DE10356063B4 - Method for interference suppression of audio signals - Google Patents

Abstract

method for filtering of audio signals (1, 5.1) of a communication device, wherein the communication device Receives audio signals (1, 5.1) of different frequencies, which disturbing parts, preferably in the form of reverberation (3) and echo (2) are superimposed and the reverberation time (T60) is determined at which the audio signals (1, 5.1) divided into individual successive time blocks (= signal frames) and then transformed from the time domain to the frequency domain (6.1), suppressed and then transformed back (6.2) and to filtering the recorded and interference-prone Audio signals (1, 5.1) the interference components (2, 3) are removed by an algorithm where the sound energy of the audio signals (1, 5.1) the square of the signal amplitudes of the audio signals (1, 5.1), characterized in that the sound energy the disturbing Shares (2, 3) of a respective subsequent signal frame in dependence the reverberation time (T60) is determined and further a signal amplitude ratio RD (7.1) of disturbing Shares (2, 3) undisturbed Shares (1) is determined (7), whereby the sound energy of the disturbing shares (2, 3) with ...

Description

The The invention relates to a method for interference suppression of audio signals of a Communications equipment the communication device Receives audio signals of different frequencies, which disturbing parts, preferably in the form of reverberation and echo, are superimposed and the reverberation time is determined, in which the audio signals in each successive blocks of time (= Signal frames) and then from the time domain into the frequency domain transformed, suppressed and then transformed back become and to the suppression the recorded and interference-prone Audio signals the interference components be removed by an algorithm where the sound energy the audio signals the square of the signal amplitudes of the audio signals equivalent.

Such a known method is for example in the DE 694 20 027 T2 as well as in the technical article "Spectral Subtraction Based on Minimum Statistics", by R. Martin, EUSIPCO-94, Edinburgh, Scotland, September 13-16, 1994, pp. 1182-1185.

It is well known that over Speaker radiated or from a sound source, for example the human voice, generated audio signals in rooms in the Generally falsified by echoes and reverberations. The type and duration of reverberation is determined by the geometry of the room and the texture the walls decisively influenced. While the echoes from direct, simple and early reflections of the audio signal on one of the boundary surfaces of space, for example the walls, the floor or the ceiling, the reverberation is diffuse and becomes by multiple, only by statistic methods estimable reflections at the boundary surfaces generated.

For example in the communication of two mobile subscribers calling in hands-free mode, leads the Reverberation often to a significant deterioration of voice quality and speech intelligibility for the remote mobile subscribers.

remedy can here in addition to the use of directed microphones method for reverberation of the audio signal.

It There are already some algorithms and algorithms for rewriting Audio signals known from the literature. For example, the following Cited references, which are incorporated in full in this application become:

• [1] Louis D. Fielder: "Analysis of Traditional and Reconversion-Reducing Methods of Room Equalization "; AES Journal Vol. 51 Number 1/2; p. 3-26; Jan. 2003.
• [2] Yegnanarayana and Murthy: "Enhancement of Reverberant Speech Using LP Residual Signal "; IEEE Transactions on Speech and Audio Processing; Vol. 8, no. 3; may 2000th
• [3] Wu and Wang: "A One-Microphone Algorithm for Reverberant Speech Enhancement "; Proc. IEEE Int. Conference on acoustics, speech and signal processing; Hong Kong; April 6-10 2003
• [4] Gillespie, Malvar and Florencio: "Speech Dereverberation Via Maximum Kurtosis Subband Adaptive Filtering "; Proc. IEEE Int. Conference on Acoustics, Speech and Signal Processing; Salt Lake City, Utah, USA; May 2001.

Some On the one hand, these methods use statistical properties of Speech signals to reduce the reverberation. Some of the procedures of Digital voice and audio signal processing require one appraisal the so-called reverberation time or lead this under a Algorithm through.

For example represents the algorithm described by Wu and Wang [3] a general Approach, which examines the influence of energy in a certain frequency band within a speech signal section (Frame) on the energy in the following frames in each one Frequency band has.

A Another group of methods are noise suppression based on spectral Subtraction work. These methods capture and estimate via a longer Period the spectral structure of noise and then reduce in the frequency domain the audio signal so that it is exactly the shares of the interference signal is reduced.

