CN100524464C - Adaptive filter method and apparatus for improving speech quality of mobile communication apparatus - Google Patents
Adaptive filter method and apparatus for improving speech quality of mobile communication apparatus Download PDFInfo
- Publication number
- CN100524464C CN100524464C CNB2005101351203A CN200510135120A CN100524464C CN 100524464 C CN100524464 C CN 100524464C CN B2005101351203 A CNB2005101351203 A CN B2005101351203A CN 200510135120 A CN200510135120 A CN 200510135120A CN 100524464 C CN100524464 C CN 100524464C
- Authority
- CN
- China
- Prior art keywords
- noise
- vector
- voice
- probability
- fft
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 32
- 230000003044 adaptive effect Effects 0.000 title claims description 107
- 238000010295 mobile communication Methods 0.000 title claims description 19
- 238000001914 filtration Methods 0.000 claims description 38
- 230000000694 effects Effects 0.000 claims description 32
- 238000002789 length control Methods 0.000 claims description 24
- 238000004364 calculation method Methods 0.000 claims description 12
- 230000004044 response Effects 0.000 claims description 12
- 230000035939 shock Effects 0.000 claims description 12
- 230000008676 import Effects 0.000 claims description 10
- 230000021615 conjugation Effects 0.000 claims description 7
- 230000015572 biosynthetic process Effects 0.000 claims description 3
- 238000004458 analytical method Methods 0.000 claims description 2
- 238000010586 diagram Methods 0.000 description 9
- 238000011946 reduction process Methods 0.000 description 3
- 238000005728 strengthening Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 230000006978 adaptation Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 239000004615 ingredient Substances 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
Images
Landscapes
- Noise Elimination (AREA)
- Filters That Use Time-Delay Elements (AREA)
Abstract
The invention relates to a self-adapting wave filter method which adjusts the self-adapting step length by computing the probability of the vice. It also relates to the self-adapting method application in the FD-NLMS analog of the mobile communicating device.
Description
Technical field
The present invention relates to the auto-adaptive filtering technique in the communications field, relate in particular to improved NLMS (the Frequency-Domain Normalized Least Mean Squares) algorithm that is used for mobile communication equipment.
Background technology
NLMS (Normalized Least Mean Squares) algorithm is a kind of common adaptive algorithm.Fig. 1 carries out the block diagram of noise reduction process for using the NLMS wave filter.Need two microphones when using the NLMS wave filter to carry out noise reduction process, one of them microphone is away from the speaker, and near noise source, the signal of being gathered is a reference noise signal; Another microphone is gathered by the voice signal of noise pollution as far as possible near the speaker.Yet in mobile communcations system, this requirement often is false.Because mobile communication equipment itself is size-constrained, often still can gather into more voice away from speaker's microphone.If the NLMS adaptive filter algorithm is not carried out suitable change, the voice after this situation will cause handling have tangible echo, and voice quality is descended.
FD-NLMS is frequency domain (Frequency-Domain) NLMS algorithm, it is a kind of form of NLMS algorithm, its purpose mainly is in order to save the calculated amount of NLMS auto adapted filtering, to replace the convolution on the time domain and ask relevant calculating in the method for using frequency domain to multiply each other.Its detailed principle document " sef-adapting filter theory " (Adaptive Filter Theory) that sees reference, Fourth Edition, SimonHaykin.
Fig. 2 has provided and has utilized FD-NLMS to eliminate the algorithm block diagram of noise processed in the prior art.
The step of the FD-NLMS algorithm of prior art (referring to) as follows:
1) N new reference noise sample v[0 of reference noise signal end input] ..., v[N-1], and and N the sample v ' [0] of last input,, v ' [N-1] forms the vector V of a new 2N length, V={v ' [0] ... v ' [N-1], v[0] ... v[N-1] };
2) V is carried out the FFT computing, obtain its FFT coefficient vector U, U=FFT (V)=u[0], u[1] ... u[2N-1] }, the length of U is 2N;
3) the FFT coefficient F of U and sef-adapting filter shock response
WCarry out dot product, product is carried out IFFT, obtain Y '.Y ' is real number vector (imaginary part is 0), abandons the top n value of Y ', makes it become a new vector Y, Y={y[0] ..., y[N-1] }, its length is N;
4) the new noisy speech s[0 of noisy speech end input] ..., s[N-1], obtain E, E={e[0 after S and Y subtract each other] ..., e[N-1] }=s[0]-y[0] ..., s[N-1]-y[N-1], E is the voice signal after strengthening;
5) insert N 0 value in the front of E and form vector E ', E '=0 ..., and 0, e[0] ..., e[N-1] }, E ' is carried out the FFT computing obtain vector F, F=FFT (E ');
6) vector U is got conjugation and obtain U
H, U
H=U, U
HCarry out dot product with F and obtain vector G '.G ' is carried out the IFFT computing obtain vector H, H=IFFT (G '), H are that length is the real number vector (imaginary part is 0) of 2N, the back N dimension value zero setting of H, and carry out the FFT computing, obtain vector G, and G is that length is the complex vector of 2N;
7) upgrade vector F
W, F
W=F
W+ μ G, μ are constant, perhaps a μ=c/E
n, c is a constant, E
n(its method of estimation does not here describe in detail, asks for an interview list of references " AdaptiveFilter Theory ", 4 for the current energy that estimates
ThEdith, Simon Haykin work);
8) get back to step 1, till phonetic entry is intact.
In said method, when calculating clean speech, in fact utilized following formula:
In formula (1), s[n] represent that noisy speech is at n value constantly, u[n] represent that reference noise is at n value constantly, w[i] and be the coefficient of sef-adapting filter, sef-adapting filter is the N-1 rank.Formula (1) shows, if contain strong phonetic element among the reference noise u, this phonetic element is relevant with phonetic element among the noisy speech S, may cause the estimated value of auto adapted filtering
Principal ingredient is not a noise, but the linear combination and the clean speech of current voice and former voice had correlationship.E[n like this] might concerning original pure voice, obtain a less value earlier, and also because e[n] may contain the linear combination of former phonetic element, cause a kind of echo phenomenon.
