EP1202253B1 - Adaptive noise level estimator - Google Patents

Abstract

The method involves adopting a definable initializing noise level estimate, determining further estimate values from the minimum of the maximum value of an input signal (x(k)) sampled at preferably equidistant intervals over a short period, adopting the value if the input signal dynamic variation is below a definable threshold (epsilon) and adopting the value as the new noise level estimate n(x) if the dynamic variations exceed a threshold. Independent claims are also included for: a processor unit for supporting the method; and a programmable gate-array unit for supporting the method.

Description

The invention relates to a method for determining an estimated value for the noise level n of a background noise which is superimposed on an acoustic useful signal, in particular a human voice signal, transmitted via a telecommunication (TK) system, wherein in a first step a specifiable initialization value n0 is used as estimated value n (x) is assumed for a current noise level n;
in the next and optionally in further steps, the estimated value n (x) of the noise level n for an input signal x (k) sampled at preferably equidistant time steps T at times k at a sampling frequency fs = 1 / T is determined as a value n1 (x) the value n1 (x) is adopted as the estimated value n (x) for the current noise level n if the dynamic variations of the input signal x (k) fall below a predefinable threshold value; and wherein the estimated value n (x) determined in the preceding step is taken over unchanged as a new estimated value n (x) for the current noise level n if the dynamic variations of the input signal x (k) exceed a predefinable threshold value ε.
Furthermore, the invention also relates to computer programs and devices for supporting and carrying out such a method, in particular suitable server units, signaling devices, processor modules and programmable gate array modules.

Methods for noise estimation of background noise are known. For example, noise estimators are used where for the estimation the noise level of a signal of the averaged in a short time interval Value of the signal (SAM = short average magnitude) is used.

In other methods, the so-called. At longer time intervals MAM (= medium average magnitude) value of an input signal measured. To achieve a reliable result of the estimation, measurement times are up to required for 500 ms. Often the MAM value too high Noise level compared with the actual noise level before.

Generally, the value of the noise level of a signal is common to many algorithms for signal processing as a threshold or control value of great importance. The reliability and temporal behavior of a noise estimator have a big impact on the achievable quality of a signal processing algorithm. This is especially true in the field of speech recognition, to improve the recognition rate in the field of echo cancellation and for noise reduction. Application areas for noise estimators are, for example, exchanges, conference facilities, but also conventional phones or cell phones.

In Technical disclosure bulletin, Volume 29, No. 12 (1987), pages 5606 to 5609 An algorithm describing a noise wave is described automatically updated. A sound signal is low-pass filtered and a Power of the microphone signal is constantly determined. Remains this power constant over a longer period of time, the signal becomes background noise classified, and the noise wave is at a value just above the maximum measured microphone power set. However, the changes Microphone power strong, it is assumed that the sound signal off Speech exists and the noise wave is left unchanged.

A disadvantage of known estimation methods is the relatively slow behavior at averaging in noise estimator. Especially with voice activity with only short Speech pauses in periods of <100 ms often are not enough time to to grasp the "noise floor".

According to the ITU-T guideline G.168 so-called composite signals are used which consists of a sequence of signal bursts with a pause time of exist for about 100 ms. Again, with the previously known methods no exact noise estimation possible.

Another problem with the noise threshold is that of successful speech level estimation performed noise update with temporally changing Environmental conditions. The estimated noise value thus fluctuates in certain, sometimes relatively large, limits.

Object of the present invention is in contrast, a method of initially described type with the simplest possible means to the effect to further develop that as accurate as possible determination of the current noise level is achieved with the fastest possible adaptation times, the significant are lower than in known methods, and that to the least possible Calculation effort is required.

According to the invention this object is as surprisingly simple as effectively solved by the value n1 (x) by the Minimum value from the set of all successive each within a Short-term interval with a time length ts ≥ 1ms, preferably ts ≥ 3ms found maximum values of the input signal x (k) is determined; and that ts = 1 / fug, where fug is the lower limit frequency of the transmitting TK system is.

The method according to the invention thus in each case in a short-term interval the length ts is a maximum value from the samples of the input signal x (k) and for the estimation of the current noise level from the set of several maximum values found one after the other, in each case the Minimum n1 (x) used as an estimate n (x) for the current noise level n. In order to provide an estimated value n (x) before the first measuring period, an initialization value n0 is specified.

If the dynamic variations of the input signal, in particular caused due to major changes in the background noise, such as slamming a door, passing a lorry, etc. exceeds certain predetermined threshold ε, the estimation method in a sense "stopped" and in each case the last estimate taken, where the dynamics of the input signal x (k) below the given Threshold ε lag. This prevents erratic estimates Reason faster fluctuations in the signal come about. The Namely, the method according to the invention achieves an extremely fast adaptation to the current noise level in periods of about 10 ms, in contrast to the above-mentioned known methods, the times in the Order of magnitude of 500 ms.