Of the Disadvantage of the methods that the statistical properties of Using voice signals to reduce reverberation is that the Reduction of reverberation, depending from the statistical fluctuations, is inaccurate. Also the noise reduction methods, which still work on the basis of spectral subtraction a certain residual disturbance on.

It It is therefore an object of the invention to provide an improved method for suppression of audio signals, which compared to the previous ones known method allows a more accurate interference suppression of the audio signals.

These The object of the invention is achieved by a method having the features of claim 1. Advantageous developments of the invention are subordinate to the subject Claims.

The Inventors have recognized that disturbing parts, like reverberation or echo, significantly reduced from an audio signal can be if a medium drop in energy or sound energy of the audio signal the time is determined and the audio signal to be included, according to this thus determined reverberation curve, in certain areas in terms the sound energy is reduced.

Corresponding In accordance with this inventive concept, the inventors propose the known one Method for filtering of audio signals of a communication device, wherein the communication device audio signals different frequencies, which disturbing shares, preferably in the form of reverberation and echo, superimposed and the reverberation time is determined at which the audio signals divided into individual consecutive time blocks (= signal frames) and then transformed from the time domain to the frequency domain, suppressed and then transformed back and to suppress the recorded and noisy Audio signals the interference components be removed by an algorithm where the sound energy the audio signals correspond to the square of the signal amplitudes of the audio signals, to improve the sound energy or the signal amplitude the disturbing Shares of a respective subsequent signal frame in dependence the reverberation time is determined and further a signal amplitude or signal energy ratio RD of disturbing Shares to undisturbed Shares is determined, the signal amplitude or the sound energy the disturbing Shares are weighted with RD and the so determined and weighted interference components be removed from the subsequent signal frame.

hereby let yourself the noisy one To reduce the audio signal by the portions that are not from the interference-free Direct sound, but come from the reverberation. The reduction of the reverb content takes place in blocks in the desired Frequency range, similar like the reduction of background noise in the known noise reduction methods. When noise reduction is Reference is made to the following literature, the entire contents of which are incorporated herein by reference accepted this application will: Rainer Martin: "Spectral Subtraction Based on Minimum Statistics "in Signal Processing VII: Theories and Applications M. Holt, C. Cowan, P. Grant, W. Sandham (Eds.), (c) 1994 European Association for Signal Processing) The procedure is especially good for Communication terminals, like for example Radios, mobile devices or Landline telephones. But also for Stand-alone devices that Audio signals via output a speaker and also audio signals via a Can record a microphone, the method is suitable, which is determined by the ambient conditions of the Stand-alone device resulting interference components to suppress.

The Reverberation time can be determined frequency-dependent in the new process. Analogous For this purpose, the signal amplitude ratio RD can be determined as a function of frequency become.

The frequency-dependent Determination of reverberation time and / or signal amplitude ratio allows a improved determination of the noise components, for example, the reverberation, especially when the reverberation times of the room for different frequencies vary greatly. This then leads to an improved noise reduction and thus to a higher one quality of the disturbed one Output signal.

Of the Algorithm for filtering the recorded and interference-prone Audio signals is used, can spectrally subtract the noise components. Thus, you can Energy shares occur by subtracting negative values and set to zero by the algorithm. alternative can The algorithm initiates the audio signals in a frequency-dependent manner multiplied by a factor. This factor is in the range between zero and one. The factor is close to zero, if in the associated frequency range of the audio signal the noise component a similar one high or even higher Energy possesses as the audio signal. The factor is one, if in the associated frequency range almost or no noise component is available.

in the new method can be used to transform the audio signals from the time domain in the frequency domain a Fourier transform or a discrete Cosine Transformation (DCT) or a Modulated Complex Lapped Transformation (MCLT) can be used. Especially for the MCLT will open following literature: Henrique S. Malvar: "A modulated complex lapped transform and its application to audio processing ", Proc. ICASSP, pp. 1421-1424, 1999.

It is cheap, when the audio signals in a magnitude portion and a phase portion be split. From the amount portion of the audio signal can be relative simply the signal energy can be determined, which accordingly a frequency-dependent energy attenuation is enabled. The phase content remains unchanged.

The time course of the sound energy of the interference signal E (f, n) can be modeled according to the following formula: E (f, n) = E (f, n-1) e ^{-A (f)} + S (f, n), where E (f, n-1) is the frequency-dependent energy of the signal frame having the index n-1 at a certain time, and A (f) is a factor determined by the reverberation time and S (f, n) is the frequency-dependent energy of the noise suppressed Signal frames with the index n corresponds.