Summary of the invention
The purpose of this invention is to provide the adaptive filter method and the device that can improve the mobile communication equipment voice quality.
To achieve these goals, the invention provides a kind of adaptive filter method, it comprises the steps:
1) gathers one section Noisy Speech Signal;
2) gather one section reference noise signal;
3), according to current auto adapted filtering coefficient described Noisy Speech Signal is carried out filtering and obtain noise in this Noisy Speech Signal with described reference noise input adaptive wave filter;
4) noise of getting rid of in Noisy Speech Signal in the described Noisy Speech Signal obtains the clean speech signal;
5) calculate the probability that contains voice in adaptive step and the noisy speech, adjust this adaptive step according to the probability that contains voice in the current reference noise;
6) according to clean speech signal, reference noise, adjusted auto adapted filtering step-length to current auto adapted filtering coefficient update.
According to a further aspect in the invention, also provide a kind of improved FD-NLMS algorithm that is applicable to mobile communication equipment, it comprises the steps:
1) N new reference noise sample v[0 of reference noise signal end input] ..., v[N-1], and and N the reference noise sample v ' [0] of last input,, v ' [N-1] forms the vector V of a new 2N length, V={v ' [0] ... v ' [N-1], v[0] ... v[N-1] };
2) V is carried out the FFT computing, obtain its FFT coefficient vector U, U=FFT (V)=u[0], u[1] ... u[2N-1] }, the length of U is 2N;
3) the FFT coefficient F of U and sef-adapting filter shock response
WCarry out dot product, product is carried out IFFT, obtain Y ', wherein Y ' is the real number vector, and its imaginary part is 0, abandons the top n value of Y ', makes it become a new vector Y, Y={y[0] ..., y[N-1] }, its length is N;
4) the new noisy speech s[0 of noisy speech end input] ..., s[N-1], obtain E, E={e[0 after S and Y subtract each other] ..., e[N-1] }=s[0]-y[0] ..., s[N-1]-y[N-1], E is the language signal after strengthening;
5) insert N 0 value in the front of E and form vector E ', E '=0 ..., and 0, e[0] ..., e[N-1] }, E ' is carried out the FFT computing obtain vector F, F=FFT (E ');
6) vector U is got conjugation and obtain U
H, U
H=U, U
HCarry out dot product with F and obtain vector G '.G ' is carried out the IFFT computing obtain vector H, H=IFFT (G '), H are that length is the real number vector of 2N, and its imaginary part is 0, the back N dimension value zero setting of H, and carries out the FFT computing, obtains vector G, and G is that length is the complex vector of 2N;
7) analytical procedure 2) in FFT coefficient vector U, calculate the probability that contains voice in adaptive step μ and the current reference noise, wherein μ is constant, perhaps μ=c/E
n, c is a constant, En is the current energy that estimates; Adjust adaptive step μ according to the probability that contains voice in the current reference noise, when the probability that contains voice is high, reduce step size mu, when the probability that contains voice is low, increase step size mu;
8) upgrade vector F
W, F
W=F
W+ μ G, F
WUse when the filtering next time as the FFT coefficient of new sef-adapting filter shock response;
9) get back to step 1), till phonetic entry is intact.
In said method of the present invention, on the one hand, in above-mentioned steps 7) in, can carry out the voice posterior probability estimation to current FFT coefficient U with reference to noise, output vector p={p[0], ..., p[N-1] }, p[i wherein] be illustrated in the voice posterior probability that exists on each FFT coefficient, like this, it is p '=1-p that there is the posterior probability vector of noise in each FFT coefficient, p ' and step size mu multiply each other then, obtain new adaptive step vector μ=μ [0] ..., μ [N-1] }, it has provided the adaptive step on each FFT coefficient.
Wherein, above-mentioned new adaptive step μ is multiplied by rule of thumb determined constant C 1 again and obtains, so that new adaptive step μ has suitable size to guarantee convergence of algorithm after multiplying each other with step size mu by the posterior probability p ' that makes noise.
On the other hand, in above-mentioned steps 7) in, can find the device processing through speech activity to current FFT coefficient U with reference to noise, the output probability α that voice exist in N reference noise sample of current input, calculate the probability β=1-α of noise in current existence, β and step size mu multiply each other then, obtain new adaptive step μ.When speech activity finds that device adopts the mode of hard decision, its α=0 o'clock, expression does not have voice, α=1 o'clock, expression has voice; When speech activity found that device adopts the mode of soft-decision, its α was likelihood probability 0≤α≤1, the probability that the expression voice exist.
The present invention also provides a kind of adaptive filter device, and it comprises: sef-adapting filter is used for estimating according to reference noise signal the noise of voice signal; Subtracter is used for deducting the estimated noise that comes out of sef-adapting filter from Noisy Speech Signal, obtains and voice signal behind the noise is eliminated in output, and the voice signal that will eliminate simultaneously behind the noise feeds back to sef-adapting filter as feedback signal; The step-length control module, it calculates the probability that contains voice in the current reference noise, thereby the probability that has noise in the acquisition reference noise. described sef-adapting filter calculates adaptive step μ, exist the probability of noise to adjust adaptive step μ in the reference noise that is obtained according to the step-length control module, upgrade the auto adapted filtering coefficient, feed back to wherein feedback signal according to described subtracter then and reappraise noise in the voice signal.
