A kind of method for eliminating millimeter wave bioradar noise in voice
【Technical field】
The present invention relates to make an uproar in radar field of voice signal, more particularly to a kind of elimination millimeter wave bioradar voice
The method of sound.
【Background technology】
There is the strong speech detection method of noncontact, remote, acoustic resistive noise jamming ability, bioradar as one
Speech detection technological break-through microphone is easily by acoustic noise interference and other speech detections for needing with contact human skin
The limitation of device, is gradually applied and is developed in speech detection field.At present, although millimeter wave bioradar can have
Effect obtain 20 meters outside human body voice signal, but obtain voice signal often by electromagnetism harmonic noise, circuit noise and
Ambient noise etc. is disturbed, and the presence of these noises reduces to a certain extent the quality of radar voice, or even have impact on language
The intelligibility of sound, therefore, how the effective noise content eliminated in radar voice exists to millimeter wave bioradar technology
Further apply significant in speech detection field.
【The content of the invention】
It is an object of the invention to:A kind of method for eliminating noise in millimeter wave bioradar is provided, the method is according to Jing
Radar voice noise characteristic distributions after empirical mode decomposition, on the premise of voice signal distortion and excessive damage is not caused,
Effectively remove the noise content in radar voice.For bioradar speech detection technology development provide speech de-noising in terms of
Technical support.
Effectively to realize above-mentioned purpose, the technical solution used in the present invention is:
A kind of method for eliminating millimeter wave bioradar noise in voice, comprises the following steps:
1) bioradar voice is sampled, obtains grandfather tape and make an uproar radar voice signal;
2) radar voice signal of making an uproar to the grandfather tape of collection carries out empirical mode decomposition for intrinsic mode function IMF;
3) noise content analysis is carried out to the intrinsic mode function after empirical mode decomposition;
4) according to radar voice noise characteristic distributions, the selection of noise component(s) number is carried out using Mutual information entropy, to distinguish
High frequency and intermediate frequency separation, using given threshold intermediate frequency and low frequency separation are distinguished;
5) intrinsic mode function Jing after self-adaptive solution is reconstructed and obtains enhanced radar voice.
As a further improvement on the present invention, step 2) specifically include following steps:
Assume that given voice signal is x (t), empirical mode decomposition method is adaptive by the original letter by screening process
Number intrinsic mode function is decomposed into, and each intrinsic mode function IMF is a sub- frequency content of primary signal, its tool
The screening process of body is as follows:
2.1) all of maximum of signal x (t) and minimum point are found out;
2.2) maximum and minimum point are fitted respectively by three sample interpolations, respectively obtain maximum envelope eu(t)
With minimum envelope ed(t);
2.3) average m of maximum envelope and minimum envelope is calculated1(t)=(eu(t)+ed(t))/2, and with original
Signal deducts the average and can obtain h1(t)=x (t)-m1(t);
2.4) h is judged1T whether () meet two screening primary conditions of intrinsic mode function IMF:
A () in all of signal length, the zero passage of each intrinsic mode function IMF points must count out phase with extreme value
Same, or both number is more or less the same in 1;
B () is zero in the average of any sample point, maximum envelope and minimum envelope, that is, meet IMF with regard to the time
Axle is geometrically symmetric;
Now, if h1T () meets (a) and (b) two primary conditions of intrinsic mode function IMF, then IMF1(t)=h1
(t);
If 2.5) h1T () is unsatisfactory for two primary conditions of IMF, then make h1T () is used as new primary signal repeat step
2.1), 2.2), 2.3) h is obtained2(t)=h1(t)-m2(t);If h2T () meets IMF and sieves condition then IMF substantially1(t)=h2
(t), if now h2T () is unsatisfactory for IMF screenings condition and then performs stopping screening threshold value, the stopping criterion such as following formula:
Stopping screening threshold value SD span should meet between 0.2 and 0.3;
If h2T () meets stopping criterion, then IMF1(t)=h2(t);
If h2T () is unsatisfactory for, then by h2(t) repeat step 2.1) -2.4), under continuation h is goti(t), until hi(t)
Meet two primary conditions or meet stopping criterion, then IMF1(t)=hi(t);
2.6) when getting IMF1After (t), it is removed from primary signal and obtains residual signals r1(t):r1(t)=x
(t)-IMF1(t), using residual signals as a new primary signal, repeat step 2.1) -2.5), until obtaining next residual
Difference signal, then decompose finally, then residual signals are represented with following formula:rn(t)=rn-1(t)-IMFn(t);
Now, rnT () is a monotonic sequence, after the completion of screening, primary signal is broken down into multiple intrinsic mode function IMF1
(t)、IMF2(t)…IMFn(t) and residual sequence rn(t);N is positive integer;Therefore, primary signal is represented by:
As a further improvement on the present invention, by original radar voice signal Jing empirical mode decompositions, it is decomposed into radio-frequency head
Point, intermediate-frequency section, low frequency part, noise concentrates on HFS, and intermediate-frequency section is useful primary speech signal, and low frequency is language
The detailed information of message number.