Farther is it cheap when the reverberation time from the energetic fall of the reverberation is calculated, the information about the speed of this Waste is included in the room impulse response. The room impulse response is for example a dirac-like audio signal (= peak-like and narrowly limited impulse), the sound pressure curve over time in a certain space. This can be a more accurate Determination of the reverberation time compared to the statistically based estimation of the Reverberation time.

The Method for interference suppression, preferably for reverberation reduction, can be particularly favorable by a level to be determined RD (f) are controlled. The level ratio RD (f) gives the ratio between disturbing Proportion of audio signal and undisturbed audio signal again. It is advantageous, this level ratio RD (f) by a suitable Algorithm, which information from the amount or the energy of the suppressed Signal frames used to adapt. In the case of spectral subtraction For example, the value RD (f) is increased as long as no or less than 50 percent, preferably less than 25 percent, of the energy or Amount values of the suppressed Signal frames have negative values. This will ensure that not too much damping of the audio signal.

Vice versa should the level ratio RD (f) are lowered as soon as negative values for the amount the disturbed Signal frames occur. This will ensure that a sufficient damping the disturbance parts of the audio signal.

In the following the invention will be described with reference to preferred embodiments with the aid of 1 to 4 The following abbreviations are used in the figures: 1 : Direct sound of a pulse-like audio signal; 2 : Echoes; 3 : Diffuse reverberation / reverberation curve; 4 : Shift of the reverberation curve; 5.1 : Time-dependent input signal; 5.2 : Amount portion of the input signal that has been transformed into the frequency domain; 5.2 ' : Suppressed amount portion of the input signal; 5.3 : Phase component of the input signal transformed into the frequency domain; 5.4 : Time-dependent (interference-suppressed) output signal; 6.1 : Fast Fourier Transformation; 6.2 : Inverse Fast Fourier Transformation; 7 : Determination and adjustment of the level ratio RD (f) of the echoes to the direct sound; 7.1 : Specific and adjusted level ratio RD (f); 8th : Squaring; 8.1 : Adding the energy value; 9 : Sound energy in the room; 10 : Calculation of the energetic decrease of the sound energy; 11 : Rooting of the sound energy; 11.1 : Adjusted amplitude value = root value of the adjusted sound energy; 12 : Multiplication; 12.1 : Noise signal estimation; 13 : Subtraction; A60: energy range in which the energy drops by 60 decibels; T60: reverberation time.

1 : Temporal energy distribution of an originally pulsed audio signal in a room;

2 : Temporal energy distribution of the same audio signal 1 , wherein additionally the displacement of the reverberation curve is shown as a function of the level ratio between direct sound and reflections;

3 : Block diagram which explains the method for reverberating audio signals;

4 : Block diagram, which explains a further implementation of the new method for reverberation of audio signals.

The 1 shows by way of example in a diagram the temporal energy distribution of an audio signal which can be recorded and measured in a room with a microphone when a pulse-like audio signal, for example a bang, occurs. This energy distribution is obtained by forming the absolute square of a spatial impulse response and logarithmizing it.

On the ordinate of the diagram are the logarithmized signal amplitudes, as a measure of the (sound) energy of the audio signal, and on the abscissa the time, each in arbitrary Units, applied.

You can see the so-called direct sound as the first line on the far left of the diagram 1 , which passes directly from the sound source to the measuring microphone. Then follow in this example energetically weaker echoes 2 that have been reflected once or twice on walls of the room before they reach the measuring microphone.

Finally, followed by a triangle, the diffuse reverberation follows 3 , which consists of quasi-chaotic superimposed reflections of the audio signal in the room and diffusely arrives at the measuring microphone.

The reverberation time, usually designated T60, which represents a characteristic in acoustics size, is the energy distribution of the audio signal in the 1 refer to.

To determine the reverberation time, see 1 determined on the ordinate of the diagram, an energy range A60, which is in the range of diffuse reverberation 3 in which the energy of the reverberation drops by 60 decibels. The upper and lower limits of the energy range A60 are first applied to the reverberation curve 3 and then projected onto the abscissa, ie the time axis. The projected area on the time axis corresponds to the reverberation time T60.

