CN100466061C - Broadband wave beam forming method and apparatus - Google Patents

Abstract

The disclosed forming method for broadband waveform comprises: determining the sub-band signal opposite to the microphone signal, as well as the signal frequency-domain correlation matrix; according to 3D space transmission vector of signal source and former matrix, determining the weight vector for every sub-band signal; then deciding the output signal. This invention combines frequency and space domain for speech process, and improves SNR for wide application.

Description

A kind of broadband beams forms method and apparatus

Technical field

The present invention relates to audio signal processing technique, be specifically related to a kind of broadband beams and form method and apparatus.

Background technology

Flourish along with modern science, communication or message exchange have become the necessary condition that human society exists, and voice are as the acoustics performance of language, are that human exchange of information is the most natural, the most effective, one of the means of most convenient.

In voice communication course, can be subjected to interference inevitably from noise, communication facilities internal electrical noise and even other talkers of surrounding environment, transmission medium introducing.These disturb the voice that the voice recipient is received no longer is pure raw tone, but the noisy speech of being crossed by noise pollution.For example, the phone in automobile, street, the airport, regular meeting be subjected to strong background noise in disturbing, thereby had a strong impact on speech quality.The pollution of neighbourhood noise also can make the performance rapid deterioration of many speech processing systems.For example, speech recognition system makes substantial progress, just steps into the practical stage, but, present speech recognition system is mostly worked in quiet environment, and especially in strong noise environment, the discrimination of speech recognition system will be had a strong impact in noise circumstance.Low rate voice coding, particularly parameter coding also run into similar problem.Because speech production model is the basis of low rate coding, when the extraction of model parameter is subjected to being mingled in when ground unrest seriously disturbs in the voice, the quality of reconstructed speech is rapid deterioration, even becomes and can not understand fully.

Speech enhancement technique can effectively suppress ground unrest, improves voice communication quality, improves the antijamming capability of speech processing system, keeps the performance of speech processing system.Therefore, the research speech enhancement technique has significant values in actual applications.Oneself has obtained application more and more widely in fields such as speech processing system, communication system, multimedia technology, digitizing household electrical appliances speech enhancement technique.

The fundamental purpose of speech enhancement technique is: extract pure as far as possible raw tone from Noisy Speech Signal.Yet, all be at random owing to disturb usually, extracting fully from noisy speech, pure voice almost are impossible.In the case, the fundamental purpose that voice strengthen is: by noisy speech is handled, to eliminate ground unrest, improve voice quality, improve sharpness, intelligibility and the comfort level of voice, improve the performance of speech processing system.These purposes often can not get both, and need decide according to the concrete needs of speech processing system usually.

The research of speech enhancement technique starts from 20th century the mid-1970s.Along with the maturation of digital signal processing theory, voice strengthen an important branch that has developed into field of voice signal.1978, Lim and Oppenheim proposed the Wiener filtering method that voice strengthen.1979, Boll proposed the spectrum subtraction method and has suppressed noise.1980, Maulay and Malpass proposed the soft-decision noise suppressing method.1984, Ephraim and Malah proposed the sound enhancement method based on MMSE short-time spectrum amplitude Estimation.1987, Paliwal was incorporated into voice to Kalman filtering and strengthens the field.In nearly 30 years research, various sound enhancement methods constantly are suggested, and it has been established the basis of voice enhancing theory and has made it to move to maturity gradually.

In the last few years, along with the development of VLSI (VLSI (very large scale integrated circuit)) technology and the appearance of high-speed dsp (digital signal processing) chip, it is practical that speech enhancement technique is progressively moved towards, and emerges in large numbers in succession again with stylish speech enhancement technique.

Voice strengthen, the method for denoising can simply be divided into the filtering technique based on time domain, frequency domain and spatial domain, as the speech enhancement technique of Wiener filtering, based on the spectrum cancellation technology of frequency domain etc.In recent years, the ARRAY PROCESSING technology also is incorporated in the speech processes, has formed the airspace filter technology based on wave beam, as time delay summation beam-forming technology (DSB) etc.

MVDR (Minim Variance Distortion Response, the arrowband minimum variance is undistorted) beam-forming technology is mainly used in traditional narrow band signal processing procedure.

Be set with M sensor composition aerial array as shown in Figure 1, receive from direction

Narrow band signal s (t), the key step of carrying out airspace filter with the MVDR beam-forming technology to received signal is as follows:

Step 1, the analog signal conversion that each sensor is received are digital signal, and digital signal is formed input data matrix X (n):

X(n)＝[x ₁(n)x ₂(n)…x _M(n)] ^T (1)

Wherein, [] ^TExpression is made transposition computing, x to matrix or vector _i(n) expression n constantly i sensor receive pass through digital signal and i=1 after the AD conversion ..., M.

To step 2, the L point data of getting L snap, promptly get n constantly, n-1 constantly ..., data that n-L+1 gathers on each sensor constantly, and ask the frequency domain correlation matrix R of input signal according to formula (2):

$R=\frac{1}{L}\underset{l=1}{\overset{L}{\mathrm{\Σ}}}\{X(n-l+1)*X^H(n-l+1)\}---\left(2\right)$

Wherein, [] ^HExpression is made transposition and each element is got common volume computing matrix or vector.For example:

$A=\left[\begin{array}{cc}1+2i& 2+4i\\ 4-4i& 5-8i\end{array}\right],$ Then $A^H=\left[\begin{array}{cc}1-2i& 4+4i\\ 2-4i& 5+8i\end{array}\right].$

To step 3, according to the direction of signal source Obtain the direction vector a of signal source with array topology.When obtaining the direction vector a of signal source, array topology is not limit, as can being uniform circular array, uniform straight line array or other array structure, and the direction of signal source

Preparation method do not limit.