The present invention also provides a kind of frequency domain adaptive filtering device that improves the mobile communication equipment voice quality, and described mobile communication equipment comprises at least two Mikes, and one of them Mike is used to import reference noise signal, and another Mike is used to import noisy speech.Described frequency domain adaptive filtering device comprises: input block, be used for N new reference noise sample v[0 of input],, v[N-1] and last N the reference noise sample v ' [0] that imports ... v ' [N-1] forms the vector V of a new 2N length, V={v ' [0] ..., v ' [N-1], v[0] ... v[N-1] }; The FFT arithmetic element is used for vector V is carried out the FFT computing, obtains its FFT coefficient vector U, U=FFT (V)=u[0], u[1] ... u[2N-1] }, the length of U is 2N; Filter unit is used for the FFT coefficient F with vector U and sef-adapting filter shock response
WCarry out dot product; The IFFT arithmetic element is used for the resultant product of filter unit is carried out IFFT, obtains Y ', and wherein Y ' is the real number vector, and its imaginary part is 0, abandons the top n value of Y ', makes it become a new vector Y, Y={y[0] ..., y[N-1] }, its length is N; Output unit is used for the new noisy speech vector S=s[0 with noisy speech end input] ..., s[N-1] obtain vector E after subtracting each other with vector Y, E={e[0] ..., e[N-1] }=s[0]-y[0],, s[N-1]-y[N-1], the voice signal E after output strengthens; Auto adapted filtering coefficient update unit, the front of the voice signal vector E that it is exported at output unit 5 insert N 0 value formation vector E '=0 ..., 0, e[0] ..., e[N-1] }, E ' is carried out the FFT computing obtains vector F, F=FFT (E '); Vector U is got conjugation obtain U
H, U
H=U, U
HCarry out dot product with F and obtain vector G '; G ' is carried out the IFFT computing obtain vector H, H=IFFT (G '), H are that length is the real number vector of 2N, and its imaginary part is 0, the back N dimension value zero setting of H, and carries out the FFT computing, obtains vector G, and G is that length is the complex vector of 2N; Analyze the resulting FFT coefficient vector of FFT arithmetic element U and obtain upgrading step size mu, pass through F
W=F
W+ μ G upgrades the FFT coefficient vector F of sef-adapting filter shock response
WWherein said auto adapted filtering updating block comprises the step-length control module, and it analyzes the resulting FFT coefficient vector of FFT arithmetic element U, calculates the probability that contains voice in adaptive step μ and the current reference noise, and wherein μ is constant, perhaps μ=c/E
n, c is a constant, En is the current energy that estimates; Adjust adaptive step μ according to the probability that contains voice in the current reference noise, when the probability that contains voice is high, reduce step size mu, when the probability that contains voice is low, increase step size mu.
According to said apparatus of the present invention, described step-length control module comprises: voice posterior probability estimation unit, current FFT coefficient U with reference to noise is carried out the voice posterior probability estimation, output vector p={p[0] ..., p[N-1], p[i wherein] be illustrated in the voice posterior probability that exists on each FFT coefficient, i=0 ..., N-1; Noise probability calculation unit, according to voice posterior probability estimation unit resultant voice posterior probability p[i], obtaining the posterior probability vector that there is noise in each FFT coefficient is p '=1-p, p ' and step size mu are multiplied each other, obtain new adaptive step vector μ={ μ [0], ..., μ [N-1] }, μ [i] has provided the adaptive step on each FFT coefficient.
According to said apparatus of the present invention, described step-length control module comprises: speech activity is found device, current FFT coefficient U with reference to noise is handled the output probability α that voice exist in current N reference noise sample; Noise probability calculation unit according to finding the resultant probability α of device at speech activity, calculates the probability β=1-α of noise in current existence, and β and step size mu multiply each other then, obtain new adaptive step μ.Described speech activity finds that device adopts the mode of hard decision, and when α=0, expression does not have voice, α=1 o'clock, and expression has voice.
Utilize the present invention, can avoid the echo phenomenon, and when effectively eliminating noise, do not cause the reduction of voice quality.
Description of drawings
Fig. 1 carries out the block diagram of noise reduction process for using the NLMS wave filter;
Fig. 2 has provided and has utilized FD-NLMS to eliminate the algorithm block diagram of noise processed in the prior art;
Fig. 3 has provided the synoptic diagram of adaptive filter method of the present invention;
Fig. 4 has provided the algorithm block diagram that the FD-NLMS of utilization of the present invention eliminates noise processed;
Fig. 5 has provided a kind of implementation of step-length control of the present invention;
Fig. 6 has provided the another kind of implementation of step-length control of the present invention;
Fig. 7 has provided the synoptic diagram of adaptive filter device of the present invention;
Fig. 8 has provided the adaptive filter device structural representation of raising mobile communication equipment voice quality of the present invention;
Fig. 9 has provided the synoptic diagram of a kind of step-length control module of the present invention;
Figure 10 has provided the synoptic diagram of another kind of step-length control module of the present invention.
Embodiment
Describe the specific embodiment of the present invention in detail below in conjunction with accompanying drawing, but these embodiments are not limitation of the present invention.
Because the mobile communication equipment size is little, so when the small type mobile devices that adopts the diamylose gram, as mobile phone, when having two adjacent speech sources and noise source, the diamylose gram, promptly Mike A and Mike B record speech source, suppose that Mike A is near and far away from noise source from speech source, Mike B is near and far away from speech source from noise source, at this moment, can be recorded to the voice signal of a two-channel by Mike A and B.Yet because the existence of noise source, Mike A and B can be recorded to noise, thereby cause both poor sound quality.Be example just more than with two adjacent speech sources and noise source, in fact, multiple often speech source and noise source coexistence, at this moment, noise will become more serious to the influence of voice.
In this case,, reduce to export the echo in the voice, improve voice quality, can adopt following adaptive filter method in order to reduce or to eliminate noise.As shown in Figure 3, suppose that the signal that Mike A is recorded is a noisy speech, the signal that Mike B is recorded is a reference noise.Adaptive filter method of the present invention comprises the steps:
Gather one section Noisy Speech Signal and one section reference noise signal;
With the reference noise signal input adaptive wave filter of being gathered 11;
Sef-adapting filter 11 is according to its auto adapted filtering coefficient, utilizes reference noise signal to estimate noise in the Noisy Speech Signal;
Utilize subtracter 12, deduct the estimated noise that goes out of sef-adapting filter 11 in the Noisy Speech Signal from channel A, obtain the clean speech signal;
The clean speech signal is offered sef-adapting filter 11 as feedback signal;
Calculate the probability that contains voice in adaptive step μ and the noisy speech, wherein μ=c/E
n, c is a constant, E
nBe the energy of the current reference noise signal that estimates, adjust adaptive step μ, when the probability that contains voice is high, reduce adaptive step μ, when the probability that contains voice is low, increase adaptive step μ according to the probability that contains voice in the current reference noise;
According to clean speech signal, reference noise signal, adjusted auto adapted filtering step size mu, upgrade current auto adapted filtering coefficient according to following formula:
U(kL+i)＝{u(kL+i)，u(kL+i-1)……u(kL+i-N+1)}
Here, suppose that sef-adapting filter is the N rank, W[k+1] be the k+1 time auto adapted filtering coefficient, L upgrades an auto adapted filtering coefficient W[k after representing L reference noise sample], u (kL+i) represents kL+i reference noise value constantly, e[kL+i] represent the kL+i clean noise figure in the moment.Wherein L can be for greater than 1 natural number, and L equals N usually.
