FIELD OF THE INVENTION

The present invention relates to audio noise reduction in conference equipment. More specifically, the invention relates to a method and a device for reducing keyboard noise in a conferencing equipment which includes a microphone and a keyboard, such as a videoconferencing endpoint, in particular a desktop videoconferencing endpoint, or a conference telephone.

BACKGROUND OF THE INVENTION

Many communication appliances as well as recording equipment have keyboards for control in the same physical enclosure as a microphone capturing the sound. Examples of such appliances are desktop videoconferencing equipment, audio conference phones, mobile telephones, other types of telephones, MP3 recorders, tape recorders or similar.

When pressing a key on the keyboard, an acoustic sound (keyboard noise) is created. Normally, this sound is unwanted. In the local room, the keyboard noise is usually of such low level that it doesn't disturb the user very much, but when the audio is captured at a microphone, for communication or archiving, keyboard noise may be disturbing. Noise from the keyboard may transfer to the microphone both as sound propagating through the air, and as sound propagating through the physical structure of an equipment enclosure. Keyboard noise usually lasts for a very short time, but are often picked up relatively strong at the microphone, and it is usually broadband noise.

Such keyboard noise has previously been handled in various ways. The simplest one is accepting the noise. Another previous approach is known as masking, wherein a masking tone is added at the same time as the occurrence of the keyboard noise Such masking may make the keyboard noise inaudible or neglectable, or it may make the user focusing on something else, or at least make the user aware that something is happening and therefore more accepting for the audible noise. Another suggested approach is muting the entire audio signal (including both the keyboard noise and the wanted signal) when a key is pressed on the keyboard. Spectral subtraction is a widely used approach for removal of relatively low level, stationary broadband noise. Based on an assumption that noise is stationary, it calculates an estimate of the noise, from which it defines a linear time invariant filter, which is applied to the wanted signal including unwanted noise. The result is an output signal with a magnitude spectrum which is fairly equal to the magnitude spectrum of the wanted signal, but with a phase equalling the wanted signal plus unwanted noise phase. As long as the noise level is moderate, the phase error is small, yielding a well sounding result. When noise level increases, artefacts become more and more audible and annoying. People skilled in the art will consider spectral subtraction useful for stationary noise, not for transient noise such as keyboard noise.

SUMMARY OF THE INVENTION

A basic object of the invention is to provide a method and a device for reducing keyboard noise in a conferencing equipment.

The invention provides a method and a device as set forth in the appended claims.

BRIEF DESCRIPTION OF THE DRAWINGS In order to make the invention more readily understandable, the discussion that follows will refer to the accompanying drawings, wherein

Fig. 1 is a schematic block diagram illustrating basic principles of a device for reducing keyboard noise in a conferencing equipment,

Fig. 2 is a schematic block diagram illustrating further possible principles of a device for reducing keyboard noise in a conferencing equipment,

Fig. 3 is a schematic flow chart illustrating the principles of a method for reducing keyboard noise in a conferencing equipment.

DETAILED DESCRIPTION OF THE INVENTION

In the following, the present invention will be discussed by describing various embodiments, and by referring to the accompanying drawings. However, people skilled in the art will realize other applications and modifications within the scope of the invention as defined in the enclosed independent claims.

Fig. 1 is a schematic block diagram illustrating basic principles of a device for reducing keyboard noise in a conferencing equipment. The conference equipment includes a microphone 110 and a keyboard 160. The microphone and keyboard 160 are arranged in such a way that the microphone picks up both wanted audio, typically speech from a speaker (a conference participant), and unwanted audio noise, which may be generated by manual operation of the keyboard 160. In addition, the microphone may pick up background noise from the environment, in particular the conference room.

The microphone 110 is connected to an amplifier 120, which may also advantageously include a low pass filter. The amplifier is further connected to the quantizer/digitizer 130, which includes circuits for sampling and analog-digital conversion, resulting in a digital audio signal at the output of element 130.

This digital audio signal is fed to a frequency band divider 140, or frequency band demultiplexer, which is further explained in detail below with reference to fig. 2. In fig. 1, only one output of the frequency band divider 140 has been shown for simplicity. The output from the frequency band divider 140 represents a time- varying portion of the audio signal in a frequency sub band defined by characteristics of the frequency band divider.