Setting M sensor is reference point with spacing d composition uniform straight line array row and with first sensor, and then the direction vector a of signal source is:

a＝[1?e ^-jφ…e ^-j(M-1)φ] ^T (3)

In formula (3), φ be space phase and

Wherein, λ is the wavelength of incoming signal, and d is an array pitch,

Incident angle for incoming signal.To step 4, ask optimal weight vector W according to the direction vector a and the frequency domain correlation matrix R of signal source _Opt:

$W_{\mathrm{opt}}=\frac{R^{-1}a}{a^H{R}^{-1}a}---\left(5\right)$

To step 5, input signal is carried out spatial filtering, obtain output signal y (n) according to optimal weight vector:

$y\left(n\right)=W_{\mathrm{opt}}^H*X\left(n\right)---\left(6\right)$

Then, converting digital signal y (n) to simulating signal gets final product.

Above-mentioned MVDR beam-forming technology can only be applicable to the narrow band signal source, when this method is used for wideband signal source, its voice are strengthened the property and can be descended significantly, and, this technology can only be applicable to the far-field signal source, and promptly incoming signal is a plane wave, when this technology is applicable to the near-field signals source, be incoming signal when being spherical wave, voice are strengthened the property and can be descended significantly equally.

Summary of the invention

The objective of the invention is to, provide a kind of broadband beams to form method and apparatus, by voice signal being handled, to realize improving the purpose that voice are strengthened the property in conjunction with frequency domain and spatial domain.

For achieving the above object, a kind of broadband beams formation method provided by the invention comprises:

Each subband signal that the signal of a, definite each microphone of input is corresponding respectively;

B, determine the frequency domain correlation matrix of described each subband signal;

C, determine the weight vector of each subband signal according to the three dimensions used for vector transfer of signal source, described each frequency domain correlation matrix;

D, determine the signal of output according to the weight vector of described each subband signal and each subband signal.

Described step a specifically comprises:

A1, the signal of importing each microphone is carried out speech detection, and definite speech frame;

A2, determine each subband signal of described speech frame correspondence.

The signal of setting each microphone of input is: F (t)=[f _i(t) ... f _i(t) ... f _M(t)] ^T

Wherein: f _i(t) i signal that microphone receives of expression, i=1 ..., M, M are the quantity of microphone, [] ^TExpression is done the transposition computing to matrix or vector; And described step a1 specifically comprises the steps:

A11, the signal of importing each microphone is carried out AD conversion according to predetermined sampling frequency:

F (n)=[f ₁(n) ... f _i(n) ... f _M(n)] ^TWherein, n is a discrete time;

Choose signal frame a12, the signal after described AD changes and carry out short time discrete Fourier transform:

$F\left(\mathrm{\ω}\right)=\underset{m=1}{\overset{N}{\mathrm{\Σ}}}F\left(n\right)w(n-m)\exp (-\mathrm{j\ωm})=\left[\begin{array}{c}\underset{m=1}{\overset{N}{\mathrm{\Σ}}}{f}_1\left(n\right)w(n-m)\exp (-\mathrm{j\ωm})\\ \·\·\·\\ \underset{m=1}{\overset{N}{\mathrm{\Σ}}}{f}_M\left(n\right)w(n-m)\exp (-\mathrm{j\ωm})\end{array}\right];$ Wherein, w (n) is a window function, and n, m are discrete time;

A13, the signal frame behind the Fourier transform is carried out speech detection, determine speech frame.

Described step a13 specifically comprises the steps:

Signal frame behind the Fourier transform is carried out speech detection;

When definite signal frame is not speech frame, this signal frame is stored as current estimating noise spectrum;

When definite signal frame is speech frame, according to current estimating noise spectrum described speech frame is composed the counteracting denoising, determine that the speech frame S (ω) after the spectrum counteracting denoising is:

$S\left(\mathrm{\ω}\right)=F\left(\mathrm{\ω}\right)-N\left(\mathrm{\ω}\right)={\left[\begin{array}{ccc}{s}_1\left(1\right)& \·\·\·& s_1\left(\mathrm{NFFT}\right)\\ \·\·\·& \·\·\·& \·\·\·\\ s_M\left(1\right)& \·\·\·& s_M\left(\mathrm{NFFT}\right)\end{array}\right]}_{M\×\mathrm{NFFT}};$

Wherein: $N\left(\mathrm{\ω}\right)={\left[\begin{array}{ccc}{n}_1\left(1\right)& \·\·\·& n_1\left(\mathrm{NFFT}\right)\\ \·\·\·& \·\·\·& \·\·\·\\ n_M\left(1\right)& \·\·\·& n_M\left(\mathrm{NFFT}\right)\end{array}\right]}_{M\×\mathrm{NFFT}}$ Be current estimating noise spectrum, NFFT is that the frequency sampling of short time discrete Fourier transform is counted, and F (ω) is the signal frame behind the short time discrete Fourier transform, and M is the quantity of microphone.

