NZ743390B2 - Estimation of background noise in audio signals - Google Patents
Estimation of background noise in audio signals Download PDFInfo
- Publication number
- NZ743390B2 NZ743390B2 NZ743390A NZ74339015A NZ743390B2 NZ 743390 B2 NZ743390 B2 NZ 743390B2 NZ 743390 A NZ743390 A NZ 743390A NZ 74339015 A NZ74339015 A NZ 74339015A NZ 743390 B2 NZ743390 B2 NZ 743390B2
- Authority
- NZ
- New Zealand
- Prior art keywords
- audio signal
- linear prediction
- background noise
- signal segment
- energy
- Prior art date
Links
- 230000005236 sound signal Effects 0.000 title claims abstract description 171
- 238000000034 method Methods 0.000 claims abstract description 59
- 230000007774 longterm Effects 0.000 claims description 65
- 230000000694 effects Effects 0.000 claims description 28
- 230000003595 spectral effect Effects 0.000 claims description 28
- 238000004891 communication Methods 0.000 claims description 12
- 238000001914 filtration Methods 0.000 claims description 12
- 238000004590 computer program Methods 0.000 claims description 9
- 230000000670 limiting effect Effects 0.000 claims description 6
- 230000005540 biological transmission Effects 0.000 claims description 3
- 238000010586 diagram Methods 0.000 description 22
- 238000012545 processing Methods 0.000 description 20
- 230000006870 function Effects 0.000 description 17
- 238000005516 engineering process Methods 0.000 description 12
- 238000001514 detection method Methods 0.000 description 11
- 101100355940 Xenopus laevis rcor1 gene Proteins 0.000 description 8
- 238000004458 analytical method Methods 0.000 description 7
- 230000002829 reductive effect Effects 0.000 description 7
- 238000004364 calculation method Methods 0.000 description 6
- 230000014509 gene expression Effects 0.000 description 6
- 230000009467 reduction Effects 0.000 description 6
- 206010019133 Hangover Diseases 0.000 description 5
- 101150014198 epsP gene Proteins 0.000 description 5
- GHOKWGTUZJEAQD-ZETCQYMHSA-N (D)-(+)-Pantothenic acid Chemical compound OCC(C)(C)[C@@H](O)C(=O)NCCC(O)=O GHOKWGTUZJEAQD-ZETCQYMHSA-N 0.000 description 4
- 230000008901 benefit Effects 0.000 description 4
- 230000000873 masking effect Effects 0.000 description 4
- 238000011084 recovery Methods 0.000 description 4
- 101000712600 Homo sapiens Thyroid hormone receptor beta Proteins 0.000 description 3
- 102100033451 Thyroid hormone receptor beta Human genes 0.000 description 3
- 230000003247 decreasing effect Effects 0.000 description 3
- 238000004422 calculation algorithm Methods 0.000 description 2
- 230000015556 catabolic process Effects 0.000 description 2
- 238000006731 degradation reaction Methods 0.000 description 2
- 238000009432 framing Methods 0.000 description 2
- 239000000203 mixture Substances 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000003491 array Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 239000006227 byproduct Substances 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 235000019800 disodium phosphate Nutrition 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000007429 general method Methods 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000001373 regressive effect Effects 0.000 description 1
- 230000008672 reprogramming Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 230000002441 reversible effect Effects 0.000 description 1
- 230000009291 secondary effect Effects 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 238000010183 spectrum analysis Methods 0.000 description 1
- 230000001629 suppression Effects 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
- G10L19/012—Comfort noise or silence coding
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
- G10L19/02—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis using spectral analysis, e.g. transform vocoders or subband vocoders
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
- G10L19/02—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis using spectral analysis, e.g. transform vocoders or subband vocoders
- G10L19/0204—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis using spectral analysis, e.g. transform vocoders or subband vocoders using subband decomposition
- G10L19/0208—Subband vocoders
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
- G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
- G10L21/0208—Noise filtering
- G10L21/0216—Noise filtering characterised by the method used for estimating noise
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
- G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
- G10L21/0316—Speech enhancement, e.g. noise reduction or echo cancellation by changing the amplitude
- G10L21/0324—Details of processing therefor
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
- G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
- G10L21/038—Speech enhancement, e.g. noise reduction or echo cancellation using band spreading techniques
- G10L21/0388—Details of processing therefor
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
- G10L25/03—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
- G10L25/03—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters
- G10L25/12—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters the extracted parameters being prediction coefficients
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
- G10L25/78—Detection of presence or absence of voice signals
Abstract
The invention relates to a background noise estimator and a method therein, for estimation of background noise in an audio signal. The method comprises computing at least one parameter associated with an audio signal segment that is among the audio signal segments, based on both of: a first linear prediction gain calculated as a quotient between an energy of the input signal and a residual signal energy from a first linear prediction for the audio signal segment; and a second linear prediction gain calculated as a quotient between the residual signal energy from the first linear prediction and a residual signal energy from a second linear prediction for the audio signal segment. The method further comprises determining whether the audio signal segment comprises a pause based at least on the obtained at least one parameter; and, updating a background noise estimate based on the audio signal segment when the audio signal segment comprises a pause. rediction gain calculated as a quotient between an energy of the input signal and a residual signal energy from a first linear prediction for the audio signal segment; and a second linear prediction gain calculated as a quotient between the residual signal energy from the first linear prediction and a residual signal energy from a second linear prediction for the audio signal segment. The method further comprises determining whether the audio signal segment comprises a pause based at least on the obtained at least one parameter; and, updating a background noise estimate based on the audio signal segment when the audio signal segment comprises a pause.
Description
ESTIMATION OF BACKGROUND NOISE IN AUDIO SIGNALS
This application has been divided out of New Zealand patent application 728080
(NZ 728080).
NZ 728080 is the national phase entry in New Zealand of PCT international
application (published as A1). The full
disclosure of A1 is incorporated herein by reference.
TECHNICAL FIELD
The embodiments of the present invention relate to audio signal processing, and in
particular to estimation of background noise, e.g. for supporting a sound activity
decision.
BACKGROUND
In communication systems utilizing discontinuous transmission (DTX) it is important
to find a balance between efficiency and not reducing quality. In such systems an
activity detector is used to indicate active signals, e.g. speech or music, which are to
be actively coded, and segments with background signals which can be replaced
with comfort noise generated at the receiver side. If the activity detector is too
efficient in detecting non-activity, it will introduce clipping in the active signal, which is
then perceived as subjective quality degradation when the clipped active segment is
replaced with comfort noise. At the same time, the efficiency of the DTX is reduced if
the activity detector is not efficient enough and classifies background noise segments
as active and then actively encodes the background noise instead of entering a DTX
mode with comfort noise. In most cases the clipping problem is considered worse.
Figure 1 shows an overview block diagram of a generalized sound activity detector,
SAD or voice activity detector, VAD, which takes an audio signal as input and
produces an activity decision as output. The input signal is divided into data frames,
i.e. audio signal segments of e.g. 5-30 ms, depending on the implementation, and
one activity decision per frame is produced as output.
A primary decision, "prim", is made by the primary detector illustrated in figure 1. The
primary decision is basically just a comparison of the features of a current frame with
background features, which are estimated from previous input frames. A difference
between the features of the current frame and the background features which is
larger than a threshold causes an active primary decision. The hangover addition
block is used to extend the primary decision based on past primary decisions to form
the final decision, "flag". The reason for using hangover is mainly to reduce/remove
the risk of mid and backend clipping of burst of activity. As indicated in the figure, an
operation controller may adjust the threshold(s) for the primary detector and the
length of the hangover addition according to the characteristics of the input signal.
The background estimator block is used for estimating the background noise in the
input signal. The background noise may also be referred to as “the background” or
“the background feature” herein.
Estimation of the background feature can be done according to two basically different
principles, either by using the primary decision, i.e. with decision or decision metric
feedback, which is indicated by dash-dotted line in figure 1, or by using some other
characteristics of the input signal, i.e. without decision feedback. It is also possible to
use combinations of the two strategies.
An example of a codec using decision feedback for background estimation is AMR-
NB (Adaptive Multi-Rate Narrowband) and examples of codecs where decision
feedback is not used are EVRC (Enhanced Variable Rate CODEC) and G.718.
There are a number of different signal features or characteristics that can be used,
but one common feature utilized in VADs is the frequency characteristics of the input
signal. A commonly used type of frequency characteristics is the sub-band frame
energy, due to its low complexity and reliable operation in low SNR. It is therefore
assumed that the input signal is split into different frequency sub-bands and the
background level is estimated for each of the sub-bands. In this way, one of the
background noise features is the vector with the energy values for each sub-band,
These are values that characterize the background noise in the input signal in the
frequency domain.
To achieve tracking of the background noise, the actual background noise estimate
update can be made in at least three different ways. One way is to use an Auto
Regressive, AR,-process per frequency bin to handle the update. Examples of such
codecs are AMR-NB and G.718. Basically, for this type of update, the step size of the
update is proportional to the observed difference between current input and the
current background estimate. Another way is to use multiplicative scaling of a current
estimate with the restriction that the estimate never can be bigger than the current
input or smaller than a minimum value. This means that the estimate is increased
each frame until it is higher than the current input. In that situation the current input is
used as estimate. EVRC is an example of a codec using this technique for updating
the background estimate for the VAD function. Note that EVRC uses different
background estimates for VAD and noise suppression. It should be noted that a VAD
may be used in other contexts than DTX. For example, in variable rate codecs, such
as EVRC, the VAD may be used as part of a rate determining function.
A third way is to use a so-called minimum technique where the estimate is the
minimum value during a sliding time window of prior frames. This basically gives a
minimum estimate which is scaled, using a compensation factor, to get and
approximate average estimate for stationary noise.
In high SNR cases, where the signal level of the active signal is much higher than the
background signal, it may be quite easy to make a decision of whether an input audio
signal is active or non-active. However, to separate active and non-active signals in
low SNR cases, and in particular when the background is non-stationary or even
similar to the active signal in its characteristics, is very difficult.
The performance of the VAD depends on the ability of the background noise
estimator to track the characteristics of the background – in particular when it comes
to non-stationary backgrounds. With better tracking it is possible to make the VAD
more efficient without increasing the risk of speech clipping.
While correlation is an important feature that is used to detect speech, mainly the
voiced part of the speech, there are also noise signals that show high correlation. In
these cases the noise with correlation will prevent update of background noise
estimates. The result is a high activity as both speech and background noise is
coded as active content. While for high SNRs (approximately >20dB) it would be
possible to reduce the problem using energy based pause detection, this is not
reliable for the SNR range 20dB down to 10dB or possibly 5dB. It is in this range that
the solution described herein makes a difference.
SUMMARY
A first aspect of the present invention provides a method for a background noise
estimator for estimation of background noise in an audio signal, wherein the audio
signal comprises a plurality of audio signal segments, the method comprising:
computing at least one parameter associated with an audio signal segment that is
among the audio signal segments, based on both of: a first linear prediction gain
calculated as a quotient between an energy of the input signal and a residual signal
energy from a first linear prediction for the audio signal segment; and a second linear
prediction gain calculated as a quotient between the residual signal energy from the
first linear prediction and a residual signal energy from a second linear prediction for
the audio signal segment; determining whether the audio signal segment comprises
a pause free of speech and music, based at least on the at least one parameter; and
responsive to when the audio signal segment is determined to comprise a pause,
updating to obtain an updated background noise estimate based on the audio signal
segment. Computing the at least one parameter comprises determining a difference
between two long term estimates associated with one of the linear prediction gains.