The frequency sub bands, in particular the distribution of the frequency limits determining the frequency sub bands, may be configured in various ways. In one possible aspect, the distribution of the frequency limits is configured in accordance with a uniform scale (i.e. in a linear manner). In other possible aspects, the frequency limits are configured in accordance with a logarithmic scale, or with a special scale of psychoacoustic type, such as a Mel scale or Bark scale. The output from the frequency band divider 140 is connected to the power calculation circuit 150, which calculates the power of the signal provided by the frequency band divider 140. In alternative aspects, the amplitude or any other measure indicating the magnitude of the signal could be calculated in the calculation circuit 150. The keyboard 160 is connected to a keyboard operation detecting and noise estimating circuit 170. In this circuit 170, the operation of the keyboard is detected and possibly identified. Upon the detection of a keyboard operation, the circuit causes en estimate of a keyboard audio noise, i.e. an estimate of the noise that results from the operation of the keyboard, to be provided. In an aspect, the noise estimating circuit 170 may provide the estimate of the keyboard noise as a pre-stored estimate of noise power in a frequency sub band with respect to time. The pre-stored estimate may be selected as a power signal sample which is read from a pre-stored look-up table. The selection of the power signal sample may be based on an identification of the key that has actually been operated on said keyboard. This identification may be derived by the keyboard operation detecting and noise estimating circuit 170, which is connected to the keyboard and thus may obtain an identification of the pressed keys in addition to the mere detection of keyboard activity.

Such a pre-stored table may be pre-generated once, in a static fashion, or it may be established by training, e.g. during the operation of the keyboard in a noise-free or low-noise environment.

The table may also be adaptive, as mentioned further below with reference to fig. 2. The output of the power calculation circuit 150 and the output of the keyboard operation detecting and noise estimating circuit 170 are fed to the gain calculation circuit 180. A possible operation of the gain calculation circuit is explained in closer detail below with reference to figure 2. The gain value provided by the gain calculator is fed to a gain control input of the gain controlled amplifier 190. The signal input of the amplifier 190 is connected to the output of the frequency band divider 140.

This results in that the output signal of the amplifier 190 is considerably attenuated in case of a prevailing estimated keyboard noise in the frequency sub band selected by the frequency band divider, while the output signal is less attenuated, or not attenuated at all, in case of a small or zero estimated keyboard noise in the selected frequency sub band.

The output of the amplifier 190 is fed to an input a frequency band combiner circuit 200 which combines said input with corresponding input signals relating to other frequency sub bands. This results in that a combined digital audio signal is generated as the output of the circuit 200.

In an aspect, the output signal generated by the circuit 200 is further modified by circuits and means which are not shown in fig. 1. Instead, such additional circuits and means have been illustrated and explained with reference to fig. 2. More specifically, in this aspect, a comfort noise signal is generated and added to the noise reduced output signal. The adding of a comfort noise signal may comprise calculating a comfort noise gain which is based on the audio signal in the frequency sub band, i.e. the output of circuit 140, and the generated estimate of the keyboard audio noise, i.e. the output of the circuit 170. Fig. 2 is a schematic block diagram illustrating further possible principles of a device for reducing keyboard noise in a conferencing equipment.

Many of the elements illustrated in fig. 2 correspond to similar elements with similar reference numerals on fig. 1. Where further optional details are presented in the detailed description in figure 2, it should be understood that such optional details may be combined individually or in any combination, with the device described above with reference to figure 1.

The microphone 110 and the keyboard 160 are included in a conference equipment in the same way as explained with reference to fig. 1. The user (speaker, conference participant) 102 provides a wanted acoustic speech signal to the microphone, while unwanted audio noise generated by the operation of the keyboard 160 is also picked up by the microphone 110. The microphone 110 is connected to an amplifier 120, which may also advantageously include a low pass filter. The amplifier is further connected to the quantizer/digitizer 130, which includes circuits for sampling and analog-digital conversion, resulting in a digital audio signal at the output of element 130. This digital audio signal is fed to a frequency band divider 140, or frequency band demultiplexer.

As illustrated in fig. 2, a plurality of outputs of the frequency band divider 140 has been shown, but the further handling of only one of the output signals has been shown for simplicity. The output from the frequency band divider 140 represents a time-varying portion of the audio signal in a frequency sub band defined by characteristics of the frequency band divider.

Each frequency sub band represents a small fraction of the complete frequency spectrum of the full band signal.