It turns out that with the method according to the invention in particular also a correct calculation using the above-mentioned G168 "composite signals" with exact determination of the noise level and very fast adaptation times with extremely low computational complexity becomes.

The Erfinding is as shown in claims 1 and 8. In the method according to the invention, the time interval ts = 1 / fug is selected, where fug is the lower limit frequency of the transmitting TK system. In order to the envelope curve of the input signals can be optimally followed.

In particular, the time length ts should each be chosen such that a Adaptation of low-frequency signals in the range <100 Hz excluded becomes. Usually, the lower limit frequencies are in a range fug ≦ 500 Hz. In conventional telephony systems, for example, the lower Cutoff frequency 330 Hz. A value of about 10 Hz as lower limit for the lower Limit frequency fug corresponds to the value of a conventional hi-fi amplifier and is therefore reasonable.

An advantage for carrying out the method according to the invention is a Variant in which the initial value n0 is the maximum representable value of the Target system for the signal transmission within the telecommunications system is selected.

A further advantageous variant of the process according to the invention is characterized, is that n for the determination of the estimated value n (x) is the value n1 (x) at a predeterminable or fixed lower limit value set _min if a value n1 (x) <n _min is determined. As a result, misjudgements are reliably prevented in a simple manner, and thus a higher accuracy of the estimated value due to the range restriction is achieved.

This also applies to an upper limit that should be introduced to ensure distortion-free signal transmission. Accordingly, in a further variant of the method according to the invention, the value n1 (x) is set to a predefinable or fixed upper limit value n _max for the determination of the estimated value n (x), if a value n1 (x)> n _{max is} determined becomes.

Particularly preferred is a further development of this method variant, in which the upper limit value n _{max is} chosen to be smaller than or equal to the initialization value n 0, preferably n _max ≦ n 0 - 16 dB. For linear, distortion-free signal transmission in the relevant TC system, this upper limit value is predetermined by the statistically determined speech dynamics of human speech.

A further advantageous embodiment of the method according to the invention provides that the maximum values found within the short-term intervals of the input signal x (k) multiplied by a scaling factor S <1 Determine the value n1 (x). The majority of the actual level values that is actually below the maximum value determined in each case within the relevant short-term interval.

If the scaling factor S ≅ 0.5 is selected, this roughly corresponds to the position the maximum value of a statistical distribution, for example a Gaussian distribution the samples relative to the location of the found maximum Level value. As a result, the actual current noise level n becomes average Much better hit than by using the unscaled Maximum value.

For applications of the method according to the invention for secure speech pause detection it is advantageous if the estimated value n (x) as a measure of a currently estimated noise level is scaled by a factor D> 1.

By simulation, the most favorable values for factor D became application-dependent Values in the range 2 ≦ D ≦ 5, preferably 3 ≦ D ≦ 4 found. In order to Incidentally, there is a distance of about 6 dB between the speech signal and the statistically averaged noise signal, which is generally considered to be an acceptable signal-to-noise ratio applies.

Particularly preferred is also an embodiment of the invention Method in which a fixed threshold ε = const. is set, preferably ε ≈ 12dB. With this value obtained by simulations can be cover most practical applications well.

Alternatively to the introduction of a fixed threshold value ε can be at another advantageous method variant of the threshold ε = ε (x) adaptive with the Roughness of the level of the input signal x (k) to be changed. Leave it an optimal and extremely fast update and adaptation of the reach estimated level value to the actual noise conditions.

Advantageously, in a development of this method variant for the threshold value ε (x) to be determined adaptively selected a starting value ε0 = 12dB be as in the alternative process variant described above as constant fixed value is proposed.

The scope of the present invention also includes a server unit, a Processor assembly and a gate array assembly to support the method described above and a computer program to carry out the process. The method can be used both as a hardware circuit, as well as in the form of a computer program. Nowadays, software programming for powerful DSP's preferred because new knowledge and additional functions easier by a Modification of the software can be implemented on an existing hardware basis are. However, methods can also be used as hardware components, for example in IP or TK terminals or conventional telephone systems implemented become.

Further advantages of the invention will become apparent from the description and the Drawing. Likewise, the above and those listed further Features according to the invention each individually or for several find use in any combination. The shown and described Embodiments are not meant to be exhaustive. but rather have exemplary character for the description the invention.

The invention is illustrated in the drawing and is based on embodiments explained in more detail.

The figure shows a highly schematic diagram of the functioning of a Estimation device for carrying out the method according to the invention.

Starting from an initialization value n0, in a first short-term interval of the time length ts ≥ 1 ms, a first estimated value n1 (x) for the noise level n from a sampled input signal x (k) becomes the background noise superimposed on a useful signal in the input signal x (k) calculated according to the following equation:

Where K = fs / fug is the quotient of the sampling frequency of the sampled input signal x (k) and the lower limit frequency fug of the transmitting TK system. The length of the short-term interval is ts = 1 / fug. This will be represented over the running index k the smallest time interval that are observed so as not to adapt to low-frequency signals.