A second aspect of the present invention provides a background noise estimator, for
estimating background noise in an audio signal comprising a plurality of audio signal
segments, the background noise estimator comprising: at least one processor; and at
least one memory storing computer readable instructions executed by the at least
one processor to perform operations comprising: compute at least one parameter
based on both of: a first linear prediction gain calculated as a quotient between an
energy of the input signal and a residual signal energy from a first linear prediction for
the audio signal segment; and a second linear prediction gain calculated as a
quotient between the residual signal energy from the first linear prediction and a
residual signal energy from a second linear prediction for the audio signal segment;
determine whether the audio signal segment comprises a pause free of speech and
music, based at least on the at least one parameter; and responsive to when the
audio signal segment is determined to comprise a pause, updating to obtain an
updated a background noise estimate based on the audio signal segment.
Computing the at least one parameter comprises determining a difference between
two long term estimates associated with one of the linear prediction gains.
A third aspect of the present invention provides a Sound Activity Detector (SAD)
comprising a background noise estimator according to the second aspect of the
present invention.
A fourth aspect of the present invention provides a codec comprising a background
noise estimator according to the second aspect of the present invention.
A fifth aspect of the present invention provides a computer program product
comprising a non-transitory computer readable storage medium storing instructions
which, when executed on at least one processor, cause the at least one processor to
perform operations comprising: computing at least one parameter associated with an
audio signal segment that is among the audio signal segments, based on both of: a
first linear prediction gain calculated as a quotient between an energy of the input
signal and a residual signal energy from a first linear prediction for the audio signal
segment; and a second linear prediction gain calculated as a quotient between the
residual signal energy from the first linear prediction and a residual signal energy
from a second linear prediction for the audio signal segment; determining whether
the audio signal segment comprises a pause free of speech and music, based at
least on the at least one parameter; and responsive to when the audio signal
segment is determined to comprise a pause, updating to obtain an updated
background noise estimate based on the audio signal segment. Computing the at
least one parameter comprises determining a difference between two long term
estimates associated with one of the linear prediction gains.
It would be desirable to achieve improved estimation of background noise in audio
signals. “Improved” may here imply making more correct decision in regard of
whether an audio signal comprises active speech or music or not, and thus more
often estimating, e.g. updating a previous estimate, the background noise in audio
signal segments actually being free from active content, such as speech and/or
music. Herein, an improved method for generating a background noise estimate is
provided, which may enable e.g. a sound activity detector to make more adequate
decisions.
For background noise estimation in audio signals, it is important to be able to find
reliable features to identify the characteristics of a background noise signal also
when an input signal comprises an unknown mixture of active and background
signals, where the active signals can comprise speech and/or music.
The inventor has realized that features related to residual energies for different linear
prediction model orders may be utilized for detecting pauses in audio signals. These
residual energies may be extracted e.g. from a linear prediction analysis, which is
common in speech codecs. The features may be filtered and combined to make a set
of features or parameters that can be used to detect background noise, which makes
the solution suitable for use in noise estimation. The solution described herein is
particularly efficient for the conditions when an SNR is in the range of 10 to 20 dB.
Another feature provided herein is a measure of spectral closeness to background,
which may be made e.g. by using the frequency domain sub-band energies which
are used e.g. in a sub-band SAD. The spectral closeness measure may also be used
for making a decision of whether an audio signal comprises a pause or not.
According to a first aspect of the present disclosure, a method for background noise
estimation is provided. The method comprises obtaining at least one parameter
associated with an audio signal segment, such as a frame or part of a frame, based
on a first linear prediction gain, calculated as a quotient between a residual signal
from a 0th-order linear prediction and a residual signal from a 2nd-order linear
prediction for the audio signal segment; and, a second linear prediction gain
calculated as a quotient between a residual signal from a 2nd-order linear prediction
and a residual signal from a 16th-order linear prediction for the audio signal segment.
The method further comprises determining whether the audio signal segment
comprises a pause based at least on the obtained at least one parameter; and,
updating a background noise estimate based on the audio signal segment when the
audio signal segment comprises a pause.
According to a second aspect of the present disclosure, a background noise
estimator is provided. The background noise estimator is configured to obtain at least
one parameter associated with an audio signal segment based on a first linear
prediction gain, calculated as a quotient between a residual signal from a 0th-order
linear prediction and a residual signal from a 2nd-order linear prediction for the audio
signal segment; and, a second linear prediction gain calculated as a quotient
between a residual signal from a 2nd-order linear prediction and a residual signal
from a 16th-order linear prediction for the audio signal segment. The background
noise estimator is further configured to determine whether the audio signal segment
comprises a pause based at least on the obtained at least one parameter; and, to
update a background noise estimate based on the audio signal segment when the
audio signal segment comprises a pause.
According to a third aspect of the present disclosure, a SAD is provided, which
comprises a background noise estimator according to the second aspect of the
present disclosure.
According to a fourth aspect of the present disclosure, a codec is provided, which
comprises a background noise estimator according to the second aspect of the
present disclosure.
According to a fifth aspect of the present disclosure, a communication device is
provided, which comprises a background noise estimator according to the second
aspect of the present disclosure.
According to a sixth aspect of the present disclosure, a network node is provided,
which comprises a background noise estimator according to the second aspect of the
present disclosure.
According to a seventh aspect of the present disclosure, a computer program is
provided, comprising instructions which, when executed on at least one processor,
cause the at least one processor to carry out the method according to the first aspect
of the present disclosure.
According to an eighth aspect of the present disclosure, a carrier is provided, which
contains a computer program according to the seventh aspect of the present
disclosure.
BRIEF DESCRIPTION OF DRAWINGS
The foregoing and other objects, features, and advantages of the technology
disclosed herein will be apparent from the following more particular description of
embodiments as illustrated in the accompanying drawings. The drawings are not
necessarily to scale, emphasis instead being placed upon illustrating the principles of
the technology disclosed herein.
Figure 1 is a block diagram illustrating an activity detector and hangover
determination logic.
Figure 2 is a flow chart illustrating a method for estimation of background noise,
according to an exemplifying embodiment.
Figure 3 is a block diagram illustrating calculation of features related to the residual
energies for linear prediction of order 0 and 2 according to an exemplifying
embodiment.
Figure 4 is a block diagram illustrating calculation of features related to the residual
energies for linear prediction of order 2 and 16 according to an exemplifying
embodiment.
Figure 5 is a block diagram illustrating calculation of features related to a spectral
closeness measure according to an exemplifying embodiment.
Figure 6 is a block diagram illustrating a sub-band energy background estimator.
Figure 7 is a flow chart illustrating a background update decision logic from the
solution described in Annex A.
Figures 8-10 are diagrams illustrating the behaviour of different parameters
presented herein when calculated for an audio signal comprising two speech bursts.
Figures 11a-11c and 12-13 are block diagrams illustrating different implementations
of a background noise estimator according to exemplifying embodiments.
Figures 14-21 on figure pages marked “Annex A” are associated with Annex A.
DETAILED DESCRIPTION
The solution disclosed herein relates to estimation of background noise in audio
signals. In the generalized activity detector illustrated in figure 1, the function of
estimating background noise is performed by the block denoted “background
estimator”. Some embodiments of the solution described herein may be seen in
relation to solutions previously disclosed in W02011/049514, W02011/049515, which
are incorporated herein by reference, and also in Annex A (Appendix A). The solution
disclosed herein will be compared to implementations of these previously disclosed
solutions. Even though the solutions disclosed in W02011/049514, W02011/049515
and Annex A are good solutions, the solution presented herein still has advantages in
relation to these solutions. For example, the solution presented herein is even more
adequate in its tracking of background noise.
The performance of a VAD depends on the ability of the background noise estimator
to track the characteristics of the background — in particular when it comes to non-
stationary backgrounds. With better tracking it is possible to make the VAD more
efficient without increasing the risk of speech clipping.
One problem with current noise estimation methods is that to achieve good tracking
of the background noise in low SNR, a reliable pause detector is needed. For speech
only input, it is possible to utilize the syllabic rate or the fact that a person cannot talk
all the time to find pauses in the speech. Such solutions could involve that after a
sufficient time of not making background updates, the requirements for pause
detection are “relaxed”, such that it is more probable to detect a pause in the speech.
This allows for responding to abrupt changes in the noise characteristics or level.
Some examples of such noise recovery logics are: 1) As speech utterances contain
segments with high correlation, it is usually safe to assume that there is a pause in
the speech after a sufficient number of frames without correlation. 2) When the
Signal to Noise Ratio, SNR >0, the speech energy is higher than the background
noise, so if the frame energy is close to the minimum energy over a longer time, e.g.
1-5 seconds, it is also safe to assume that one is in a speech pause. While the
previous techniques work well with speech only input they are not sufficient when
music is considered an active input. In music there can be long segments with low
correlation that still are music. Further, the dynamics of the energy in music can also
trigger false pause detection, which may result in unwanted, erroneous updates of
the background noise estimate.
Ideally, an inverse function of an activity detector, or what would be called a “pause
occurrence detector”, would be needed for controlling the noise estimation. This
would ensure that the update of the background noise characteristics is done only
when there is no active signal in the current frame. However, as indicated above, it is
not an easy task to determine whether an audio signal segment comprises an active
signal or not.
Traditionally, when the active signal was known to be a speech signal, the activity
detector was called Voice Activity Detector (VAD). The term VAD for activity
detectors is often used also when the input signal may comprise music. However, in
modern codecs, it is also common to refer to the activity detector as a Sound Activity
Detector (SAD) when also music is to be detected as an active signal.
The background estimator illustrated in figure 1 utilizes feedback from the primary
detector and/or the hangover block to localize inactive audio signal segments. When
developing the technology described herein, it has been a desire to remove, or at
least reduce the dependency on such feedback. For the herein disclosed background
estimation it has therefore been identified by the inventor as important to be able to
find reliable features to identify the background signals characteristics when only an
input signal with an unknown mixture of active and background signal is available.
The inventor has further realized that it cannot be assumed that the input signal
starts with a noise segment, or even that the input signal is speech mixed with noise,
as it may be that the active signal is music.
One aspect is that even though the current frame may have the same energy level as
the current noise estimate, the frequency characteristics may be very different, which
makes it undesirable to perform an update of the noise estimate using the current
frame. The introduced closeness feature relative background noise update can be
used to prevent updates in these cases.
Further, during initialization it is desirable to allow the noise estimation to start as
soon as possible while avoiding wrong decisions as this potentially could result in
clipping from the SAD if the background noise update is made using active content.
Using an initialization specific version of the closeness feature during initialization
can at least partly solve this problem.
The solution described herein relates to a method for background noise estimation, in
particular to a method for detecting pauses in an audio signal which performs well in
difficult SNR situations. The solution will be described below with reference to figures
2-5.
In the field of speech coding, it is common to use so-called linear prediction to
analyze the spectral shape of an input signal. The analysis is typically made two
times per frame, and for improved temporal accuracy the results are then
interpolated such that there is a filter generated for each 5 ms block of the input
signal.
Linear prediction is a mathematical operation, where future values of a discrete-time
signal are estimated as a linear function of previous samples. In digital signal
processing, linear prediction is often called linear predictive coding (LPC) and can
thus be viewed as a subset of filter theory. In linear prediction in a speech coder, a
linear prediction filter A(z) is applied to an input speech signal. A(z) is an all zero filter
that when applied to the input signal removes the redundancy that can be modeled
using the filter A(z) from the input signal. Therefore the output signal from the filter
has lower energy than the input signal when the filter is successful in modelling some
aspect or aspects of the input signal. This output signal is denoted “the residual”, “the
residual energy” or “the residual signal”. Such linear prediction filters, alternatively
denoted residual filters, may be of different model order having different number of
filter coefficients. For example, in order to properly model speech, a linear prediction
filter of model order 16 may be required. Thus, in a speech coder, a linear prediction
filter A(z) of model order 16 may be used.
