CN104637494A - Double-microphone mobile equipment voice signal enhancing method based on blind source separation - Google Patents

Abstract

The invention discloses a double-microphone mobile equipment voice signal enhancing method based on blind source separation. The method comprises the following steps: receiving a voice signal through two microphones which are kept a distance d away from each other on an identical horizontal line at the bottom end of communication equipment to obtain an observation signal; performing mean-removing processing on the observation signal; performing de-correlation processing on the mean-removed signal; separating a mixed signal from the de-correlated signal by adopting an underdetermined blind separation method to obtain the estimated value of a mixture matrix and each sound source estimation; selecting a sound source with maximum energy from each sound source estimation for serving as a target human voice signal. By adopting the double-microphone mobile equipment voice signal enhancing method, the signal to noise ratio of a transmitted voice signal can be increased, and the intelligibility of the voice signal is enhanced.

Description

Based on two microphone mobile device voice signals enhancement methods of blind source separating

Technical field

The invention belongs to voice signals enhancement field, particularly relate to a kind of signal transmission signal to noise ratio (S/N ratio) that can improve, based on two microphone mobile device voice signals enhancement methods of blind source separating.

Background technology

More and more higher along with people's living standard, mobile device is as panel computer, phone has entered among daily life, become a kind of necessity in daily life, the purposes of current flat panel computer is not only and is seen document, video, photo, play games, since micro-letter, the application software such as qq are with phonetic function, the role of dull and stereotyped also performer's speech communication gradually, although the function of phone is more next various at present, but the major function of her speech communication be do not have vicissitudinous, and mobile phone incorporated micro-letter, these softwares with voice transmission capabilities of qq, therefore before Signal transmissions, we just wish there is high intelligibility, the voice signal of high validity, this just will inevitably use voice signals enhancement technology.

Traditional speech enhancement technique is some filtering techniques, as medium filtering, Wiener filtering, auto-adaptive filtering technique, but these filtering methods can only eliminate the noise of some special frequency channel, as the power frequency component of 50Hz, but these methods can not remove the noise of the long-time acoustic background existed.

Summary of the invention

The object of this invention is to provide a kind of signal transmission signal to noise ratio (S/N ratio) that can improve, based on two microphone mobile device voice signals enhancement methods of blind source separating.

The present invention is achieved by the following technical solutions:

Based on two microphone mobile device voice signals enhancement methods of blind source separating, comprise following step,

Step one: two microphones being arranged on d apart on the same level line of communication apparatus bottom, for received speech signal, obtain observation signal x=[x ₁; x ₂], x ₁the voice signal in a microphone, x ₂it is the voice signal in another microphone;

Step 2: average value processing is gone to observation signal;

Step 3: carry out decorrelative transformation to removing the signal after average value processing;

Step 4: owe to determine blind separating method to the signal employing after decorrelative transformation and carry out separating mixture of source signals, estimated value and each sound source of obtaining Mixture matrix are estimated;

Step 5: estimate that the sound source selecting energy maximum is as target person acoustical signal from each sound source.

The present invention is based on two microphone mobile device voice signals enhancement methods of blind source separating, can also comprise:

1, average value processing is gone to be to observation signal:

$v_i\left(t\right)=x_i\left(t\right)-\frac{1}{N}\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{x}_i\left(t\right);k$

Wherein, v _it (), for removing the signal after average value processing, N is the number of sampled point, i is the i-th road signal, is a kth sampled point.2, the method obtaining Mixture matrix estimated value is,

Step one: signal z (t) after decorrelative transformation is expressed as polar form is:

$\left\{\begin{array}{c}{l}_t=\sqrt{{\left(z_1^t\right)}^2+{\left(z_2^t\right)}^2}\\ {\mathrm{\θ}}_t={\tan }^{-1}(z_2^t/z_1^t)\end{array}\right.$

Wherein, signal z (t) after decorrelative transformation is 2D signal, and t is l _tand θ _trepresent radius and angle respectively;

Step 2: definition basis function φ:

Wherein, α is any direction and θ _tbetween angular deviation,

Obtain the overall potential function φ (θ, λ) about angle θ:

$\mathrm{\φ}(\mathrm{\θ},\mathrm{\λ})=\underset{t}{\mathrm{\Σ}}{l}_t\mathrm{\φ}\left(\mathrm{\λ}(\mathrm{\θ}-{\mathrm{\θ}}_t)\right)$

Wherein, λ is parameter;

Step 3: to the smoothing process of overall potential function, obtains new potential function φ ' (θ, λ), and each peak value of this potential function is a column vector a of hybrid matrix _i,

a _i＝[sin(π/2+θ _i),cos(π/2+θ _i)] ^T

Wherein, i=1,2 ..., npeak, npeak are total number of peak value;

Obtain the estimated value of hybrid matrix

$\tilde{A}={[a}_1,a_2,...,a_{\mathrm{npeak}}].$

3, the method obtaining the estimation of each sound source is:

Using all sample points as input, under As (t)=x (t) condition, ask and minimize l1 norm | s (t) | ₁,

$\min \underset{t=1}{\overset{T}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}\left|s_j\left(t\right)\right|,\mathrm{As}\left(t\right)=x\left(t\right)$

Obtain each sound source to estimate.

