Audio enhancement system and method

TECHNICAL FIELD

This invention relates to a method for improving the sound reproduced by an audio system in a listening environment such as a vehicle and, in particular, to a system which compensates for noise outside the audio system.

BACKGROUND

When music or speech is reproduced, for example, in a vehicle, the signal is corrupted by external acoustic noise present in the vehicle. This noise may result from and is dependent upon vehicle speed, road condition, weather and condition of the vehicle. The presence of such noise results in a situation where soft sounds of interest are hidden, the perceived loudness is reduced and the intelligibility of the signal is lessened. The vehicle's driver and/or passengers may compensate for increased external noise by increasing the volume of the audio system. However, when the vehicle speed decreases or another source of external noise is alleviated, the audio system's volume will be too high, requiring the user to reduce it. To overcome this, <patcit id="pcit0001" dnum="US5434922A"><text>US Patents 5,434,922</text></patcit> and <patcit id="pcit0002" dnum="US6529605B"><text>6,529,605</text></patcit> propose an enhanced dynamic volume control (DVC) system which extracts the noise signal from a signal provided by a sensor, e. g. a sensing microphone, in the listening environment and calculates a control signal therefrom. This control signal is used to control the volume and/or dynamics of the desired signal (i.e., music).

A DVC system extracts the noise signal from mixed signals derived from a sensor, e. g., a microphone. The mixed signals comprise music components (i.e., the desired signal), voice components and noise components. The noise component is intended to be used solely for obtaining the control signal for the volume or dynamics of the desired signal. The other components are not wanted to have any effect on the derivation of the control signal as otherwise the system would respond to voice signals or control itself through the music, which would end in a so-called gain chase situation (i.e., direct feedback). Such gain chase situation could lead to instability in the entire audio system.

The music signal is extracted from the sensor signal using an adaptive filter. The voice signal left in the remaining signal mixture is then masked out using a "voice activity detector (VAD)". The VAD operates continuously in the time range - i.e., in a broadband manner - and is implemented by an averaging unit with permanently configures increments and decrements. With other words, as long as the input signal is larger than the output signal, the VAD increases its output signal by a fixed increment, or reduces it by a fixed decrement whenever the input signal is less than the VAD's output signal. In this way, the VAD utilizes the different stationary properties of the voice and noise signals. The noise signal is strongly (broadband) smoothed so that the VAD output signal (i.e., the control signal), even if somewhat delayed, reaches a stationary final value, which approximately corresponds to the average power of the noise signal in the sensor signal. Depending on the configured volume or selected equalizing, which mainly refers in this instance to the bass setting, the music signal more or less penetrates the noise signal - i.e., the louder the desired signal (music signal) is played or the higher the bass controller is set, the greater the share of the music signal that passes unfiltered through the adaptive filter. This can lead to the known gain chase situation described above, which is imperative to prevent.

It has been found that the adaptive filter works better (i.e., permits less of the desired signal to pass through it) if the signals have a narrower bandwidth. For this reason, the DVC system mostly works with strongly undersampled signals, which, on the one hand, reduces the implementation complexity, but, on the other hand, leads to a control signal which is derived solely from the low-frequency noise component and then applied in a broadband manner to control the volume or dynamics. Since low-frequency noise signals dominate in vehicles - the field for which most of the DVC system are primarily designed - the spectral limitation described above can only actually be considered in this context. Nonetheless, the solution is not fully ideal and could lead to overlapping effects in certain circumstances, which is why a broadband solution is preferable. Although the risk of gain chase is reduced through limiting the bandwidth, it is not fully eliminated.

One way of completely avoiding gain chase is to upward-limit the control signal in accordance with the existing power of the desired signal which is implemented in common systems in the form of an anti-gain chase function. This function permits the control signal, provided the desired signal is below a specific minimum threshold value, to pass through the filter without being changed, but limits it to a maximum value specified by another function if the power of the desired signal rises above the threshold, and blocks further control once the desired signal has exceeded a maximum threshold - i.e., the control signal is then replaced by zero. The control signal modified in this way can then be used to alter the volume and/or dynamics of the desired signal using a compressor. However, the control signal is, nevertheless, fully dependent on the mean power of the currently existing noise signal but does not consider its spectral distribution or colouring.

In this regard, systems known as "dynamic equalizer control (DEC) systems" are considered successors to DVC systems. However, one aspect that hinders the transition from DVC to DEC systems is the limited bandwidth with which DVC systems work. The reason why the bandwidth is limited is primarily to lessen the risk of gain chase and additionally to reduce the implementation costs.

It is already known from the DVC system how to securely calculate the required signal from a microphone signal to use it as a loudness control value (volume). However, it is also evident that a straightforward volume control does not represent an optimum solution because it does not consider the spectral distribution of the noise signal. Methods in both the time and spectral domains with which the spectral coloring of the noise can be integrated into the processing of the required signal using adaptive equalizing as illustrated above in the DEC system. In particular, the problems associated with noisy voice signal components in the microphone signal were given special attention. However, other difficulties relating to both the DVC and DEC systems were also disregarded as before.

SUMMARY

It is an object of the present invention to create an improved system and method which automatically compensate for the noise level in a listening environment in the time domain.

According to the invention, a system and method for enhancing the sound signal produced by an audio system in a listening environment by compensating for ambient noise in a listening environment are provided, which produce an electrical sound signal and generating a sound output from said electrical sound signal; obtain a total sound signal representative of the total sound level in said environment, wherein said total sound level comprises both said sound output from said audio means and the ambient noise within said environment; extract an ambient noise signal representative of the ambient noise in said environment from said total sound signal in response to said total sound signal and to a reference signal derived from said electrical sound signal, said extraction step comprising an adaptive filtering with an adaptive adaptation step size; generate a control signal based on the outcome of said analysis in response to said ambient noise signal; adjust the sound output of said audio means to compensate for said ambient noise level in response to said control signal; and a step size calculation step for controlling the adaptive adaption step size of said adaptive filtering.