The inventor has realized that features related to linear prediction may be used for
detecting pauses in audio signals in an SNR range of 20dB down to 10dB or possibly
5dB. According to embodiments of the solution described herein, a relation between
residual energies for different model orders for an audio signal is utilized for detecting
pauses in the audio signal. The relation used is the quotient between the residual
energy of a lower model order and a higher model order. The quotient between
residual energies may be referred to as the “linear prediction gain”, since it is an
indicator of how much of the signal energy that the linear prediction filter has been
able to model, or remove, between one model order and another model order.
The residual energy will depend on the model order M of the linear prediction filter
A(z). A common way of calculating the filter coefficients for a linear prediction filter is
the Levinson-Durbin algorithm. This algorithm is recursive and will in the process of
creating a prediction filter A(z) of order M also, as a “by-product”, produce the
residual energies of the lower model orders. This fact may be utilized according to
embodiments of the invention.
Figure 2 shows an exemplifying general method for estimation of background noise
in an audio signal. The method may be performed by a background noise estimator.
The method comprises obtaining 201 at least one parameter associated with an
audio signal segment, such as a frame or part of a frame, based on a first linear
prediction gain, calculated as a quotient between a residual signal from a 0th-order
linear prediction and a residual signal from a 2nd-order linear prediction for the audio
signal segment; and, a second linear prediction gain calculated as a quotient
between a residual signal from a 2nd-order linear prediction and a residual signal
from a 16th-order linear prediction for the audio signal segment.
The method further comprises determining 202 whether the audio signal segment
comprises a pause, i.e. is free from active content such as speech and music, based
at least on the obtained at least one parameter; and, updating 203 a background
noise estimate based on the audio signal segment when the audio signal segment
comprises a pause. That is, the method comprises updating of a background noise
estimate when a pause is detected in the audio signal segment based at least on the
obtained at least one parameter.
The linear prediction gains could be described as a first linear prediction gain related
to going from 0th-order to 2nd-order linear prediction for the audio signal segment;
and a second linear prediction gain related to going from 2nd-order to 16th-order
linear prediction for the audio signal segment. Further, the obtaining of the at least
one parameter could alternatively be described as determining, calculating, deriving
or creating. The residual energies related to linear predictions of model order 0, 2
and 16 may be obtained, received or retrieved from, i.e. somehow provided by, a part
of the encoder where linear prediction is performed as part of a regular encoding
process. Thereby, the computational complexity of the solution described herein may
be reduced, as compared to when the residual energies need to be derived
especially for the estimation of background noise.
The at least one parameter obtained based on the linear prediction features may
provide a level independent analysis of the input signal that improves the decision for
whether to perform a background noise update or not. The solution is particularly
useful in the SNR range 10 to 20dB, where energy based SADs have limited
performance due to the normal dynamic range of speech signals.
Herein, among others, the variables E(0), …,E(m), …, E(M) represent the residual
energies for model orders 0 to M of the M+1 filters Am(z). Note that E(0) is just the
input energy. An audio signal analysis according to the solution described herein
provides several new features or parameters by analyzing the linear prediction gain
calculated as a quotient between a residual signal from a 0th-order linear prediction
and a residual signal from a 2nd-order linear prediction, and the linear prediction gain
calculated as a quotient between a residual signal from a 2nd-order linear prediction
and a residual signal from a 16th-order linear prediction. That is, the linear prediction
gain for going from 0th-order to 2nd-order linear prediction is the same thing as the
“residual energy” E(0) (for a 0th model order) divided by the residual energy E(2) (for
a 2nd model order). Correspondingly, the linear prediction gain for going from 2nd-
order linear prediction to the 16th order linear prediction is the same thing as the
residual energy E(2) (for a 2nd model order) divided by the residual energy E(16) (for
a 16th model order). Examples of parameters and the determining of parameters
based on the prediction gains will be described in more detail further below. The at
least one parameter obtained according to the general embodiment described above
may form a part of a decision criterion used for evaluating whether to update the
background noise estimate or not.
In order to improve a long-term stability of the at least one parameter or feature, a
limited version of the predictions gain can be calculated. That is, the obtaining of the
at least one parameter may comprise limiting the linear prediction gains, related to
going from 0th-order to 2nd-order and from 2nd-order to 16th-order linear prediction,
to take on values in a predefined interval. For example, the linear prediction gains
may be limited to take on values between 0 and 8, as illustrated e.g. in Eq.1 and Eq.6
below.
The obtaining of the at least one parameter may further comprise creating at least
one long term estimate of each of the first and second linear prediction gain, e.g. by
means of low pass filtering. Such at least one long term estimate would then be
further based on corresponding linear prediction gains associated with at least one
preceding audio signal segment. More than one long term estimate could be created,
where e.g. a first and a second long term estimate related to a linear prediction gain
react differently on changes in the audio signal. For example a first long term
estimate may react faster on changes than a second long term estimate. Such a first
long term estimate may alternatively be denoted a short term estimate.
The obtaining of the at least one parameter may further comprise determining a
difference, such as the absolute difference Gd_0_2 (Eq.3) described below, between
one of the linear prediction gains associated with the audio signal segment, and a
long term estimate of said linear prediction gain. Alternatively or in addition, a
difference between two long term estimates could be determined, such as in Eq.9
below. The term determining could alternatively be exchanged for calculating,
creating or deriving.
The obtaining of the at least one parameter may as indicated above comprise low
pass filtering of the linear prediction gains, thus deriving long term estimates, of
which some may alternatively be denoted short term estimates, depending on how
many segments that are taken into consideration in the estimate The filter
coefficients of at least one low pass filter may depend on a relation between a linear
prediction gain related, e.g. only, to the current audio signal segment and an
average, denoted e.g. long term average, or long term estimate, of a corresponding
prediction gain obtained based on a plurality of preceding audio signal segments.
This may be performed to create, e.g. further, long term estimates of the prediction
gains. The low pass filtering may be performed in two or more steps, where each
step may result in a parameter, or estimate, that is used for making a decision in
regard of the presence of a pause in the audio signal segment. For example, different
long term estimates (such as G1_0_2 (Eq.2) and Gad_0_2 (Eq.4), and/or, G1_2_16
(Eq.7), G2_2_16 (Eq.8) and Gad_2_16 (Eq.10) described below) which reflect
changes in the audio signal in different ways, may be analyzed or compared in order
to detect a pause in a current audio signal segment.
The determining 202 of whether the audio signal segment comprises a pause or not
may further be based on a spectral closeness measure associated with the audio
signal segment. The spectral closeness measure will indicate how close the “per
frequency band” energy level of the currently processed audio signal segment is to
the “per frequency band” energy level of the current background noise estimate, e.g.
an initial value or an estimate which is the result of a previous update made before
the analysis of the current audio signal segment. An example of determining or
deriving of a spectral closeness measure is given below in equations Eq.12 and
Eq.13. The spectral closeness measure can be used to prevent noise updates based
on low energy frames with a large difference in frequency characteristics, as
compared to the current background estimate. For example, the average energy over
the frequency bands could be equally low for the current signal segment and the
current background noise estimate, but the spectral closeness measure would reveal
if the energy is differently distributed over the frequency bands. Such a difference in
energy distribution could suggest that the current signal segment, e.g. frame, may be
low level active content and an update of the background noise estimate based on
the frame could e.g. prevent detection of future frames with similar content. As the
sub-band SNR is most sensitive to increases of energy using even low level active
content can result in a large update of the background estimate if that particular
frequency range is non-existent in the background noise, such as the high frequency
part of speech compared to low frequency car noise. After such an update it will be
more difficult to detect the speech.
As already suggested above, the spectral closeness measure may be derived,
obtained or calculated based on energies for a set of frequency bands, alternatively
denoted sub-bands, of the currently analyzed audio signal segment and current
background noise estimates corresponding to the set of frequency bands. This will
also be exemplified and described in more detail further below, and is illustrated in
figure 5.
As indicated above, the spectral closeness measure may be derived obtained or
calculated by comparing a current per frequency band energy level of the currently
processed audio signal segment with a per frequency band energy level of a current
background noise estimate. However, to start with, i.e. during a first period or a first
number of frames in the beginning of analyzing an audio signal, there may be no
reliable background noise estimate, e.g. since no reliable update of a background
noise estimate will have been performed yet. Therefore, an initialization period may
be applied for determining the spectral closeness value. During such an initialization
period, the per frequency band energy levels of the current audio signal segment will
instead be compared with an initial background estimate, which may be e.g. a
configurable constant value. In the examples further below, this initial background
noise estimate is set to the exemplifying value Emin=0,0035. After the initialization
period the procedure may switch to normal operation, and compare the current per
frequency band energy level of the currently processed audio signal segment with a
per frequency band energy level of a current background noise estimate. The length
of the initialization period may be configured e.g. based on simulations or tests
indicating the time it takes before an, e.g. reliable and/or satisfying, background noise
estimate is provided. An example used below, the comparison with an initial
background noise estimate (instead of with a “real” estimate derived based on the
current audio signal) is performed during the first 150 frames.
The at least one parameter may be the parameter exemplified in code further below,
denoted NEW_POS_BG, and/or one or more of the plurality of parameters described
further below, leading to the forming of a decision criterion or a component in a
decision criterion for pause detection. In other words, the at least one parameter, or
feature, obtained 201 based on the linear prediction gains may be one or more of the
parameters described below, may comprise one or more of the parameters described
below and/or be based on one or more of the parameters described below.
Features or parameters related to the residual energies E(0) and E(2)
Figure 3 shows an overview block diagram of the deriving of features or parameters
related to E(0) and E(2), according to an exemplifying embodiment. As can be seen
in figure 3, the prediction gain is first calculated as E(0)/E(2). A limited version of the
predictions gain is calculated as
G_0_2=max(0,min(8,E(0)/E(2))) (Eq 1)
where E(0) represents the energy of the input signal and E(2) is the residual energy
after a 2nd order linear prediction. The expression in equation 1 limits the prediction
gain to an interval between 0 and 8. The prediction gain should for normal cases be
larger than zero, but anomalies may occur e.g. for values close to zero, and therefore
a “larger than zero” limitation (0<) may be useful. The reason for limiting the
prediction gain to a maximum of 8 is that, for the purpose of the solution described
herein, it is sufficient to know that the prediction gain is about 8 or larger than 8,
which indicates a significant linear prediction gain. It should be noted that when there
is no difference between the residual energy between two different model orders, the
linear prediction gain will be 1, which indicates that the filter of a higher model order
is not more successful in modelling the audio signal than the filter of a lower model
order. Further, if the prediction gain G_0_2 would take on too large values in the
following expressions it may risk the stability of the derived parameters. It should be
noted that 8 is just an example value, which has been selected for a specific
embodiment. The parameter G_0_2 could alternatively be denoted e.g. epsP_0_2, or
The limited prediction gain is then filtered in two steps to create long term estimates
of this gain. The first low pass filtering and thus the deriving of a first long term
feature or parameter is made as:
G1_0_2=0.85 G1_0_2 + 0.15 G_0_2, (Eq. 2)
Where the second “G1_0_2” in the expression should be read as the value from a
preceding audio signal segment. This parameter will typically be either 0 or 8,
depending on the type of background noise in the input once there is a segment of
background-only input. The parameter G1_0_2 could alternatively be denoted e.g.
epsP_0_2_lp or � ̅ . Another feature or parameter may then be created or
calculated using the difference between the first long term feature G1_0_2 and the
frame by frame limited prediction gain G_0_2, according to:
Gd_0_2=abs(G1_0_2-G_0_2) (Eq. 3)
This will give an indication of the current frame’s prediction gain as compared to the
long term estimate of the prediction gain. The parameter Gd_0_2 could alternatively
be denoted e.g. epsP_0_2_ad or � . In figure 4, this difference is used to
create a second long term estimate or feature Gad_0_2. This is done using a filter
applying different filter coefficients depending on if the long term difference is higher
or lower than the currently estimated average difference according to:
Gad_0_2 = (1-a) Gad_0_2 + a Gd_0_2 (Eq. 4)
where, if Gd_0_2 < Gad_0_2 then a=0.1 else a=0.2
Where the second “Gad_0_2” in the expression should be read as the value from a
preceding audio signal segment.