Described step a2 specifically comprises:

According to K predetermined frequency band speech frame is divided into K subband signal, and with K preset frequency ω _i, i=1 ... K is defined as the centre frequency of each subband;

Determine the component of signal S (ω of i subband _i) be: $S\left({\mathrm{\ω}}_i\right)=\left[\begin{array}{c}{S}_1\left(i\right)\\ \·\·\·\\ S_M\left(i\right)\end{array}\right];$

Wherein: M is the number of microphone of microphone array, i=1 ... K, K are number of sub-bands.

Described step b specifically comprises:

Determine the frequency domain correlation matrix R (i) of each subband signal:

R (i)=E{S (ω _i) S ^H(ω _i); Wherein: ^HExpression is carried out transposition and conjugate operation to matrix or vector, S (ω _i) be the component of signal of i subband, and $S\left({\mathrm{\ω}}_i\right)=\left[\begin{array}{c}{S}_1\left(i\right)\\ \·\·\·\\ S_M\left(i\right)\end{array}\right].$

The three dimensions used for vector transfer of signal source can obtain as follows among the described step c:

C1, obtain source location (r ₀, θ ₀, φ ₀) coordinate vector S=r ₀* [sin θ ₀Cos φ ₀Sin θ ₀Sin φ ₀Cos θ ₀];

C2, obtain the coordinate vector P of each microphone _i=r _i* [sin θ _iCos φ _iSin θ _iSin φ _iCos θ _i];

C3, determine source location (r ₀, θ ₀, φ ₀) to the relative amplitude decay factor of i microphone

For: ${\∂}_i=\frac{\left|\right|S\left|\right|}{\left|\right|P_i-S\left|\right|};$ Wherein: ‖ * ‖ represents the norm of vector *;

C4, determine source location (r ₀, θ ₀, φ ₀) to the relative time delay factor τ of i microphone _iFor:

${\mathrm{\τ}}_i=\frac{\left|\right|S-P_i\left|\right|-\left|\right|S\left|\right|}{c};$

Wherein: c is the aerial velocity of propagation of sound, and ‖ * ‖ represents the norm of vector *;

C5, determine source location (r ₀, θ ₀, φ ₀) three dimensions used for vector transfer a (r ₀, θ ₀, φ ₀) be:

$a({\mathrm{<figure-callout\; id="0"\; label="r"\; filenames="C200510090740E00211.png"\; state="\{\{state\}\}">r</figure-callout>}}_0,{\mathrm{\&theta;}}_0,{\mathrm{\&phi;}}_0)=\left[\begin{array}{ccc}{\&PartialD;}_1{e}^{-j{\mathrm{\&omega;}}_i{\mathrm{\&tau;}}_1}& \&CenterDot;\&CenterDot;\&CenterDot;{\&PartialD;}_m{e}^{-j{\mathrm{\&omega;}}_i{\mathrm{\&tau;}}_m}\&CenterDot;\&CenterDot;\&CenterDot;& {\&PartialD;}_M{e}^{-j{\mathrm{\&omega;}}_i{\mathrm{\&tau;}}_M}\end{array}\right];$

Wherein: ω _iBe the centre frequency of each subband, r ₀Be the distance of signal source to true origin, θ ₀Be the angle of the Z axle of signal source and three-dimensional coordinate, φ ₀Be the projection of signal source on XOY plane and the angle of X-axis.

Described step c specifically comprises: the optimal weight vector of determining i subband

For:

$W_{\mathrm{opt}}^i=\frac{R{\left(i\right)}^{-1}a}{a^{H}R{\left(i\right)}^{-1}a};$

Wherein: R (i) is the frequency domain correlation matrix of i subband signal, and a is described source location (r ₀, θ ₀, φ ₀) the three dimensions used for vector transfer.

Described steps d comprises: the optimal weight vector according to each subband signal carries out the subband spatial filtering to each subband signal, obtains the frequency domain output signal y (ω of i subband _i):

$y\left({\mathrm{\&omega;}}_i\right)={\left(W_{\mathrm{opt}}^i\right)}^H*S\left({\mathrm{\&omega;}}_i\right);$ Wherein: ^HExpression is carried out transposition and conjugate operation to vector or matrix,

Be the optimal weight vector of i subband, S (ω _i) be the component of signal of i subband;

The frequency domain output signal of each subband is combined as Y (ω): Y (ω)=[y (ω ₁) y (ω ₂) ... y (ω _K)] ^T

Frequency domain output signal Y (ω) after the combination is carried out contrary fast fourier transform obtain output signal Y (n);

Convert described Y (n) to simulating signal y (t), and y (t) is carried out the voice signal that signal after the low-pass filtering needing to be defined as output.

The present invention also provides a kind of broadband beams to form device, comprising:

Divide the subband signal module: determine each corresponding respectively subband signal of signal of each microphone of input, and each subband signal is transferred to frequency domain correlation matrix module;

Frequency domain correlation matrix module: determine the frequency domain correlation matrix of described each subband signal, and transmit it to the weight vector module;

Weight vector module: determine the weight vector of each subband signal, and transmit it to output module according to the three dimensions used for vector transfer of signal source, described each frequency domain correlation matrix;

Output module: according to the weight vector of described each subband signal, the signal that each subband signal is determined output.