BRIEF DESCRIPTION OF THE DRAWINGS

The invention will be more readily understood from the following description taken in connection with the accompanying drawings forming part thereof, in which: <dl id="dl0001"><dt>FIG 1</dt><dd>is a schematic diagram of an adaptive FIR filter with an adaptation step size;</dd><dt>FIG 2</dt><dd>is an example of a four-speaker/one-microphone arrangement;</dd><dt>FIG 3</dt><dd>illustrates a DVC system with an improved composition of the additive reference signal;</dd><dt>FIG 4</dt><dd>is a schematic diagram of a DVC system to include all room transfer functions;</dd><dt>FIG 5</dt><dd>illustrates a mixture of the two systems shown in FIGs 3 and 4;</dd><dt>FIG 6</dt><dd>illustrates the use of a whitening filter in place of a simple high-pass filter for spectral smoothing of the noise signal;</dd><dt>FIG 7</dt><dd>shows the signal flow diagram for a voice activity detection model according to the invention;</dd><dt>FIG 8</dt><dd>is an example of an arrangement of a multi-channel DVC/DEC system;</dd><dt>FIG 9</dt><dd>illustrates the structure of an equalizing filter for use with the present invention;</dd><dt>FIG 10</dt><dd>shows a simple exemplary DVC/DEC system according to the invention;</dd><dt>FIG 11</dt><dd>shows another simple exemplary DVC/DEC alternative system according to the invention;</dd><dt>FIG 12</dt><dd>shows a complex exemplary DVC/DEC system according to the invention;</dd><dt>FIG 13</dt><dd>illustrates the generation of the reference signal and reference volume;</dd><dt>FIG 14</dt><dd>shows the characteristics of the gain correction function;</dd><dt>FIG 15</dt><dd>illustrates an exemplary implementation of the gain correction function;</dd><dt>FIG 16</dt><dd>illustrates the characteristics of an activation threshold; and</dd><dt>FIG 17</dt><dd>illustrates an implementation of the minimum noise activation threshold function.</dd></dl>

DETAILED DESCRIPTION

In all the adaptive filters presented thus far, which have the effect of removing the desired signal (in the present case: the source signal, e.g. music, speech etc.) as much as possible from the microphone signal, only a fixed adaptation step size expressed by p was used. This is by no means optimum, but instead reflects a compromise between stability and adaptation speed. Provided it can be assured that the desired signal is above the noise signal, using a fixed adaptation step size is perfectly acceptable. However, should the opposite situation apply to the signals, even if only for a brief period, the estimated loudspeaker/room microphone (LRM) transfer function is always corrupted.

As a consequence, the estimated noise level incorrectly rises due to the desired signal components that are now considered. This situation is unacceptable because it can lead to problems in unfavorable circumstances. The reason for the exhibited behavior is clear: the desired signal (the sound signal) serves as a test (or reference) signal with which the LRM transfer function is estimated using the adaptive filter. If the noise signal component is now greater than the desired signal component in the microphone signal, the room can no longer be correctly analyzed. The result is an LRM model comprising fully randomly assembled filter coefficients that cannot be iteratively reordered until the signal for measuring the room is above the noise signal, which logically considered requires some time to occur, during which the estimated transfer function is incorrect, and which more or less constitutes a major error in the removal of the desired signal component from the microphone signal.

The behavior of the adaptive filter in such situations should be as follows: Once the noise component becomes greater than the required signal component, the current LRM estimation - which exists as a set of filter coefficients - would have to be frozen until the correct relationships apply once more. It is implicitly assumed that the LRM system changes only very slightly or not at all during this time, an assumption that is borne out in practice. Moreover, the current adaptation success, as represented by the value of the so-called system distance, has to be considered along with the present signal-to-noise ratio (SNR) value.

The system distance indicates the extent to which the current model differs from the one that actually exists, i.e., it gives a measure of how much the adaptive filter has already approached the optimum value. If the distance between the estimated and existing acoustic models is small, then the system distance is likewise small. Consequently, the adaptation step size must be reduced, on the one hand, in order to approach the optimum model even more closely, and on the other hand, to increasingly consolidate the discovered model in order to be more resistant to brief occurrences of noise.

To estimate the system distance, a defined delay is introduced, for example, into the microphone signal branch. The delay is estimated by the adaptive filter using so-called delayed (filter) coefficients. If the adaptive filter is perfectly configured, the coefficients assume the value zero. If the filter is imperfect, the coefficients assume a non-zero value that varies in a direct relationship to the existing noise, thereby enabling conclusions regarding both the success of the adaptation and the system distance to be made. Generally, a small number of coefficients whose sum provides a value for the current system distance is all that is needed.

FIG 1 is a schematic diagram of an adaptive finite impulse response (FIR) filter with an adaptation step size µ[n]. Said filter comprises a finite impulse response (FIR) filter core 1 which is controllable by a Least Mean Square (LMS) adaption unit 2. The LMS unit 2 receives as an input signal a source signal x[n] from a signal source (e. g., a CD player, radio, etc.), an error signal e[n] from a subtraction unit 4, and an adaption step size µ[n] from a step size calculation unit 5. The step size calculation unit 5 may be supplied with a mean source signal Mean X[n], an estimated system distance Dist[n] supplied by the FIR filter core 1, and a mean error signal Mean E[n]. The signals Mean X[n] and Mean E[n] are provided each by a averaging unit 6 or 7, respectively. The substraction unit 4 further receives via a delay unit 8 a signal from a microphone 9 which is part of a loudspeaker-room-microphone (LRM) system 10. Said LRM system further comprises a room 11 having a transfer function H_Room(Z), in which the microphone 9 and a loudspeaker 12 supplied with source signal x[n] are arranged. The microphone 9 receives acoustic signals from the loudspeaker 12 filtered by the transfer function H_Room(z) and noise signals from at least one noise source 13 outside (as shown) and/or inside (not shown) the room 11.

In order to calculate the SNR, both the source signal, x[n], and the resulting output signal of the adaptive filter - i.e., the error signal e[n], are recursively determined using a simple 1st order Infinite Impulse Response (IIR) low-pass filter.