The parameter Gad_0_2 could alternatively be denoted e.g. Glp_0_2,
epsP_0_2_ad_lp or � ̅ In order to prevent the filtering from masking occasional
high frame differences another parameter may be derived, which is not shown in the
figure. That is, the second long term feature Gad_0_2 may be combined with the
frame difference in order to prevent such masking. This parameter may be derived by
taking the maximum of the frame version Gd_0_2 and the long term version
Gad_0_2 of the prediction gain feature, as:
Gmax_0_2 = max(Gad_0_2,Gd_0_2) (Eq. 5)
The parameter Gmax_0_2 could alternatively be denoted e.g. epsP_0_2_ad_lp_max
or � .
Features or parameters related to the residual energies E(2) and E(16)
Figure 4 shows an overview block diagram of the deriving of features or parameters
related to E(2) and E(16), according to an exemplifying embodiment. As can be seen
in figure 4, the prediction gain is first calculated as E(2)/E(16). The features or
parameters created using the difference or relation between the 2 order residual
energy and the 16th order residual energy is derived slightly differently than the ones
described above related to the relation between the 0th and 2nd order residual
energies.
Here, as well, a limited prediction gain is calculated as
G_2_16 = max(0,min(8,E(2)/E(16))) (Eq. 6)
where E(2) represents the residual energy after a 2nd order linear prediction and
E(16) represents the residual energy after a 16th order linear prediction. The
parameter G_2_16 could alternatively be denoted e.g. epsP_2_16 or � . This
limited prediction gain is then used for creating two long term estimates of this gain:
one where the filter coefficient differs if the long term estimate is to be increased or
not as shown in:
G1_2_16=(1-a) G1_2_16 + a G_2_16 (Eq. 7)
where if G_2_16 > G1_2_16 then a=0.2 else a=0.03
The parameter G1_2_16 could alternatively be denoted e.g. epsP_2_16_lp or
� ̅ .
The second long term estimate uses a constant filter coefficient as according to:
G2_2_16=(1-b) G2_2_16 + b G_2_16, where b=0.02 (Eq. 8)
The parameter G2_2_16 could alternatively be denoted e.g. epsP_2_16_lp2 or
� ̅ .
For most types of background signals, both G1_2_16 and G2_2_16 will be close to 0,
but they will have different responses to content where the 16th order linear
prediction is needed, which is typically for speech and other active content. The first
long term estimate, G1_2_16, will usually be higher than the second long term
estimate G2_2_16. This difference between the long term features is measured
according to:
Gd_2_16 = G1_2_16 - G2_2_16 (Eq. 9)
The parameter Gd_2_16 could alternatively be denoted epsP_2_16_dlp or
Gd_2_16 may then be used as an input to a filter which creates a third long term
feature according to:
Gad_2_16 = (1-c) Gad_2_16 + c Gd_2_16 (Eq. 10)
where if Gd_2_16 < Gad_2_16 then c=0.02 else c=0.05
This filter applies different filter coefficients depending on if the third long term signal
is to be increased or not. The parameter Gad_2_16 may alternatively be denoted e.g.
epsP_2_16_dlp_lp2 or � ̅ . Also here, the long term signal Gad_2_16 may be
combined with the filter input signal Gd_2_16 to prevent the filtering from masking
occasional high inputs for the current frame. The final parameter is then the
maximum of the frame or segment and the long term version of the feature
Gmax_2_16 = max(Gad_2_16, Gd_2_16) (Eq. 11)
The parameter Gmax_2_16 could alternatively be denoted e.g. epsP_2_16_dlp_max
or �
Spectral closeness/difference measure
A spectral closeness feature uses the frequency analysis of the current input frame
or segment where sub-band energy is calculated and compared to the sub-band
background estimate. A spectral closeness parameter or feature may be used in
combination with a parameter related to the linear prediction gains described above
e.g. to make sure that the current segment or frame is relatively close to, or at least
not too far from, a previous background estimate.
Figure 5 shows a block diagram of the calculation of a spectral closeness or
difference measure. During the initialization period, e.g. the 150 first frames, the
comparison is made with a constant corresponding to the initial background estimate.
After the initialization it goes to normal operation and compares with the background
estimate. Note that while the spectral analysis produces sub-band energies for 20
sub-bands, the calculation of nonstaB here only uses sub-bands i=2, … 16, since it is
mainly in these bands that speech energy is located. Here nonstaB reflects the non-
stationarity.
So, during initialization, nonstaB is calculated using an Emin, which here is set to
Emin=0.0035 as:
nonstaB = sum(abs(log(Ecb(i)+1)-log(Emin+1))) (Eq. 12)
where sum is made over i=2…16.
This is done to reduce the effect of decision errors in the background noise
estimation during initialization. After the initialization period the calculation is made
using the current background noise estimate of the respective sub-band, according
nonstaB = sum(abs(log(Ecb(i)+1)-log(Ncb(i)+1))) (Eq. 13)
where sum is made over i=2...16
The addition of the constant 1 to each sub-band energy before the logarithm reduces
the sensitivity for the spectral difference for low energy frames. The parameter
nonstaB could alternatively be denoted e.g. non_staB or � �� .
A block diagram illustrating an exemplifying embodiment of a background estimator
is shown in figure 6. The embodiment in figure 6 comprises a block for Input Framing
601, which divides the input audio signal into frames or segments of suitable length,
e.g. 5-30 ms. The embodiment further comprises a block for Feature Extraction 602
that calculates the features, also denoted parameters herein, for each frame or
segment of the input signal. The embodiment further comprises a block for Update
Decision Logic 603, for determining whether or not a background estimate may be
updated based on the signal in the current frame, i.e. whether the signal segment is
free from active content such as speech and music. The embodiment further
comprises a Background Updater 604, for updating the background noise estimate
when the update decision logic indicates that it is adequate to do so. In the illustrated
embodiment, a background noise estimate may be derived per sub-band, i.e. for a
number of frequency bands.
The solution described herein may be used to improve a previous solution for
background noise estimation, described in Annex A herein, and also in the document
WO2011/049514. Below, the solution described herein will be described in the
context of this previously described solution. Code examples from a code
implementation of an embodiment of a background noise estimator will be given.
Below, actual implementation details are described for an embodiment of the
invention in a G.718 based encoder. This implementation uses many of the energy
features described in the solution in Annex A and WO2011/049514 incorporated
����
herein by reference. For further details than presented below, we refer to Annex A
and WO2011/049514.
The following energy features are defined in W02011/049514:
Etot;
Etot_l_lp;
Etot_v_h;
totalNoise;
sign_dyn_lp;
The following correlation features are defined in W02011/049514:
aEn;
harm_cor_cnt
act_pred
cor_est
The following features were defined in the solution given in Annex A:
Etot_v_h;
lt_cor_est = 0.01f*cor_est + 0.99f*lt_cor_est;
lt_tn_track = 0.03f* (Etot - totalNoise < 10) + 0.97f*lt_tn_track;
lt_tn_dist = 0.03f* (Etot - totalNoise) + 0.97f*lt_tn_dist;
lt_Ellp_dist = 0.03f* (Etot - Etot_l_lp) + 0.97f*lt_Ellp_dist;
harm_cor_cnt
low_tn_track_cnt
The noise update logic from the solution given in Annex A is shown in figure 7. The
improvements, related to the solution described herein, of the noise estimator of
Annex A are mainly related to the part 701 where features are calculated; the part
702, where pause decisions are made based on different parameters; and further to
the part 703, where different actions are taken based on whether a pause is detected
or not. Further, the improvements may have an effect on the updating 704 of the
background noise estimate, which could e.g. be updated when a pause is detected
based on the new features, which would not have been detected before introducing
the solution described herein. In the exemplifying implementation described here, the
new features introduced herein are calculated as follows, starting with non_staB,
which is determined using the current frame’s sub-band energies enr[i], which
corresponds to Ecb(i) above and in figure 6, and the current background noise
estimate bckr[i], which corresponds to Ncb(i) above and in figure 6. The first part of
the first code section below is related to a special initial procedure for the first 150
frames of an audio signal, before a proper background estimate has been derived.
/* calculate non-stationarity feature relative background (spectral closeness feature non_staB */
if (ini_frame < 150)
{
/* During init don't include updates */
if ( i >= 2 && i <= 16 )
non_staB += (float)fabs(log(enr[i] + 1.0f) -
log(E_MIN + 1.0f));
else
/* After init compare with background estimate */
if ( i >= 2 && i <= 16 )
non_staB += (float)fabs(log(enr[i] + 1.0f) -
log(bckr[i] + 1.0f));
}
if (non_staB >= 128)
non_staB = 32767.0/256.0f;
The code sections below show how the new features for the linear prediction residual
energies, i.e. the for the linear prediction gain, are calculated. Here the residual
energies are named epsP[m] (cf. E(m) used previously).
/*-----------------------------------------------------------------*
* Linear prediction efficiency 0 to 2 order
th nd
*(linear prediction gain going from 0 to 2 order model of linear prediction filter)
*-----------------------------------------------------------------*/
epsP_0_2 = max(0 , min(8, epsP[0] / epsP[2]));
epsP_0_2_lp = 0.15f * epsP_0_2 + (1.0f-0.15f) * st->epsP_0_2_lp;
epsP_0_2_ad = (float) fabs(epsP_0_2 - epsP_0_2_lp );
if (epsP_0_2_ad < epsP_0_2_ad_lp)
40 epsP_0_2_ad_lp = 0.1f * epsP_0_2_ad + (1.0f - 0.1f) * epsP_0_2_ad_lp;
else
epsP_0_2_ad_lp = 0.2f * epsP_0_2_ad + (1.0f - 0.2f) * epsP_0_2_ad_lp;
45 }
epsP_0_2_ad_lp_max = max(epsP_0_2_ad,st->epsP_0_2_ad_lp);
/*-----------------------------------------------------------------*
* Linear predition efficiency 2 to 16 order
*(linear prediction gain going from 2 to 16th order model of linear prediction filter)
*-----------------------------------------------------------------*/
epsP_2_16 = max(0 , min(8, epsP[2] / epsP[16]));
if (epsP_2_16 > epsP_2_16_lp)
epsP_2_16_lp = 0.2f * epsP_2_16 + (1.0f-0.2f) * epsP_2_16_lp;
else
epsP_2_16_lp = 0.03f * epsP_2_16 + (1.0f-0.03f) * epsP_2_16_lp;
epsP_2_16_lp2 = 0.02f * epsP_2_16 + (1.0f-0.02f) * epsP_2_16_lp2;
epsP_2_16_dlp = epsP_2_16_lp-epsP_2_16_lp2;
if (epsP_2_16_dlp < epsP_2_16_dlp_lp2 )
epsP_2_16_dlp_lp2 = 0.02f * epsP_2_16_dlp + (1.0f-0.02f) * epsP_2_16_dlp_lp2;
else
epsP_2_16_dlp_lp2 = 0.05f * epsP_2_16_dlp + (1.0f-0.05f) * epsP_2_16_dlp_lp2;
}
epsP_2_16_dlp_max = max(epsP_2_16_dlp,epsP_2_16_dlp_lp2);
The code below illustrates the creation of combined metrics, thresholds and flags
used for the actual update decision, i.e. the determining of whether to update the
background noise estimate or not. At least some of the parameters related to linear
prediction gains and/or spectral closeness are indicated in bold text.