Described division subband signal module comprises:

Sampling submodule: according to predetermined sampling frequency the signal of importing each microphone is carried out the AD conversion, and from the signal after the described AD conversion, choose signal frame and carry out short time discrete Fourier transform;

Speech detection submodule: the signal frame behind the Fourier transform is carried out speech detection, when definite signal frame is not speech frame, this signal frame is stored as current estimating noise spectrum, when definite signal frame is speech frame, this speech frame is transferred to spectrum offset the denoising submodule;

Spectrum is offset the denoising submodule: according to current estimating noise spectrum the speech frame of its reception is composed the counteracting denoising, and transfer to division subband signal submodule;

Divide the subband signal submodule: according to predetermined frequency band the speech frame of its reception is divided into a plurality of subband signals, and each subband signal is transferred to frequency domain correlation matrix module.

Description by technique scheme as can be known, the present invention is by adopting three dimensions propagation vector a (r to the source location vector ₀, θ ₀, φ ₀), solved the three dimensions filtering problem, suppressed spatial interference signal and noise, improved the signal to noise ratio (S/N ratio) of output signal; By voice signal being divided into a plurality of subbands, each subband is carried out three dimensions filtering respectively, the present invention can be combined frequency domain and spatial domain voice signal is handled, make the present invention can be good at being applicable to wideband signal source, near-field signals source; By adopting speech detection technology such as zero-crossing rate, short-time energy combine to determine speech frame, to have avoided when not having the voice signal input, the phenomenon of consume system resources has improved accuracy and the stability of exporting voice signal; By adopting the spectrum cancellation technology to remove system noise, avoided of the influence of non-white Gauss noise to system, effectively improved the filtering performance of voice signal; At definite source location vector a (r ₀, θ ₀, φ ₀) in the process, by adopting relative amplitude decay factor and the relative time delay factor of source location to each microphone, and the amplitude fading factor adopts ratio, the time delay factor of signal source to the distance of each microphone and signal source to the distance of reference microphone to adopt the poor of the time delay of signal source to the time delay of each microphone and signal source to reference microphone, make the present invention consistent with the model hypothesis of subspace theory, reduce model error, improved the three dimensions filtering performance; By characteristics according to narrow band signal assumed condition and voice signal, the voice signal of input microphone is divided into several subbands, determines the frequency domain correlation matrix of each subband to have significantly reduced operand according to the component of signal of each subband, improve the real-time of system, saved hardware cost; Thereby realized that by technical scheme provided by the invention the raising voice strengthen the property, improved the purpose of voice system practicality.

Description of drawings

Fig. 1 is even straight line microphone array synoptic diagram;

Fig. 2 is that broadband beams of the present invention forms method flow diagram;

Fig. 3 is a near-field signals model synoptic diagram.

Embodiment

The core of method and apparatus of the present invention is: each subband signal of determining the signal correspondence of each microphone of input, determine the frequency domain correlation matrix of each subband signal, according to the three dimensions used for vector transfer of signal source, the weight vector that each frequency domain correlation matrix is determined each subband signal, according to the weight vector of each subband signal, the signal that each subband signal is determined output.

Based on core concept of the present invention technical scheme provided by the invention is further described below.

Microphone type among the present invention is an omnidirectional microphone, the pickup distance of microphone can be determined according to concrete applied environment, be long 5 meters, wide 10 meters, high 4 meters as room-size, if require all sound in this room are handled, then the pickup of microphone distance is at least 10 meters.The present invention does not limit the timbering material that constitutes microphone, and still, the physical dimension of timbering material is the smaller the better, to reduce the reflection of support to sound, reduces multipath effect.

The topological structure of the microphone array among the present invention can be arbitrary form, as ULA (uniform straight line array row), UCA (evenly circle ring array) etc.

The present invention is based on microphone array wideband signal source wave beam formation method process flow diagram as shown in Figure 2.

In Fig. 2, method of the present invention mainly comprises three parts, i.e. signals collecting preprocessing part, The Wideband Signal Processing part and output signal processing section.

Detailed implementation procedure below in conjunction with 2 pairs in the accompanying drawing broadband beams formation method based on microphone array of the present invention is described.

The signals collecting preprocessing part mainly comprises following 5 steps:

Step 1, set M common omnidirectional microphone and form microphone array according to certain topological structure, the voice signal that send in microphone array pickoff signals source, and other be in all voice signals in the microphone range of receiving.

The signal that microphone array picks up can be expressed as with mathematical formulae:

F(t)＝[f ₁(t)…f _i(t)…f _M(t)] ^T (7)

In the formula (7): f _i(t) i voice signal that microphone receives of expression, i=1 ..., M, M are the quantity of microphone in the microphone array, [] ^TThe transposition computing of representing matrix.

The coordinate vector of setting i microphone is:

In the formula (8): r _iBe the distance of i microphone apart from the microphone array center, the microphone array center is a true origin, and true origin can be the center of microphone array, also can be the position of any one microphone in the microphone array, or other positions; θ _iBe the coordinate vector of i microphone and the angle of Z axle positive dirction,

Be that i microphone coordinate vector is at the projection of XOY plane and the angle of X-axis positive dirction.

The coordinates matrix that the coordinate vector of M microphone is formed whole microphone array is:

Step 2, the signal that each microphone is received carry out the AD conversion.

When carrying out the AD conversion, can be according to sample frequency and sampling precision be chosen in the requirement of sound quality, can be 16KHz, 22KHz or 44Khz etc. as sample frequency, sampling precision can be 8bit, 16bit or 32bit etc.The present invention does not limit the technology and the chip of sampling.

The sampling back forms the multi-path digital voice signal, that is:

F(n)＝[f ₁(n)…f _i(n)…f _M(n)] ^T (10)

In the formula (10): i=1 ..., M, M are the quantity of microphone in the microphone array.