The resulting estimated optimum adaptation step size, µ[n], can be calculated as follows since the values for all relevant components are known: <maths id="math0001" num="(1)"><math display="block"><mi>μ</mi><mfenced open="[" close="]"><mi>n</mi></mfenced><mo>=</mo><mfrac><mrow><mi mathvariant="italic">Mean</mi><mi>X</mi><mfenced open="[" close="]"><mi>n</mi></mfenced><mo>∗</mo><mi mathvariant="italic">Dist</mi><mfenced open="[" close="]"><mi>n</mi></mfenced></mrow><mrow><mi mathvariant="italic">Mean</mi><mi>E</mi><mfenced open="[" close="]"><mi>n</mi></mfenced></mrow></mfrac></math><img file="EP1720249A1_D0001.tif" /></maths> where the estimated system distance is given by the following equation (estimated system distance): <maths id="math0002" num="(2)"><math display="block"><mi mathvariant="italic">Dist</mi><mfenced open="[" close="]"><mi>n</mi></mfenced><mo>=</mo><mfrac><mn>1</mn><msub><mi>N</mi><mi>t</mi></msub></mfrac><mo>∗</mo><mstyle displaystyle="true"><munderover><mo>∑</mo><mrow><mi>n</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>N</mi><mi>t</mi></msub></munderover><mfenced open="|" close="|" separators=""><mi>w</mi><mfenced open="[" close="]"><mi>n</mi></mfenced></mfenced></mstyle></math><img file="EP1720249A1_D0002.tif" /></maths>

The expression <maths id="math0003" num=""><math display="inline"><mfrac><mrow><mi mathvariant="italic">Mean</mi><mi>X</mi><mfenced open="[" close="]"><mi>n</mi></mfenced></mrow><mrow><mi mathvariant="italic">Mean</mi><mi>E</mi><mfenced open="[" close="]"><mi>n</mi></mfenced></mrow></mfrac></math><img file="EP1720249A1_D0003.tif" /></maths> in equation 1 corresponds to the ratio of the signal-to-noise ratio (SNR) to the sampling time n. Thus, we can also express equation 1 as follows: <maths id="math0004" num="(3)"><math display="block"><mi>μ</mi><mfenced open="[" close="]"><mi>n</mi></mfenced><mo>=</mo><mi mathvariant="italic">Dist</mi><mfenced open="[" close="]"><mi>n</mi></mfenced><mo>∗</mo><mi mathvariant="italic">SNR</mi><mfenced open="[" close="]"><mi>n</mi></mfenced></math><img file="EP1720249A1_D0004.tif" /></maths> which represents the optimum adaptation step size at the sampling time n.

The use of the DVC system is usually restricted solely to the lower spectral range, but even this range is generally not emitted from just one, but from a number of loudspeakers located in different positions within the listening room. As a consequence, the source signal played through the considered loudspeaker, representing the reference signal by <i>x</i>_<i>x</i>[<i>n</i>]<i>,</i> which is folded with the room impulse response between the source signal and the microphone, affects the composite signal recorded by the microphone.

FIG 2 is an example of a four-speaker/one-microphone arrangement. The arrangement comprises a microphone 14 and four loudspeakers 15,16,17,18 arranged in the front left (FL), front right (FR), rear left (RL), and the rear right (RR) of a room 19 which is in the present exemplary case the interior of a vehicle. Between each loudspeaker 15, 16, 17, 18 and the microphone 14 there is a room transfer function H₁(2), H₂(2), H₃(2), H₄(2) respectively. Although in the known DVC alternatives the different source signals were computed through simple addition and the resulting composite signal used as a reference signal, x[n], for the adaptive filter, the effects of the various room transfer functions still remain unconsidered. For example, if a system consisting of four loudspeakers and one microphone as in the present example is considered, all four room transfer functions H₁(2), H₂(2), H₃(2), H₄(2) between the loudspeakers 15, 16, 17, 18 and the microphone 14 are to be considered. Until now, only a single adaptive filter was used to remove the required (additive) signal from the microphone signal. Here, the adaptive filter could do no more than reflect the mean value of the four existing transfer functions.

The transfer functions with the strongest signal component, which normally stem from the nearest loudspeakers, have greater impact on the mean calculation than the weaker functions, derived mostly from loudspeakers positioned further away. This system functions quite well on average, but that it still appears to be inadequate in terms of a broader range of applications. To improve the system, the system of FIG 2 consisting of four different loudspeakers and a microphone is considered. The composition of the resulting composite reference signal, x[n], can be improved by considering in the calculation of the additive reference signal at least the different signal delay times and amplitudes resulting from the different distances of the loudspeakers from the microphone. Beginning with the nearest loudspeaker (in the present case loudspeaker 15), which is taken as a reference point, the additional (acoustic) signal delays and attenuations in relation to the reference point are determined.

FIG 3 illustrates a system using an improved composition of the composite reference signal. The system of FIG 3 comprises the microphone 14 (as in FIG 2) connected to an analog-to-digital (AD) converter 20 for converting an analog signal from the microphone 14 into a digital signal which is fed into a low pass (LP) filter 21 serving as an anti-aliasing filter for the subsequent sample rate conversion unit 22 which changes the sample rate by, e.g., 16. Further the composite reference signal x[n], which is the sum of individual source (reference) signals x₁[n], x₂[n], x₃[n], x₄[n] each supplied to the loudspeaker 15, 16, 17, 18 respectively, is provided by an adder unit 23 wherein the source signals x₂[n], x₃[n], and x₄[n] each are supplied to the adder unit 23 via a delay unit 24 ('Delay 2'), 25 ('Delay 3'), and 26 ('Delay 4') respectively and an attenuation unit 28 (Mag2), 29 (Mag3), and 30 (Mag4). The source signal x₁[n] is supplied to the adder unit 23 via an attenuation unit 27 only.

An analog-to-digital (AD) converter 28 is connected to the adder unit 28 for converting the analog additive reference signal x[n] into a digital signal which is fed into low pass (LP) filter 29 serving as an anti-aliasing filter for a subsequent sample rate conversion unit 30. The output signal of sample rate conversion unit 30 is fed via a delay unit 31 into a finite impulse response (FIR) filter core 32 and a least mean square (LMS) filter control unit 34 for controlling the FIR filter core 32. The output of FIR filter core 32 is connected to one input of a subtraction unit 33 wherein a second input of subtraction 33 is connected to the output of sampling rate conversion unit 22. Subtraction unit 33 outputs the error signal e[n] which is fed back to the filter control unit 34 for a comparison with the delayed reference signal x[n]. Error signal e[n] may also be supplied to other units as, for example, a predictive error filter as shown in FIG 6.

