CN102682781A - Self-adaptive suppression method of time varying sinusoidal interference in audio signals - Google Patents

Abstract

The invention discloses a self-adaptive suppression method of time varying sinusoidal interference in audio signals. The method includes the following steps: step one, inputting original audio signals and performing AD conversion on the original audio signals; step two, calculating preliminary frequency of sinusoidal interference signals in the audio signals through a self-adaptive notch filter algorithm; step three, calculating amplitude, phase and accurate frequency of the sinusoidal interference signals in the audio signals through a least mean square (LMS) gradient descent algorithm; step four, deciding credibility of the sinusoidal interference signals through a decision algorithm, turning to step five if the credibility is true, and outputting the audio signals directly if the credibility is false; and step five, subtracting the calculated sinusoidal interference signals from the audio signals so as to obtain and output the audio signals after interference suppression. The self-adaptive suppression method of the time varying sinusoidal interference in the audio signals has the advantages of being capable of suppressing the time varying sinusoidal interference, insensitive to preliminary value setting, high in convergence rate, simple in calculation, high in quality, good in accuracy and stability, and suitable for processing all audio signals.

Description

Become the Adaptive Suppression method of sinusoidal interference in the time of in a kind of sound signal

Technical field

The present invention relates to the digital audio processing field, become the Adaptive Suppression method of sinusoidal interference when especially relating in a kind of sound signal.

Background technology

The audio pickup process becomes the interference of sine or cosine signal when regular meeting runs into.For example, the power frequency interference phenomenon, because the self-sustained oscillation phenomenon that acoustic feedback produces, and receive interference (like automobile, aircraft roar) of specific noise frequency or the like.Sinusoidal interference and cosine disturb and just on phase place, differ pi/2, so be referred to as sinusoidal interference.In most cases, sinusoidal interference has time variation, belongs to the non-stationary signal category.The energy of undesired signal or amplitude are bigger usually, and signal to noise ratio (S/N ratio) is very low, and signal is submerged among the interference usually.

One of key that becomes sinusoidal interference when inhibition and elimination becomes the frequency of sinusoidal signal when being estimation.The algorithm that becomes sinusoidal frequency when from non-stationary signal, extracting has adaptive resistance-trap wave technology, multiband search technique and LMS gradient prompt drop method etc.

The adaptive resistance-trap wave technology becomes sinusoidal signal when following the tracks of with an arrowband second order trapper, target be make through the error energy behind the trap minimum, the time to become sinusoidal frequency be the only parameter of its estimation.The initial value of single adaptive notch algorithm offset of sinusoidal Frequency Estimation is provided with very responsive, and different initial values can have a strong impact on speed of convergence, even can cause dispersing.

The basis of multiband search technique is the DFT algorithm, and computational complexity is high.When signal belonged to the non-stationary category, the estimated quality of this algorithm can obvious variation.

The target of LMS gradient prompt drop method is that the square error that makes input signal and become the difference of sinusoidal signal when estimative minimizes.Through gradient search, become frequency, amplitude and the phase place of sinusoidal signal in the time of can estimating simultaneously.In order to simplify LMS gradient prompt drop algorithm computation complicacy, LMS gradient prompt drop method often will be simplified the recursion formula of Frequency Estimation, and its Frequency Estimation precision and stability are lower than adaptive notch filter.

Because people's ear is very high to the susceptibility of voice signal, if estimate false sinusoidal interference, result can cause audio distortions; Otherwise, if because computational complexity causes time-delay to estimate sinusoidal interference, then again can be by the perception of people's ear.When especially when sinusoidal interference, becoming, when the algorithm robustness is bad,, can influence sound quality on the contrary owing to become sinusoidal signal can not in time follow the tracks of the time.

Summary of the invention

The present invention solves that existing in prior technology is difficult to fast, the accurate technical matters of the sinusoidal interference in the sound-inhibiting signal, provide a kind of to initial value insensitive, fast convergence rate is set, calculating is simple, quality is high, the Adaptive Suppression method of change sinusoidal interference in the sound signal of precision and good stability the time.

The present invention is directed to above-mentioned technical matters mainly is able to solve through following technical proposals: become the Adaptive Suppression method of sinusoidal interference in the time of in a kind of sound signal, may further comprise the steps:

Step 1, input original audio signal carry out the AD conversion to original audio signal, obtain digitized sound signal;

The first synchronizing frequency of sinusoidal interference signal in step 2, the use adaptive notch algorithm computation sound signal;

Amplitude, phase place and the precise frequency of sinusoidal interference signal in step 2, the use LMS gradient prompt drop algorithm computation sound signal;

Step 3, use decision algorithm are judged the confidence level of sinusoidal interference signal: if confidence level is true, then forward step 4 to; If confidence level is false, then direct output audio signal;

Step 4, from sound signal, deduct the sinusoidal interference signal that calculates, sound signal and output after disturbing are inhibited.