Step 3, choose the signal frame that frame length is 32ms from every road sampled signal of formula (10), carry out short time discrete Fourier transform, short time discrete Fourier transform can be selected Hamming window or other window function for use.

Can realize short time discrete Fourier transform with 512 the FFT (fast fourier transform) of NFFT generally speaking.That is:

$F\left(\mathrm{\&omega;}\right)=\underset{m=1}{\overset{N}{\mathrm{\&Sigma;}}}F\left(n\right)w(n-m)\exp (-\mathrm{j\&omega;m})=\left[\begin{array}{c}\underset{m=1}{\overset{N}{\mathrm{\&Sigma;}}}{f}_1\left(n\right)w(n-m)\exp (-\mathrm{j\&omega;m})\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ \underset{m=1}{\overset{N}{\mathrm{\&Sigma;}}}{f}_M\left(n\right)w(n-m)\exp (-\mathrm{j\&omega;m})\end{array}\right]---\left(11\right)$

Step 4, to carrying out speech detection through the signal frame behind the short time discrete Fourier transform in the formula (11), the speech detection technology can combine etc. for zero-crossing rate, short-time energy, the present invention does not limit the speech detection technology that adopts.

According to the speech detection technology, when definite signal frame is non-speech frame, non-speech frame is stored as current estimating noise spectrum, the initial value of current estimating noise spectrum can be set to 0 matrix; When definite signal frame is speech frame, this speech frame is carried out the processing of following step 5.

Step 5, adopt the spectrum counteracting method promptly to subtract spectrometry to speech frame to compose the counteracting denoising.

Setting current estimating noise spectrum is: $N\left(\mathrm{\&omega;}\right)={\left[\begin{array}{ccc}{n}_1\left(1\right)& \&CenterDot;\&CenterDot;\&CenterDot;& n_1\left(\mathrm{NFFT}\right)\\ \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;\\ n_M\left(1\right)& \&CenterDot;\&CenterDot;\&CenterDot;& n_M\left(\mathrm{NFFT}\right)\end{array}\right]}_{M\&times;\mathrm{NFFT}}---\left(12\right)$

Speech frame is composed the speech frame of offsetting after the denoising is:

$S\left(\mathrm{\&omega;}\right)=F\left(\mathrm{\&omega;}\right)-N\left(\mathrm{\&omega;}\right)={\left[\begin{array}{ccc}{s}_1\left(1\right)& \&CenterDot;\&CenterDot;\&CenterDot;& s_1\left(\mathrm{NFFT}\right)\\ \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;\\ s_M\left(1\right)& \&CenterDot;\&CenterDot;\&CenterDot;& s_M\left(\mathrm{NFFT}\right)\end{array}\right]}_{M\&times;\mathrm{NFFT}}---\left(13\right)$

To step 5, finished the collection preprocessing process of the present invention through above-mentioned steps 1 to signal.Following step 6 is for having realized The Wideband Signal Processing process of the present invention.

Step 6, according to the frequency characteristics of signal the S (ω) in the formula (13) is divided into each and does subband, then, from each subband, choose an interested frequencies omega respectively _i, wherein: i=1 ... K, K are the quantity of subband.With frequencies omega _iCentre frequency as its corresponding subband.

If the component of signal S (ω of i subband signal _i) be:

$S\left({\mathrm{\&omega;}}_i\right)=\left[\begin{array}{c}{S}_1\left(i\right)\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ S_M\left(i\right)\end{array}\right]---\left(14\right)$

In the formula (14): M is the quantity of microphone in the microphone array.

Each subband of formula (14) is done the processing of following four aspects:

1, obtains the frequency domain correlation matrix R (i) of speech frame: R (i)=E{S (ω _i) S ^H(ω _i) (15)

2, picked up signal source location vectors a (r ₀, θ ₀, φ ₀):

The coordinate vector of setting i microphone is P _i, source location (r ₀, θ ₀, φ ₀) coordinate vector be S, as shown in Figure 3, r ₀Be the distance of signal source to true origin, θ ₀Be the angle of the Z axle of signal source and three-dimensional coordinate, φ ₀Be the projection of signal source on XOY plane and the angle of X-axis.That is:

S＝r ₀*[sinθ ₀cosφ ₀ sinθ ₀sinφ ₀ cosθ ₀] (16)

P _i＝r _i*[sinθ _icosφ _i?sinθ _isinφ _i?cosθ _i] (17)

From source location (r ₀, θ ₀, φ ₀) to the relative amplitude decay factor of i microphone

For:

${\&PartialD;}_i=\frac{\left|\right|S\left|\right|}{\left|\right|P_i-S\left|\right|}---\left(18\right)$

From source location (r ₀, θ ₀, φ ₀) to the relative time delay factor τ of i microphone _iFor:

${\mathrm{\&tau;}}_i=\frac{\left|\right|S-P_i\left|\right|-\left|\right|S\left|\right|}{c}---\left(19\right)$

C is the aerial velocity of propagation of sound in the formula (19), can get 340 meter per seconds when room temperature, and ‖ ‖ represents to ask the norm of vector, as vector a=[x y z], then $\left|\right|a\left|\right|=\sqrt{x^2+y^2+z^2}.$

Source location (r ₀, θ ₀, φ ₀) position vector a (r ₀, θ ₀, φ ₀) be:

$a({\mathrm{<figure-callout\; id="0"\; label="r"\; filenames="C200510090740E00211.png"\; state="\{\{state\}\}">r</figure-callout>}}_0,{\mathrm{\&theta;}}_0,{\mathrm{\&phi;}}_0)=\left[\begin{array}{ccc}{\&PartialD;}_1{e}^{-j{\mathrm{\&omega;}}_i{\mathrm{\&tau;}}_1}& \&CenterDot;\&CenterDot;\&CenterDot;{\&PartialD;}_m{e}^{-j{\mathrm{\&omega;}}_i{\mathrm{\&tau;}}_m}\&CenterDot;\&CenterDot;\&CenterDot;& {\&PartialD;}_M{e}^{-j{\mathrm{\&omega;}}_i{\mathrm{\&tau;}}_M}\end{array}\right]---\left(20\right)$

3, obtain the optimal weight vector of i subband

Position vector a (the r of the signal source that obtains according to formula (20) ₀, θ ₀, φ ₀) and the frequency domain frequency domain correlation matrix R (i) that obtains of formula (15), obtain the optimal weight vector of i subband

The optimal weight vector of i subband

For:

$W_{\mathrm{opt}}^i=\frac{R{\left(i\right)}^{-1}a}{a^{H}R{\left(i\right)}^{-1}a}---\left(21\right)$

4, utilize optimal weight vector and subband signal to carry out the subband spatial filtering, obtain the frequency domain output signal of i subband:

$y\left({\mathrm{\&omega;}}_i\right)={\left(W_{\mathrm{opt}}^i\right)}^H*S\left({\mathrm{\&omega;}}_i\right)---\left(22\right)$

In above-mentioned each formula, represent the position and the source location of microphone by adopting three dimensional space coordinate, make method of the present invention can be used for randomly topologically structured microphone array, microphone array of the present invention is not limited to circle battle array, linear array etc.Because position information such as the microphone among the present invention, signal source all are three-dimensional, so filtering technique of the present invention belongs to the three dimensions filtering technique, still, when microphone uses one-dimensional array such as ULA, uniform straight line array row etc., the three-dimensional filtering characteristic disappears.

The direction vector method for solving that uses among the present invention is applicable to randomly topologically structured microphone array.

Following step 7, step 8 are the output signal processing section.

Step 7, the frequency domain output signal of K subband is combined into frequency domain output signal Y (ω):

Y(ω)＝[y(ω ₁)y(ω ₂)…y(ω _K)] ^T (23)

Step 8, Y (ω) is carried out contrary FFT, obtain output signal Y (n), then Y (n) is converted to simulating signal y (t), y (t) is carried out low-pass filtering, obtain speech output signal.

Broadband beams based on microphone array provided by the invention forms device and mainly comprises: divide subband signal module, frequency domain correlation matrix module, weight vector module and output module.The function of dividing the subband signal module is realized by sampling submodule, speech detection submodule, spectrum counteracting denoising submodule and division subband signal submodule.

The sampling submodule is mainly used in according to predetermined sampling frequency the signal of importing each microphone is carried out the AD conversion, then, chooses signal frame and carry out short time discrete Fourier transform from the signal after the AD conversion.Above-mentioned sample frequency can be 16KHz, 22KHz or 44Khz etc., sampling precision can for: 8bit, 16bit or 32bit etc., short time discrete Fourier transform can realize by 512 the FFT of NFFT and short time discrete Fourier transform can be selected Hamming window or other window function etc. for use.The description of F in specific implementation process such as the method (n) and F (ω).

The speech detection submodule is mainly used in signal frame that the sampling submodule is handled, behind the Fourier transform and carries out speech detection, when definite signal frame is not speech frame, this signal frame is stored as current estimating noise spectrum, when definite signal frame is speech frame, this speech frame is transferred to spectrum offset the denoising submodule.The speech detection technology that the speech detection submodule adopts can combine etc. for zero-crossing rate, short-time energy.

Spectrum counteracting denoising submodule is mainly used in the current estimating noise spectrum of storing according to the speech detection submodule, and the speech frame that the transmission of speech detection submodule comes is composed the counteracting denoising, and the speech frame that will compose after the counteracting denoising transfers to division subband signal submodule.The description of S (ω) in signal after the spectrum counteracting denoising after process and the spectrum counteracting denoising such as the above-mentioned method.

Dividing the subband signal submodule is mainly used in and will compose the speech frame of offsetting the transmission of denoising submodule according to predetermined frequency band and be divided into a plurality of subband signals, and each subband signal transferred to frequency domain correlation matrix module and output module, S (ω in each subband signal of speech frame such as the above-mentioned method _i) description.

Frequency domain correlation matrix module is mainly used in the frequency domain correlation matrix of each subband signal of determining its reception, and transmits it to the weight vector module.Obtain the method for frequency domain correlation matrix R (i) such as the description in the above-mentioned method.

The weight vector module is mainly used in the three dimensions used for vector transfer a (r according to signal source ₀, θ ₀, φ ₀), each frequency domain correlation matrix R (i) of its reception determines the optimal weight vector of each subband signal

And

Transfer to output module.The three dimensions used for vector transfer a (r of signal source ₀, θ ₀, φ ₀), optimal weight vector

Acquisition methods such as the description in the above-mentioned method.