In FIG 3, each component of the reference signal is considered as delay or amplification or attenuation factor in the composition of the reference signal. Subsequently, the basic delay inherent in the path from the reference loudspeaker 15 the microphone 14, which is referred to as 'Delayl' or ('bulk delay'), can be considered within the undersampled signal processing range in order to increase the effective length of the adaptive filter. In doing so, an attempt is made to move the long delays into the undersampled range as much as possible to reduce memory consumption and to equalize only the differences in delay at the full sampling frequency using delay lines.

At the same time, existing symmetries can be utilized - for example, if the paths of the two front (FL, FR) and rear (RL, RR) loudspeakers, are identical. In such cases, first the symmetrical signals could be added together and then the composite reference signal is adapted to the corresponding second adaptive signal in terms of delay time and amplitude. The effective length of the adaptive filter is increased by including the bulk delay in the calculation since the adaptive filter would otherwise equalize the acoustic signal delay time associated with the loudspeaker nearest to the microphone (e.g., loudspeaker 15 in FIG 2). This intrinsic, acoustic signal delay time can also be used advantageously to determine the delayed coefficients. Using this method, the delay line on the microphone signal path can be fully omitted under certain circumstances. Nonetheless, even this enhancement does not yet grant an optimum system, to do so, all input signals must be separately considered, resulting in the system illustrated in FIG 4.

FIG 4 is a schematic diagram of a DVC system to include all room transfer functions. As in FIG 3, the system of FIG 4 comprises a signal path (microphone path) established by the microphone 14, the AD converter 20, the low-pas filter 21, and the sampling rate conversion unit 22. Further, the signal path of FIG 3 established by AD converter 28, low-pass filter 29, sampling rate conversion unit 30, delay unit 31, FIR filter core 32, filter control unit 34, and subtraction unit 33 is realized in FIG 4 by a path (reference path 1) comprising an AD converter 128, a low-pass filter 129, a sampling rate conversion unit 130, a delay unit 131, a FIR filter core 132, a filter control unit 134, and a subtraction unit 133, wherein the AD converter 128 receives the reference signal x₁[n] (instead of signal x[n] of FIG 3) and the subtraction unit 133 provides an error signal e₁[n] from the processed microphone signal and the processed reference signal x₁[n].

In contrast to the system of FIG 3, the system of FIG 4 comprises three further reference paths (reference paths 2, 3, 4) established each by AD converters 228, 328, 428; low pass filters 229, 329, 429; sampling rate conversion units 230, 330, 430; delay units 231, 331, 431; FIR filter cores 232, 332, 432; filter control units 234, 334, 434; and subtraction units 233, 333, 433 wherein each AD converter 28 receives the respective reference signal x₂[n], x₃[n], x₄[n] and the subtraction units 233, 333, 433, provide error signals e₂[n], e₃[n], and e₄[n] from the error signals e₁[n], e₂[n], and e₃[n] and the processed reference signals x₂[n], x₃[n], and x₄[n] respectively. The reference paths 1, 2, 3, and 4 are identical at least in structure.

The reason why the individual room transfer functions have to be used as shown in the arrangement in FIG 4, i.e., with linking of the individual adaptive filters, and not (as one might think) in the conventional, parallel fashion, can be explained by the strong correlation of the input signals. The individual room transfer functions could only be estimated in parallel independently of each other if the interrelations of the input signals x₁[n],.., x₄[n] were non-correlated.

However, since there is a strong correlation between the input signals, the system functions are as follows: the first adaptive filter (in reference path 1), which is fed with the undersampled reference signal (=x₁[n]) through the delay line 131 (Delay1), removes all signal components from the microphone signal that are correlated with x₁[n]. Ideally, the error or output signal e₁[n] of the first adaptive filter will only contain signal components that are no longer correlated with x₁[n]. The second adaptive filter (in reference path 2) can adapt only if there are components in e₁[n] that are correlated with x₂[n]. Otherwise, the filter transfers the signal that is also not correlated with x₂[n] as its output signal, and so forth. At the end of the chain, a signal is correspondingly output that is correlated with neither x₁[n], x₂[n], x₃[n] nor x₄[n] - i.e., all that remains is a signal mixture comprising the speech signal and background noise, represented in FIG 4 as 'Noise[n]'.

Normally when using music as a reference signal source, a strong correlation can be noted between the individual reference signals x₁[n], ..., x₄[n]. For this reason, hardly any signal is left remaining after the first adaptive filter for which it would be worthwhile analyzing the remaining reference signals for correlations using adaptive filters. It is therefore questionable as to whether the entire effort is worthwhile in practical terms. In terms of signal theory, the system shown in FIG 4 is indubitably valid. However, when viewed economically, the system illustrated in FIG 3 is likely to be adopted. What is also conceivable is a mixture of the two systems shown in FIGs 3 and 4, which would represent a viable compromise.

FIG 5 illustrates such a mixture of the two systems shown in FIGs 3 and 4. The system of FIG 5 has a microphone path and the reference paths 1, 2 as shown in the system of FIG 4. However, path 2 (i.e. AD converter 228) is not supplied with reference signal x₂[n] as in FIG 4 but is supplied with a mean reference signal similar to the one of FIG 3. In the system of FIG 5, said mean reference signal is provided by an adder unit 123 receiving the processed reference signals x₂[n], x₃[n], x₄[n]. The processing is such that reference signal x₂[n] is provided to the adder unit 123 via an attenuation unit 127 and reference signals x₃[n], x₄[n] via attenuation units 29, 30 and delay units 25, 26 respectively. The additive error signal Noise[n] is output by subtraction unit 233.

One of the DVC alternative models known has, at least optionally, a high-pass filter arranged after the adaptive filter. The function of the high-pass filter is to smooth the noise spectrum - which is known to have very low frequencies in vehicles - to such an extent that the DVC system can respond not only to the predominant engine and movement noises but also to high-frequency wind noise, which has much weaker signal levels. Even though wind noise has considerably weaker levels than those of the low-frequency components, it is found in a spectral region in which human hearing is particularly sensitive, and therefore even a low noise signal level requires a control mechanism.