The frequency sub bands, in particular the distribution of the frequency limits determining the frequency sub bands, and the width of each band, may be configured in various ways. In one possible aspect, the distribution of the frequency limits is configured in accordance with a uniform scale (i.e. in a linear manner). In other possible aspects, the frequency limits are configured in accordance with a logarithmic scale, or with a special scale of psychoacoustic type, such as a Mel scale or Bark scale. Further, each frequency sub band signal undergo a processing before all sub bands are merged together to a keyboard noise reduced full band signal.

A sub band type realization may be advantageous since filter banks, i.e. the frequency band divider and frequency band combiner, may already be present for other purposes, echo cancelling, stationary noise reduction, etc. In such a case, the sub band type realization of the invention does not represent very much additional complexity to the system.

Also, the sub band realization implies that all necessary spectrum calculations and filter calculations/operations are reduced to simple power calculations and gains, respectively.

In an alternative aspect, all processing may be performed directly on the full band signal. In this case, the frequency band divider 140 and the frequency band combiner 200 are not necessary. However, higher calculation capabilities will be needed for the further signal processing in this case. The output from the frequency band divider 140 is connected to the power calculation circuit 150, which calculates the power of the signal provided by the frequency band divider 140. In alternative aspects, the amplitude or any other measure indicating the magnitude of the signal could be calculated in the calculation circuit 150.

The power calculation circuit 150 may be configured to calculate the power of the signal in each frequency sub band, e.g. following the equation siglev(k,n) = siglev(k,n-l) + r(abs(sigsample(k,n)) ^κ — siglev(k,n-l)) wherein k is the sub band index, n is the time index, sigsample is the signal sample value of the sub band signal, and siglev denotes a smoothed estimate of the mean absolute value (for κ=l) or mean square (for κ=2) of consecutive signal samples. In the following discussion and parameter setting, κ=l is assumed, but any value K C [1,2] gives satisfactory results. r is a value defining a time constant, and its value will depend on the sampling rate in the frequency sub band. A F giving a time constant (standard analog definition, siglev falling from 1/e ~ 37% of its original value in one time constant if no input are applied, i.e. sigsamples equals zero) of 30 ms is a preferred choice. Siglev includes both wanted audio and unwanted keyboard noise.

Similarly, if implicit known a keyboard noise level could be estimated: keylev(k,n) = keylev(k,n-l) + r(abs(keysample(k,n)) ^κ — keylev(k,n-l))

In this equation, k, n, F and K is the same as in the previous equation, and keylev is a smoothed estimate of the mean absolute value of the keyboard noise. According to the equation above, the keyboard noise sample, denoted keysample, the above equation, will have to be known or estimated in order to calculate the estimated keyboard noise. Since the keyboard noise sample keylev(k,n) is generally not known, it is estimated in the noise estimating circuit 170. In an exemplary aspect, the keyboard noise estimating circuit 170 is a keyboard noise level lookup table, and the keyboard noise sample keylev(k, n) is tabulated in the keyboard noise level lookup table 170, in fig. 2 also denoted KNL. The idea behind the use of a keyboard noise estimating circuit 170, in particular a keyboard noise level lookup table, is that even though sample values of the keyboard noise will have very random characteristics, and therefore are not feasible to tabulate, the keyboard noise level as a function of frequency (i.e. sub band index k) and time (i.e. time index n) do have an adequate repeatability between different presses of a key.

The keyboard noise level lookup table is trained either once, during design of the product, or during production, or it may even be adaptively trained during use of the product, by averaging keylev calculations made from multiple instances of samples recorded using the keyboard/appliance in a noise free environment. Which learning strategy which should be used, may e.g. depend on the quality of the keyboard. For high quality appliances, keyboard noise will approximately be similar between different units in production, and thus a design phase estimation can be used. For lower quality appliances, keylev estimation might be necessary to measure for each different unit. Such predefmitions could be replaced or combined with an adaptive learning during use, by updating the lookup table for key presses done without any or neglectable additional sound.

Due to different physics (for example keys), distance from keyboard to microphone or similar, it might be necessary to tabulate different keylev for different keys or set of keys. Also, if the sampling rate in the sub bands is very low, i.e. time between samples comparable to the time constant of signal level, it may be necessary to tabulate interpolated versions of keylevs, or to interpolate the levels runtime.

As will be understood from the above explanation, the keyboard 160 is connected to the keyboard operation detecting and noise estimating circuit 170. In this circuit 170, the operation of the keyboard is detected and possibly identified. Upon the detection of a keyboard operation, the circuit 170 causes en estimate of a keyboard audio noise, i.e. an estimate of the noise that results from the operation of the keyboard, to be provided.