In 2 Again, the temporal energy distribution of a pulse-like audio signal, for example a bang, is shown as an example. Analogous to 1 On the ordinate of the diagram are the logarithmized signal amplitudes, as a measure of the (sound) energy of the audio signal, and on the abscissa the time, in arbitrary units, is plotted. Accordingly, in the diagram of 2 the direct sound 1 and the echoes 2 shown.

In the diagram of 2 It now becomes clear how, depending on the specific situation, for example, depending on the space and the distance between the sound source and microphone, the level ratio or signal amplitude ratio of direct sound 1 to reflections 2 and reverberation 3 changed. The reflections tend to be in only slightly reverberant rooms and with a small distance between the sound source and the microphone 2 and the reverberation 3 a low level or a low amplitude compared to the direct sound 1 so that the reverberation curve is more like the white dashed line ( 2 ) runs. Conversely, in strongly reverberant rooms and long distances between the sound source and the microphone reflections 2 and the reverberation 3 be stronger, so that the reverberation curve rather corresponds to a course indicated by the black dashed line.

This frequency-dependent level ratio or signal amplitude ratio of the direct sound 1 to the reverberation 3 , usually abbreviated to RD (f) (Reverberation to Direct Sound Ratio), affects the reverberation curve 3 crucial. This fact is exploited in the new method of reverberating an audio signal.

The 3 shows a block diagram, which explains the new method for reverberation of audio signals. A time-dependent input signal 5.1 is split in blocks and in the block 6.1 transformed to Fourier. That is, there is a transformation of the signal blocks or signal frames called from the time domain in the frequency domain instead. The input signal 5.1 is in a frequency-dependent amount portion 5.2 and a frequency-dependent phase component 5.3 split. The transformed frequency-dependent phase component 5.3 of the input signal 5.1 stays unchanged. Only on the transformed frequency-dependent amount portion 5.2 the input signal 5.1 the new suppression method is used, which will be explained below. The suppressed amount share 5.2 ' and the unchanged phase proportion 5.3 are merged again and in the block 6.2 subjected to an inverse Fourier transform to obtain a back transformation from the frequency domain to the time domain. The thus interference-suppressed signal, ie at least substantially reduced signal from echo and reverberation, again becomes a time-dependent (interference-suppressed) output signal 5.4 output.

In the new suppression method, the interference components of the signals are reduced as a function of the frequency-dependent energy. The new interference suppression method is based on the frequency-dependent amount as follows 5.2 applied: In the block with the reference numeral 8th is the suppressed amount share 5.2 ' squared to get a conversion of the amplitude signals into energy signals.

The after squaring 8th obtained energy value is - by the reference numeral 8.1 expressed - to the sound energy 9 added.

In the block 10 becomes the frequency-dependent sound energy 9 according to the formula: E (f, n) = E (f, n-1) e ^{-A (f)} where n is the index of the signal frame and f is the frequency index and the factor A (f) is inversely proportional to the reverberation time and thus includes the temporal decline of energy.

In the block 11 the adjusted energy value thus obtained is converted into an adjusted amplitude value by root extraction 11.1 transformed.

In the block 7 is from the suppressed amount share 5.2 ' of the input signal 5.1 the level or amplitude ratio RD (f) 7.1 determined and adapted between echo and direct sound.

The level or amplitude ratio RD (f) 7.1 and the adjusted amplitude value 11.1 be in the block 12 multiplied by each other. The signal frame calculated in this way 12.1 corresponds to the estimated noise component. This piece of noise is finally in the block 13 from the frequency dependent amount share 5.2 of the input signal 5.1 subtracted. As a result, the interference components, such as reverberation and echo, largely from the Einganssignals 5.1 removed and thus achieved a filtering.

Join after the subtraction 13 negative values for the amount share 5.2 the signal frames on, that is the algorithm subtracts too much energy, the factor RD (f) 7.1 decreases, otherwise a slow increase is performed.

The block diagram in the 4 describes another way of implementing the new method for reverberating audio signals. In contrast to 3 is already after the split of the input signal 5.1 in the amount share 5.2 and the phase proportion 5.3 a squaring 8th the amount share 5.2 of the input signal. So it will be the signal amplitudes of the amount share 5.2 converted into sound energies.