Until now, a high-pass filter was used to smooth the noise spectrum as much as possible so that the system can respond practically equally well to both low-frequency noise and the high-frequency wind noise signals just described. The high-pass filter was set to a fixed value and then left unchanged. This setting is adequate for the majority of cases, as demonstrated in many tests, but there are certain noise situations in which the high-pass filter may act poorly - in other words, the filter does not appreciably smooth the spectrum of the noise signal, with the effect that there are different responses to noise signals with differing spectral distribution.

To overcome this, the spectral smoothing filter may be set adaptively in response to the currently applicable noise situation, wherein linear predictive coding (LPC) analysis is utilized by feeding a predictor error filter (also known as a pre-whitening filter) with the iteratively calculated reflection coefficient. The system is also able to configure the equalizing filter according to the pertaining noise situation in such a way that the output signal of the predictor error filter is distributed across the spectrum as white noise as much as possible.

FIG 6 illustrates the use of a whitening filter in place of a simple high-pass filter for spectral smoothing of a noise signal, wherein only just one reference path (reference path 1 of FIG 5) is used for explanation purposes in FIG 6 but can easily adapted for the other paths. In the system of FIG 6, high-pass filters 35 and 135 are included into the microphone path and the reference path (reference path 1) after the sample rate conversion units respectively. Further, the error signal e[n] provided by the subtraction unit 133 is fed into a predictive error filter 36 which is controlled by a linear predictive coding (LPC) unit 37 receiving the error signal e[n], too. An effective means of doing this in the time range is according to the invention to compute the linear predictive coding (LPC) coefficients using LPC analysis, and to input the coefficients into a predictor filter with which the error signal is weighted.

It is known from other applications - e.g., acoustic echo cancellation (AEC) - that echo coefficients are calculated from the voice signal in those cases using LPC analysis, and that the coefficients are then fed into a predictor error filter for filtering the microphone signal. The voice signal recorded by the sensing microphone is spectrally smoothed so that the AEC system can work more efficiently. Said AEC application therefore refers to a so-called pre-whitening filter whose effects can be described as follows: the LPC analysis supplies coefficients that are used in a corresponding predictor filter that calls a transfer function corresponding to the power spectral density (PSD) of the signal (i.e., voice signal) examined by the analysis. When used in a predictor error filter, its transfer function does not follow the PSD of the analysis signal, but rather its inverse pattern. For this reason, the voice component in the microphone signal on which the LPC analysis was based is spectrally smoothed after the predictor error filter - i.e., it is whitened (thus giving rise to the term "pre-whitening filter").

The analysis/synthesis method known as linear predictive coding (LPC) is an efficient and effective mean, for example, to achieve synthetic speech and speech signal communication. The efficiency of the method is due to the speed of the analysis algorithm and to the low bandwidth required for the encoded signals. The effectiveness is related to the intelligibility of the decoded vocal signal. LPC implements a type of vocoder which is an analysis/synthesis scheme where the spectrum of a source signal is weighted by the spectral components of the target signal that is being analyzed. The phase vocoder is, for example, a special kind of vocoder where amplitude and phase information of the analysis channels is retained and can be used as coefficients for an error filter.

In standard LPC, the source signals are either white noise or a pulse train, thus resembling voiced or unvoiced excitations of the vocal tract, respectively. The basic assumption behind LPC is the correlation between the n-th sample and the p previous samples of the target signal. Namely, the n-th signal sample is represented as a linear combination of the previous P samples, plus a residual representing the prediction error. The analysis problem is equivalent to the identification of the coefficients a of an all-pole filter. If minimizing the error in a mean square sense, the problem translates into a set of P equations which, in the z domain (frequency domain), reduce to E(z)=A(z)X(z), where A(z) is the polynomial with coefficients a and E(z) is the prediction error. In the case of voice signal analysis, the filter 1/A(z) is called the all-pole formant filter because, if the proper order p is chosen, its magnitude frequency response follows the envelope of the signal spectrum, with its broad resonances called formants. The filter A(z) is called the inverse formant filter because it extracts from the voice signal a residual resembling the vocal tract excitation. A(z) is also called a whitening filter because it produces a residual having a flat spectrum.

However, there are two kinds of residuals, both having a flat spectrum: the pulse train and the white noise, the first being the idealized vocal-fold excitation for voiced speech, the second being the idealized excitation for unvoiced speech. In reality, the residual is neither one of the two idealized excitations. At the re-synthesis stage the choice is either to use an encoded residual or to choose one of the two idealized excitations according to a voiced/unvoiced decision made by the analysis stage.

When the target signal is periodic (voiced speech), a pitch detector can be added to the analysis stage, so that the re-synthesis can be driven by periodic replicas of a basic pulse, with the correct inter-pulse period. Several techniques are available for pitch detection, either using the residual or the target signal. Although not particularly efficient, one possibility is to do a Fourier analysis of the residual and estimate the fundamental frequency by the techniques of section. For example, the information extracted by the analysis stage are the prediction coefficients a₁,...,a_p; the residual e; pitch of the excitation residual; voiced/unvoiced information; and signal energy (RMS amplitude). These parameters, possibly modified, are used in the re-synthesis stage, for example, by means of the well-known Levinson-Durbin recursion, which provides the reflection coefficients of the lattice realization of the filter 1/A(z).

An effective means of doing this is to compute the linear predictive coding (LPC) coefficients using LPC analysis, and to input the coefficients into a predictor filter with which the music signal is weighted. Known DVC systems always use a simple voice activity detection (VAD) function that makes use of the long-term steady state of the background noise signal and the short-term steady state of the human voice with the aid of two mean value calculation units. This VAD variant is nonetheless extremely difficult to control, and for this reason a new VAD model that is easier to control and also even simpler to implement is preferred.

FIG 7 shows the signal flow diagram for a new voice activity detection (VAD) model. The new VAD model simply examines whether the current noise value (Noise[n]) is greater than the previous noise level (NoiseLevel[n]). If this is the case, the new noise level (NoiseLevel[n+1]) is calculated from the previous noise level in such a manner that this noise level increases by a permanently defined increment, regardless of how much the current noise signal value is greater than the previous noise level.