In an aspect, the noise estimating circuit 170 may provide the estimate of the keyboard noise as a pre-stored estimate of noise power in a frequency sub band with respect to time. The pre-stored estimate may be selected as a power signal sample which is read from a pre-stored look-up table. The selection of the power signal sample may be based on an identification of the key that has actually been operated on said keyboard. This identification may be derived by the keyboard operation detecting and noise estimating circuit 170, which is connected to the keyboard and thus may obtain an identification of the pressed keys in addition to the mere detection of keyboard activity.

Such a pre-stored table may be pre-generated once, in a static fashion, or it may be established by training, e.g. during the operation of the keyboard in a noise-free or low-noise environment. The table may also be adaptive. The output of the power calculation circuit 150 and the output of the keyboard operation detecting and noise estimating circuit 170 are fed to the gain calculation circuit 180.

The actual noise removal process defined herein may in some cases be relatively forgiving for deviations in keyboard noise levels, i.e. keylev. Thus, for a given time, the key level lookup table may output the key noise level estimate keylev for a given key and time, based on a control input from the keyboard, indicating time of pressure and which key is pressed. If no key has been pressed, keylev may be zero. The gain value provided by the gain calculating circuit 180 is fed to a gain control input of the gain controlled amplifier 190. The signal input of the amplifier 190 is connected to the output of the frequency band divider 140.

The gain calculation circuit 180, in fig. 2 also denoted GNC, is configured to calculate a signal gain for the sub band, based on the following equation: keynoisegain(k ₅ n) = max(φ,(siglev(k,n) ^λ - (l+δ)*keylev(k,n) ^λ )/(siglev(k,n) ^λ ))

In this equation, δ is a key noise level overestimation factor. It may be chosen according to possible deviation in the keyboard noise level from the actual level. It may be set higher for lower quality appliances/keyboards than for high quality appliance/keyboards. Normally, it is preferable to overestimate the key noise than to underestimate it.

The constant φ defines the wanted attenuation of keyboard noise, and is preferably set to zero.

The constant λ usually equals 1/κ. If κ=2, λ=l/κ and δ=0, the formula is very similar to the formula used in spectral subtraction for stationary noise removal, referred to above. However, κ=2, λ=l and δ=0 could be chosen, and then the formula is very similar to Wiener filtering.

In a particular embodiment, κ=l. In another particular embodiment, λ=l. In another particular embodiment, δ=l/3. In still another particular embodiment, κ=l, λ=l and δ=l/3. This gives a particularly good result for a high quality keyboard, and represents a modified spectral subtraction formula.

This results in that the output signal of the amplifier 190 is considerably attenuated in case of a prevailing estimated keyboard noise in the frequency sub band selected by the frequency band divider, while the output signal is less attenuated, or not attenuated at all, in case of a small or zero estimated keyboard noise in the selected frequency sub band.

The output of the amplifier 190 is fed to an input a frequency band combiner circuit 200 which combines said input with corresponding input signals relating to other frequency sub bands. This results in that a combined digital audio signal is generated as the output of the circuit 200.

The amplifier 190 may have a gain Gn, applying keynoisegain to the signal samples including key noise. noiseredsample(k,n) = keynoisegain(k,n)*sigsample(k,n)

This works very well removing noise presence of speech. The speech signal passes the gain without substantial audible attenuation of any of the speech signal. However, when the input signal is low level background noise only, the algorithm may reduce gain so much that it sounds muted. This is not important from an informational theory perspective, as background noise usually is unwanted, but in practical use, loss of noise may be perceived as a loss of connectivity, in a conference or similar.

Similar effects are also present in half duplex communication system, and are often compensated for by adding so called comfort noise. It may be advantageous to use the same technique after key noise removal as well. Very often, a comfort noise generator already exists in the appliance. Background noise is estimated by the background noise estimator BNE. Many different techniques are possible, the minimum statistics being the most well known: noilev(k,n)= noilev(k,n-l)+max(0,αu _* (siglev(k,n)-noilev(k,n-l)))+min(0,α _D «(siglev(k,n)-noilev(k,n-l))) The constant αu bay be chosen small and represents a long time constant (typically seconds, e.g. 10 seconds), whereas the constant αp represents a shorter time constant (typically a small fraction of a second, e.g. 0,01 seconds). The result is that the noise level estimate falls fast to the actual noise level in time with only noise present, whereas the noise level rises slow, avoiding that the noise estimate increases in periods with speech.