As a further improvement on the present invention, step 4) in, bioradar voice intrinsic mode function calculates mutual trust entropy side
Method is as follows:
The mutual trust entropy of two variable Xs and Y can be calculated according to the following formula:
Wherein, p (x, y) is the Joint Distribution of variable X and Y, and p (x) and p (y) is respectively marginal distribution, or Mutual information entropy
Following formula is expressed as with the form of entropy:
I(X;Y)=H (X)-H (X | Y) (4)
Wherein:
H (X) is the intrinsic entropy of variable X, and H (X | Y) it is the entropy of variable X on the premise of variable Y occurs.
As a further improvement on the present invention, step 4) in, high frequency, intermediate frequency separation, intermediate frequency, low frequency separation according to
Lower step determines:
The mutual trust entropy between adjacent IMF is calculated, according to mutual trust entropy theory, adjacent Mutual information entropy there will be following rule:
First reduce and increase afterwards, then in increase:
According to IMF during formula (7) selection information entropy minimum of a value as the separation for distinguishing high frequency and intermediate frequency, and distinguish intermediate frequency
With the separation of low frequency by carrying out to radar voice signal after empirical mode decomposition, according to noise in each intrinsic mode function point
Cloth situation finally determines that fixed threshold FT is 10-1;If the amplitude of the maximum of IMFs is less than threshold value FT, then IMFs is as low
Frequency part.
As a further improvement on the present invention, step 5) in, to containing noisy intrinsic mode function Jing adaptive thresholds
Denoising, is that the function of wherein adaptive threshold is to radar speech de-noising using empirical mode decomposition method:
Wherein, N is the length of signal, and σ is the noise variance that each rank IMF estimates:
To HFS using formula (8) threshold function table, the collapse threshold that intermediate-frequency section is adopted for:
Soft threshold method is used to eliminate the noise signal of high frequency and intermediate-frequency section:
Wherein, sign is sign function, and i represents the i-th rank intrinsic mode function.
As a further improvement on the present invention, step 5) in, the high frequency with remaining low frequency component and Jing after threshold denoising
Voice signal is reconstructed with intermediate frequency component, reconstruction signal is represented with following formula:
Wherein, the high fdrequency component number that k is to determine, the intrinsic mode number of components for distinguishing high frequency and intermediate frequency that l is to determine, n
It is the intrinsic mode number after empirical modal boundary.
The present invention is due to using above-mentioned technology, being allowed to the good effect having compared with existing radar speech enhancement technique
It is:
The present invention is by the millimeter radar voice signal to obtaining Jing after empirical mode decomposition, in intrinsic mode function
On the basis of noise content is analyzed, according in intrinsic mode function the characteristics of noise profile, using adaptive threshold to height
Frequently, intermediate frequency, low frequency part carry out denoising such that it is able to according to noise content number targetedly eliminate voice in noise
Content, with stronger adaptability and validity, particularly can be good at retaining the detailed information of voice, contribute in radar
Voice quality is significantly improved.Show this can have with targetedly empirical mode decomposition method using the example of the method
Effect eliminates the noise content in millimetre-wave radar voice, compared with traditional sound enhancement method, with stronger adaptability, and
The intelligibility of voice can be significantly improved on the basis of radar noise in voice content is eliminated.Do not causing radar voice signal
On the premise of distortion, radar noise in voice is efficiently removed.The present invention can be for later using bioradar detection human body language
Sound aspect provides effective technical support.Therefore, the present invention has stronger use value in terms of radar voice noise is eliminated
And application prospect.