As preferably, the preliminary frequencies omega of sinusoidal interference signal in the said use adaptive notch algorithm computation sound signal ₀(n), the preliminary frequencies omega of sinusoidal interference signal ₀(n) with preliminary frequency correlation coefficient k ₀(n) relation is:

|k ₀(n)|＝|-cos(ω ₀(n))|

W (n) shown in Figure of description 2,

w(n)＝u(n)-k ₀(n-1)[1+α]w(n-1)-αw(n-2)

Definition:

$A\left(n\right)=\underset{k=0}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\λ}}^{n-k}{\left[2w(n-1)\right]}^2$

$B\left(n\right)=\underset{k=0}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\λ}}^{n-k}(w\left(n\right)+w(n-2))\·w(n-1)$

Preliminary frequency correlation coefficient k ₀(n) recursion formula is:

k ₀(n)＝-B(n)/A(n)

A(n)＝λA(n-1)+2w(n-1)·w(n-1)

B(n)＝λB(n-1)+w(n-1)·[w(n)+w(n-2)]

In the formula, the value of forgetting factor λ is more than or equal to 0.95 and smaller or equal to 0.99.

As preferably, if preliminary frequency correlation coefficient k ₀(n) absolute value then uses following formula that it is carried out Nonlinear Processing greater than 1:

As preferably, use the amplitude of sinusoidal interference signal in the LMS gradient prompt drop algorithm computation sound signal

Phase place

And precise frequency

Adopt following recursion formula (T in the formula _sBe the sampling period):

$\hat{P}(n+1)=\hat{P}\left(n\right)+2T_s{\mathrm{\μ}}_{P}e\left(n\right)\sin \widehat{\mathrm{\φ}}\left(n\right)$

$\widehat{\mathrm{\φ}}(n+1)=\widehat{\mathrm{\φ}}\left(n\right)+T_s\widehat{\mathrm{\ω}}\left(n\right)+2T_s{\mathrm{\μ}}_{\mathrm{\ω}}^{\′}{\mathrm{\μ}}_{\mathrm{\δ}}^{\′}e\left(n\right)\hat{P}\left(n\right)\cos \widehat{\mathrm{\φ}}\left(n\right)$

$\widehat{\mathrm{\ω}}(n+1)=\widehat{\mathrm{\ω}}\left(n\right)+2T_s{\mathrm{\μ}}_{\mathrm{\ω}}^{\′}e\left(n\right)\hat{P}\left(n\right)\cos \widehat{\mathrm{\φ}}\left(n\right)$

μ′ _ω＝μ _ωnT _s μ′ _δ＝μ _δ/μ′ _ω

Iterative formula is insensitive to the initial value of and

; Initial value can be set to 1.0,0.2 and 300 π respectively.

In the formula, definition error signal e (n) is:

$e\left(n\right)=u\left(n\right)-\widehat{\mathrm{\η}}\left(n\right)$

In the formula, u (n) is the input that comprises interference,

Be that just (surplus) string disturbs recursion result, μ _A, μ _ω, μ _δBe respectively 100,1000 and 0.02.As preferably; If the amplitude of undesired signal

then uses following formula to carry out Nonlinear Processing less than 0:

$\hat{P}(n+1)=\hat{P}\left(n\right),\mathrm{if}\hat{P}(n+1)<0.$

As preferably; Use decision algorithm to judge that the method for the confidence level of sinusoidal interference signal does, in 3 to 8 sinusoidal signal window ranges, calculate the square error of precise frequency and amplitude; If the square error of precise frequency is less than or equal to defined threshold (suggestion be defined as signal averaging energy mean square value 0.1%); And the amplitude square error is within specialized range or be forward recursive, and then confidence level is true, otherwise confidence level is false.

The substantial effect that the present invention brings is; With adaptive notch algorithm and LMS gradient prompt drop algorithm set; Have the advantage that response speed is fast, calculated amount is little, have simultaneously, avoided the defective of adaptive notch algorithm through Nonlinear Processing recursion amplitude, phase place and the insensitive characteristic of angular frequency initial value; Reduce the interdependency of algorithm convergence to iterative initial value; The initial value setting of revising the back algorithm influences Algorithm Convergence hardly, can judge the true and false of sinusoidal interference, has good inhibition effect and accuracy.