In above-mentioned adaptive filter method of the present invention, adjusted auto adapted filtering step size mu can be undertaken by following dual mode:
A kind of mode is, carries out the voice posterior probability estimation to current with reference to noise, obtains to exist the posterior probability of noise, and this probability and adaptive step multiply each other then, obtain new adaptive step.Particularly, utilize existing NMLS algorithm computation to go out self-adaptation step size mu=c/E
n, c is a constant, E
nEnergy for the current reference noise signal that estimates; Import L reference noise signal, this L reference noise signal and before this L reference noise signal form the vector of length 2L, be transformed in the frequency domain by FFT, estimate the probability P that voice exist on each Frequency point then, P=P={p[0], P[1] ... p[2L-1] }.About the method for estimation of P, at document " Speech enhancement for non-stationary noise environments ", IsraelCohen
*, Baruch Berdugo, Signal Processing (2001), Elsevier discloses a kind of.Calculate the adaptive step μ on each Frequency point:
μ＝u·c
1·(1-P)＝{u·c
1·(1-p[0])，u·c
1·(1-p[1])，...，u·c
1·p[2L-1]}
Wherein, the constant of C1 for rule of thumb drawing.
Another kind of mode is, finds that through speech activity device handles with reference to noise to current, calculates the probability of the current existence of noise, and this probability and adaptive step multiply each other then, obtain new adaptive step.Particularly, utilize existing NMLS algorithm computation to go out self-adaptation step size mu=c/E
n, c is a constant, E
nEnergy for the current reference noise signal that estimates; Find the individual reference noise signal of device input L (L 〉=1) to speech activity; Speech activity is found the value of device according to a current L reference noise signal, estimates the Probability p that voice exist in a current L reference noise, and p is a scalar; Calculate the adaptive step μ of this L reference noise signal correspondence, μ=uc
2(1-p).C wherein
2Be the constant that rule of thumb draws.
Said method of the present invention can be applied in the existing NLMS adaptive filter algorithm, to make improvements.As shown in Figure 4, it has provided the frequency domain adaptive filtering method of raising mobile communication equipment voice quality of the present invention, it is used to comprise at least two Mikes' mobile communication equipment, one of them Mike is the reference noise signal input end, another Mike is the noisy speech input end, here, the phonetic element of signals with noise end is stronger than the phonetic element of reference noise input end, a little less than the noise contribution of the noise contribution of signals with noise end than reference noise input end.The frequency domain adaptive filtering method of raising mobile communication equipment voice quality of the present invention comprises the steps:
1) N new reference noise sample v[0 of reference noise signal end input] ..., v[N-1], and and N the reference noise sample v ' [0] of last input,, v ' [N-1] forms the vector V of a new 2N length, V={v ' [0] ... v ' [N-1], v[0] ... v[N-1] };
2) V is carried out the FFT computing, obtain its FFT coefficient vector U, U=FFT (V)=u[0] ..., u[2N-1] }, the length of U is 2N;
3) the FFT coefficient FW of U and sef-adapting filter shock response carries out dot product, and product is carried out IFFT, obtains Y ', and wherein Y ' is real number vector (imaginary part is 0), abandon the top n value of Y ', make it become a new vector Y, Y={y[0],, y[N-1] }, its length is N;
4) the new noisy speech s[0 of noisy speech end input] ..., s[N-1], obtain E, E={e[0 after S and Y subtract each other] ..., e[N-1] }=s[0]-y[0] ..., s[N-1]-y[N-1], the voice signal E after output strengthens;
5) insert N 0 value in the front of E and form vector E ', E '=0 ..., and 0, e[0] ..., e[N-1] }, E ' is carried out the FFT computing obtain vector F, F=FFT (E ');
6) vector U is got conjugation and obtain U
H, U
H=U, U
HCarry out dot product with F and obtain vector G '.G ' is carried out the IFFT computing obtain vector H, H=IFFT (G '), H are that length is the real number vector (imaginary part is 0) of 2N, the back N dimension value zero setting of H, and carry out the FFT computing, obtain vector G, and G is that length is the complex vector of 2N;
7) analyze in step 2) in FFT coefficient vector U, calculate the probability that contains voice in step size mu and the current reference noise, wherein μ is constant, perhaps μ=c/E
n, c is a constant, En is the energy of the current reference noise signal that estimates; Adjusting adaptive step μ according to the probability that contains voice in the current reference noise is μ, when the probability that contains voice is high, reduces step size mu, when the probability that contains voice is low, increases step size mu; Wherein the computing method of step size mu are identical with the computing method of step size mu in the existing FD-NLMS algorithm;
8) upgrade vector F
W, F
W=F
W+ μ G.
9) get back to step 1), till phonetic entry is intact.