Further, the intrinsic mode function to decomposing is divided into high frequency, intermediate frequency and the part of low frequency three, by observing eigen mode
Noise profile situation in state function, be by radar noise in voice characteristic distributions:Noise focuses primarily upon HFS, intermediate frequency portion
Divide predominantly useful primary speech signal, but still contain a small amount of noise signal, it is thin that low frequency part is mainly voice signal
Section signal.On this basis high frequency and intermediate frequency component number are selected by Mutual information entropy, wherein, HFS contains
Much noise, arranges higher threshold value.And the separation for distinguishing intermediate frequency and low frequency finds out fixed threshold 10 according to experiment-1.If
The amplitude of the maximum of IMFs is less than threshold value FT, then these IMFs are low frequency part.For intermediate-frequency section arranges higher-frequency
The less adaptive threshold in part, low frequency part does not do denoising.
【Description of the drawings】
Fig. 1 is a kind of method flow diagram of elimination millimeter wave bioradar noise in voice;
Fig. 2 is one section of radar phonetic material Jing after empirical mode decomposition, wherein (a) is containing noisy original radar language
Message number;B () is each rank intrinsic mode function after empirical mode decomposition;
Fig. 3 is the original radar voice signal of radar system collection, wherein, when (a) being millimetre-wave radar voice signal
Domain waveform;B () is the sound spectrograph of millimetre-wave radar voice signal;
Fig. 4 is that traditional algorithm spectrum-subtraction carries out the voice signal after de-noising, wherein, (a) it is millimetre-wave radar voice signal
Time domain waveform;B () is the sound spectrograph of millimetre-wave radar voice signal;
Fig. 5 is that the empirical mode decomposition described in this patent and Mutual information entropy algorithm eliminate the radar voice signal after noise,
Wherein, (a) be millimetre-wave radar voice signal time domain waveform;B () is the sound spectrograph of millimetre-wave radar voice signal.
【Specific embodiment】
Below in conjunction with the accompanying drawings, the specific embodiment of the present invention is described in detail, but the invention is not restricted to the enforcement
Example.In order that the public has to the present invention thoroughly understanding, in present invention below preferably applies example concrete details is described in detail.
Referring to Fig. 1, the present invention eliminates the general principle of the method for millimeter wave bioradar noise in voice and is:First to original
Beginning radar voice signal carries out empirical mode decomposition;Noise profile in intrinsic mode function Jing after empirical mode decomposition is carried out
Divide;The separation of HFS and intermediate-frequency section is divided using Mutual information entropy, each rank intrinsic mode function is carried out
Adaptive Wavelet Thrinkage;Radar voice signal after process is reconstructed.
A kind of method of elimination millimeter wave bioradar noise in voice of the present invention, it is comprised the following steps that:
First, bioradar voice is sampled, obtains grandfather tape and make an uproar radar voice signal;
Secondly, the original radar voice signal to gathering carries out empirical mode decomposition;
It is as follows that millimetre-wave radar voice signal of making an uproar to described band carries out empirical mode decomposition step:
Assume that given voice signal is x (t), empirical mode decomposition method is original by this by " screening " process adaptive
Signal decomposition is intrinsic mode function (IMFs), and each IMF is a sub- frequency content of primary signal.Its is specific
Screening process is as follows:
1) all of maximum of signal x (t) and minimum point are found out.
2) maximum and minimum point are fitted respectively by three sample interpolations, respectively obtain maximum envelope eu(t) and
Minimum envelope ed(t)。
3) average m of maximum envelope and minimum envelope is calculated1(t)=(eu(t)+ed(t))/2, and use original letter
Number deducting the average can obtain h1(t)=x (t)-m1(t).
4) h is judged1T whether () meet two screening primary conditions of IMF:
(a) in all of signal length, the zero passage of each IMF points must count out with extreme value it is identical, or at most
Difference one.
B () is necessary for zero in the average of any sample point, maximum envelope and minimum envelope, that is to say, that IMF is
It is geometrically symmetric with regard to time shaft.