comb_ahc_epsP = max(max(act_pred,lt_haco_ev),epsP_2_16_dlp);
comb_hcm_epsP = max(max(lt_haco_ev,epsP_2_16_dlp_max),epsP_0_2_ad_lp_max);
haco_ev_max = max(st_harm_cor_cnt==0,>lt_haco_ev);
Etot_l_lp_thr = st->Etot_l_lp + (1.5f + 1.5f * (Etot_lp<50.0f))*Etot_v_h2;
40 enr_bgd = Etot < Etot_l_lp_thr;
cns_bgd = (epsP_0_2 > 7.95f) && (non_sta< 1e3f);
lp_bgd = epsP_2_16_dlp_max < 0.10f;
ns_mask = non_sta < 1e5f;
lt_haco_mask = lt_haco_ev < 0.5f;
45 bg_haco_mask = haco_ev_max < 0.4f;
SD_1 = ( (epsP_0_2_ad > 0.5f) && (epsP_0_2 > 7.95f) );
bg_bgd3 = enr_bgd || ( ( cns_bgd || lp_bgd ) && ns_mask && lt_haco_mask && SD_1==0 );
PD_1 = (epsP_2_16_dlp_max < 0.10f ) ;
PD_2 = (epsP_0_2_ad_lp_max < 0.10f ) ;
PD_3 = (comb_ahc_epsP < 0.85f );
PD_4 = comb_ahc_epsP < 0.15f;
PD_5 = comb_hcm_epsP < 0.30f;
BG_1 = ( (SD_1==0) || (Etot < Etot_l_lp_thr) ) && bg_haco_mask && (act_pred < 0.85f) && (Etot_lp < 50.0f);
PAU = (aEn==0) || ( (Etot < 55.0f) && (SD_1==0) && ( ( PD_3 && (PD_1 || PD_2 ) ) || ( PD_4 || PD_5 ) ) );
NEW_POS_BG = (PAU | BG_1) & bg_bgd3;
/* Original silence detector works in most cases */
aE_bgd = aEn == 0;
/* When the signal dynamics is high and the energy is close to the background estimate */
sd1_bgd = (st->sign_dyn_lp > 15) && (Etot - st->Etot_l_lp ) < 2*st->Etot_v_h2 && st->harm_cor_cnt > 20;
/* init conditions steadily dropping act_pred and/or lt_haco_ev */
tn_ini = ini_frame < 150 && harm_cor_cnt > 5 &&
( (st->act_pred < 0.59f && st->lt_haco_ev <0.23f ) ||
st->act_pred < 0.38f ||
st->lt_haco_ev < 0.15f ||
non_staB < 50.0f ||
aE_bgd );
/* Energy close to the background estimate serves as a mask for other background detectors */
bg_bgd2 = Etot < Etot_l_lp_thr || tn_ini ;
As it is important not to do an update of the background noise estimate when a
current frame or segment comprises active content, several conditions are evaluated
in order to decide if an update is to be made. The major decision step in the noise
update logic is whether an update is to be made or not, and this is formed by
evaluation of a logical expression, which is underlined below. The new parameter
NEW_POS_BG (new in relation to the solution in Annex A and WO2011/049514) is a
pause detector, and is obtained based on the linear prediction gains going from 0th
nd nd th
to 2 , and from 2 to 16 order model of a linear prediction filter, and tn_ini is
obtained based on features related to spectral closeness. Here follows a decision
logic using the new features, according to the exemplifying embodiment.
40 updt_step=0.0f;
if (( bg_bgd2 && ( aE_bgd || sd1_bgd || lt_tn_track >0.90f || NEW_POS_BG ) ) ||
tn_ini )
if( ( ( act_pred < 0.85f ) &&
45 aE_bgd &&
( lt_Ellp_dist < 10 || sd1_bgd ) && lt_tn_dist<40 &&
( ( Etot - totalNoise ) < 10.0f ) ) ||
( st->first_noise_updt == 0 && st->harm_cor_cnt > 80 && aE_bgd && st->lt_aEn_zero > 0.5f ) ||
( tn_ini && ( aE_bgd || non_staB < 10.0 || st->harm_cor_cnt > 80 ) )
updt_step=1.0f;
st->first_noise_updt = 1;
for( i=0; i< NB_BANDS; i++ )
st->bckr[i] = tmpN[i];
}
else if ( ( ( st->act_pred < 0.80f ) && ( aE_bgd || PAU ) && st->lt_haco_ev < 0.10f ) ||
( ( st->act_pred < 0.70f ) && ( aE_bgd || non_staB < 17.0f ) && PAU && st->lt_haco_ev < 0.15f ) ||
( st->harm_cor_cnt > 80 && st->totalNoise > 5.0f && Etot < max(1.0f,Etot_l_lp + 1.5f* st->Etot_v_h2) ) ||
( st->harm_cor_cnt > 50 && st->first_noise_updt > 30 && aE_bgd && st->lt_aEn_zero>0.5f ) ||
tn_ini
updt_step=0.1f;
if ( !aE_bgd &&
st->harm_cor_cnt < 50 &&
( st->act_pred > 0.6f ||
( !tn_ini && Etot_l_lp - st->totalNoise < 10.0f && non_staB > 8.0f ) ) )
updt_step=0.01f;
if (updt_step > 0.0f )
st->first_noise_updt = 1;
for( i=0; i< NB_BANDS; i++ )
st->bckr[i] = st->bckr[i] + updt_step * (tmpN[i]-st->bckr[i]);
}
else if (aE_bgd || st->harm_cor_cnt > 100 )
( st->first_noise_updt) += 1;
40 }
else
/* If in music lower bckr to drop further */
if ( st->low_tn_track_cnt > 300 && st->lt_haco_ev >0.9f && st->totalNoise > 0.0f)
45 {
updt_step=-0.02f;
for( i=0; i< NB_BANDS; i++ )
if (st->bckr[i] > 2*E_MIN)
50 {
st->bckr[i] = 0.98f*st->bckr[i];
55 }
st->lt_aEn_zero = 0.2f * (st->aEn==0) + (1-0.2f)*st->lt_aEn_zero;
As previously indicated, the features from the linear prediction provide level
independent analysis of the input signal that improves the decision for background
noise update which is particularly useful in the SNR range 10 to 20dB, where energy
based SAD’s have limited performance due to the normal dynamic range of speech
signals
The background closeness features also improves background noise estimation as it
can be used both for initialization and normal operation. During initialization, it can
allow quick initialization for (lower level) background noise with mainly low frequency
content, common for car noise. Also the features can be used to prevent noise
updates of using low energy frames with a large difference in frequency
characteristics compared to the current background estimate, suggesting that the
current frame may be low level active content and an update could prevent detection
of future frames with similar content.
Figures 8-10 show how the respective parameters or metrics behave for speech in
background at 10dB SNR car noise. In the figures 8-10 the dots, “•”, each represent
the frame energy. For the figures 8 and 9a-c, the energy has been divided by 10 to
be more comparable for the G_0_2 and G_2_16 based features. The diagrams
correspond to an audio signal comprising two utterances, where the approximate
position for the first utterance is in frames 1310 - 1420 and for the second utterance,
in frames 1500 – 1610.
Figure 8 shows the frame energy (/10) (dot, “•”) and the features G_0_2 (circle, “○”)
and Gmax_0_2 (plus, “+”), for 10dB SNR speech with car noise. Note that the G_0_2
is 8 during the car noise as there is some correlation in the signal that can be
modeled using linear prediction with model order 2. During utterances the feature
Gmax_0_2 becomes over 1.5 (in this case) and after the speech burst it drops to 0.
In a specific implementation of a decision logic, the Gmax_0_2 needs to be below 0.1
to allow noise updates using this feature.
Figure 9a shows the frame energy (/10) (dot, “•”) and the features G_2_16 (circle,
“○”), G1_2_16 (cross, “x”), G2_2_16 (plus, “+”). Figure 9b shows the frame energy
(/10) (dot, “•”), and the features G_2_16 (circle, “○”) Gd_2_16 (cross, “x”), and
Gad_2_16 (plus, “+”). Figure 9c shows the frame energy (/10) (dot, “•”) and the
features G_2_16 (circle, “○”) and Gmax_2_16 (plus, “+”).The diagrams shown in
figures 9a-c also relate to 10dB SNR speech with car noise. The features are shown
in three diagrams in order to make it easier to see each parameter. Note that the
G_2_16 (circle, “○”) is just above 1 during the car noise (i.e. outside utterances)
indicting that the gain from the higher model order is low for this type of noise. During
utterances the feature Gmax_2_16 (plus, “+” in figure 9c) increases, and then start to
drop back to 0. In a specific implementation of a decision logic the feature
Gmax_2_16,also has to become lower than 0.1 to allow noise updates. In this
particular audio signal sample, this does not occur.
Figure 10 shows the frame energy (dot, “•”) (not divided by 10 this time) and the
feature nonstaB (plus, “+”) for 10dB SNR speech with car noise. The feature nonstaB
is in the range 0-10 during noise-only segments, and for utterances, it becomes
much larger (as the frequency characteristics is different for speech). It should be
noted, though, that even during the utterances there are frames where the feature
nonstaB falls in the range 0 – 10. For these frames there might be a possibility to
make background noise updates and thereby better track the background noise.
The solution disclosed herein also relates to a background noise estimator
implemented in hardware and/or software.
Background noise estimator, figures 11a-11c
An exemplifying embodiment of a background noise estimator is illustrated in a
general manner in figure 11a. By background noise estimator it is referred a module
or entity configured for estimating background noise in audio signals comprising e.g.
speech and/or music. The encoder 1100 is configured to perform at least one method
corresponding to the methods described above with reference e.g. to figures 2 and 7.
The encoder 1100 is associated with the same technical features, objects and
advantages as the previously described method embodiments. The background
noise estimator will be described in brief in order to avoid unnecessary repetition.
The background noise estimator may be implemented and/or described as follows:
The background noise estimator 1100 is configured for estimating a background
noise of an audio signal. The background noise estimator 1100 comprises
processing circuitry, or processing means 1101 and a communication interface 1102.
The processing circuitry 1101 is configured to cause the encoder 1100 to obtain, e.g.
determine or calculate, at least one parameter, e.g. NEW_POS_BG, based on a first
linear prediction gain calculated as a quotient between a residual signal from a 0th-
order linear prediction and a residual signal from a 2nd-order linear prediction for the
audio signal segment; and a second linear prediction gain calculated as a quotient
between a residual signal from a 2nd-order linear prediction and a residual signal
from a 16th-order linear prediction for the audio signal segment.
The processing circuitry 1101 is further configured to cause the background noise
estimator to determine whether the audio signal segment comprises a pause, i.e. is
free from active content such as speech and music, based on the at least one
parameter. The processing circuitry 1101 is further configured to cause the
background noise estimator to update a background noise estimate based on the
audio signal segment when the audio signal segment comprises a pause.
The communication interface 1102, which may also be denoted e.g. Input/Output
(I/O) interface, includes an interface for sending data to and receiving data from other
entities or modules. For example, the residual signals related to the linear prediction
model orders 0, 2 and 16 may be obtained, e.g. received, via the I/O interface from
an audio signal encoder performing linear predictive coding.
