EP1211671A2 - Automatic gain control with noise suppression - Google Patents

Abstract

Audio processing apparatus (40) includes a a noisesuppression stage (42), which applids a variable level ofnoise suppression to an input audio signal, so as togenerate a noise-suppressed signal, and an automatic gaincontrol (AGC) stage (44), coupled to apply a variablegain to the noise-suppressed signal, responsive to alevel of the signal. A noise controller (46) receives anindication of the gain from the AGC stage and determinesthe level of noise suppression to be applied by the noisesuppression stage responsive to the gain.

Description

CROSS-REFERENCE TO RELATED APPLICATION

This application claims the benefit of U.S.Provisional Patent Application No. 60/249,388, filedNovember 16, 2000, which is incorporated herein byreference.

FIELD OF THE INVENTION

The present invention relates generally toprocessing of audio signals, and specifically to noisereduction and automatic gain control in processing ofsuch signals.

BACKGROUND OF THE INVENTION

Automatic gain control (AGC) is used in voicecommunications to compensate for differences in signallevel. Such difference may arise, for example, inspeakerphone applications due to the differences indistance between the microphone and several speakersparticipating in a teleconference. Assuming the AGC isworking perfectly, the output level of the audioprocessing circuits should remain constant even for largevariations in the input signal level received by themicrophone. Unfortunately, microphones used in realenvironments pick up background noises. Since the levelof background noise remains roughly constant, while thelevel of the signal varies depending on the distance ofthe speaker from the microphone, the signal-to-noise(s/N) level varies accordingly. When the signal isamplified or reduced by AGC to compensate for variationsin the signal level, the noise level in the output signalis affected accordingly. This variable amplification ofthe noise level leads to an annoying effect known as"noise modulation."

Fig. 1 is a plot that schematically illustratessignals received by a microphone, representing the voicesof two speakers. The first speaker (who speaks duringintervals marked "A" in the figure) is about 50 cm fromto the microphone, while the second speaker (speakingduring intervals marked "B") is about 2 m from themicrophone. Since the sound pressure level is inverselyproportional to the distance from the microphone, theinput level of an audio signal 20 received during the Aintervals is about four times (12 dB) greater than asignal 22 during the B intervals. Abackground noiselevel 24 remains roughly constant.

Fig. 2 is a plot that schematically illustrates theresult of applying AGC to the signals of Fig. 1. The AGCcauses anoutput signal 32 during the B intervals to havea level that is roughly equal to that of anoutput signal30 during the A intervals. Anoise level 34 during the Aintervals remains reasonably low. Strong amplificationof the weak signal in the B intervals, however, causescorresponding amplification of anoise level 36 duringthese intervals. As a result, while the signals fromboth speakers are heard at approximately the same outputsignal level, the noise level has sharp and noticeablevariations.

Digital noise suppression techniques can be used toreduce the background noise level before AGCamplification of the signal. (Noise suppression mustprecede AGC, since if the order of operation is reversed,variations in the AGC gain will confuse the noisesuppressor's estimate of the noise level.) Common noisesuppression techniques typically involve determining thenoise spectrum and filtering the signal based on thisspectrum in order to remove the noise components insofaras possible. Such techniques are commonly referred to as methods of "spectral attenuation" or "spectralsubtraction." They are described, for example, by Bollin an article entitled "Suppression of Acoustic Noise inSpeech Using Spectral Subtraction," published inIEEETransactions on Acoustics,Speech and Signal Processing,ASSP-27, No. 2 (April, 1979), which is incorporated hereinby reference.

A variety of methods of noise suppression aredescribed in the patent literature. For example, U.S.Patent 4,185,168, to Graupe et al., whose disclosure isincorporated herein by reference, describes a system foradaptively filtering near-stationary noise from aninformation bearing signal. An input signal containinginformation as well as near-stationary noise is appliedto a noise-analysis circuit and simultaneously to anoise-reduction circuit, each of which circuits comprisesa plurality of bandpass filters. The background noisepower is estimated by measuring an average of successiveminima in each of the filters during times whensubstantially only noise is present. Several methods aredescribed for determining the gain of each filter,responsive to the measured successive averaged minima andthe size of the signal.