Fig. 5 has provided a kind of implementation method of step-length control of the present invention.As shown in Figure 5, current FFT coefficient vector U with reference to noise is carried out the voice posterior probability estimation, exports a vector p={p[0] ..., p[N-1], p[i] be illustrated in the posterior probability that has voice on each FFT coefficient.The voice posterior probability estimation is a kind of existing technology, about the visible list of references of its specific algorithm " Speechenhancement for non-stationary noise environments ", Israel Cohen
*, BaruchBerdugo, Signal Processing (2001), Elsevier.After voice posterior probability p calculated, each FFT coefficient existed the posterior probability p ' of noise to draw by p '=1-p.P ' and step size mu (scalar) multiply each other, obtain new adaptive step μ=μ [0] ..., μ [N-1].Wherein μ adapts to the step-length that step size computation draws with existing FD-NLMS algorithm, about the concrete mode of its calculating referring to document " Adaptive Filter Theory ", 4
ThEdition, Simon Haykin work.μ=and μ [0] ..., μ [N-1] } provided the adaptive step on each FFT coefficient.In order to make adaptive step be more suitable for the present situation, new adaptive step μ can be multiplied by rule of thumb determined constant C 1 and obtain by after p ' and step size mu (scalar) are multiplied each other again.
Fig. 6 has provided another kind of step-length control implementation method of the present invention.Speech activity among Fig. 6 finds that device (Voice Activity Detector is called for short VAD) is a prior art.The FFT coefficient vector U of reference noise is input to speech activity and finds device VAD, and VAD is output as factor alpha.α is illustrated in the probability that voice exist in N the reference noise sample of current input.VAD has hard decision or two kinds of forms of soft-decision.When adopting hard decision, speech activity is found device output α=0 (not having voice) or α=1 (voice are arranged), and when adopting soft-decision, speech activity finds that device is output as likelihood probability 0≤α≤1, the probability that the expression voice exist.Use β=1-α to calculate the probability β of noise in current existence, β and step size mu multiply each other then, obtain new adaptive step μ.μ is a scalar in such cases, promptly all adopts identical adaptation step size mu on all FFT coefficient points.
The present invention also provides a kind of adaptive filter device, and as shown in Figure 7, it comprises: sef-adapting filter is used for estimating according to reference noise signal the noise of voice signal; Subtracter is used for deducting the estimated noise that comes out of sef-adapting filter from Noisy Speech Signal, obtains and voice signal behind the noise is eliminated in output, and the voice signal that will eliminate simultaneously behind the noise feeds back to sef-adapting filter as feedback signal; The step-length control module, it calculates the probability that contains voice in the reference noise, thereby obtains to exist in the reference noise probability of noise; Wherein, described sef-adapting filter calculates adaptive step μ, exist the probability of noise to adjust adaptive step μ in the reference noise that is obtained according to the step-length control module, upgrade the auto adapted filtering coefficient, feed back to wherein feedback signal according to described subtracter then and reappraise noise in the voice signal.
Wherein, described step-length control module can be realized by following scheme: it comprises: voice posterior probability estimation unit, carry out the voice posterior probability estimation to current with reference to noise, the voice posterior probability that obtains existing; Noise probability calculation unit according to the resulting voice posterior probability in voice posterior probability estimation unit, obtains existing the posterior probability of noise.Described step-length control module can also be realized by following proposal: it comprises speech activity discovery device, handles the probability α that voice exist in the output reference noise sample with reference to noise to current; Noise probability calculation unit according to finding the resultant probability α of device at speech activity, calculates the probability β=1-α of noise in current existence, and β and adaptive step μ multiply each other then, obtain new adaptive step μ.Wherein, described speech activity finds that device can adopt the mode of hard decision, and when α=0, expression does not have voice, α=1 o'clock, and expression has voice.Described speech activity finds that device also can adopt the mode of soft-decision, and α is likelihood probability 0≤α≤1.
Step-length control module wherein can be independent of the sef-adapting filter setting, as shown in Figure 7; Also can be arranged within the sef-adapting filter. when the step-length control module was arranged in the sef-adapting filter, can also combine with the adaptive step computing function constituted a step size computation and control module.
Fig. 8 has provided the frequency domain adaptive filtering device of raising mobile communication equipment voice quality of the present invention.Described mobile communication equipment comprises at least two microphones, and a Mike is used to import reference noise signal, and a Mike is used to import noisy speech.The frequency domain adaptive filtering device comprises input block 1, FFT arithmetic element 2, auto adapted filtering coefficient update unit 5, filter unit 3, IFFT arithmetic element 4, output unit 6.Wherein input block 1 will be imported N new reference noise sample v[0] ..., v[N-1] and last N the reference noise sample v ' [0] that imports ... v ' [N-1] forms the vector V of a new 2N length, V={v ' [0] ... v ' [N-1], v[0] ... v[N-1] }; 2 pairs of vector V of FFT arithmetic element are carried out the FFT computing, obtain its FFT coefficient vector U, U=FFT (V)=u[0], u[1] ... u[2N-1] }, the length of U is 2N; Filter unit 3 is with the FFT coefficient F of vector U and sef-adapting filter shock response
WCarry out dot product; 4 pairs of resultant products of filter unit of IFFT arithmetic element carry out IFFT, obtain Y ', and wherein Y ' is the real number vector, and its imaginary part is 0, abandons the top n value of Y ', make it become a new vector Y, Y={y[0] ..., y[N-1] }, its length is N; Output unit 6 is used for noisy speech vector S=s[0 that noisy speech end input is new] ..., s[N-1] obtain vector E after subtracting each other with vector Y, E={e[0] ..., e[N-1] }=s[0]-y[0],, s[N-I]-y[N-1], E is the voice signal after strengthening.The front of the voice signal vector E that it is exported at output unit 5 insert N 0 value formation vector E '=0 ..., 0, e[0] ..., e[N-1] }, E ' is carried out the FFT computing obtains vector F, F=FFT (E '); Vector U is got conjugation obtain U
H, U
H=U, U
HCarry out dot product with F and obtain vector G '; G ' is carried out the IFFT computing obtain vector H, H=IFFT (G '), H are that length is the real number vector of 2N, and its imaginary part is 0, the back N dimension value zero setting of H, and carries out the FFT computing, obtains vector G, and G is that length is the complex vector of 2N; Analyze the resulting FFT coefficient vector of FFT arithmetic element U, pass through F
W=F
W+ μ G upgrades the FFT coefficient vector F of sef-adapting filter shock response
W, wherein μ is the adaptive step of control renewal speed; Wherein said auto adapted filtering updating block 5 comprises step size computation and control module 51, it is used to analyze FFT arithmetic element 2 resulting FFT coefficient vector U, calculate the probability that contains voice in adaptive step μ and the current reference noise, wherein μ is constant, perhaps μ=c/E
n, c is a constant, En is the current energy that estimates; Adjusting adaptive step μ according to the probability that contains voice in the current reference noise is μ, when the probability that contains voice is high, reduces step size mu, when the probability that contains voice is low, increases step size mu.
