value of the estimated short-term power density spectrum for the transient noise component P _GI (f). The following queries are carried out:

L _GI <L _SW ? (4a)

If (4a) is fulfilled, the value of the estimated short-term power density spectrum for the transient noise component P _GI (f) is used as the noise estimate P _G (t) and the selection is completed. Otherwise, past values of the estimated short-term power density spectrum for the transient noise component P _GI (f) are used (delayed by means of stage TG). In this case, the value of the estimated power of the past transient noise signal L _{GI, TG is used} for the comparison. Is the condition within the search area SBS

L _{GI, TG} <L _SW ? (4b)

fulfilled, the time-delayed estimated short-term power density spectrum for the transient noise component P _{GI, TG} (f) is used as the noise estimate P _G (f) and the selection is completed. If condition (4b) is also not met, the estimated short-term power spectrum for the stationary noise component P _GS (f) is used as the final noise estimate P _G (f). The switch S of the multiplier MU shown in FIG. 2 is then set in accordance with the decision made in the selection stage AS for forwarding the noise estimate P _G (f).

The short-term power density spectrum of the noise estimate P _G (f) can still contain slight speech components. In order to reduce this residual influence, the bypass control BS is used to adaptively set a weighting factor b (0 <b <1) as a multiplier for P _G (f). At high power of the input signal L _x , small values of b are set (with the consequence of a weakened noise estimate), so that a noise filter GF with a reduced filter effect is calculated. The influence of the weighting factor b thus acts as an additional bypass for the optimal noise filter. Since at high power of the input signal L _{x there is} probably an energetic speech and the value of the estimated short-term power density spectrum for the transient noise component P _GI (f) could still contain speech components, the influence of such speech components on the filter adaptation is thus mitigated. At low power of the input signal L _x , the noise _component dominates; high values of the weighting factor b are then set (b → 1) and the adaptive noise filter GF has its full effect.

An expedient embodiment of the bypass function b (L _x ) is outlined in FIG. 4. In addition to the power of the input signal L _x , the power maximum of the input signal L _XM according to (2) and the estimated power of the stationary noise signal L _GS are evaluated in the bypass control BS. The estimated power of the stationary noise signal L _GS can be determined directly from the value for the estimated short-term power density spectrum for the stationary noise component P _GS (f). The dynamic corner points L _y and L _{z of} the characteristic curve b (L _x ) in FIG. 4 are determined by

where y, z and c are constant factors.

The design of the characteristic curve b (L _x ) according to FIG. 4 ensures that in speech-free fall, that is to say with small values of L _x , uniformly strong noise damping with the bypass factor b = b _max is always achieved. With large values of L _x , that is, if speech is likely to be present, the corner point L _z with b = b _{min is} always adapted to the current power maximum L _XM and thus to the power range of the high-energy vowel components.

For the sake of clarity, both frequency-dependent variables were only described in the frequency domain in FIGS. 1 and 2 and in the text. It is of course possible to implement parts or the entire process in the time domain. Instead of the short-term power density spectra P _GS (f), P _GI (f) and P _G (f), the corresponding short-term autocorrelation functions can also be determined. The adaptive noise filter GF can then, for. B. can be calculated from these prepared correlation functions as Wiener filters. The input signal x can also be filtered with the adaptive noise filter GF in the time domain. For this purpose, the noise filter is calculated as a transversal filter with a finite number of coefficients, so that the filtering is implemented as a convolution operation with the input signal x.

The expansion of the single-channel noise reduction process by the level of unsteady noise estimation GSI in combination with the selection level AS enables effective noise reduction even in a transient noise environment. An essential condition is that the noise spectrum is somewhat slower than that Language spectrum changed. Beyond that there are no other conditions such as the number or structure of the noise sources.

The additional stage for time delay of the input signal TX, which is before the input of the noise filter GF is arranged, extends the analysis range for the transient Noise estimation also on the "future" signal range and thus reduces the undesirable influence of speech components in the noise estimation. By the The influence of the time delay stage for the input signal TX increases the total Signal throughput time only marginal. In combination with picture codecs this will increased signal throughput time even meaningless, because picture codecs are usually one have significantly longer signal throughput times.

The inserted bypass control BS reduces the influence of unwanted ones Speech components from the transient noise estimate when calculating the  Noise filter GF. This prevents this undesirable influence from occurring Artifacts occur in the noise-filtered speech signal.  