This approach suppresses strong effects of voice signal components within the noise signal on the noise level, and even without any delay since the new VAD system described here has no 'memory' function, unlike the older model. If the current noise signal is less than the previous noise level, the new noise level value is reduced by a fixed decrement. The increment is generally defined to be less than the decrement in order that there is no agitated response to brief energy spikes in the noise signal. The noise level is restricted to a low limit value using the low threshold value of the noise level, 'MinNoiseLevel'. The noise level is prevented from falling below a specific level in this way. Without a low threshold value of this kind, the system would at times respond much too slowly to a rising noise signal because the rising speed (which is governed by the increment) is slower than the falling speed.

The DVC systems previously described were always operated with just one microphone. The noise situation was only recorded at the mounting point of the microphone. This information was then used to develop a global controlled variable for the entire vehicle. However, a number of aspects were not considered, which can lead to errors. For example, if the microphone is positioned where standing waves are located in the interior of the vehicle, the system can respond only inadequately or not at all to noise signals at exactly the same frequencies. Ideally, the microphones should always be positioned where the passenger's head currently is, because this noise situation at every position affects the passenger's auditory perception, which is intended to be enhanced using DVC and DEC systems.

Thus, it would make sense to integrate the microphone in the headrest, for example, to ensure that it is always positioned as close as possible to the best point of auditory perception. The same challenge applies, of course, to all other passengers in the vehicle. Even if it could be guaranteed that a microphone is placed in the ideal position for the driver, this does not automatically mean that this arrangement would produce good results at the remaining positions in the interior of the vehicle. Each listening position would need its own microphone for the system to provide equally good results. However, it is extremely complicated to obtain an effective control system as each loudspeaker would have to provide a contribution for each microphone position in order to acquire the desired control effect at the same time at every listening point.

FIG 8 is an example of an arrangement of a multi-channel DVC/DEC system in a room 19. The simplest method of obtaining an individually adapted control system is to deploy the loudspeakers 15, 16, 17, 18 located nearest to associated microphones 515, 516, 517, 518 for control of the amplitude through the DVC system and control of the equalization through the DEC system. Only the spectral range of a subwoofer 500 would then need to be controlled globally. This is however readily acceptable as the range cannot generally be localized, and the human ear is in any case quite insensitive in this frequency range to variations in signal level.

The method using linear predictive coding (LPC) analysis, which can be efficiently implemented - for example, using the gradient adaptive lattice (GAL) algorithm - is shown in principle to be an effective means of implementing dynamic equalization with adaptation to the power density of the noise signal. The conjecture made at that time that any voice signal components present in the noise signal can for the most part be excluded using a very slow update speed for the GAL algorithm has also proven to be correct in principle. However, still some problems remain which shall be illustrated together with their associated solution.

If the LPC analysis is applied to the noise signal isolated by the adaptive filter using the GAL algorithm with a correspondingly slow update speed, the reflection coefficients deriving from the analysis are set in such a way that when included in a predictor filter (all-pole filter), whose transfer function more or less exactly matches the characteristics of the noise signal's power spectrum density (PSD), depending on the predictor order selected. The LPC analysis returns reflection coefficients whose values are based exclusively on the PSD development of the analyzed noise signal, but that do not provide any information on the signal's amplitude relationships. Furthermore, the amplitude of the predictor filter is much too high to be ever directly used.

In order to solve these problems, the transfer function of the predictor filter was continually scaled in such a way that its maximum value is as close to 0 dB as possible. From the known scaling standards, i.e., the L1, L2 and L∞ standards finally the L1 standard was chosen since it is, on the one hand, easier to calculate and, on the other hand, it represents the hardest scaling type. <maths id="math0005" num="(4)"><math display="block"><mi mathvariant="italic">Scale</mi><mo>=</mo><mstyle displaystyle="true"><munderover><mo>∑</mo><mrow><mi>n</mi><mo>=</mo><mn>0</mn></mrow><mi>∞</mi></munderover><mfenced open="|" close="|" separators=""><mi>f</mi><mfenced open="[" close="]"><mi>n</mi></mfenced></mfenced></mstyle></math><img file="EP1720249A1_D0005.tif" /></maths>

Equation 4 is a formula for calculating the scaling factor according to the L1 standard. As can be seen from equation 4, an infinite number of values of the filter's impulse response is required in theory to calculate the scaling factor exactly according to the L1 standard, which is however impossible in practical terms. In view of this, restriction was made to the assumption of a limited impulse response, which is then used to calculate the scaling factor ('scale') according to the L1 standard. The length of the impulse response is defined on the basis of both the error that is made when calculating the scaling factor and of the update time with which is wanted (or needed) to generate a scaling factor applicable to the next impulse response interval. To keep the update interval as short as possible, it is advisable to perform the calculation (at least for the impulse response) using the original sampling frequency, f_s, and not with the reduced sampling frequency, f_s/R.

Once it has been ensured that the predictor filter's maximum value is 0 dB (i.e., it is scaled to 0 dB), it is still to decide to what extent the equalizing, in other words, predictor filter is permitted to influence the required useful signal. This decision is mainly dependent on the current noise level that we have already determined using the DVC system. This implies that it is practical to couple the DEC and DVC systems together. The question is still posed as to how the equalizing filter can be controlled on the basis of the existing noise level or in what structure it can be applied.

FIG 9 illustrates the structure of the equalizing filter. For the implementation of the equalizing (EQ) filter it is referred back to that of peaking filters. In peaking filters, a gain control can be used to set how far the implemented EQ filter is to extend beyond the reference amplitude, which is also normally at 0 dB. This structure is chosen for the EQ filter - i.e., the gain from the DVC system indicates how much the maximum spectral share of the EQ filter scaled (by 'Scale') to 0 dB should be higher than the reference amplitude. In this way, only the spectral components that are particularly strongly concealed by background noise are amplified along with the DVC gain. This has the same effect as dynamic equalizing - i.e., a DEC system. The gain derived from the DVC system does not have to be accepted on a 1-to-1 basis, rather it can be scaled at will for the DEC system or changed in any other way to obtain an satisfactory control effect. The filter illustrated in FIG 9 comprises a source 40 connected to a subsequent adder unit 41 as well as a scaling unit 42. A equalizing (EQ) filter 43 and a subsequent gain adjusting unit 44 are connected downstream to the scaling unit 42. The output signal provided by the gain adjusting unit 44 is fed into the adder unit 41, whose output signal is (via units not shown) to a loudspeaker 45.