The value n1 (x) thus becomes the minimum of a previous value n1 (x) or an initialization value n0 and the maximum value of the one Scaling factor S ≈ 0.5 scaled amounts of the input signal x (k) in the interval k = 0 to k = K won.

In the event that voice activity is present in the input signal x (k), then Value n1 (x) assumes a value dependent on the speech level, since the Speech level is louder than the noise. For example, one acceptable Signal to noise ratio of 6 dB.

The value n1 (x) found in this way still changes with the language, responds but on noise reduction and during voice pauses with extremely short Adaptation time.

As the actual estimated value n (x) for the current noise level n, the value n1 (x) described above is adopted only if the dynamic variations of the input signal x (k) fall below a predefinable threshold value ε, that is, if dx (i) ... dx (i-ts) <ε

This condition controls dynamic level fluctuations of the examined Signal. For example, with a value ε = 12 dB becomes an update of the noise signal with level fluctuations> 12 dB prevented. In In this case, simply the previous estimate is unchanged for the current noise level n taken over. For example, this is the case when the background noise suddenly increases or decreases, so that the speech level estimator must become active. This can be ruled out that noise or speech peaks erroneously approximate the estimate n (x) change at short intervals.

The dynamic level fluctuations dx (i) described above can be determined, for example, from the difference successive successive short-term mean values sam (i) according to FIG dx (i) = sam (i) - sam (i-1)

If now the envelope of the incoming input signals x (i) is "stable", ie with almost certainly no speech signals, The current level values can be assigned directly to the background noise become. Otherwise, if the envelope "wobbles", it is very likely Language, ie predominant useful signal in the input signal x (i) before, so that the peaks of the input signal are not for the estimation of the noise background can be used. In this case, then, as above described, obtained from the speech signal itself a scaled noise value become.

The drawing now shows this process in a schematic way, in particular the maximum formation from the input signal x (k), the scaling with a Scaling factor S and the minimum formation for obtaining the value n1 (x), the assumption of this value as a function of a speech pause detector (SPD) whose output value may be an application-dependent one Factor D is scaled, as well as the threshold estimate of the dynamic Variations of the input signal x (k), in the example shown the temporal change of the short term mean dsam (x) / dt are obtained.

The output signal of this method is then the desired updated Estimated value n (x) for an actual noise level n.

Claims (9)

1. A process for determining an estimated value for the noise level n of a background noise which is superimposed on an acoustic useful signal, in particular a human speech signal, transmitted over a telecommunications (= TC) system, in which process
   in a first step a predeterminable initialisation value n0 is adopted as estimated value n(x) for a current noise level n;
   in the next step and optionally in further steps the estimated value n(x) of the noise level n for an input signal x(k), sampled in preferably equidistant time steps T in each case at times k with a sampling frequency fs = 1/T, is defined as a value n1(x);
   the value n1(x) is adopted as estimated value n(x) for the current noise level n when the dynamic variations of the input signal x(k) undershoot a predeterminable threshold value ε;
   and in which process the estimated value n(x) determined in the preceding step is adopted unchanged as new estimated value n(x) for the current noise level n when the dynamic variations of the input signal x(k) exceed a predeterminable threshold value ε,
   characterised in that
   the value n1(x) is determined by means of the minimum value of the quantity of all the successive maximum values of the input signal x(k) in each case found within a short time interval with a time length ts ≥ 1ms, preferably ts ≥ 3 ms; and that ts = 1/fug, where fug is the lower limit frequency of the transmitting TC system, and where fug ≤ 1000 Hz.
2. A process according to Claim 1, characterised in that fug ≤ 500 Hz, preferably fug ≤ 330 Hz and fug ≥ 10 Hz.
3. A process according to Claim 1, characterised in that the maximum representable value of the destination system for the signal transmission within the TC system is selected as initialisation value n0.
4. A process according to Claim 1, characterised in that for the determination of the estimated value n(x), the value n1(x) is set at a predeterminable or fixed lower limit value nmin if a value n1(x) < nmin is determined.
5. A process according to Claim 1, characterised in that for the determination of the estimated value n(x), the value n1(x) is set at a predeterminable or fixed upper limit value nmax if a value n1(x) > nmax is determined.
6. A process according to Claim 1, characterised in that the maximum values, found within the short time intervals, of the input signal x(k), multiplied by a scaling factor S < 1, enter into the determination of the value n1(x).
7. A process according to Claim 1, characterised in that a threshold value ε = ε(x) is changed adaptively with the roughness of the level of the input signal x(k).
8. Processor module, in particular digital signal processor (=DSP) with means for executing each individual step of the process according to one of the preceding claims.
9. Processor module according to Claim 8, characterised in that the processor module consists of a programmable gate array.

Images