4, the energy that each sound source is estimated is:

$\mathrm{LE}=\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{\left(s_i\right)}^2,$

Wherein, n represents the length of whole signal, and j represents that sound source estimates s _ia jth point.

5, distance d is greater than or equal to 4.25cm, and the length of each microphone received speech signal is less than or equal to 30ms.

Beneficial effect

Because blind source separate technology is exactly solve in a kind of method not knowing to obtain under source signal and these prioris of aliasing information information source, profit in this way can background sound and noise remove, so the present invention is under the prerequisite not changing device hardware structure in a large number, just increase microphone, then according to current demand signal disposal route after two-way Speech processing, then use the method for blind source separating to voice signals enhancement.

The hardware cost of the increase communication apparatus that the present invention does not need a large amount of increases extra, outward appearance and volume, mainly improve intelligibility and the signal to noise ratio (S/N ratio) of communication input speech signal, make the other end can receive voice signal more clearly from computing method; The Enhancement Method of classic method voice signal just improves the sharpness of voice signal, but the noisy speech signals mixed in voice signal can not be removed, object of the present invention solves regular speech signal exactly can not remove acoustic background, substantially increase the signal to noise ratio (S/N ratio) of input voice, substantially increase the intelligibility of voice signal.

Accompanying drawing explanation

Fig. 1 is the position relationship of two microphones of communication apparatus;

Fig. 2 is the digital process of voice digital signal;

Fig. 3 is critical path method (CPM) schematic diagram.

Embodiment

Below in conjunction with accompanying drawing, the present invention is described in further details.

For the noise cannot removing voice in current communication apparatus, on system hardware structure, first do change a little, after then utilizing system self processes voice signals, we take following method to realize the noise remove of acoustic background in computing method.

Based on two microphone mobile device voice signals enhancement methods of blind source separating, comprise following components:

(1) microphone of mobile device has two, and two microphones all must be arranged on bottom, front, and distance between the two must be more than or equal to 4.25cm;

(2) when digitization of speech signals, the length of every section of voice must be not more than 30ms, and the voice signal after digitizing must go average and decorrelation to voice signal, to reduce next step iterations;

(3) signal after pre-service carry out based on shortest path owe determine blind separating method and calculate the voice number of aliasing and each road voice signal, the maximum voice signal of energy is exactly required voice signal.

The method of owing the potential function of determining in blind separation is adopted to estimate Mixture matrix.Potential function is l _tand θ _trepresent respectively in plane a bit to initial point distance and the angle of this straight line with x positive axis, then new potential function φ ' (θ is smoothly obtained to potential function, λ), try to achieve potential function and just can try to achieve its peak value, each peak value is exactly a column vector of hybrid matrix, so there are the computing method of hybrid matrix column vector:

a _i＝[sin(π/2+θ _i),cos(π/2+θ _i)] ^T

θ _irepresent the angle at i-th peak value place, so there is the estimation of hybrid matrix A:

At Mixture matrix A, under having observation signal x and supposing the condition that source signal is sparse, use l1-norm as sparse tolerance, source signal estimation procedure is exactly for all sample points x (t), under the condition of As (t)=x (t) after trying to achieve each information source, try to achieve the signal that energy is maximum, be transmission of speech signals.

Below in conjunction with accompanying drawing, introduce step of the present invention in detail:

The first step: install another one microphone in communication apparatus bottom, two microphones will be in the same horizontal line, and distance d is therebetween greater than 4.25cm, two microphone simulating signals in mobile phone are gathered by digital signal processor wherein simultaneously, and its schematic diagram is as Fig. 1.

Second step: the processing procedure of digital signal:

The process of digital signal processing as shown in Figure 2, observation signal matrix x=[x ₁; x ₂], wherein x ₁the signal in a microphone, x ₂be the signal in another microphone, in each signals collecting microphone, the length of every segment signal is within 10 to 30ms.

Step1: go average

Go average to make each road observation signal be all zero-mean, namely deduct its mean value vector E (x) in x, in the calculating of reality, the mathematical expectation of each road signal x adopts arithmetic mean to replace, the i-th road signal go average as shown in the formula:

$v_i\left(t\right)=x_t\left(t\right)-\frac{1}{N}\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{x}_i\left(t\right)---\left(1\right)$

N represents the number of sampled point, and i represents the i-th road signal, and k represents a kth sampled point.