In frequency domain adaptive filtering device shown in Figure 8, described step size computation and control module 51 can adopt structure shown in Figure 9.As shown in Figure 9, step size computation and control module 51 comprise: voice posterior probability estimation unit 511, current FFT coefficient U with reference to noise is carried out the voice posterior probability estimation, output vector p={p[0] ..., p[N-1], p[i wherein] be illustrated in the voice posterior probability that exists on each FFT coefficient, i=0 ..., N-1; Noise probability calculation unit 512, according to voice posterior probability estimation unit 511 resultant voice posterior probability p[i], obtaining the posterior probability vector that there is noise in each FFT coefficient is p '=1-p, p ' and step size mu are multiplied each other, obtain new adaptive step vector μ={ μ [0], ..., μ [N-1] }, μ [i] has provided the adaptive step on each FFT coefficient.Wherein, after the posterior probability p ' that makes noise multiplies each other with step size mu, can be multiplied by rule of thumb determined constant C 1 again in order to make new adaptive step μ more be applicable to current environment.
In frequency domain adaptive filtering device shown in Figure 8, described step size computation and control module 51 also can adopt structure shown in Figure 10.As shown in figure 10, step size computation and control module 51 comprise: speech activity is found device 513, current FFT coefficient U with reference to noise is handled the output probability α that voice exist in current N reference noise sample; Noise probability calculation unit 512 according to finding the resultant probability α of device at speech activity, calculates the probability β=1-α of noise in current existence, and β and step size mu multiply each other then, obtain new adaptive step μ.Described speech activity finds that device 513 can adopt the mode of hard decision also can adopt the mode of soft-decision.When speech activity found that device 513 adopts the mode of hard decisions, in α=0 o'clock, expression did not have voice, α=1 o'clock, and expression has voice; When adopting the mode of soft-decision, α is likelihood probability 0≤α≤1.
In said apparatus, described step size computation and control module 51 also can be divided into the step size computation module and these two functional modules of step-length control module realize according to function.
The present invention can be applied to mobile communication equipment, for example in the mobile phone etc., can improve voice quality effectively, reduces noise, avoids echo.
Foregoing is not to be used for limiting the specific embodiment of the present invention, and all modification and change or combinations of carrying out according to the main inventive concept of this method all should belong to protection domain of the presently claimed invention.
Claims (19)
1. an adaptive filter method is characterized in that, comprises the steps:
1) gathers one section Noisy Speech Signal;
2) gather one section reference noise signal;
3), according to current auto adapted filtering coefficient described Noisy Speech Signal is carried out filtering and obtain noise in this Noisy Speech Signal with described reference noise input adaptive wave filter;
4) noise of getting rid of from Noisy Speech Signal in the described Noisy Speech Signal obtains the clean speech signal;
5) calculate the probability that contains voice in adaptive step and the reference noise, adjust this adaptive step according to the probability that contains voice in the current reference noise;
6) according to clean speech signal, reference noise, adjusted adaptive step to current auto adapted filtering coefficient update.
2. method according to claim 1 is characterized in that,
In described step 5), when the probability that contains voice in the reference noise is high, reduce adaptive step, when the probability that contains voice in the reference noise is low, increase adaptive step.
3. method according to claim 1 and 2 is characterized in that,
In described step 5), carry out the voice posterior probability estimation to current with reference to noise, obtain to exist the posterior probability of noise, this probability and adaptive step multiply each other then, obtain new adaptive step.
4. method according to claim 1 and 2 is characterized in that,
In described step 5), find device processing with reference to noise through speech activity to current, calculate the probability of the current existence of noise, this probability and adaptive step multiply each other then, obtain new adaptive step.
5. a frequency domain adaptive filtering method that improves the mobile communication equipment voice quality comprises the steps:
1) N new reference noise sample v[0 of reference noise signal end input] ..., v[N-1], and and N the reference noise sample v ' [0] of last input,, v ' [N-1] forms the vector V of a new 2N length, V={v ' [0] ... v ' [N-1], v[0] ... v[N-1] };
2) vector V is carried out the FFT computing, obtain its FFT coefficient vector U, U=FFT (V)=u[0], u[1] ... u[2N-1] }, the length of U is 2N;
3) the FFT coefficient F of vector U and sef-adapting filter shock response
WCarry out dot product, product is carried out IFFT, obtain Y ', wherein Y ' is the real number vector, and its imaginary part is 0, abandons the top n value of Y ', makes it become a new vector Y, Y={y[0] ..., y[N-1] }, its length is N;
4) the new noisy speech S=s[0 of noisy speech end input] ..., s[N-1] obtain E, E={e[0 after subtracting each other with Y] ..., e[N-1] }=s[0]-y[0] ..., s[N-1]-y[N-1], the voice signal vector E after output strengthens;
5) insert N 0 value in the front of voice signal vector E and form vector E ', E '=0 ..., and 0, e[0] ..., e[N-1] }, E ' is carried out the FFT computing obtain vector F, F=FFT (E ');
6) vector U is got conjugation and obtain U
H, U
H=U, U
HCarry out dot product with F and obtain vector G '; G ' is carried out the IFFT computing obtain vector H, H=IFFT (G '), H are that length is the real number vector of 2N, and its imaginary part is 0, the back N dimension value zero setting of H, and carries out the FFT computing, obtains vector G, and G is that length is the complex vector of 2N;
7) analytical procedure 2) in FFT coefficient vector U, calculate the probability that contains voice in adaptive step μ and the current reference noise; Adjusting adaptive step μ according to the probability that contains voice in the current reference noise is μ;
8) upgrade vector F
W, F
W=F
W+ μ G, F
WUse when the filtering next time as the FFT coefficient of new sef-adapting filter shock response;
9) get back to step 1), till phonetic entry is intact.