List of reference symbols

E Entrance of the nearby participant
A Exit to the far party
x input signal
s voice signal
n Interference signal
reconstructed speech signal
M microphone
GSS level for stationary noise estimation
GF adaptive noise filter
LM stage for calculating the power and the maximum power of the input signal
AS selection level
BS bypass control
GSI level for transient noise estimation
MU multiplier
TX time delay stage for the input signal
TG time delay stage for the transient noise component
P _G

(f) Estimated short-term power density spectrum for the current noise component
P _GS

(f) Estimated short-term power density spectrum for the stationary noise component
P _GI

(f) Estimated short-term power density spectrum for the transient noise component
P _{GI, TG}

(f) Estimated time-delayed short-term power density spectrum for the transient noise component
Lx power of the input signal
L _XM

Maximum power of the input signal
L _SW

Power threshold
L _GS

Estimated power of the stationary noise signal
L _GI

Estimated power of the transient noise signal
L _{GI / TG}

Estimated power of the past transient noise signal
Ly; Lz dynamic corner points of the characteristic curve b (Lx)
µ decay factor
SBS search area for the stationary noise component (for level GSS)
SBI search area for the transient noise component (for level GSI)
b weighting factor
y, zc constant factors

Claims (5)

1.Procedure for single-channel noise reduction of disturbed speech signals, the noise of which changes more slowly than the speech signal, the signal-to-noise ratio at the input of the microphone (M) being at least 6 dB, and in which the estimated short-term power density spectrum for the stationary noise component ( P _GS (f)) is determined, characterized in that
in addition to the estimated short-term power density spectrum for the stationary noise component (P _GS (f)) the estimated short-term power density spectrum for the current unsteady noise component (P _GI (f)) is determined that
A permanently active selection function is used to determine a power threshold value (L _SW ) from the power of the sliding power maximum of the input signal (L _XM ), which successively corresponds to the power of the current estimated short-term power density spectrum for the transient noise component (P _GI (f)) , and the power of the time-delayed estimated short-term power density spectrum for the unsteady noise component P _{GI / TG} (f) is compared, starting from the most current estimate for the short-term power density spectrum of the noise component, the estimated value for further processing is always selected, its associated power First the comparison criterion power <power _threshold value (L _SW ) is fulfilled and its power is therefore significantly smaller than the power maximum of the input signal (L _XM ), and that in the event that no assigned power value meets the comparison criterion, the estimate for the short-term power density spectrum of the stationary ren noise component (P _GS (f)) is selected for further processing, and
that the time-delayed power of the input signal (L _x ), the power maximum of the input signal (L _XM ) and the estimated power of the stationary noise signal (L _GS ) are evaluated using a bypass function b (L), the estimated power of the stationary noise signal (L _GS ) is derived directly from the estimated short-term power density spectrum for the stationary noise component (P _GS (f)), and that

• a) at a signal level of the power of the input signal (L _X ), which is equal to or in the range of 3-6 dB above the level of the estimated stationary noise component (L _GS ), the maximum value of the bypass factor b _{max is} set and thus a maximum effect of the noise filter (GF) is achieved and that
• b) when the power of the input signal (L _X ) increases, the bypass factor (b) and thus the effect of the noise filter (GF) is reduced, the strength of the reduction being additionally determined by the size of the maximum power of the input signal (L _XM ) .

2. The method according to claim 1, characterized in that to determine the sliding power maximum (L _XM ), the power (L _X ) of the input signal (x) is determined continuously as a short-term square mean over a period of time that corresponds to the block length of a signal block. and that from the relationship L _XM.neu = max {L _x ; µL _{XM, old} }; µ <1 the sliding maximum power (L _XM ) is continuously determined. 3. The method according to claim 1, characterized in that the power threshold (L _SW ) according to the relationship L _SW = cL _XM ; c <1 is determined so that it is significantly below the performance of the current vocal language areas. 4. The method according to claim 1, characterized in that the time delay of the Power of the input signal over the time delay stage for the input signal (TX) is set so that it corresponds to the block length of a signal block, and that the previous signal block as well as the current and the future signal block for the evaluation of the search area for the transient Noise component (SBI) can be provided. 5. Arrangement for single-channel noise reduction for disturbed speech signals from a known stage for calculating the power and the Power maximum of the input signal (LM), a stage to the stationary  Noise Estimation (GSS), a known bypass control (BS) and one Noise filter (GF) with filter adaptation, characterized in that between the near participant's input (E) and the adaptive noise filter (GF) a time delay stage for the input signal (TX) is arranged that the output of the time delay stage for the input signal (TX) with a Input of the bypass control (BS), an input of the adaptive noise filter (GF) and an input of the filter adaptation of the adaptive noise filter (GF) is connected that the input (E) of the nearby participant additionally with a step is connected to the transient noise estimation (GSI), which on the one hand is direct and on the other hand via a time delay stage for the transient noise component (TG) is connected to a selection level (AS) that the selection level (AS) via a multiplier (MU) with the filter adaptation of the adaptive noise filter (GF) and that the multiplier (MU) is connected to the output of the bypass Control (BS), the output of the stationary noise estimation stage (GSS), the output of the transient noise estimation (GSI) stage and the output the time delay stage for the transient noise component (TG) is connected.