CN106842112A - Sound localization method based on parametrization Bayes's dictionary learning under strong reverberant ambiance - Google Patents

Abstract

The invention discloses the sound localization method based on parametrization Bayes's dictionary learning, the technical problem for solving existing sound localization method poor accuracy under a kind of strong reverberant ambiance.Technical scheme is to simplify dictionary based on microphone mirror image, under the framework of sparse Bayesian, realizes that auditory localization and the synchronous of room reflections coefficient are estimated to solve.Dictionary is precalculated out using the inventive method, the efficiency for completing auditory localization and reflectance factor estimation can get a promotion, and so in the case where amount of calculation is not obviously improved, improve the precision of sound localization method.The present invention has very big potential practical value to the auditory localization under complex indoor environment, the method that can also estimate as parameters,acoustic in closing space.

Description

Sound localization method based on parametrization Bayes's dictionary learning under strong reverberant ambiance

Technical field

The present invention relates to a kind of sound localization method, based on parametrization Bayes's word under more particularly to a kind of strong reverberant ambiance The sound localization method of allusion quotation study.

Background technology

Document " Structured sparsity models for reverberant speech separation, Audio,Speech,and Language Processing,IEEE/ACM Transactions on,vol.22,pp.620- 633,2014 " the multiple microphone signals gathered by house interior, set multiple grid point estimation room shapes inside and outside room Shape and multistage imaginary source, secondly estimate the acoustic absorptivity in room using the false reputation source model in room, finally use convex optimization Data algorithm carries out the research of indoor multi-source separation.However, when room reverberation is stronger, mesh point number meeting is set in space Greatly increase, and sound absorption unknown parameters, the method is involved that predictor is more, and computational efficiency is low, derivation algorithm essence Degree is not high, thus cannot accurately calculate the position for obtaining indoor sound source.Therefore, for indoor auditory localization, to unknown room It is the precondition for rationally estimating room sound source position information that parameters,acoustic carries out accurate estimation.

The content of the invention

In order to overcome the shortcomings of that existing sound localization method poor accuracy, the present invention are based under providing a kind of strong reverberant ambiance Parameterize the sound localization method of Bayes's dictionary learning.The method is based on microphone mirror image and simplifies dictionary, in sparse Bayesian Framework under, realize that the synchronous of auditory localization and room reflections coefficient is estimated to solve.Dictionary is counted in advance using the inventive method Calculate, the efficiency for completing auditory localization and reflectance factor estimation can get a promotion, and so not be obviously improved amount of calculation In the case of, improve the precision of sound localization method.The present invention has very big latent to the auditory localization under complex indoor environment In practical value, the method that can also estimate as parameters,acoustic in closing space.

The technical solution adopted for the present invention to solve the technical problems：Based on parametrization Bayes under a kind of strong reverberant ambiance The sound localization method of dictionary learning, is characterized in comprising the following steps：

Step one, localizing environment are chosen and can be used in rectangle or the room according to shape direct access virtual source.Selected room Interior certain one-dimensional or two dimensional surface carries out indoor mesh generation, it is assumed that indoor mesh point number is G.

Step 2, structure signal model：It is sound source position in a room that the target information for obtaining is needed during auditory localization, The quantization of this information is discrete to turn to limited mesh point for being distributed in the interior space and outdoor mirror image space, each mesh point Sound source amplitude be x_g.Assuming that sound source is present in any one mesh point, reception signal y (r, f) of optional position is expressed as Formula (1)

In formula, f is current frequency, and c is the velocity of sound in air, and G is indoor mesh point number, β_gIt is the reflection of correspondence virtual source Coefficient, x_gTo reflect the vector of sound source position.On all of G mesh point, x_gSome only little non-zero entry numbers, correspondence Actual sound source position and the false reputation source position in mirror image space in room space.When dictionary is constructed, only in house interior setting Mesh point, each dictionary element is that the in-house network lattice point that sound source is present is superimposed with the passage between receiving point mirror image.Thus obtain

y_f=A_fx_f (4)