Description of drawings

Fig. 1 is the adaptive notch algorithm block diagram;

Fig. 2 is the lattice shape implementation structure of IIR second order trapper;

Fig. 3 is the ADAPTIVE MIXED algorithm block diagram;

Fig. 4 becomes (sampled point) in time variation diagram of sinusoidal interference signal for the test time spent;

Fig. 5 is the comparison of sinusoidal interference signal among Fig. 4 with the sinusoidal interference signal that calculates;

Fig. 6 is the mixed signal after common music signal stacking diagram 4 sinusoidal interference;

Fig. 7 is the estimation convergence process of the sinusoidal interference in Fig. 6 mixed signal;

Fig. 8 suppresses the complete audio frequency time-domain signal after the sinusoidal interference for Fig. 6 mixed signal;

Fig. 9 suppresses the local audio frequency time-domain signal after the sinusoidal interference for Fig. 6 mixed signal;

Among the figure: 201, IIR notch filter, 202, A (n) recursion, 203, B (n) recursion, 204, k ₀(n) recursion, 205, k ₀(n) Nonlinear Processing, 601, the time domain subtraction, 602, the adaptive notch frequency computation part, 603, LMS gradient prompt drop algorithm, 604, disturb confidence level to judge, 605, subtraction, 606, the phase place recursion, 607, the accompanying drawing recursion, 608, amplitude non-linearity handles.

Embodiment

Pass through embodiment below, and combine accompanying drawing, do further bright specifically technical scheme of the present invention.

Embodiment: the Adaptive Suppression method that becomes sinusoidal interference in the time of in a kind of sound signal of present embodiment; At first the sound signal of input is carried out the AD conversion; The sound signal of output numeral then is in interference calculation and inhibition module; Interference calculation with suppress module in the sound signal the time become sinusoidal signal and calculate and suppress, the sound signal after output is handled is amplified to amplifier, plays through loudspeaker at last.

The first step is the sinusoidal signal frequency that adopts in the adaptive notch algorithm estimated signal of revising.

Shown in Figure 1, the present invention utilizes the adaptive notch algorithm to estimate sinusoidal interference, and trapper adopts the second order iir digital filter, and its pulsed transfer function is:

$H\left(z\right)=\frac{N\left(z\right)}{N(z/\mathrm{\&alpha;})}=\frac{1+2k_0{z}^{-1}+k_1{z}^{-2}}{1+k_0(1+\mathrm{\&alpha;})z^{-1}+\mathrm{\&alpha;}{z}^{-2}},{\mathrm{<figure-callout\; id="0"\; label="k"\; filenames="120331090519.png,120331090525.png"\; state="\{\{state\}\}">k</figure-callout>}}_0=-\cos \left({\mathrm{\&omega;}}_0\right),0<\mathrm{\&alpha;}\&le;1$

Calculate intermediate variable A (n) according to recursion formula, calculate intermediate variable B (n), calculate intermediate variable k according to recursion formula according to recursion formula ₀(n).

Adaptive targets is that output energy mean square value is minimized:

$e\left(n\right)=\underset{k=0}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}^{n-k}{\left[\hat{s}\left(n\right)\right]}^2$

$=\underset{k=0}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}^{n-k}{[2w(n-1){\mathrm{<figure-callout\; id="0"\; label="k"\; filenames="120331090519.png,120331090525.png"\; state="\{\{state\}\}">k</figure-callout>}}_0+(w\left(n\right)+w(n-2))]}^2$

λ is a forgetting factor in the following formula, the frequency correlation coefficient k that obtains thus ₀(n) recursion formula is:

k ₀(n)＝-B(n)/A(n)

A(n)＝λA(n-1)+2w(n-1)·w(n-1)

B(n)＝λB(n-1)+w(n-1)·[w(n)+w(n-2)]

k ₀(n) with the relation of disturbing orthodox frequency be:

k ₀＝|-cos(ω ₀)|

In order to prevent to occur under the specific starting condition | k ₀|＞1, thus cause iteration not restrain, to k ₀(n) carry out Nonlinear Processing, the Nonlinear Processing process is:

Utilize the adaptive notch algorithm of above-mentioned correction can follow the tracks of the variation of sinusoidal interference frequency fast, especially its speed of convergence does not receive the influence of initial value, phenomenon more can not occur dispersing.

Fig. 2 is the implementation structure of second order iir digital filter.Lattice shape implementation structure has increased the numerical evaluation stability of IIR filtering.

Fig. 4 is a test signal.The frequency of test signal is the time varying signal by varies with cosine, and this time varying signal is polluted by 1 white Gaussian noise by the zero-mean variance.