U.S. Patent 5,550,924, to Helf et al., whosedisclosure is incorporated herein by reference, describesa method for reducing background noise in order toenhance speech. Properties of human audio perception areused to perform spectral and time masking to reduceperceived loudness of noise added to the speech signal. Asignal is divided temporally into blocks which are thenpassed through a plurality of filters to remove narrowfrequency band components of the noise. An estimate ofthe noise level in each of the filters is made byaveraging measured noise powers. A FFT (Fast Fourier Transform) is performed on the blocks to determine theaverage noise power. Responsive to the determined noisepower, a noise-reduced signal is recovered using aninverse FFT.

U.S. Patent 5,768,473, to Eatwell et al., whosedisclosure is incorporated herein by reference, describesan adaptive speech filter. The filter is a modifiedversion of that described in U.S. Patent 4,185,168, usinga noise power estimate of an average of the power. Thefilter implements an improved adaptive spectral estimatorfor estimating the spectral components in a signalcontaining both an information signal, such as speech,and noise. Improvements over 4,185,168 relate to a noisepower estimator and a computationally-efficient gaincalculation method. The adaptive spectral estimator issaid to be particularly suited to implementation usingdigital signal processing and can be used to provideimproved spectral estimates of the information signal.

Generally, the amount of noise suppressed by a givennoise suppressor is adjustable over a certain range.Suppression technologies known in the art can typicallyprovide up to 8-10 dB of noise suppression with asignificant improvement in sound quality. When noisesuppression is increased above this level, however,noticeable distortion may be introduced in the speechsignals. Therefore, in noisy environments, finding theoptimal level of noise suppression involves trading offbackground noise against speech distortion. Referringback to the example of Figs. 1 and 2, it will be seenthat if sufficient noise suppression is applied in orderto eliminate the annoying noise modulation effect in theB intervals, the result will likely be undesireddistortion in the audio signals in both the A and Bintervals. On the other hand, if only mild noise suppression is applied as indicated by the A intervalsignals, noticeable noise modulation will remain,

SUMMARY OF THE INVENTION

It is an object of the present invention to provideimproved methods and devices for processing of audiosignals in the presence of amplitude variations andnoise.

It is a further object of some aspects of thepresent invention to provide methods and devices foraudio signal processing that reduce or eliminate noisemodulation without introducing excessive distortion.

In preferred embodiments of the present invention,an audio processor comprises a noise suppression stageand an AGC stage. The amount of noise suppressed isadjusted continually according to the current AGC gain.Thus, if greater signal amplification is necessary tocompensate for a drop in the signal level, more noise issuppressed compensate for residual noise amplification inthe output signal from the audio processor. On the otherhand, when the signal level increases, the noisesuppression is reduced in order to eliminate possibledistortion. The audio processor can thus be adjusted togive optimal audio quality, balancing noise modulationagainst signal distortion, over a range of differentsignal levels.

There is therefore provided, in accordance with apreferred embodiment of the present invention, audioprocessing apparatus, including:

a noise suppression stage, adapted to apply avariable level of noise suppression to an input audiosignal, so as to generate a noise-suppressed signal;

an automatic gain control (AGC) stage, coupled todetermine a variable gain responsive to a level of thenoise-suppressed signal, and to apply the gain to the noise-suppressed signal so as to generate an amplifiedoutput signal; and

a noise controller, coupled to receive an indicationof the gain from the AGC stage and to determine the levelof noise suppression to be applied by the noisesuppression stage responsive to the gain.

Preferably, the noise suppression stage is adaptedto apply spectral compression to the input audio signal.

Additionally or alternatively, the noise controlleris adapted to determine the level of noise suppression asa monotonically-increasing function of the gain.Preferably, the level of noise suppression determined bythe noise controller increases in proportion to a powerof the gain, wherein the power is less than or equal toone. Most preferably, the level of noise suppressionL(t) is given substantially by an expression of the formL(t) =L_B + (G(t)^x, whereinG(t) is the gain,L_B is anadditive factor, andx is a number less than or equal toone.

In a preferred embodiment, the AGC stage is adaptedto increase and decrease the gain in alternation inresponse to alternations in the level of thenoise-suppressed signal due to receiving the input audiosignal from alternating weak and strong audio sources,respectively, and the noise Controller is adapted todecrease and increase the level of noise suppression,responsive respectively to the gain increasing anddecreasing.

There is also provided, in accordance with apreferred embodiment of the present invention, a methodfor audio processing, including:

suppressing noise in applying an input audio signalusing a variable level of noise suppression, so as togenerate a noise-suppressed signal;

determining a variable gain responsive to a level ofthe noise-suppressed signal;

applying the gain to the noise-suppressed signal soas to generate an amplified output signal; and

determining the level of noise suppression to beapplied to the input audio signal responsive to the gain.