Output module is mainly used in the weight vector of each subband signal that transmission comes according to the weight vector module and carries out the subband spatial filtering to dividing each next subband signal of subband signal submodule transmission, obtain the frequency domain output signal of each subband, the frequency domain output signal of K subband is combined into the frequency domain output signal, and the frequency domain output signal that is combined into carried out contrary FFT, then, convert simulating signal to, this simulating signal is carried out the voice signal that signal after the low-pass filtering needing to be defined as output.

Though described the present invention by embodiment, those of ordinary skills know, the present invention has many distortion and variation and do not break away from spirit of the present invention, and the claim of application documents of the present invention comprises these distortion and variation.

Claims (11)

1, a kind of broadband beams formation method is characterized in that, comprises step:

Each subband signal that the signal of a, definite each microphone of input is corresponding respectively;

B, determine the frequency domain correlation matrix of described each subband signal;

C, determine the weight vector of each subband signal according to the three dimensions used for vector transfer of signal source, described each frequency domain correlation matrix;

D, determine the signal of output according to the weight vector of described each subband signal and each subband signal.

2, a kind of broadband beams formation method as claimed in claim 1 is characterized in that described step a specifically comprises the steps:

A1, the signal of importing each microphone is carried out speech detection, and definite speech frame;

A2, determine each subband signal of described speech frame correspondence.

3, a kind of broadband beams formation method as claimed in claim 2 is characterized in that:

The signal of setting each microphone of input is: F (t)=[f ₁(t) ... f _i(t) ... f _M(t)] ^T

Wherein: f _i(t) i signal that microphone receives of expression, i=1 ..., M, M are the quantity of microphone, [] ^TExpression is done the transposition computing to matrix or vector; And described step a1 specifically comprises the steps:

A11, the signal of importing each microphone is carried out AD conversion according to predetermined sampling frequency:

F (n)=[f ₁(n) ... f _i(n) ... f _M(n)] ^TWherein, n is a discrete time;

Choose signal frame a12, the signal after described AD changes and carry out short time discrete Fourier transform:

$F\left(\mathrm{\&omega;}\right)=\underset{m=1}{\overset{N}{\mathrm{\&Sigma;}}}F\left(n\right)w(n-m)\exp (-\mathrm{j\&omega;m})=\left[\begin{array}{c}\underset{m=1}{\overset{N}{\mathrm{\&Sigma;}}}{f}_1\left(n\right)w(n-m)\exp (-\mathrm{j\&omega;m})\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ \underset{m=1}{\overset{N}{\mathrm{\&Sigma;}}}{f}_M\left(n\right)w(n-m)\exp (-\mathrm{j\&omega;m})\end{array}\right];$ Wherein, w (n) is a window function, and n, m are discrete time;

A13, the signal frame behind the Fourier transform is carried out speech detection, determine speech frame.

4, a kind of broadband beams formation method as claimed in claim 3 is characterized in that described step a13 specifically comprises the steps:

Signal frame behind the Fourier transform is carried out speech detection;

When definite signal frame is not speech frame, this signal frame is stored as current estimating noise spectrum;

When definite signal frame is speech frame, according to current estimating noise spectrum described speech frame is composed the counteracting denoising, determine that the speech frame S (ω) after the spectrum counteracting denoising is:

$S\left(\mathrm{\&omega;}\right)=F\left(\mathrm{\&omega;}\right)-N\left(\mathrm{\&omega;}\right)={\left[\begin{array}{ccc}{s}_1\left(1\right)& \&CenterDot;\&CenterDot;\&CenterDot;& s_1\left(\mathrm{NFFT}\right)\\ \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;\\ s_M\left(1\right)& \&CenterDot;\&CenterDot;\&CenterDot;& s_M\left(\mathrm{NFFT}\right)\end{array}\right]}_{M\&times;\mathrm{NFFT}};$

Wherein: $N\left(\mathrm{\&omega;}\right)={\left[\begin{array}{ccc}{n}_1\left(1\right)& \&CenterDot;\&CenterDot;\&CenterDot;& n_1\left(\mathrm{NFFT}\right)\\ \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;\\ n_M\left(1\right)& \&CenterDot;\&CenterDot;\&CenterDot;& n_M\left(\mathrm{NFFT}\right)\end{array}\right]}_{M\&times;\mathrm{NFFT}}$ Be current estimating noise spectrum, NFFT is that the frequency sampling of short time discrete Fourier transform is counted, and F (ω) is the signal frame behind the short time discrete Fourier transform, and M is the quantity of microphone.

5, as claim 2,3 or 4 described a kind of broadband beams formation methods, it is characterized in that described step a2 specifically comprises the steps:

According to K predetermined frequency band speech frame is divided into K subband signal, and with K preset frequency ω _i, i=1 ... K is defined as the centre frequency of each subband;

Determine the component of signal S (ω of i subband _i) be: $S\left({\mathrm{\&omega;}}_i\right)=\left[\begin{array}{c}{S}_1\left(i\right)\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ S_M\left(i\right)\end{array}\right];$

Wherein: M is the number of microphone of microphone array, i=1 ... K, K are number of sub-bands.