Fig. 5 is the frequency detecting tracking results of Fig. 4.The result shows except that having approximately estimated time the delay of 100 sampled points, and estimated frequency has been followed the tracks of well and become sinusoidal signal frequency when real-time.

Second step was to adopt the LMS-gradient to estimate the amplitude and the phase place of sinusoidal interference signal.

Fig. 3 is the ADAPTIVE MIXED algorithm block diagram, and the Frequency Estimation result that the LMS-gradient is estimated the adaptive notch algorithm of application first step correction accurately estimates amplitude and the phase place disturbed.

Suppose estimation just (surplus) string undesired signal be:

$\widehat{\mathrm{\&eta;}}\left(n\right)=\hat{P}\left(n\right)\sin (\widehat{\mathrm{\&phi;}}\left(n\right)+\widehat{\mathrm{\&delta;}}\left(n\right))$

Definition error signal e (n):

e(n)＝d(u(n)， $\widehat{\mathrm{\&eta;}}\left(n\right))=u\left(n\right)-\widehat{\mathrm{\&eta;}}\left(n\right)$

The discrete recursion formula that obtains frequency, amplitude and phase place after the square error of error signal e (n) minimizes is:

$\hat{P}(n+1)=\hat{P}\left(n\right)+2T_s{\mathrm{\&mu;}}_{A}e\left(n\right)\sin \widehat{\mathrm{\&phi;}}\left(n\right)$

$\widehat{\mathrm{\&omega;}}(n+1)=\widehat{\mathrm{\&omega;}}\left(n\right)+2T_s{\mathrm{\&mu;}}_{\mathrm{\&omega;}}^{\&prime;}e\left(n\right)\hat{P}\left(n\right)\cos \widehat{\mathrm{\&phi;}}\left(n\right)$

$\widehat{\mathrm{\&phi;}}(n+1)=\widehat{\mathrm{\&phi;}}\left(n\right)+T_s\widehat{\mathrm{\&omega;}}\left(n\right)+2T_s{\mathrm{\&mu;}}_{\mathrm{\&omega;}}^{\&prime;}{\mathrm{\&mu;}}_{\mathrm{\&delta;}}^{\&prime;}e\left(n\right)\hat{P}\left(n\right)\cos \widehat{\mathrm{\&phi;}}\left(n\right)$

$e\left(n\right)=u\left(n\right)-\widehat{\mathrm{\&eta;}}\left(n\right)$

Wherein,

Directly adopt the result of the adaptive notch algorithm of revising (602), μ _A, μ ' _ω, μ ' _δBe respectively

Prompt drop coefficient μ _A, μ _ω, μ _δ(603) be LMS-gradient algorithm for estimating, (605) will be imported to subtract each other in time domain with output and obtain error signal; (606) the phase place recursion is calculated, and (607) amplitude recursion is calculated.

For preventing iteration diverges, amplitude Estimation is made following Nonlinear Processing:

$\hat{P}(n+1)=\hat{P}\left(n\right),$ $\mathrm{if}\hat{P}(n+1)<0.$

In the 3rd step, the sequence estimation of frequency and amplitude is further handled.

In order to prevent to detect false sinusoidal signal, Interference Estimation is carried out confidence level judge.Estimate in the sinusoidal signal window ranges square error of calculated rate estimation and amplitude Estimation at 3～8.When implicit true sinusoidal interference signal, the square error of frequency should be greater than prescribed threshold, and the amplitude variance or is forward recursive (like the feedback self-exciting process) within specialized range.

In the 4th step, suppress the sinusoidal interference signal

With sampled signal time domain with estimate sinusoidal interference signal subtraction, the useful signal behind the sinusoidal interference that the is inhibited signal.

Fig. 6～Fig. 9 shows the mixed signal after the typical music signal stack sinusoidal interference, eliminates the time domain waveform and the estimated result of sinusoidal interference signal through algorithm of the present invention.

Fig. 6 is an original signal, and Fig. 7 is a Frequency Estimation, and the estimated result explanation fast, accurately, has stably detected fixedly interference sinusoidal signal.

Fig. 8 and Fig. 9 are the output behind the inhibition sinusoidal interference signal, and result and test tone music signal are very approaching, and sinusoidal interference signal suppressing effect reaches about 20dB.

Specific embodiment described herein only is that the present invention's spirit is illustrated.Person of ordinary skill in the field of the present invention can make various modifications or replenishes or adopt similar mode to substitute described specific embodiment, but can't depart from spirit of the present invention or surmount the defined scope of appended claims.