Now, if h1T () meets two primary conditions of IMF, then IMF1(t)=h1(t)
If 5) h1T () is unsatisfactory for two primary conditions of IMF, then h1T () is used as new primary signal repeat step 1),
2), 3) h is obtained2(t)=h1(t)-m2(t), if h2T () meets IMF and sieves condition then IMF substantially1(t)=h2(t).Now such as
Fruit h2T () is unsatisfactory for IMF screenings condition then a new stopping screening threshold value is performed.The stopping criterion such as following formula:
Generally, SD spans are between 0.2 and 0.3.If h2T () meets stopping criterion, then IMF1(t)=
h2(t).If h2T () is unsatisfactory for, then by h2(t) repeat step 1) -4), under continuation h is goti(t), until hiT () meets two
Individual stop condition meets stopping criterion.Then IMF1(t)=hi(t)。
6) when getting IMF1After (t), it is removed from primary signal and obtains residual signals r1(t):r1(t)=x (t)-
IMF1(t). now, residual signals are used as a new primary signal, repeat step 1-5, until obtaining next residual signals.
So decompose finally, then residual signals can be represented with following formula:rn(t)=rn-1(t)-IMFn(t)。
Now, rnT () is a monotonic sequence, after the completion of screening, primary signal is broken down into multiple intrinsic mode function IMF1
(t),IMF2(t),…IMFn(t) and residual sequence rn(t). therefore, primary signal is represented by:
It is intrinsic mode function by original radar voice signal Jing empirical mode decompositions.Fig. 2 is one section of Jing empirical modal point
Radar phonetic material after solution.Fig. 2 (a) is containing noisy original radar voice signal.Fig. 2 (b) is after empirical mode decomposition
Each rank intrinsic mode function, from top to bottom, the frequency of the intrinsic mode function is gradually lowered.Due to radar voice it is main
Noise source be white Gaussian noise, therefore, in general, noise can be distributed in whole intrinsic mode function, from figure (b) in can
To find out, voice signal waveform, predominantly noise signal are substantially not visible in first three rank intrinsic mode function, and from 4-9 ranks
Intrinsic mode starts, mainly voice signal waveform, and there is partial noise signal.From the 10th rank to last single order, eigen mode
The frequency of state function is low and amplitude very little, but the wherein detailed information comprising some voices, therefore, we can make such
Assume:By original radar voice signal Jing empirical mode decompositions, HFS can be decomposed into, intermediate-frequency section, low frequency part is made an uproar
Sound focuses primarily upon HFS, and intermediate-frequency section is mainly useful primary speech signal, and low frequency is believed for the details of voice signal
Breath.
Again, noise content analysis is carried out to the intrinsic mode function after empirical mode decomposition.
Then, according to radar voice noise characteristic distributions, the selection of noise component(s) number is carried out using Mutual information entropy, with area
Divide high frequency and intermediate frequency separation, using given threshold intermediate frequency and low frequency separation are distinguished.
It is as follows that bioradar voice intrinsic mode function calculates mutual trust entropy method:
Mutual information entropy is all generally that the mutual trust entropy of nonnegative number, two variable Xs and Y can be calculated according to the following formula:
Wherein, p (x, y) is the Joint Distribution of variable X and Y, and p (x) and p (y) are respectively marginal distributions.Or Mutual information entropy
Following formula is expressed as with the form of entropy:
I(X;Y)=H (X)-H (X | Y) (4)
Wherein:
H (X) is the intrinsic entropy of variable X, and H (X | Y) it is the entropy of variable X on the premise of variable Y occurs.
When event X is more uncertain, entropy is bigger. and in general, the correlation between two variables is stronger, then mutual between variable
Information entropy is bigger.According to above feature, the mutual trust entropy between adjacent IMF is calculated, according to mutual trust entropy theory, adjacent mutual trust entropy will
Have following rule:First reduce and increase afterwards, then in increase:
According to IMF during formula (7) selection information entropy minimum of a value as the separation for distinguishing high frequency and intermediate frequency, and distinguish intermediate frequency
With the separation of low frequency by carrying out to radar voice signal after empirical mode decomposition, according to noise in each intrinsic mode function point
Cloth situation finally determines that fixed threshold (FT) is 10-1;If the amplitude of the maximum of IMFs is less than threshold value FT, then IMFs is
Low frequency part.
Finally, the intrinsic mode function Jing after self-adaptive solution is reconstructed and obtains enhanced radar voice.