6, a kind of broadband beams formation method as claimed in claim 5 is characterized in that described step b specifically comprises:

Determine the frequency domain correlation matrix R (i) of each subband signal:

R (i)=E{S (ω _i) S ^H(ω _i); Wherein: ^HExpression is carried out transposition and conjugate operation to matrix or vector, S (ω _i) be the component of signal of i subband, and $S\left({\mathrm{\&omega;}}_i\right)=\left[\begin{array}{c}{S}_1\left(i\right)\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ S_M\left(i\right)\end{array}\right].$

7, a kind of broadband beams formation method as claimed in claim 5 is characterized in that, the three dimensions used for vector transfer of signal source can obtain as follows among the described step c:

C1, obtain source location (r ₀, θ ₀, φ ₀) coordinate vector S=r ₀* [sin θ ₀Cos φ ₀Sin θ ₀Sin φ ₀Cos θ ₀];

C2, obtain the coordinate vector P of each microphone _i=r _i* [sin θ _iCos φ _iSin θ _iSin φ _iCos θ _i];

C3, determine source location (r ₀, θ ₀, φ ₀) to the relative amplitude decay factor of i microphone

For:

${\&PartialD;}_i=\frac{\left|\right|S\left|\right|}{\left|\right|P_i-S\left|\right|};$

Wherein: ‖ * ‖ represents the norm of vector *;

C4, determine source location (r ₀, θ ₀, φ ₀) to the relative time delay factor τ of i microphone _iFor:

${\mathrm{\&tau;}}_i=\frac{\left|\right|S-P_i\left|\right|-\left|\right|S\left|\right|}{c};$

Wherein: c is the aerial velocity of propagation of sound, and ‖ * ‖ represents the norm of vector *;

C5, determine source location (r ₀, θ ₀, φ ₀) three dimensions used for vector transfer a (r ₀, θ ₀, φ ₀) be:

$a(r_0,{\mathrm{\&theta;}}_0,{\mathrm{\&phi;}}_0)=\left[\begin{array}{ccc}{\&PartialD;}_1{e}^{-j{\mathrm{\&omega;}}_1{\mathrm{\&tau;}}_1}& \&CenterDot;\&CenterDot;\&CenterDot;{\&PartialD;}_m{e}^{-j{\mathrm{\&omega;}}_i{\mathrm{\&tau;}}_m}\&CenterDot;\&CenterDot;\&CenterDot;& {\&PartialD;}_M{e}^{-j{\mathrm{\&omega;}}_i{\mathrm{\&tau;}}_M}\end{array}\right];$

Wherein: ω _iBe the centre frequency of each subband, r ₀Be the distance of signal source to true origin, θ ₀Be the angle of the Z axle of signal source and three-dimensional coordinate, φ ₀Be the projection of signal source on XOY plane and the angle of X-axis.

8, as claim 1,2,3 or 4 described a kind of broadband beams formation methods, it is characterized in that described step c specifically comprises:

Determine the optimal weight vector of i subband

For:

$W_{\mathrm{opt}}^i=\frac{R{\left(i\right)}^{-1}a}{a^{H}R{\left(i\right)}^{-1}a};$

Wherein: R (i) is the frequency domain correlation matrix of i subband signal, and a is described source location (r ₀, θ ₀, φ ₀) the three dimensions used for vector transfer.

9, as claim 1,2,3 or 4 described a kind of broadband beams formation methods, it is characterized in that described steps d specifically comprises the steps:

Optimal weight vector according to each subband signal carries out the subband spatial filtering to each subband signal, obtains the frequency domain output signal y (ω of i subband _i):

$y\left({\mathrm{\&omega;}}_i\right)={\left(W_{\mathrm{opt}}^i\right)}^H*S\left({\mathrm{\&omega;}}_i\right);$

Wherein: ^HExpression is carried out transposition and conjugate operation to vector or matrix, Be the optimal weight vector of i subband, S (ω _i) be the component of signal of i subband;

The frequency domain output signal of each subband is combined as Y (ω): Y (ω)=[y (ω ₁) y (ω ₂) ... y (ω _K)] ^T

Frequency domain output signal Y (ω) after the combination is carried out contrary fast fourier transform obtain output signal Y (n);

Convert described Y (n) to simulating signal y (t), and y (t) is carried out the voice signal that signal after the low-pass filtering needing to be defined as output.

10, a kind of broadband beams forms device, it is characterized in that, comprising:

Divide the subband signal module: determine each corresponding respectively subband signal of signal of each microphone of input, and each subband signal is transferred to frequency domain correlation matrix module;

Frequency domain correlation matrix module: determine the frequency domain correlation matrix of described each subband signal, and transmit it to the weight vector module;

Weight vector module: determine the weight vector of each subband signal, and transmit it to output module according to the three dimensions used for vector transfer of signal source, described each frequency domain correlation matrix;

Output module: according to the weight vector of described each subband signal, the signal that each subband signal is determined output.

11, a kind of broadband beams as claimed in claim 10 forms device, it is characterized in that described division subband signal module comprises:

Sampling submodule: according to predetermined sampling frequency the signal of importing each microphone is carried out the AD conversion, and from the signal after the described AD conversion, choose signal frame and carry out short time discrete Fourier transform;

Speech detection submodule: the signal frame behind the Fourier transform is carried out speech detection, when definite signal frame is not speech frame, this signal frame is stored as current estimating noise spectrum, when definite signal frame is speech frame, this speech frame is transferred to spectrum offset the denoising submodule;

Spectrum is offset the denoising submodule: according to current estimating noise spectrum the speech frame of its reception is composed the counteracting denoising, and transfer to division subband signal submodule;

Divide the subband signal submodule: according to predetermined frequency band the speech frame of its reception is divided into a plurality of subband signals, and each subband signal is transferred to frequency domain correlation matrix module.

Images