Although this paper has used terms such as sinusoidal interference, adaptive notch algorithm morely, do not get rid of the possibility of using other term.Using these terms only is in order to describe and explain essence of the present invention more easily; It all is contrary with spirit of the present invention being construed to any additional restriction to them.

Claims (6)

1. become the Adaptive Suppression method of sinusoidal interference in the time of in the sound signal, it is characterized in that, may further comprise the steps:

Step 1, input original audio signal carry out the AD conversion to original audio signal, obtain digitized sound signal;

The first synchronizing frequency of sinusoidal interference signal in step 2, the use adaptive notch algorithm computation sound signal;

Amplitude, phase place and the precise frequency of sinusoidal interference signal in step 3, the use LMS gradient prompt drop algorithm computation sound signal;

Step 4, use decision algorithm are judged the confidence level of sinusoidal interference signal: if confidence level is true, then forward step 5 to; If confidence level is false, then direct output audio signal;

Step 5, from sound signal, deduct the sinusoidal interference signal that calculates, sound signal and output after disturbing are inhibited.

2. become the Adaptive Suppression method of sinusoidal interference in the time of in a kind of sound signal according to claim 1, it is characterized in that the preliminary frequencies omega of sinusoidal interference signal in the said use adaptive notch algorithm computation sound signal ₀(n), the preliminary frequencies omega of sinusoidal interference signal ₀(n) with preliminary frequency correlation coefficient k ₀(n) relation is:

|k ₀(n)|＝|-cos(ω ₀(n))|

Preliminary frequency correlation coefficient k ₀(n) recursion formula is:

k ₀(n)＝-B(n)/A(n)

A(n)＝λA(n-1)+2w(n-1)·w(n-1)

B(n)＝λB(n-1)+w(n-1)·[w(n)+w(n-2)]

In the formula, the value of forgetting factor λ is more than or equal to 0.95 and smaller or equal to 0.99.

3. become the Adaptive Suppression method of sinusoidal interference in the time of in a kind of sound signal according to claim 2, it is characterized in that, if preliminary frequency correlation coefficient k ₀(n) absolute value then uses following formula that it is carried out Nonlinear Processing greater than 1:

4. become the Adaptive Suppression method of sinusoidal interference in the time of in a kind of sound signal according to claim 1; It is characterized in that, use amplitude

phase place

and the precise frequency

of sinusoidal interference signal in the LMS gradient prompt drop algorithm computation sound signal to adopt following recursion formula:

$\hat{P}(n+1)=\hat{P}\left(n\right)+2T_s{\mathrm{\&mu;}}_{A}e\left(n\right)\sin \widehat{\mathrm{\&phi;}}\left(n\right)$

$\widehat{\mathrm{\&phi;}}(n+1)=\widehat{\mathrm{\&phi;}}\left(n\right)+T_s\widehat{\mathrm{\&omega;}}\left(n\right)+2T_s{\mathrm{\&mu;}}_{\mathrm{\&omega;}}^{\&prime;}{\mathrm{\&mu;}}_{\mathrm{\&delta;}}^{\&prime;}e\left(n\right)\hat{P}\left(n\right)\cos \widehat{\mathrm{\&phi;}}\left(n\right)$

$\widehat{\mathrm{\&omega;}}(n+1)=\widehat{\mathrm{\&omega;}}\left(n\right)+2T_s{\mathrm{\&mu;}}_{\mathrm{\&omega;}}^{\&prime;}e\left(n\right)\hat{P}\left(n\right)\cos \widehat{\mathrm{\&phi;}}\left(n\right)$

In the formula, definition error signal e (n) is:

$e\left(n\right)=u\left(n\right)-\widehat{\mathrm{\&eta;}}\left(n\right)$

In the formula, μ _A, μ _ω, μ _δBe respectively 100,1000 and 0.02.

5. become the Adaptive Suppression method of sinusoidal interference in the time of in a kind of sound signal according to claim 4; It is characterized in that; If the amplitude of undesired signal then uses following formula to carry out Nonlinear Processing less than 0:

$\hat{P}(n+1)=\hat{P}\left(n\right),\mathrm{if}\hat{P}(n+1)<0.$

6. become the Adaptive Suppression method of sinusoidal interference according to time in claim 1 or 2 or 3 or the 4 or 5 described a kind of sound signals, it is characterized in that, use decision algorithm to judge that the method for the confidence level of sinusoidal interference signal does; In 3 to 8 sinusoidal signal window ranges; Calculate the square error of precise frequency and amplitude, if the square error of precise frequency is less than or equal to defined threshold, and the amplitude square error is within specialized range or be forward recursive; Then confidence level is true, otherwise confidence level is false.

Images