The processing circuitry 1101 could, as illustrated in figure 11b, comprise processing
means, such as a processor 1103, e.g. a CPU, and a memory 1104 for storing or
holding instructions. The memory would then comprise instructions, e.g. in form of a
computer program 1105, which when executed by the processing means 1103
causes the encoder 1100 to perform the actions described above.
An alternative implementation of the processing circuitry 1101 is shown in figure 11c.
The processing circuitry here comprises an obtaining or determining unit or module
1106, configured to cause the background noise estimator 1100 to obtain, e.g.
determine or calculate, at least one parameter, e.g. NEW_POS_BG, based on a first
linear prediction gain calculated as a quotient between a residual signal from a 0th-
order linear prediction and a residual signal from a 2nd-order linear prediction for the
audio signal segment; and a second linear prediction gain calculated as a quotient
between a residual signal from a 2nd-order linear prediction and a residual signal
from a 16th-order linear prediction for the audio signal segment. The processing
circuitry further comprises a determining unit or module 1107, configured to cause
the background noise estimator 1100 to determine whether the audio signal segment
comprises a pause, i.e. is free from active content such as speech and music, based
at least on the at least one parameter. The processing circuitry 1101 further
comprises an updating or estimating unit or module 1110, configured to cause the
background noise estimator to update a background noise estimate based on the
audio signal segment when the audio signal segment comprises a pause.
The processing circuitry 1101 could comprise more units, such as a filter unit or
module configured to cause the background noise estimator to low pass filter the
linear prediction gains, thus creating one or more long term estimates of the linear
prediction gains Actions such as low pass filtering may otherwise be performed e.g.
by the determining unit or module 1107.
The embodiments of a background noise estimator described above could be
configured for the different method embodiments described herein, such as limiting
and low pass filtering the linear prediction gains; determining a difference between
linear prediction gains and long term estimates and between long term estimates;
and/or obtaining and using a spectral closeness measure, etc.
The background noise estimator 1100 may be assumed to comprise further
functionality, for carrying out background noise estimation, such as e.g. functionality
exemplified in Appendix A.
Figure 12 illustrates a background estimator 1200 according to an exemplifying
embodiment. The background estimator 1200 comprises an input unit e.g. for
receiving residual energies for model orders 0, 2 and 16. The background estimator
further comprises a processor and a memory, said memory containing instructions
executable by said processor, whereby said background estimator is operative for:
performing a method according an embodiment described herein.
Accordingly, the background estimator may comprise, as illustrated in figure 13, an
input/output unit 1301, a calculator 1302 for calculating the first two sets of features
from the residual energies for model orders 0, 2 and 16 and a frequency analyzer
1303 for calculating the spectral closeness feature.
A background noise estimator as the ones described above may be comprised e.g. in
a VAD or SAD, an encoder and/or a decoder, i.e. a codec, and/or in a device, such
as a communication device. The communication device may be a user equipment
(UE) in the form of a mobile phone, video camera, sound recorder, tablet, desktop,
laptop, TV set-top box or home server/home gateway/home access point/home
router. The communication device may in some embodiments be a communications
network device adapted for coding and/or transcoding of audio signals. Examples of
such communications network devices are servers, such as media servers,
application servers, routers, gateways and radio base stations. The communication
device may also be adapted to be positioned in, i.e. being embedded in, a vessel,
such as a ship, flying drone, airplane and a road vehicle, such as a car, bus or lorry.
Such an embedded device would typically belong to a vehicle telematics unit or
vehicle infotainment system.
The steps, functions, procedures, modules, units and/or blocks described herein may
be implemented in hardware using any conventional technology, such as discrete
circuit or integrated circuit technology, including both general-purpose electronic
circuitry and application-specific circuitry.
Particular examples include one or more suitably configured digital signal processors
and other known electronic circuits, e.g. discrete logic gates interconnected to
perform a specialized function, or Application Specific Integrated Circuits (ASICs).
Alternatively, at least some of the steps, functions, procedures, modules, units and/or
blocks described above may be implemented in software such as a computer
program for execution by suitable processing circuitry including one or more
processing units. The software could be carried by a carrier, such as an electronic
signal, an optical signal, a radio signal, or a computer readable storage medium
before and/or during the use of the computer program in the network nodes.
The flow diagram or diagrams presented herein may be regarded as a computer flow
diagram or diagrams, when performed by one or more processors. A corresponding
apparatus may be defined as a group of function modules, where each step
performed by the processor corresponds to a function module. In this case, the
function modules are implemented as a computer program running on the processor.
Examples of processing circuitry includes, but is not limited to, one or more
microprocessors, one or more Digital Signal Processors, DSPs, one or more Central
Processing Units, CPUs, and/or any suitable programmable logic circuitry such as
one or more Field Programmable Gate Arrays, FPGAs, or one or more
Programmable Logic Controllers, PLCs. That is, the units or modules in the
arrangements in the different nodes described above could be implemented by a
combination of analog and digital circuits, and/or one or more processors configured
with software and/or firmware, e.g. stored in a memory. One or more of these
processors, as well as the other digital hardware, may be included in a single
application-specific integrated circuitry, ASIC, or several processors and various
digital hardware may be distributed among several separate components, whether
individually packaged or assembled into a system-on-a-chip, SoC.
It should also be understood that it may be possible to re-use the general processing
capabilities of any conventional device or unit in which the proposed technology is
implemented. It may also be possible to re-use existing software, e.g. by
reprogramming of the existing software or by adding new software components.
The embodiments described above are merely given as examples, and it should be
understood that the proposed technology is not limited thereto. It will be understood
by those skilled in the art that various modifications, combinations and changes may
be made to the embodiments without departing from the present scope. In particular,
different part solutions in the different embodiments can be combined in other
configurations, where technically possible.
When using the word "comprise" or “comprising” it shall be interpreted as non-
limiting, i.e. meaning "consist at least of".
It should also be noted that in some alternate implementations, the functions/acts
noted in the blocks may occur out of the order noted in the flowcharts. For example,
two blocks shown in succession may in fact be executed substantially concurrently or
the blocks may sometimes be executed in the reverse order, depending upon the
functionality/acts involved. Moreover, the functionality of a given block of the
flowcharts and/or block diagrams may be separated into multiple blocks and/or the
functionality of two or more blocks of the flowcharts and/or block diagrams may be at
least partially integrated. Finally, other blocks may be added/inserted between the
blocks that are illustrated, and/or blocks/operations may be omitted without departing
from the scope of inventive concepts.
It is to be understood that the choice of interacting units, as well as the naming of the
units within this disclosure are only for exemplifying purpose, and nodes suitable to
execute any of the methods described above may be configured in a plurality of
alternative ways in order to be able to execute the suggested procedure actions.
It should also be noted that the units described in this disclosure are to be regarded
as logical entities and not with necessity as separate physical entities.
Reference to an element in the singular is not intended to mean "one and only one"
unless explicitly so stated, but rather "one or more." All structural and functional
equivalents to the elements of the above-described embodiments that are known to
those of ordinary skill in the art are expressly incorporated herein by reference and
are intended to be encompassed hereby. Moreover, it is not necessary for a device
or method to address each and every problem sought to be solved by the technology
disclosed herein, for it to be encompassed hereby.
In some instances herein, detailed descriptions of well-known devices, circuits, and
methods are omitted so as not to obscure the description of the disclosed technology
with unnecessary detail. All statements herein reciting principles, aspects, and
embodiments of the disclosed technology, as well as specific examples thereof, are
intended to encompass both structural and functional equivalents thereof.
Additionally, it is intended that such equivalents include both currently known
equivalents as well as equivalents developed in the future, e.g. any elements
developed that perform the same function, regardless of structure.
ANNEX A
Figure 14 is a flow chart illustrating an exemplifying embodiment of a method for
background noise estimation according to the herein proposed technology. The
method is intended to be performed by a background noise estimator, which may be
part of a SAD. The background noise estimator, and the SAD, may further be
comprised in an audio encoder, which may in its turn be comprised in a wireless
device or a network node. For the described background noise estimator, adjusting
the noise estimate down, is not restricted. For each frame a possible new sub-band
noise estimate is calculated, regardless if the frame is background or active content,
if the new value is lower than the current it is used directly as it most likely would be
from a background frame. The following noise estimation logic is a second step
where it is decided if the sub-band noise estimate can be increased and if so how
much, the increase is based on the previously calculated possible new sub-band
noise estimate. Basically this logic forms the decision of the current frame is a
background frame and if it is not sure it may allow a smaller increase compared to
what was originally estimated.
The method illustrated in figure 14 comprises: when an energy level of an audio
signal segment is more than a threshold higher 202:1 than a long term minimum
energy level, lt_min, or, when the energy level of the audio signal segment is less
than a threshold higher 202:2 than lt_min, but no pause is detected 204:1 in the
audio signal segment:
-reducing 206 a current background noise estimate when the audio signal segment is
determined 203:2 to comprise music and the current background noise estimate
exceeds a minimum value 205:1, denoted “T” in figure 14, and further exemplified
e.g. as 2*E_MIN in code below.
By performing the above, and providing the background noise estimate to a SAD, the
SAD is enabled to perform more adequate sound activity detection. Further, recovery
from erroneous background noise estimate updates is enabled.
The energy level of the audio signal segment used in the method described above
may alternatively be referred to e.g. as the current frame energy, Etot, or as the
energy of the signal segment, or frame, which can be calculated by summing the
sub-band energies for the current signal segment.
The other energy feature used in the method above, i.e. the long term minimum
energy level, lt_min, is an estimate, which is determined over a plurality of preceding
audio signal segments or frames. lt_min could alternatively be denoted e.g. Etot_l_lp
One basic way of deriving lt_min would be to use the minimum value of the history of
current frame energy over some number of past frames. If the value calculated as:
"current frame energy - long term minimum estimate" is below a threshold value,
denoted e.g. THR1, the current frame energy is herein said to be close to the long
term minimum energy, or to be near the long term minimum energy. That is, when
(Etot - lt_min) < THR1, the current frame energy, Etot, may be determined 202 to be
near the long term minimum energy lt_min. The case when (Etot - lt_min) = THR1
may be referred to either of the decisions, 202:1 or 202:2, depending on
implementation. The numbering 202:1 in figure 14 indicates the decision that the
current frame energy is not near lt_min, while 202:2 indicates the decision that the
current frame energy is near lt_min. Other numbering in figure 14 on the form XXX:Y
indicates corresponding decisions. The feature lt_min will be further described below.
The minimum value, which the current background noise estimate is to exceed, in
order to be reduced, may be assumed to be zero or a small positive value. For
example, as will be exemplified in code below, a current total energy of the
background estimate, which may be denoted “totalNoise” and be determined e.g. as
*log10∑backr[i], may be required to exceed a minimum value of zero in order for
the reduction to come in question. Alternatively, or in addition, each entry in a vector
backr[i] comprising the sub-band background estimates may be compared to a
minimum value, E_MIN, in order for the reduction to be performed. In the code
example below, E_MIN is a small positive value.
It should be noted that according to a preferred embodiment of the solution
suggested herein, the decision of whether the energy level of the audio signal
segment is more than a threshold higher than lt_min is based only on information
derived from the input audio signal, that is, it is not based on feedback from a sound
activity detector decision.
The determining 204 of whether a current frame comprises a pause or not may be
performed in different ways based on one or more criteria. A pause criterion may also
be referred to as a pause detector. A single pause detector could be applied, or a
combination of different pause detectors. With a combination of pause detectors
each can be used to detect pauses in different conditions. One indicator of that a
current frame may comprise a pause, or inactivity, is that a correlation feature for the
frame is low, and that a number of preceding frames also have had low correlation
features. If the current energy is close to the long term minimum energy and a pause
is detected, the background noise can be updated according to the current input, as
illustrated in figure 14. A pause may be considered to be detected when, in addition
to that the energy level of the audio signal segment is less than a threshold higher
than lt_min: a predefined number of consecutive preceding audio signal segments
have been determined not to comprise an active signal and/or a dynamic of the audio
signal exceeds a threshold. This is also illustrated in the code example further below.