6. method according to claim 5 is characterized in that,
In described step 7), current FFT coefficient U with reference to noise is carried out the voice posterior probability estimation, output vector p={p[0], ..., p[N-1] }, p[i wherein] be illustrated in the voice posterior probability that exists on each FFT coefficient, like this, it is p '=1-p that there is the posterior probability vector of noise in each FFT coefficient, p ' and step size mu multiply each other then, obtain new adaptive step vector μ=μ [0] ..., μ [N-1] }, it has provided the adaptive step on each FFT coefficient.
7. method according to claim 6 is characterized in that, new adaptive step μ is multiplied by rule of thumb determined constant C 1 again and obtains after multiplying each other with step size mu by the posterior probability p ' that makes noise.
8. method according to claim 5 is characterized in that,
In described step 7), current FFT coefficient U with reference to noise is found the device processing through speech activity, the output probability α that voice exist in N reference noise sample of current input, calculate the probability β=1-α of noise in current existence, β and step size mu multiply each other then, obtain new adaptive step μ.
9. method according to claim 8 is characterized in that,
In described step 7), speech activity finds that device adopts the mode of hard decision, and when α=0, expression does not have voice, α=1 o'clock, and expression has voice.
10. method according to claim 8 is characterized in that,
In described step 7), speech activity finds that device adopts the mode of soft-decision, and at this moment Shu Chu α is likelihood probability 0≤α≤1.
11. an adaptive filter device comprises:
Sef-adapting filter is used for estimating according to reference noise signal the noise of voice signal;
Subtracter is used for deducting the estimated noise that comes out of sef-adapting filter from Noisy Speech Signal, obtains and voice signal behind the noise is eliminated in output, and the voice signal that will eliminate simultaneously behind the noise feeds back to sef-adapting filter as feedback signal;
It is characterized in that, also comprise:
The step-length control module, it calculates the probability that contains voice in the current reference noise, thereby obtains to exist in the reference noise probability of noise,
Described sef-adapting filter calculates adaptive step μ, exist the probability of noise to adjust adaptive step μ in the reference noise that is obtained according to the step-length control module, upgrade the auto adapted filtering coefficient, feed back to wherein feedback signal according to described subtracter then and reappraise noise in the voice signal.
12. adaptive filter device according to claim 11 is characterized in that,
Described step-length control module comprises:
Voice posterior probability estimation unit carries out the voice posterior probability estimation to current with reference to noise, the voice posterior probability that obtains existing;
Noise probability calculation unit according to the resulting voice posterior probability in voice posterior probability estimation unit, obtains existing the posterior probability of noise, with existing the posterior probability and the adaptive step μ of noise to multiply each other, obtains new adaptive step μ.
13. adaptive filter device according to claim 11 is characterized in that,
Speech activity is found device, handles the probability α that voice exist in the output reference noise sample with reference to noise to current;
Noise probability calculation unit according to finding the resultant probability α of device at speech activity, calculates the probability β=1-α of noise in current existence, and β and adaptive step μ multiply each other then, obtain new adaptive step μ.
14. frequency domain adaptive filtering device that improves the mobile communication equipment voice quality, described mobile communication equipment comprises at least two Mikes, one of them Mike is used to import reference noise signal, another Mike is used to import noisy speech, it is characterized in that described frequency domain adaptive filtering device comprises:
Input block is used for N new reference noise sample v[0 of input] ..., v[N-1] and last N the reference noise sample v ' [0] that imports,, v ' [N-1] forms the vector V of a new 2N length, V={v ' [0] ... v ' [N-1], v[0] ... v[N-1] };
The FFT arithmetic element is used for vector V is carried out the FFT computing, obtains its FFT coefficient vector U, U=FFT (V)=u[0], u[1] ... u[2N-1] }, the length of U is 2N; Filter unit is used for the FFT coefficient F with vector U and sef-adapting filter shock response
WCarry out dot product;
The IFFT arithmetic element is used for the resultant product of filter unit is carried out IFFT, obtains Y ', and wherein Y ' is the real number vector, and its imaginary part is 0, abandons the top n value of Y ', makes it become a new vector Y, Y={y[0] ..., y[N-1] }, its length is N;
Output unit is used for the new noisy speech vector S=s[0 with noisy speech end input] ..., s[N-1] obtain vector E after subtracting each other with vector Y, E={e[0] ..., e[N-1] }=s[0]-y[0],, s[N-1]-y[N-1], the voice signal E after output strengthens;
Auto adapted filtering coefficient update unit, the front of the voice signal vector E that it is exported at output unit insert N 0 value formation vector E '=0 ..., 0, e[0] ..., e[N-1] }, E ' is carried out the FFT computing obtains vector F, F=FFT (E '); Vector U is got conjugation obtain U
H, U
H=U, U
HCarry out dot product with F and obtain vector G '; G ' is carried out the IFFT computing obtain vector H, H=IFFT (G '), H are that length is the real number vector of 2N, and its imaginary part is 0, the back N dimension value zero setting of H, and carries out the FFT computing, obtains vector G, and G is that length is the complex vector of 2N; U obtains upgrading step size mu by the resulting FFT coefficient vector of FFT arithmetic element, utilizes F then
W=F
W+ μ G upgrades the FFT coefficient vector F of sef-adapting filter shock response
W,
Described auto adapted filtering updating block comprises:
Step size computation and control module, it utilizes the resulting FFT coefficient vector of FFT arithmetic element U, calculates the probability that contains voice in adaptive step μ and the current reference noise; Adjusting adaptive step μ according to the probability that contains voice in the current reference noise is μ, when the probability that contains voice is high, reduces step size mu, when the probability that contains voice is low, increases step size mu.