The removal of the noisy audio signal parts is done under the blocks 7 . 9 . 10 . 12 and 13 analogous, as in 3 described.

Before merging the suppressed amount share 5.2 ' and the unchanged phase component 5.3 is through rooting 11 of the suppressed amount share 5.2 ' a re-conversion of "sound energy" in "signal amplitudes" performed. Thereafter, the merged signal amplitudes of the suppressed amount portion 5.2 ' and the unchanged phase component 5.3 in the block 6.2 subjected to an inverse Fourier transform, to then obtain a back transformation from the frequency domain to the time domain. The thus interference-suppressed signal, ie at least substantially reduced signal from echo and reverberation, again becomes a time-dependent (interference-suppressed) output signal 5.4 output.

All in all the audio signal will be much more accurate due to the new procedure as previously suppressed.

It it is understood that the above features of the invention not only in the specified combination, but also in others Combinations or alone, without the frame to leave the invention.

Claims (11)

Method for interference suppression of audio signals ( 1 . 5.1 ) of a communication device, wherein the communication device audio signals ( 1 . 5.1 ) of different frequencies, which disturbing shares, preferably in the form of reverberation ( 3 ) and echo ( 2 ) and the reverberation time (T60) is determined at which the audio signals ( 1 . 5.1 ) are divided into individual successive time blocks (= signal frames) and then transformed from the time domain into the frequency domain ( 6.1 ), and then back-transformed ( 6.2 ) and for interference suppression of recorded and noisy audio signals ( 1 . 5.1 ) the interference components ( 2 . 3 ) are removed by an algorithm whereby the sound energy of the audio signals ( 1 . 5.1 ) the square of the signal amplitudes of the audio signals ( 1 . 5.1 ), characterized in that the sound energy of the disturbing components ( 2 . 3 ) of a respective subsequent signal frame as a function of the reverberation time (T60) is determined, and further a signal amplitude ratio RD ( 7.1 ) of disturbing shares ( 2 . 3 ) to undisturbed shares ( 1 ) is determined ( 7 ), wherein the sound energy of the disturbing components ( 2 . 3 ) with RD ( 7.1 weighted ( 12 ) and the thus determined and weighted noise components ( 12.1 ) are removed from the subsequent signal frame ( 13 ). Method according to the preceding Claim 1, characterized in that the reverberation time (T60) frequency dependent is determined. Method according to one of the preceding claims 1 and 2, characterized in that the signal amplitude ratio RD ( 7 ) is determined frequency dependent. Method according to one of the preceding claims 1 to 3, characterized in that the algorithm spectrally subtracts the interference components ( 13 ). Method according to one of the preceding claims 1 to 4, characterized in that the algorithm the audio signals ( 1 . 5.1 ) Frequency-dependent multiplied by a factor. Method according to one of the preceding claims 1 to 5, characterized in that for the transformation of the audio signals ( 1 . 5.1 ) from the time domain into the frequency domain a Fourier transformation ( 6.1 ) or a Discrete Cosine Transformation (DCT) or a Modulated Complex Lapped Transformation (MCLT). Method according to one of the preceding claims 1 to 6, characterized in that the audio signals ( 1 . 5.1 ) in an amount ( 5.2 ) and a phase portion ( 5.3 ). Method according to one of the preceding claims 1 to 7, characterized in that the time profile of the sound energy E (f, n) of the interference signal is modeled (10) according to the following formula: E (f, n) = E (f, n-1 ) e ^{-A (f)} + S (f, n), where E (f, n-1) is the frequency dependent energy of the signal frame with the index n-1 at a given time and A (f) is a factor that is determined by the reverberation time (T60) is determined and S (f, n) corresponds to the frequency-dependent energy of the suppressed signal frame with the index n. Method according to one of the preceding claims 1 to 8, characterized in that the reverberation time (T60) from the energetic waste (A60) of the reverberation (A60) 3 ) of a room impulse response is calculated. A method according to claim 7, characterized in that the signal amplitude ratio RD ( 7.1 ), as long as no or less than 50 percent, preferably less than 25 percent, of the magnitude values of the suppressed signal frames ( 5.2 ' ) have negative values. Method according to claim 7, characterized in that the signal amplitude ratio RD ( 7.1 ) is lowered as soon as negative values for the amount of suppressed signal frames ( 5.2 ' ) occur.