The reduction 206 of the background noise estimate enables handling of situations
where the background noise estimate has become “too high”, i.e. in relation to a true
background noise. This could also be expressed e.g. as that the background noise
estimate deviates from the actual background noise. A too high background noise
estimate may lead to inadequate decisions by the SAD, where the current signal
segment is determined to be inactive even though it comprises active speech or
music. A reason for the background noise estimate becoming too high is e.g.
erroneous or unwanted background noise updates in music, where the noise
estimation has mistaken music for background and allowed the noise estimate to be
increased. The disclosed method allows for such an erroneously updated
background noise estimate to be adjusted e.g. when a following frame of the input
signal is determined to comprise music. This adjustment is done by a forced
reduction of the background noise estimate, where the noise estimate is scaled
down, even if the current input signal segment energy is higher than the current
background noise estimate, e.g. in a sub-band. It should be noted that the above
described logic for background noise estimation is used to control the increase of
background sub-band energy. It is always allowed to lower the sub-band energy
when the current frame sub-band energy is lower than the background noise
estimate. This function is not explicitly shown in figure 14. Such a decrease usually
has a fixed setting for the step size. However, the background noise estimate should
only be allowed to be increased in association with the decision logic according to the
method described above. When a pause is detected, the energy and correlation
features may also be used for deciding 207 how large the adjustment step size for
the background estimate increase should be before the actual background noise
update is made.
As previously mentioned, some music segments can be difficult to separate from
background noise, due to that they are very noise like. Thus, the noise update logic
may accidentally allow for increased sub-band energy estimates, even though the
input signal was an active signal. This can cause problems as the noise estimate can
become higher than they should be.
In prior art background noise estimators, the sub-band energy estimates could only
be reduced when an input sub-band energy went below a current noise estimate.
However, since some music segments can be difficult to separate from background
noise, due to that they are very noise like, the inventors have realized that a recovery
strategy for music is needed. In the embodiments described herein, such a recovery
can be done by forced noise estimate reduction when the input signal returns to
music-like characteristics. That is, when the energy and pause logic described above
prevent, 202:1, 204:1, the noise estimation from being increased, it is tested 203 if
the input is suspected to be music and if so 203:2, the sub-band energies are
reduced 206 by a small amount each frame until the noise estimates reaches a
lowest level 205:2.
A background estimator as the ones described above can be comprised or
implemented in a VAD or SAD and/or in an encoder and/or a decoder, wherein the
encoder and/or decoder can be implemented in a user device, such as a mobile
phone, a laptop, a tablet, etc. The background estimator could further be comprised
in a network node, such as a Media Gateway, e.g. as part of a codec.
Figure 17 is a block diagram schematically illustrating an implementation of a
background estimator according to an exemplifying embodiment. An input framing
block 51 first divides the input signal into frames of suitable length, e.g. 5-30 ms. For
each frame, a feature extractor 52 calculates at least the following features from the
input: 1) The feature extractor analyzes the frame in the frequency domain and the
energy for a set of sub-bands are calculated. The sub-bands are the same sub-
bands that are to be used for the background estimation. 2) The feature extractor
further analyzes the frame in the time-domain and calculates a correlation denoted
e.g. cor_est and/or lt_cor_est, which is used in determining whether the frame
comprises active content or not. 3) The feature extractor further utilizes the current
frame total energy, e.g. denoted Etot, for updating features for energy history of
current and earlier input frames, such as the long term minimum energy, lt_min. The
correlation and energy features are then fed to the Update Decision Logic block 53.
Here, a decision logic according to the herein disclosed solution is implemented in
the Update Decision Logic block 53, where the correlation and energy features are
used to form decisions on whether the current frame energy is close to a long term
minimum energy or not; on whether the current frame is part of a pause (not active
signal) or not; and whether the current frame is part of music or not. The solution
according to the embodiments described herein involves how these features and
decisions are used to update the background noise estimation in a robust way.
Below, some implementation details of embodiments of the solution disclosed herein
will be described. The implementation details below are taken from an embodiment in
a G.718 based encoder. This embodiment uses some of the features described in
W02011/049514 and W02011/049515,.
The following features are defined in the modified G.718 described in W02011/09514
Etot; The total energy for current input frame
Etot_l Tracks the miminmum energy envelope
Etot_l_lp; A Smoothed version of the mimimum energy envelope Etot_l
totalNoise; The current total energy of the background estimate
bckr[i]; The vector with the sub-band background estimates
tmpN[i]; A precalculated potential new background estimate
aEn; A background detector which uses multiple features (a counter)
harm_cor_cnt Counts the frames since the last frame with correlation or harmonic event
act_pred A prediction of activity from input frame features only
cor[i] Vector with correlation estimates for, i=0 end of current frame,
i=1 start of current frame, i=2 end of previos frame
The following features are defined in the modified G.718 described in W02011/09515
Etot_h Tracks the maximum energy envelope
sign_dyn_lp; A smoothed input signal dynamics
Also the feature Etot_v_h was defined in W02011/049514, but in this embodiment it
has been modified and is now implemented as follows:
Etot_v = (float) fabs(*Etot_last - Etot);
if( Etot_v < 7.0f) /*note that no VAD flag or similar is used here*/
*Etot_v_h -= 0.01f;
if (Etot_v > *Etot_v_h)
if ((*Etot_v -*Etot_v_h) > 0.2f)
*Etot_v_h = *Etot_v_h + 0.2f;
}
else
*Etot_v_h = Etot_v; }}}
Etot_v measures the absolute energy variation between frames, i.e. the absolute
value of the instantaneous energy variation between frames. In the example above,
the energy variation between two frames is determined to be “low” when the
difference between the last and the current frame energy is smaller than 7 units. This
is utilized as an indicator of that the current frame (and the previous frame) may be
part of a pause, i.e. comprise only background noise. However, such low variance
could alternatively be found e.g. in the middle of a speech burst. The variable
Etot_last is the energy level of the previous frame.
The above steps described in code may be performed as part of the
“calculate/update correlation and energy” steps in the flow chart in figure 14, i.e. as
part of the actions 201. In the W02011/049514 implementation, a VAD flag was used
to determine whether the current audio signal segment comprised background noise
or not. The inventors have realized that the dependency on feedback information
may be problematic. In the herein disclosed solution, the decision of whether to
update the background noise estimate or not is not dependent on a VAD (or SAD)
decision.
Further, in the herein disclosed solution, the following features, which are not part of
the W02011/049514 implementation, may be calculated/updated as part of the same
steps, i.e. the calculate/update correlation and energy steps illustrated in figure 14.
These features are also used in the decision logic of whether to update the
background estimate or not.
In order to achieve a more adequate background noise estimate, a number of
features are defined below. For example, the new correlation related features
cor_est and It_cor_est are defined. The feature cor_est is an estimate of the
correlation in the current frame, and cor_est is also used to produce It_cor_est, which
is a smoothed long-term estimate of the correlation.
cor_est = (cor[0] + cor[1] + cor[2]) / 3.0f ;
st->lt_cor_est = 0.01f*cor_est + 0.99f * st->lt_cor_est;
As defined above, cor[i] is a vector comprising correlation estimates, and cor[0]
represents the end of the current frame, cor[1] represents the start of the current
frame, and cor[2] represents the end of a previous frame.
Further, a new feature, It_tn_track, is calculated, which gives a long term estimate of
how often the background estimates are close to the current frame energy. When the
current frame energy is close enough to the current background estimate this is
registered by a condition that signals (1/0) if the background is close or not. This
signal is used to form the long-term measure It_tn_track.
st->lt_tn_track = 0,03f* (Etot - st->totalNoise < 10) + 0.97f*st->lt_tn_track;
In this example, 0,03 is added when the current frame energy is close to the
background noise estimate, and otherwise the only remaining term is 0,97 times the
previous value. In this example, “close” is defined as that the difference between the
current frame energy, Etot, and the background noise estimate, totalNoise, is less
than 10 units. Other definitions of “close” are also possible.
Further, the distance between the current background estimate, Etot, and the current
frame energy, totalNoise, is used for determining a feature, lt_tn_dist, which gives a
long term estimate of this distance. A similar feature, lt_Ellp_dist, is created for the
distance between the long term minimum energy Etot_l_lp and the current frame
energy, Etot.
st->lt_tn_dist = 0.03f* (Etot - st->totalNoise) + 0.97f*st->lt_tn_dist;
st->lt_Ellp_dist = 0.03f* (Etot - st->Etot_l_lp) + 0.97f*st->lt_Ellp_dist;
The feature harm_cor_cnt, introduced above, is used for counting the number of
frames since the last frame having a correlation or a harmonic event, i.e. since a
frame fulfilling certain criteria related to activity. That is, when the condition
harm_cor_cnt==0, this implies that the current frame most likely is an active frame, as it
shows correlation or a harmonic event. This is used to form a long term smoothed
estimate, lt_haco_ev, of how often such events occur. In this case the update is not
symmetric, that is different time constants are used if the estimate is increased or
decreased, as can be seen below.
if (st->harm_cor_cnt == 0) /*when probably active*/
st->lt_haco_ev = 0,03f + 0.97f*st->lt_haco_ev; /*increase long term estimate*/
else
st->lt_haco_ev = 0.99f*st->lt_haco_ev; /*decrease long term estimate */
A low value of the feature It_tn_track, introduced above, indicates that the input
frame energy has not been close to the background energy for some frames. This is
due to that It_tn_track is decreased for each frame where the current frame energy is
not close to the background energy estimate. It_tn_track is increased only when the
current frame energy is close to the background energy estimate as shown above.
To get a better estimate of how long this “non-tracking”, i.e. the frame energy being
far from the background estimate, has lasted, a counter, low_tn_track_cnt, for the
number of frames with this absence of tracking is formed as:
if (st->lt_tn_track<0.05f) /*when lt_tn_track is low */
st->low_tn_track_cnt++; /*add 1 to counter */
}
else
st->low_tn_track_cnt=0; /*reset counter */
In the example above, “low” is defined as below the value 0.05. This should
be seen as an exemplifying value, which could be selected differently.
For the step "Form pause and music decisions" illustrated in figure 14, the following
three code expressions are used to form pause detection, also denoted background
detection. In other embodiments and implementations, other criteria could also be
added for pause detection. The actual music decision is formed in the code using
correlation and energy features.
1: bg_bgd = Etot < Etot_l_lp + 0.6f * st->Etot_v_h;
bg_bgd will become “1” or “true” when Etot is close to the background noise
estimate. bg_bgd serves as a mask for other background detectors. That is, if
bg_bgd is not “true”, the background detectors 2 and 3 below do not need to be
evaluated. Etot_v_h is a noise variance estimate, which could alternatively be
denoted Nvar. Etot_v_h is derived from the input total energy (in log domain) using
Etot_v which measures the absolute energy variation between frames. Note that the
feature Etot_v_h is limited to only increase a maximum of a small constant value, e.g.
0.2 for each frame. Etot_l_lp is a smoothed version of the mimimum energy envelope
Etot_l.