15. device according to claim 14 is characterized in that, described step size computation and control module comprise:
The step size computation unit, it utilizes the resulting FFT coefficient vector of FFT arithmetic element U, calculates the probability that contains voice in adaptive step μ and the current reference noise;
The step-length control module, adjusting adaptive step μ according to the probability that contains voice in the current reference noise is μ, when the probability that contains voice is high, reduces step size mu, when the probability that contains voice is low, increases step size mu.
16., it is characterized in that described step size computation and control module or step-length control module comprise according to claim 14 or 15 described devices:
Voice posterior probability estimation unit carries out the voice posterior probability estimation to current with reference to noise, output vector p={p[0] ..., p[N-1], p[i wherein] be illustrated in the voice posterior probability that exists on each FFT coefficient, i=0 ..., N-1;
Noise probability calculation unit, according to voice posterior probability estimation unit resultant voice posterior probability p[i], obtaining the posterior probability vector that there is noise in each FFT coefficient is p '=1-p, p ' and step size mu are multiplied each other, obtain new adaptive step vector μ={ μ [0], ..., μ [N] }, μ [i] has provided the adaptive step on each FFT coefficient.
17., it is characterized in that described step size computation and control module or step-length control module comprise according to claim 14 or 15 described devices:
Speech activity is found device, current FFT coefficient U with reference to noise is handled the output probability α that voice exist in current N reference noise sample;
Noise probability calculation unit according to finding the resultant probability α of device at speech activity, calculates the probability β=1-α of noise in current existence, and β and step size mu multiply each other then, obtain new adaptive step μ.
18. device according to claim 17 is characterized in that, described speech activity finds that device adopts the mode of hard decision, and when α=0, expression does not have voice, α=1 o'clock, and expression has voice.
19. device according to claim 17 is characterized in that, described speech activity finds that device adopts the mode of soft-decision, and α is likelihood probability 0≤α≤1.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CNB2005101351203A CN100524464C (en) | 2005-12-23 | 2005-12-23 | Adaptive filter method and apparatus for improving speech quality of mobile communication apparatus |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CNB2005101351203A CN100524464C (en) | 2005-12-23 | 2005-12-23 | Adaptive filter method and apparatus for improving speech quality of mobile communication apparatus |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN1805011A CN1805011A (en) | 2006-07-19 |
| CN100524464C true CN100524464C (en) | 2009-08-05 |
Family
ID=36866987
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CNB2005101351203A Expired - Fee Related CN100524464C (en) | 2005-12-23 | 2005-12-23 | Adaptive filter method and apparatus for improving speech quality of mobile communication apparatus |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN100524464C (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101179294B (en) * | 2006-11-09 | 2012-07-04 | 黄山好视达通信技术有限公司 | Self-adaptive echo eliminator and echo eliminating method thereof |
| CN101119135B (en) * | 2007-07-04 | 2010-09-01 | 深圳市融创天下科技发展有限公司 | Step parameter regulation means and equipment for eliminating echo |
| CN102737646A (en) * | 2012-06-21 | 2012-10-17 | 佛山市瀚芯电子科技有限公司 | Real-time dynamic voice noise reduction method for single microphone |
| CN103730125B (en) | 2012-10-12 | 2016-12-21 | 华为技术有限公司 | A kind of echo cancelltion method and equipment |
| CN102938254B (en) * | 2012-10-24 | 2014-12-10 | 中国科学技术大学 | Voice signal enhancement system and method |
| CN106356070B (en) * | 2016-08-29 | 2019-10-29 | 广州市百果园网络科技有限公司 | A kind of acoustic signal processing method and device |
-
2005
- 2005-12-23 CN CNB2005101351203A patent/CN100524464C/en not_active Expired - Fee Related
Also Published As
| Publication number | Publication date |
|---|---|
| CN1805011A (en) | 2006-07-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| KR101225556B1 (en) | Method for determining updated filter coefficients of an adaptive filter adapted by an lms algorithm with pre-whitening | |
| CN101964670B (en) | Echo suppression method and apparatus thereof | |
| Szurley et al. | Efficient computation of microphone utility in a wireless acoustic sensor network with multi-channel Wiener filter based noise reduction | |
| CN101222555B (en) | System and method for improving audio speech quality | |
| EP3080975B1 (en) | Echo cancellation | |
| US7383178B2 (en) | System and method for speech processing using independent component analysis under stability constraints | |
| CN112017679B (en) | Method, device and equipment for updating adaptive filter coefficients | |
| JP2003534570A (en) | How to suppress noise in adaptive beamformers | |
| JP2002541753A (en) | Signal Noise Reduction by Time Domain Spectral Subtraction Using Fixed Filter | |
| CN102811267B (en) | Near-end speech interference cancelling system and mobile communication terminal | |
| CN107134281A (en) | Adaptive filter coefficient update method during a kind of adaptive echo is eliminated | |
| CN107026950B (en) | A kind of frequency domain adaptive echo cancel method | |
| US20070253565A1 (en) | Methods and systems for reducing acoustic echoes in communication systems | |
| CN100524464C (en) | Adaptive filter method and apparatus for improving speech quality of mobile communication apparatus | |
| CN112602150A (en) | Noise estimation method, noise estimation device, voice processing chip and electronic equipment | |
| CN111755020B (en) | Stereo echo cancellation method | |
| CN107071196B (en) | A kind of adaptive echo cancellation method | |
| CN115457928A (en) | Echo cancellation method and system based on neural network double-talk detection | |
| Rakesh et al. | A novel RLS based adaptive filtering method for speech enhancement | |
| JP2000323962A (en) | Adaptive identification method and device, and adaptive echo canceler | |
| CN1355916A (en) | Signal noise reduction by time-domain spectral substraction | |
| KR20170082598A (en) | Adaptive interchannel discriminitive rescaling filter | |
| JP3673727B2 (en) | Reverberation elimination method, apparatus thereof, program thereof, and recording medium thereof | |
| Ryu et al. | The performance of an adaptive noise canceller with DSP processor | |
| KR20080038714A (en) | Postprocessing method for removing cross talk |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| C17 | Cessation of patent right | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20090805 Termination date: 20111223 |