FIG 10 is an example of a simple DVC/DEC system. In addition, the new gain value can be calculated together with the previously determined scaling factor to produce a further version of the DVC/DEC system. The filter of FIG 9 is applied to system of FIG 10 such that the filter is not connected directly to the source but through a gain adjusting unit 46 which is controlled by a DVC unit 47. The DVC unit 47 is supplied with signals from the loudspeaker 45 and a microphone 48 and further controls an LPC unit 49 and a gain calculation unit 50. The DVC unit 47 and the gain calculation unit 50 are further controlled by a volume control 51. The EQ filter 43 is controlled by the LPC unit 49 which further controls via a Dirac pulsed filter 52 and scaling calculation unit 53 the scaling unit 42, wherein gain adjusting unit 44 is controlled by gain calculation unit 50.

FIG 11 is a another simple DVC/DEC system. The system of FIG 11 differs from the system of FIG 10 in that in the system of FIG 11 the scaling unit 42 is not connected between gain adjusting unit 46 and EQ filter 43 (as in FIG 10) but between the output of gain calculation unit 50 and the control input of gain adjusting unit 44.

FIG 12 is an example for an complex DVC/DEC system according to the invention. The system of FIG 12 has a microphone path and a reference path. The microphone path is established by a microphone 614, an AD converter 620, a low-pas filter 621, and a sampling rate conversion unit 622, and an A-weighting high-pass filter 623. The reference path is established by an AD converter 628, a low-pass filter 629, a sampling rate conversion unit 630, an A-weighting high-pass filter 624, a delay unit 631, wherein the AD converter 628 receives a reference signal from an adder unit 615 connected to two delay units 616, 617. Said delay units 616, 617 are each supplied with a signal provided to two loudspeakers 6 via an amplifying unit 618, 619 respectively.

The reference path further comprises an adaptive filter having a finite impulse response (FIR) filter core 641 which is controllable by a Least Mean Square (LMS) adaption unit 642. The LMS unit 642 receives as an input signal a signal from the delay unit 641, an error signal e[n] from a subtraction unit 6444, and an adaption step size µ[n] from a step size calculation unit 645. The step size calculation unit 645 is supplied with a mean source signal Mean X[n] tapped from the (down-) sampling conversion unit 631 via an average unit 650, an estimated system distance Dist[n] supplied by the FIR filter core 641, and a mean error signal Mean E[n] tapped from the subtraction unit 644 via an average unit 651. The substraction unit 644 further receives via the microphone path a signal from the microphone 614 which is part of a loudspeaker-room-microphone (LRM) system (not shown in FIG 12). Said LRM system further comprises a room having a transfer function H_Room(Z), and the two loudspeakers 612, 613.

The error signal e[n] is supplied via a scaling unit 659 to a prediction error filter core 660 serving as a whitening filter. Downstream the filter core 660 there is an averaging unit 661, a voice activity detector 662, a gain correction unit 663, a maximum gain unit 664, and a compare unit 665 evaluating whether gain below 1 or not. If so, the gain of an amplifying unit 666 is set to 1, and if not, the gain is left unchanged. The amplifying unit 666 is supplied with an electrical sound signal from a sound source (e. g. CD player, cassette player, radio and the like) and feeds the amplified sound signal into a prediction filter core 667 which is controlled by a linear predictive coding unit 658 supplied with the error signal e[n]. Downstream the filter core 667, there is a scaling unit 668 and an adder unit 669 which also receives the signal from the amplifying unit 666 input into the filter core 667. The output signal of adder unit 669 is processed by a sound signal processing unit 670 in a common way and supplied to loudspeakers 612 and 613.

The gain settings from the compare unit 665 as well as mean volume settings 671 by the listener are used to set a gain control unit 672 for controlling the gain correction unit 665. The mean volume settings 671 are further used to control a gain control unit 673 for controlling a scaling unit 674. The scaling unit 674 which controls the scaling unit 668 further receives signals from a Dirac pulsed prediction filter 675 and the VAD 662. The Dirac pulsed prediction filter 675 controls the prediction error filter 660 by copying its filter coefficients into the prediction error filter 660.

Fundamentally, the reference signal should be tapped as much as possible at the end of the signal processing chain to prevent negative effects of the equalizing and dynamic modifications (limiters, compressors, etc.) on the performance of the adaptive filter. Moreover, it was assumed that it would be advantageous to obtain the reference signal by adding all relevant signal paths. This turned out to be an error in practice however: not only was the phase of the reference signal strongly distorted, but also its amplitude, due to the phase differences between the individual signal paths, some of which are quite significant. This situation necessarily leads to an incorrectly working adaptive filter.

One way of improving the performance is to use a single, broadband output signal for which no addition is performed. This is also permissible provided all output signals are strongly correlated with each other, which is usually the case with audio signals. Unfortunately, broadband output signals of this nature are rarely found in signal flow diagrams since the output signals are generally adapted to the bandwidths of the connected loudspeakers using crossover filters before they are output, and are consequently limited in bandwidth. The optimum solution is then that using the reference signal generation as illustrated in FIG 13 because it makes use of all correlations when generating the reference signal and dispenses with all signal addition operations.

Yet another method, likewise embodies an alternative method for reference signal generation. In that method, the reference signal is generated from one or more input signals. Although this is in principle unwelcome, as stated at the beginning, it has shown itself in practice to a simple, yet efficient means of reference signal generation.

But why does the method being theoretically undesirable nonetheless work so well in practice? This question can only be answered once the entire signal processing between the input and output signals is considered as a stationary system. In this respect, the system is practically part of the LRM system, which can be reproduced by the adaptive filter. In addition to the music, which is normally available in a stereo signal, any available special channel must serve as a reference, if it is not to act as a potential noise influence. In doing so, different source signals may be mixed together in a particular relationship at the same time. The so-called mixer determines the extent to which one signal mixes with another.