2: aE_bgd = st->aEn == 0;
When aEn is zero, aE_bgd becomes “1” or “true”. aEn is a counter which is
incremented when an active signal is determined to be present in a current frame,
and decreased when the current frame is determined not to comprise an active
signal. aEn may not be incremented more than to a certain number, e.g. 6, and not
be reduced to less than zero. After a number of consecutive frames, e.g. 6, without
an active signal, aEn will be equal to zero.
sd1_bgd = (st->sign_dyn_lp > 15) && (Etot - st->Etot_l_lp ) < st->Etot_v_h && st->harm_cor_cnt > 20;
Here, sd1_bgd will be “1” or “true” when three different conditions are true: The signal
dynamics, sign_dyn_lp is high, in this example more than 15; The current frame
energy is close to the background estimate; and: A certain number of frames have
passed without correlation or harmonic events, in this example 20 frames.
The function of the bg_bgd is to be a flag for detecting that the current frame energy
is close to the long term minimum energy. The latter two, aE_bgd and sd1_bgd
represent pause or background detection in different conditions. aE_bgd is the most
general detector of the two, while sd1_bgd mainly detects speech pauses in high
SNR.
A new decision logic according to an embodiment of the technology disclosed herein,
is constructed as follows in code below. The decision logic comprises the masking
condition bg_bgd, and the two pause detectors aE_bgd and sd1_bgd. There could
also be a third pause detector, which evaluates the long term statistics for how well
the totalNoise tracks the minimum energy estimate. The conditions evaluated if the
first line is true is decision logic on how large the step size should be, updt_step and
the actual noise estimation update is the assignment of value to "st->bckr[i] =-". Note
the tmpN[i] is a previously calculated potentially new noise level calculated according
to the solution described in W02011/049514. The decision logic below follows the
part 209 of figure 14, which is partly indicated in connection with the code below
if (bg_bgd && ( aE_bgd II sd1_bgd II st->lt_tn_track >0.90f ) ) /*if 202:2 and 204:2)*/
if( (st->act_pred < 0.85f II ( aE_bgd && st->lt_haco_ev < 0.05f ) ) &&
(st->lt_Ellp_dist < 10 II sd1_bgd ) && st->lt_tn_dist<40 &&
( (Etot - st->totalNoise ) < 15.0f II st->lt_haco_ev < 0.10f ) ) /*207*/
st->first_noise_updt = 1;
for( i=0; i< NB_BANDS; i++ )
st->bckr[i] = tmpN[i) /*208*/
else if (aE_bgd && st->lt_haco_ev < 0.15f)
updt_step=0.1f;
if (st->act_pred > 0.85f )
updt_step=0.01f /*207*/
if (updt_step > 0.0f)
st->first_noise_updt = 1;
for[ i=0; i< NB_BANDS; i++ )
st->bckr[i] = st->bckr[i] + updt_step * (tmpN[i]-st->bckr[i]); /*208*/
else
(st->first_noise_updt) +=1;
}
else
/* If in music lower bckr to drop further */ /*if 203:2 and 205:1*/
If ( st->low_tn_track_cnt > 300 && st->lt_haco_ev > 0.9f && st-> totalNoise > 0.0f)
For ( i=0; i< NB_BANDS; i++)
If (st->bckr[i] > 2 * E_MIN
{
St->bckr[i] = 0.98f * st->bckr[i]; /*206*/
Else
(st->first_noise_updt) += 1;
The code segment in the last code block starting with "/* If in music ... */ contains the
forced down scaling of the background estimate which is used if it is suspected that
the current input is music. This is decided as a function: long period of poor tracking
background noise compared to the minimum energy estimate, AND, frequent
occurrences of harmonic or correlation events, AND, the last condition “totalNoise>0”
is a check that the current total energy of the background estimate is larger than
zero, which implies that a reduction of the background estimate may be considered.
Further, it is determined whether “bckr[i] > 2 * E_MIN”, where E_MIN is a small
positive value. This is a check of each entry in a vector comprising the sub-band
background estimates, such that an entry needs to exceed E_MIN in order to be
40 reduced (in the example by being multiplied by 0,98). These checks are made in
order to avoid reducing the background estimates into too small values.
The embodiments improve the background noise estimation which allows improved
performance of the SAD/VAD to achieve high efficient DTX solution and avoid the
degradation in speech quality or music caused by clipping.
With the removal of the decision feedback described in W02011/09514 from the
Etot_v_h, there is a better separation between the noise estimation and the SAD.
This has benefits as that the noise estimation is not changed if/when the SAD
function/tuning is changed. That is, the determining of a background noise estimate
becomes independent of the function of the SAD. Also the tuning of the noise
estimation logic becomes easier as one is not affected by secondary effects from the
SAD when the background estimates are changed.
Claims (27)
1. A method for a background noise estimator for estimation of background noise in an audio signal, wherein the audio signal comprises a plurality of audio signal 5 segments, the method comprising: computing at least one parameter associated with an audio signal segment that is among the audio signal segments, based on both of: a first linear prediction gain calculated as a quotient between an energy of the input signal and a residual signal energy from a first linear prediction for 10 the audio signal segment; and a second linear prediction gain calculated as a quotient between the residual signal energy from the first linear prediction and a residual signal energy from a second linear prediction for the audio signal segment; determining whether the audio signal segment comprises a pause free of 15 speech and music, based at least on the at least one parameter; and responsive to when the audio signal segment is determined to comprise a pause, updating to obtain an updated background noise estimate based on the audio signal segment; wherein the computing the at least one parameter comprises determining a 20 difference between two long term estimates associated with one of the linear prediction gains.
2. The method according to claim 1, further comprising: controlling discontinuous transmission of at least one of the audio signal 25 segments from a communication device at least partially based on the updated background noise estimate.
3. The method according to claim 1 or claim 2, wherein: the first linear prediction is a 2nd-order linear prediction; and the second linear prediction is a 16th order linear prediction. 5
4. The method according to any one of claims 1-3, wherein the method is performed by operating at least one processor of an electronic device.
5. The method according to any one of claims 1-4, wherein the computing the at least one parameter comprises: 10 limiting the first and second linear prediction gains to take on values in a predefined interval.
6. The method according to any one of claims 1-5, wherein the computing the at least one parameter comprises: 15 creating at least one long term estimate of each of the first and second linear prediction gains, wherein the long term estimate is further created based on corresponding linear prediction gains associated with at least one of the audio signal segments that precedes the audio signal segment. 20
7. The method according to any one of claims 1-6, wherein the computing the at least one parameter comprises low pass filtering the first and second linear prediction gains.
8. The method according to claim 7, wherein filter coefficients of at least one low 25 pass filter that operates to provide the low pass filtering are determined based on a relation between a linear prediction gain associated with the audio signal segment and an average of a corresponding linear prediction gain computed based on a plurality of the audio signal segments that precede the audio signal segment.
9. The method according to any one of claims 1-8, wherein the determining of 5 whether the audio signal segment comprises a pause is further based on a measure of spectral closeness associated with the audio signal segment.
10. The method according to claim 9, further comprising computing the measure of spectral closeness based on energies for a set of frequency bands of the audio 10 signal segment and background noise estimates corresponding to the set of frequency bands.
11. The method according to claim 10, wherein, during an initialization period, an initial value, Emin is used as the background noise estimates based on which the 15 measure of spectral closeness is computed.
12. A background noise estimator, for estimating background noise in an audio signal comprising a plurality of audio signal segments, the background noise estimator comprising: 20 at least one processor; and at least one memory storing computer readable instructions executed by the at least one processor to perform operations comprising: compute at least one parameter based on both of: a first linear prediction gain calculated as a quotient between an 25 energy of the input signal and a residual signal energy from a first linear prediction for the audio signal segment; and a second linear prediction gain calculated as a quotient between the residual signal energy from the first linear prediction and a residual signal energy from a second linear prediction for the audio signal segment; 5 determine whether the audio signal segment comprises a pause free of speech and music, based at least on the at least one parameter; and responsive to when the audio signal segment is determined to comprise a pause, updating to obtain an updated a background noise estimate based on the audio signal segment; 10 wherein the computing the at least one parameter comprises determining a difference between two long term estimates associated with one of the linear prediction gains.
13. The background noise estimator according to claim 12, wherein the operations 15 further comprise: controlling discontinuous transmission of at least one of the audio signal segments from a communication device at least partially based on the updated background noise estimate. 20
14. The background noise estimator according to claim 12 or claim 13, wherein: the first linear prediction is a 2nd-order linear prediction; and the second linear prediction is a 16th order linear prediction.
15. The background noise estimator according to any one of claims 12-14, 25 wherein the computing of the at least one parameter comprises limiting the first and second linear prediction gain to take on values in a predefined interval.
16. The background noise estimator according to any one of claims 12-15, wherein the computing of the at least one parameter comprises: creating at least one long term estimate of each of the first and second linear 5 prediction gains, wherein the long term estimate is further created based on corresponding linear prediction gains associated with at least one of the audio signal segments that precedes the audio signal segment.
17. The background noise estimator according to any one of claims 12-16, 10 wherein the computing of the at least one parameter comprises low pass filtering the first and second linear prediction gains.
18. The background noise estimator according to claim 17, wherein filter coefficients of at least one low pass filter that operates to provide the low pass 15 filtering are determined based on a relation between a linear prediction gain associated with the audio signal segment and an average of a corresponding linear prediction gain computed based on a plurality of the audio signal segments that precede the audio signal segment. 20
19. The background noise estimator according to any one of claims 12-18, being configured to further base the determining of whether the audio signal segment comprises a pause on a measure of spectral closeness associated with the audio signal segment. 25
20. The background noise estimator according to claim 19, being configured to compute the measure of spectral closeness based on energies for a set of frequency bands of the audio signal segment and background noise estimates corresponding to the set of frequency bands.
21. The background noise estimator according to claim 20, being configured to operate during an initialization period to use an initial value, E , as the background noise estimates based on which the measure of spectral closeness is computed.
22. A Sound Activity Detector (SAD) comprising a background noise estimator according to any one of claims 12-21.
23. A codec comprising a background noise estimator according to any one of 10 claims 12-21.
24. A computer program product comprising a non-transitory computer readable storage medium storing instructions which, when executed on at least one processor, cause the at least one processor to perform operations comprising: 15 computing at least one parameter associated with an audio signal segment that is among the audio signal segments, based on both of: a first linear prediction gain calculated as a quotient between an energy of the input signal and a residual signal energy from a first linear prediction for the audio signal segment; and 20 a second linear prediction gain calculated as a quotient between the residual signal energy from the first linear prediction and a residual signal energy from a second linear prediction for the audio signal segment; determining whether the audio signal segment comprises a pause free of speech and music, based at least on the at least one parameter; and 25 responsive to when the audio signal segment is determined to comprise a pause, updating to obtain an updated background noise estimate based on the audio signal segment; wherein the computing the at least one parameter comprises determining a difference between two long term estimates associated with one of the linear prediction gains. 5
25. The method according to claim 1 substantially as herein described with reference to figures 1-21 and/or examples.
26. The background noise estimator according to claim 12 substantially as herein described with reference to figures 1-21 and/or examples.
27. The computer program product according to claim 24 substantially as herein described with reference to figures 1-21 and/or example.
Applications Claiming Priority (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US201462030121P | 2014-07-29 | 2014-07-29 | |
US62/030,121 | 2014-07-29 | ||
NZ728080A NZ728080A (en) | 2014-07-29 | 2015-07-01 | Estimation of background noise in audio signals |
Publications (2)
Publication Number | Publication Date |
---|---|
NZ743390A NZ743390A (en) | 2021-03-26 |
NZ743390B2 true NZ743390B2 (en) | 2021-06-29 |
Family
ID=
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US11636865B2 (en) | Estimation of background noise in audio signals | |
US11164590B2 (en) | Estimation of background noise in audio signals | |
NZ743390B2 (en) | Estimation of background noise in audio signals |