Step2: decorrelation

Eigenvalues Decomposition is passed through in decorrelation _vcovariance matrix R _v=E (vv ^t)=QDQ ^t, wherein D is R _vthe diagonal matrix of eigenwert composition, Q is the matrix of the proper vector composition of character pair value, so just can obtain whitening matrix T=D ^-1/2q ^t, obtain whitened signal z by conversion z=Tv.

Step3: adopt critical path method (CPM) owe determine blind separating method and carry out separating mixture of source signals, detailed process:

A) estimation of Mixture matrix

Observation signal is two-dimentional, is transformed into the scatter diagram of two dimensional surface, and the column vector direction of Mixture matrix can represent by polar form that θ represents, t data point z (t) place coordinate points is so have

$\left\{\begin{array}{c}{l}_t=\sqrt{{\left(z_1^t\right)}^2+{\left(z_2^t\right)}^2}\\ {\mathrm{\θ}}_t={\tan }^{-1}(z_2^t/z_1^t)\end{array}\right.---\left(2\right)$

L _tand θ _trepresent radius and angle respectively.A threshold epsilon is set, if be in l in the process of signal transacting _tthen remove these points in the scope of < ε, getting α is any direction and θ _tbetween angular deviation, select one around z ^tbasis function φ, φ is the function about local angle α, has

And define an overall potential function φ about absolute angle θ:

$\mathrm{\&phi;}(\mathrm{\&theta;},\mathrm{\&lambda;})=\underset{t}{\mathrm{\&Sigma;}}{l}_t\mathrm{\&phi;}\left(\mathrm{\&lambda;}(\mathrm{\&theta;}-{\mathrm{\&theta;}}_t)\right)---\left(4\right)$

Parameter lambda is used to the width or the local resolution that regulate expected angle, and rule of thumb parameter lambda is chosen as between 0.1 to 150, l _tbe the weight of each sample point, be generally taken as the mould of sample point, the size of potential function φ (θ, λ) is exactly the size of sampled point in θ place probability density, φ (θ, λ) multiple spot is smoothly removed ghost peak, a _j-n, a _j-n+1... a _j... a _j+n-1, a _j+nfor the point on potential function φ (θ, λ), smoothing method is

b(j)＝(a _j-n+a _j-n+1+...+a _j...+a _j+n-1+a _j+n)/2n (5)

After level and smooth, new potential function is φ ' (θ, λ), and try to achieve potential function and just can try to achieve its peak value, each peak value is exactly a column vector of hybrid matrix, so there are the computing method of hybrid matrix column vector:

a _i＝[sin(π/2+θ _i),cos(π/2+θ _i)] ^T(6)

Wherein, i=1,2 ..., npeak, npeak represent total number of peak value, θ _irepresent the angle at i-th peak value place, so there is the estimation of hybrid matrix A:

$\tilde{A}={[a}_1,a_2,...,a_{\mathrm{npeak}}]---\left(7\right)$

B) estimation of each sound source

Providing under hybrid matrix A non-square matrix, observation signal x and the sparse condition of supposition source signal, asking s to become to solve the problem of linear programming, use l1-norm as sparse tolerance, then source signal estimation procedure is exactly for all sample points x (t), under the condition of As (t)=x (t), minimizes l1 norm | s (t) | ₁.

$\min \underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|s_j\left(t\right)\right|,\mathrm{As}\left(t\right)=x\left(t\right)---\left(8\right)$

May the shortest path of from initial point to x (t) be looked for solve exactly in solution at formula (8), as shown in Figure 3, postulated point x (t) is an observation signal in frequency domain, so from initial point o to x (t) shortest path be vector two nearest with counterclockwise distance vector x (t) clockwise vectorial a and b.

Suppose A _r=[a, b] ^t, it is the submatrix of 2 × 2 of any two row formations of the Mixture matrix A estimated, s _rthe t component along a and b both direction that () is x (t).

$\left\{\begin{array}{c}{s}_r=A_r^{-1}x\left(t\right)\\ s_j\left(t\right)=0,j\&NotEqual;a,b\end{array}\right.---\left(9\right)$

Set a r, if l _t< r, just the value of these points being composed is zero, and in order to obtain higher separation accuracy, it is smaller that r sets, and makes r=0.1*max (l _t).To vectorial a _ipoint around carries out pre-service, can determine in x near a _iregion be:

$\frac{1}{6}({\mathrm{\&theta;}}_{a2}-{\mathrm{\&theta;}}_{a1})\&le;\mathrm{\&theta;}\&le;\frac{1}{6}({\mathrm{\&theta;}}_{a3}-{\mathrm{\&theta;}}_{a2})---\left(10\right)$

Principle according to time-frequency masking has

$\left\{\begin{array}{c}{s}_a\left(t\right)=x_i\left(t\right)/a_{\mathrm{ii}}\\ s_j\left(t\right)=0,j\&NotEqual;a\end{array}\right.---\left(11\right)$

I=1,2 ..., the number of npeak, npeak peak value, has so just tried to achieve each original component.