The present invention will be more fully understoodfrom the following detailed description of the preferredembodiments thereof, taken together with the drawings inwhich:

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 is a plot that schematically illustratessignals received by a microphone;

Fig. 2 is a plot that schematically illustrates thesignals of Fig. 1 following AGC amplification, as isknown in the art;

Fig. 3 is a block diagram that schematicallyillustrates an audio processor, in accordance with apreferred embodiment of the present invention; and

Fig. 4 is a plot that schematically illustratessignals output by the audio processor of Fig. 3, inaccordance with a preferred embodiment of the presentinvention.

DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS

Fig. 3 is a block diagram that schematicallyillustrates anaudio processor 40, in accordance with apreferred embodiment of the present invention. The audioprocessor receives a raw audio input signalX(t), from amicrophone, for example (not shown), and outputs aprocessed audio signalY(t). Audio processor comprises anoise suppression stage 42, followed by anAGC stage 44.Preferably,X(t) andY(t) are analog signals, and stages42 and 44 are implemented using suitable analog circuitelements, such as tunable filters and variable-gainamplifiers, as are known in the art. Alternatively,X(t)andY(t) are digitized, and the processing functionsdescribed hereinbelow are implemented using digital logiccircuits. Mixed analog and digital implementations mayalso be used.

Preferably,noise suppression stage 42 implements amethod for suppressing near-stationary noises and tonesdescribed in U.S. Patent Application 09/605,174, filedJune 28, 2000, which is assigned to the assignee of thepresent patent application and whose disclosure isincorporated herein by reference. According to thismethod, the input noise is divided into multiplefrequency bands, and the maximum and minimum noise levelsin each band are determined over a period of time. Basedon these noise levels, a gain is computed in each bandusing spectral subtraction and/or spectral compression.Preferably, given an effective level of the signal a / n,and a difference between the upper and lower noiseestimates Δa /n, the noise suppression gainG_NS is given bythe following formulas:If Âan ≤Δan,then GNS =Gmin.If Âan >Δan andÂan < (S + 1)Δan,then GNS =Gmin + (1 -Gmin)Âan -ΔanS·Δan.If Âan ≥ (S + 1) · Δan,then GNS = 1,whereinG_min is a minimum value of the gainG_NS.

The gainsG_NS are applied bynoise suppression stage42 to the respective frequency bands of the input signalX(t) to generate a noise-suppressed input toAGC stage44. Alternatively, the noise suppression stage mayemploy other techniques, such as those described in theBackground of the Invention, or substantially any othersuitable noise suppression method known in the art.

As can be seen in Fig. 3,AGC stage 44 operates onthe audio signals after processing bynoise suppressionstage 42. The AGC stage determines a variable gainG_AGC(t) to be applied to the signals in order tocompensate for variations in the input signal level, Thecurrent value ofG_AGC(t) is provided to anoise controlblock 46. Based on this value, the noise control blockcomputes the amount of noise suppressionL(t) to beapplied bynoise suppression stage 42 to the input signalX(t). Preferably, the values of parameters used innoisesuppression stage 42, such asG_min, are continuallyadjusted so that the total amount of noise suppression isequal to the current value ofL(t). Althoughblock 46 is shown in the figure as a separate entity for the sakeof clarity of explanation, those skilled in the art willappreciate that the function of this block mayalternatively be integrated into eitherstages 42 orstage 44.

Preferably,L(t) is determined based on the currentAGC gainG_AGC(t) and on a basic noise suppression levelL_B, which corresponds to the amount of noise suppressedin the output signal when AGC gain is equal to unity.Various functions may be used to relateL(t) toG_AGC(t)andL_B. Preferably,L(t) increases monotonicallyrelative to bothG_AGC(t) andL_B . For example, thefollowing function provides noise suppression with fullcompensation for changes in the AGC gain:L(t)= LB +GAGC(t)(The noise suppression levels and AGC gain are specifiedhere in decibels.) It is seen that using equation (4),the noise is first suppressed byL(t) decibels and thanexpanded byG_AGC(t) decibels. Thus, the noise is alwayssuppressed by the original amount of L_B decibels.