Consequently, the output signal of the mixer, which mixes all source signals together, represents a good input signal for said purpose. The signal amplitude is still a problem though in that it is likewise modified by the mixer. Bearing in mind that the greater the amplitude of the reference signal, the better the adaptive filter works, one method of obtaining a continuously well-controlled reference signal (= mixer output signal) is to scale the reference signal in such a way that it is as fully driven as possible, regardless of the configured volume. To find the scaling factor, the maximum value of the sum of all loudness values (volumes) of the source signals is determined, and its difference to the fully driven signal is calculated - which is actually the equivalent of the scaling factor. Depending on whether the volume values are linear or logarithmic, the following formulae are used to calculate the scaling factor: <maths id="math0006" num=""><math display="block"><mi>Logarithmic</mi><mo>:</mo><msub><mi>Scale</mi><mi>dB</mi></msub><mo>=</mo><msub><mn>0</mn><mi>dB</mi></msub><mo>−</mo><msub><mi>MaxVol</mi><mi>dB</mi></msub></math><img file="EP1720249A1_D0006.tif" /></maths> or <maths id="math0007" num=""><math display="block"><mi>Linear</mi><mo>:</mo><msub><mi>Scale</mi><mi>Lin</mi></msub><mo>=</mo><mn>1</mn><mo>/</mo><msub><mi>MaxVol</mi><mi>Lin</mi></msub></math><img file="EP1720249A1_D0007.tif" /></maths>

The calculated maximum volume is then also used as the reference loudness value (=VOLUME) in the calculation of the anti-gain chase function, etc. in the DVC/DEC system so that problems due to too low volumes are avoided from the outset.

FIG 13 illustrates a model for generating the reference signal and reference volume. In the model shown in FIG 13, a mono signal is generated from the stereo music input signals 'MusicLeft' and 'MusicRight' before being fed into a mixing matrix 60. This is not absolutely necessary, but it simplifies the mixing matrix 60. The problem associated with impure phase addition normally does not occur, or only to a minor degree, in this system as no signal processing has yet been performed using the two music signals, thereby rendering the risk of reference signal distortion minimal.

Practical experimentation using the DVC system demonstrated that it functions perfectly within a specific noise level range, but that it begins to respond (at least subjectively perceived) exaggeratedly to noise levels above that range. For this reason, a gain correction (scaling unit 62 in connection with Maximum volume search 61 and Difference-to-full-scale calculation unit 63) function may be integrated to modify the noise level value ('NoiseLevel' in FIG 15) weighted with the correction factor (Corr) deriving from the anti-gain chase function. The weighted level is referred to as the ratio.

FIG 14 shows the characteristic of the gain correction function. FIG 15 illustrates the implementation of a gain correction function. The gain correction function is controlled by two parameters, namely noise threshold (NoiseTH) and NoiseSlope which is the slope of the righthand line of FIG 14. NoiseTH is used to control how a measured increase in the background noise level results in a volume increase. Below NoiseTH the ratio is about 1:1 and above it is as NoiseSlope indicates below 1:1. A reason for this may be that, as an outcome of undersampling, lower frequency components being used basically only for the calculation of the gain factors increase more than higher frequency components do with an increase of the background noise. However, as DVC gain is effective in the whole frequency range and listeners are most sensitive in the mid-range the system as to be adapted by introducing a gain correction function to address the issue outlined above. The gain correction function may be incorrect for lower frequencies but this incorrectness may be overcome by arranging the DVC gain units downstream a loudness processing unit. Alternatively, the DEC may be adapted to compensate for the incorrectness.

As known from the speech activity detection function illustrated with regard to FIG 7, sometimes the period is quite long until the genuine end noise level value is reached so that appropriate countermeasures can be taken. The waiting time is due to the low slew rate (C_INC) with which the system responds to increasing noise levels. The threshold value, 'MinNoiseLevel', is used to prevent the noise level from falling below a specific level, which in effect is identical to the activation threshold of the system. However, care must be taken to avoid setting the threshold too high as otherwise the control function could be activated although the noise level is not high enough.

By deploying the anti-gain chase function, the maximum permissible control dynamics (MaxGain) vary according to the volume. Since the activation threshold (MinNoiseLevel) depends on the maximum gain, it must likewise be modified according to the volume. This yields a characteristic for the activation threshold, a straight line whose Y axis section (t_MNTH) varies with the maximum gain (MaxGain) and the noise scaling factor (Scale), and whose slope (m_MNTH) is correlated with the slope of the anti-gain chase function (m_Corr). FIG 16 illustrates the characteristics of the activation threshold.

FIG 17 illustrates an implementation of the MinNoiseTH (activation threshold) function. The use of an adaptation step µ[n] size has shown itself to be very helpful in practice. In particular, the negative effects of strong, brief noise signals - for example, the human voice - were able to be reduced to a minimum using the appropriate step size, especially in the case of transmission of loud, bass source signals since these contribute little or nothing in the calculation of the noise level. The energy of the error signal MeanE[n](see FIG1) must be scaled according to the amplitude of the reference signal when tuning the adaptation step size (the reference signal should be driven as much as possible, as described earlier). The greater the effect of the reference signal in the calculation of the adaptation step size, the more the energy of the error signal must be scaled upwards. The converse applies equally.

If the exemplary implemented model shown in FIG 13 for generating the reference signal is used, it could be seen that the formula expressed in equation 1 was also correct. It has to be noted that this is not always the case. However, special care should be taken that no error is made since this could cripple the entire adaptive filter, which would in turn have worse consequences for the whole DVC/DEC system. It was also discovered that the adaptation step size, µ[n], may be limited to a specific range (µ[n] ∈ [0 ,.., µ_Max[n]]) so that the adaptive filter can function more reliably. In particular, the maximum permissible adaptation step size, µ_Max[n], should not be set too large as otherwise there is a risk that the adaptive filter will become unstable and start to oscillate, which would cause the DVC/DEC system to fail. A value of about 0.01 for µ_Max [n] has proven satisfying. This value is not intended to apply in general but rather should serve as a guideline figure. Its precise maximum value must be calculated for each particular case.

Although various exemplary embodiments of the invention have been disclosed, it will be apparent to those skilled in the art that various changes and modifications can be made which will achieve some of the advantages of the invention without departing from the spirit and scope of the invention. It will be obvious to those reasonably skilled in the art that other components performing the same functions may be suitably substituted. Further, the methods of the invention may be achieved in either all software implementations, using the appropriate processor instructions, or in hybrid implementations that utilize a combination of hardware logic and software logic to achieve the same results. Such modifications to the inventive concept are intended to be covered by the appended claims.