Step4: filter out target person acoustical signal from isolated signal

In step3, obtain the estimation s of information source, at mobile phone or flat board when carrying out speech communication, the voice signal gross energy of the speaker will transmitted in each section of voice is only maximum, is calculated as follows signal energy

$\mathrm{LE}=\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{\left(s_i\right)}^2---\left(12\right)$

Wherein n represents the length of whole signal, and j represents the jth point estimating information source i.The information source that in all output s, LE is maximum is exactly targeted voice signal, just can voice signal directly after transmission process.

Claims (6)

1., based on two microphone mobile device voice signals enhancement methods of blind source separating, it is characterized in that: comprise following step,

Step one: two microphones being arranged on d apart on the same level line of communication apparatus bottom, for received speech signal, obtain observation signal x=[x ₁; x ₂], x ₁the voice signal in a microphone, x ₂it is the voice signal in another microphone;

Step 2: average value processing is gone to observation signal;

Step 3: to removing the signal v after average value processing _it () carries out decorrelative transformation;

Step 4: owe to determine blind separating method to signal z (t) employing after decorrelative transformation and carry out separating mixture of source signals, estimated value and each sound source of obtaining Mixture matrix are estimated;

Step 5: estimate that the sound source selecting energy maximum is as target person acoustical signal from each sound source.

2. the two microphone mobile device voice signals enhancement methods based on blind source separating according to claim 1, is characterized in that: described go average value processing to be to observation signal:

$v_i\left(t\right)=x_i\left(t\right)-\frac{1}{N}\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{x}_i\left(t\right);k$

Wherein, v _it (), for removing the signal after average value processing, N is the number of sampled point, i is the i-th road signal, is a kth sampled point.

3. the two microphone mobile device voice signals enhancement methods based on blind source separating according to claim 1, is characterized in that: the described method obtaining Mixture matrix estimated value is,

Step one: signal z (t) after decorrelative transformation is expressed as polar form is:

$\left\{\begin{array}{c}{l}_t=\sqrt{{\left(z_1^t\right)}^2+{\left(z_2^t\right)}^2}\\ {\mathrm{\&theta;}}_t={\tan }^{-1}(z_2^t/z_1^t)\end{array}\right.$

Wherein, signal z (t) after decorrelative transformation is 2D signal, and t is l _tand θ _trepresent radius and angle respectively;

Step 2: definition basis function φ:

Wherein, α is any direction and θ _tbetween angular deviation,

Obtain the overall potential function φ (θ, λ) about angle θ:

$\mathrm{\&phi;}(\mathrm{\&theta;},\mathrm{\&lambda;})=\underset{t}{\mathrm{\&Sigma;}}{l}_t\mathrm{\&phi;}\left(\mathrm{\&lambda;}(\mathrm{\&theta;}-{\mathrm{\&theta;}}_t)\right)$

Wherein, λ is parameter;

Step 3: to the smoothing process of overall potential function, obtains new potential function φ ' (θ, λ), and each peak value of this potential function is a column vector a of hybrid matrix _i,

a _i＝[sin(π/2+θ _i),cos(π/2+θ _i)] ^T

Wherein, i=1,2 ..., npeak, npeak are total number of peak value;

Obtain the estimated value of hybrid matrix

$\tilde{A}=[a_1,a_2,...,a_{\mathrm{npeak}}].$

4. the two microphone mobile device voice signals enhancement methods based on blind source separating according to claim 1, is characterized in that: the described method obtaining the estimation of each sound source is:

Using all sample points as input, under As (t)=x (t) condition, ask and minimize l1 norm | s (t) | ₁,

$\min \underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|s_j\left(t\right)\right|,\mathrm{As}\left(t\right)=x\left(t\right)$

Obtain each sound source to estimate.

5. the two microphone mobile device voice signals enhancement methods based on blind source separating according to claim 1, is characterized in that: the energy that described each sound source is estimated is:

$\mathrm{LE}=\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{\left(s_i\right)}^2,$

Wherein, n represents the length of whole signal, and j represents that sound source estimates s _ia jth point.

6. the two microphone mobile device voice signals enhancement methods based on blind source separating according to claim 1, it is characterized in that: described distance d is greater than or equal to 4.25cm, the length of each microphone received speech signal is less than or equal to 30ms.