Of course, many other types of noise estimation techniques exist. The choice of technique may be selected by the skilled person according to circumstances.

A random generator RND generates a random, white signal of unity level. This signal is scaled by the gain function Gr, equaling noilev, to generate a noise sample estimate noisest(k,n) = noilev(k,n) * random(k,n)

The level of the noise estimates equals the level of the background noise. The noise estimate samples are used to fill in for lost noise caused by Gn. Estimated noise samples are uncorrelated with the actual noise, and random noise will therefore be added to the remaining background noise on a power basis. The noise fill gain calculator NFC calculates the appropriate gain: noisefillgain(k,n) = sqrt(l-keynoisegain(k,n) ² )

Finally, the random noise is scaled by the noise fill gain function Gf and added to the output sample: outputsample(k,n) = noiseredsample(k,n) + noisefillgain(k,n)* noisest(k,n) In the frequency band combiner circuit 200, all sub bands are merged, e.g. using a synthesize filter. Virtually, this results in a keyboard noise free audio signal, passing speech signal, with no perceptive loss of background noise. The optional, additional circuits on figure 2, included in the box "Comfort noise adding subsystem", have the effect that a comfort noise signal is generated and added to the noise reduced output signal. The adding of a comfort noise signal may comprise calculating a comfort noise gain which is based on the audio signal in the frequency sub band, i.e. the output of circuit 140, and the generated estimate of the keyboard audio noise, i.e. the output of the circuit 170.

Fig. 3 is a schematic flow chart illustrating the principles of a method for reducing keyboard noise in a conferencing equipment which includes a microphone and a keyboard.

The method starts at the initiating step 300. First, in step 310, an audio signal, originating from the microphone, is provided.

The providing step 310 may comprise steps for amplifying, filtering, sampling and digitizing, whereby the audio signal is provided as a digital audio signal.

The digital audio signal may be further split into frequency sub bands by means of a frequency band selector/frequency demultiplexer. The following processing steps may be performed on each of the frequency sub bands. The frequency sub bands may be configured in accordance with a scale selected from the set consisting of a uniform scale, a logarithmic scale, and a psychoacoustic type scale such as a Mel scale or Bark scale.

Next, in the detection step 320, an operation of the keyboard is detected. Upon the detection of keyboard activity /keyboard operation, the process continues at step 330.

Next, in the keyboard noise estimate providing step 330, an estimate of the keyboard audio noise, resulting from the detected operation of said keyboard, is provided. The estimate of keyboard noise may be provided as a pre-stored estimate, which may be selected from pre-stored data such as a lookup table. The selection of the pre-stored estimate among the pre-stored data may be based on an identification of a key operated on the keyboard. The pre-stored data may be pre-generated by training. Next, in the output signal calculating step 340, a noise reduced output signal is calculated based on the estimate of the keyboard audio noise and the audio signal.

The calculating step 340 may comprise spectral subtraction.

The calculating step 340 may comprise, for each frequency sub band, calculating a gain value which is based on the audio signal, or more particularly its power in the frequency sub band in question, and the provided estimate of the keyboard audio noise. Further, the audio signal is amplified using this gain value as a gain factor.

Further, in the optional comfort noise calculation step 350, a comfort noise signal may be generated. Further, in the optional noise adding step 360, the comfort noise signal may be added to the noise reduced output signal. The step of adding said comfort noise signal may comprise calculating a comfort noise gain value, wherein the calculating is based on the audio signal, in particular its power in the frequency sub band in question, and the generated estimate of the keyboard audio noise. It should be understood that the steps of the method illustrated in fig. 3 may be further specified in closer detail by the disclosure of figures 1 and 2 and their corresponding detailed description above, since the illustrated method and the corresponding device correspond to each other. More specifically, the device for reducing keyboard noise in a conferencing equipment may comprise processing means that are configured to perform the disclosed method. Such processing means may be distributed, i.e. as separate processing devices in each element of the device, or alternatively the processing means may be implemented as a central processing unit which performs the calculating operations of all the elements of the device or a combination of the elements included in the device. Although the detailed description specifies that a digital audio signal originating from the microphone may be provided as a digital audio signal, it should be noted that it is also possible to provide the audio signal as an analog signal.