To containing noisy intrinsic mode function Jing Adaptive Wavelet Thrinkages, its step is as follows:
Using empirical mode decomposition method in radar speech de-noising, the selection of threshold value plays vital effect.It is logical
Often the generic function of threshold value by:
Herein, N is the length of signal, and σ is the noise variance that each rank IMF estimates:
Herein, the threshold function table of formula (8) is adopted to HFS, it is relatively small for intermediate-frequency section noisiness,
To avoid voice signal distortion, the collapse threshold that intermediate-frequency section is adopted for:
Soft threshold method is used to eliminate the noise signal of high frequency and intermediate-frequency section:
Then, the high frequency and intermediate frequency component with remaining low frequency component and Jing after threshold denoising carries out weight to voice signal
Structure, reconstruction signal is represented with following formula:
Wherein, the high fdrequency component number that k is to determine, the intrinsic mode number of components for distinguishing high frequency and intermediate frequency that l is to determine, n
It is the intrinsic mode number after empirical modal boundary.
Generally speaking, the radar voice enhancement algorithm that the application is proposed is mainly following steps,
1. original radar voice signal x (t) is sieved using empirical mode decomposition.
2. the Energy-Entropy of each component is calculated by formula (5) and (6).
3. the Mutual information entropy for calculating each adjacent component presses formula (4).
4. the separation of high frequency and intermediate frequency component is determined according to formula (7).
5. the separation of intermediate frequency and low frequency component is determined using IMF fixed thresholds.
6. the noise in high frequency and intermediate frequency voice signal is leached using formula (8)-(11).
7. voice signal is reconstructed with the height after the low frequency component and formula (12) denoising not processed, intermediate frequency component.
Embodiment
In order to ensure the uniformity of sound source, the recording materials " 1-2-3-4-5-6 " of a women experimenter is selected, and by sound
Case is being played at 5 meters of radar system.
Radar voice is strengthened with following step:
1) bioradar voice is sampled, obtains grandfather tape and make an uproar radar voice signal;
2) the original radar voice signal to gathering carries out empirical mode decomposition;
3) noise content analysis is carried out to the intrinsic mode function after empirical mode decomposition;
4) according to radar voice noise characteristic distributions, using Mutual information entropy the selection of noise component(s) number is carried out;
5) intrinsic mode function Jing after self-adaptive solution is reconstructed and obtains enhanced radar voice.
Result is contrasted, and be refer to shown in Fig. 3, Fig. 4, Fig. 5, wherein (a) is millimetre-wave radar voice signal in each figure
Time domain waveform, scheme (b) be millimetre-wave radar voice signal sound spectrograph.
Fig. 2 is the original radar voice signal gathered by radar system, either from the time domain of original radar voice signal
In waveform or sound spectrograph it can be seen that, voice signal receive noise jamming, these noise sources mainly have electromagnetism harmonic noise,
Circuit noise and ambient noise.The presence of these noises often have impact on the quality of radar voice signal.
Fig. 4 is to carry out the voice signal after de-noising by traditional algorithm spectrum-subtraction, as can be seen that voice letter from time domain waveform
Noise in number is significantly eliminated, the noise especially mourned in silence in section, again it can be seen that Jing from Fig. 4 (b) sound spectrographs
After spectrum-subtraction process, the noise mourned in silence in section is significantly eliminated, but in HFS, some non-primary radar voice letters
Composition in number is introduced, and is referred to as " music noise ", and this often affects voice quality acoustically.
Fig. 5 is that the empirical mode decomposition described in this patent and Mutual information entropy algorithm eliminate the radar voice signal after noise.
It can be seen that the noise in radar voice signal has also obtained significantly suppressing, although can see in Fig. 5 (b)
Arrive, in non-speech segment, however it remains part residual noise, it is to avoid causing voice signal distortion that this is the method, using adaptive
Answer the result of threshold value.For the validity of further checking the method, Subjective audiometry experiment, experimental result table are synchronously carried out
Bright, compared with spectrum cuts algorithm, the method for proposition effectively eliminates radar voice letter on the premise of voice signal distortion is not caused
Noise in number, improves voice quality.
Although the method for eliminating the noise in millimetre-wave radar voice that the present invention is discussed has specific aim, it is especially suitable for
Millimetre-wave radar voice signal, but the scope that the present invention is used is not limited to millimetre-wave radar voice signal, for centimetre
Ripple radar voice signal, and some acoustic speech signals for gathering under equivalent environment equally have important directive significance and
Reference value.Embodiments of the present invention are explained in detail above in conjunction with accompanying drawing, not the present invention is imposed any restrictions, it is all
It is substantially above embodiment is made any simple modification, change and equivalent structure change according to the technology of the present invention,
Still fall within the protection domain of technical solution of the present invention.