The function of equation (4) may not be optimal,however, when large variations in the input signal levelcan occur, as it may lead to excessive noise suppression,with noticeable distortions in the output signalY(t).For example, if the basic noise suppression levelL_B is 5dB and AGC gain is 15 dB, then the total amount of noisesuppression will be 20 dB. Under such conditions, a milder dependence between the AGC gainG_AGC(t) and noisesuppression levelL(t) is preferable, such as adependence ofL(t) on a fractional power of the gain(G_AGC(t))^x, withx < 1. For example, the following functionprovides a good compromise between modulation of thenoise level in the output signal and the output speechquality:L(t)= LB +GAGC(t)Greater or smaller fractional powers ofG_AGC(t) may alsobe used. Alternative functions will be apparent to thoseskilled in the art.

Fig. 4 is a plot that schematically shows the outputsignalY(t) obtained by operating on the input signalX(t) shown in Fig. 1 usingaudio processor 40, inaccordance with a preferred embodiment of the presentinvention. The variable noise suppressionL(t) is givenby equation (5).Signals 50 and 52 during intervals Aand B, respectively, are amplified byAGC stage 44 togive comparable levels of perceptual loudness.Respective noise levels 54 and 56 are suppressed duringboth intervals A and B, as well. The level of noisesuppression during the B intervals is greater than thatduring the A intervals, but due to the square root factorin equation (5), there is still slightly more residualnoise in the B intervals. The parameters governing thedependence ofL(t) onG(t) are preferably chosen andadjusted based on the background noise and signalconditions so as to balance the residual noise modulationagainst distortion effects due to the noise suppression, in a way that gives the most pleasing perceived soundquality.

It will be appreciated that the preferredembodiments described above are cited by way of example,and that the present invention is not limited to what hasbeen particularly shown and described hereinabove.Rather, the scope of the present invention includes bothcombinations and subcombinations of the various featuresdescribed hereinabove, as well as variations andmodifications thereof which would occur to personsskilled in the art upon reading the foregoing descriptionand which are not disclosed in the prior art.

Claims (10)

1. Audio processing apparatus, comprising:

   a noise suppression stage, adapted to apply avariable level of noise suppression to an input audiosignal, so as to generate a noise-suppressed signal;

   an automatic gain control (AGC) stage, coupled todetermine a variable gain responsive to a level of thenoise-suppressed signal, and to apply the gain to thenoise-suppressed signal so as to generate an amplifiedoutput signal; and

   a noise controller, coupled to receive an indicationof the gain from the AGC stage and to determine the levelof noise suppression to be applied by the noisesuppression stage responsive to the gain.
2. Apparatus according to claim 1, wherein the noisecontroller is adapted to determine the level of noisesuppression as a monotonically-increasing function of thegain.
3. Apparatus according to claim 2, wherein the level ofnoise suppression determined by the noise controllerincreases in proportion to a power of the gain, whereinthe power is less than or equal to one.
4. Apparatus according to claim 3, wherein the level ofnoise suppressionL(t) is given substantially by anexpression of the formL(t) =L_B + (G(t))^x, whereinG(t) isthe gain,L_B is an additive factor, andx is a numberless than or equal to one.
5. Apparatus according to claim 1, wherein the AGCstage is adapted to increase and decrease the gain inalternation in response to alternations in the level ofthe noise-suppressed signal due to receiving the input audio signal from alternating weak and strong audiosources, respectively, and wherein the noise controlleris adapted to decrease and increase the level of noisesuppression, responsive respectively to the gainincreasing and decreasing.
6. A method for audio processing, comprising:

   suppressing noise in applying an input audio signalusing a variable level of noise suppression, so as togenerate a noise-suppressed signal;

   determining a variable gain responsive to a level ofthe noise-suppressed signal;

   applying the gain to the noise-suppressed signal soas to generate an amplified output signal; and

   determining the level of noise suppression to beapplied to the input audio signal responsive to the gain.
7. A method according to claim 6, wherein the level ofnoise suppression is determined as amonotonically-increasing function of the gain.
8. A method according to claim 7, wherein the level ofnoise suppression is proportional to a power of the gain,wherein the power is less than or equal to one.
9. A method according to claim 8, wherein the level ofnoise suppressionL(t) is given substantially by anexpression of the formL(t) =L_B + (G(t))^x, whereinG(t) isthe gain,L_B is an additive factor, andx is a numberless than or equal to one.
10. A method according to claim 6, wherein determiningthe variable gain comprises increasing and decreasing thegain in alternation in response to alternations in thelevel of the noise-suppressed signal due to receiving theinput audio signal from alternating weak and strong audio sources, and wherein determining the level of noisesuppression comprises decreasing and increasing the levelof noise suppression responsive to the increasing anddecreasing of the gain.

Images