Formula (2) right-hand member represents total superposition of microphone and its contribution of mirror image acoustic energy, and formula (3) is public for dictionary atom computing Formula, N₀It is receiving point mirror image series,It is n_s,iThe corresponding sound pressure reflection coefficient of rank mirror image.Arrangement formula (3) obtains formula (4).Formula In, y_f∈C^M×1,x_f∈C^G×1,A_f∈C^M×G。

Step 3, receiving point mirror image are solved：It is the rectangular room of lx × ly × lz for size, it is assumed that receiving point position is (x₀,y₀,z₀), certain grade of position of receiving point mirror image is (x_i,y_i,z_i), l, m, n are integer of the boundary between-∞ and+∞, then appoint The space coordinates of meaning receiving point mirror image is expressed as

x_i=2llx ± x₀

y_i=2mly ± y₀ (5)

z_i=2nlz ± z₀

The reflection series of mirror image is given by

N=| 2l- Δs_l|+|2m-Δ_m|+|2n-Δ_n| (6)

Δ l=1 is represented and is sought x_iX in formula₀Before take it is negative；Δ l=0 is represented and is sought x_iX in formula₀Before take just；Δ m=1 is represented and is sought y_iFormula Middle y₀Before take it is negative；Δ m=0 is represented and is sought y_iY in formula₀Before take just；Δ n=1 is represented and is sought z_iZ in formula₀Before take it is negative；Δ n=0 is represented and is sought z_i Z in formula₀Before take just.

Step 4, construction solve dictionary：By the dictionary A under the different quantized values of formula (3) construction reflectance factor β_fAnd its To the first derivative A' of β_fWith matrix of second derivatives A "_f。

Step 5, position vector x_fSolve：When indoor acoustic fix ranging is carried out, the mesh point number G in room is not only much More than microphone number M, and more than sound source number.That is x_fWith spatial sparsity, estimate to contain by sparse Bayesian method The x of sound source location information.Observation signal is obtained by formula (7) sampling.

I_MIt is the unit matrix of M dimensions, x_fObtained by following formula：

x_f| α~CN (0, Λ^-1I) (8)

Wherein CN tables in above formula Give instructions in reply Gaussian Profile, Gamma represents that gamma is distributed.Above-mentioned probabilistic model is solved by variation Bayes-expectation-maximization algorithm. Wherein x_fPosterior probability obtained by formula (9)：

Represent the estimate to dependent variable.

The method for providing each parameter Estimation directly below, whereinObeying average is Gamma distribution.α₀Obeying average isHeight This distribution, the mark of Tr representing matrixs.

The estimation of step 6, parameter beta：Reflectance factor β parameters are included in above-mentioned each dictionary, estimates the parameters equivalent in making formula (10) minimize：

The optimization problem is solved by Newton methods, and iterative formula is as follows：

In formula, A'_fRepresent to corresponding parameter beta derivation, A "_fRepresent second dervative.By in sparse Bayesian framework Dictionary introduces reflectance factor as parameter, and solves the problem by Newton method iteration, is solved by substep, realizes estimating Sound source position x_fWhile, estimate reflectance factor β.On the one hand the estimation for changing parameter is room acoustics parameter Estimation, while Also more accurately room acoustics information is obtained, more preferable auditory localization effect is reached.

Step 7, parameter beta quantify：In the case where room information is totally unknown, the initial value of β is set to 0.5, is joined by band Several Bayes's Sparse methods, are utilized respectively formula (9) and formula (10) to sound source position x_fEstimated with reflectance factor β, reflected Coefficient is estimated to be carried out simultaneously with the estimation of sound source position, i.e., be repeated between formula (9) and formula (10), often walks iteration all Need to estimate reflectance factor and by quantifying to find immediate dictionary and carry out next step iteration, β differences will cause dictionary many It is secondary to recalculate, bring huge amount of calculation.

In order to avoid computing repeatedly dictionary, β is quantified as some discrete data points in its codomain scope (0~1) first, Then the corresponding dictionary of the value of these discrete points is calculated before iteration, the β that will be calculated quantifies to arrive corresponding discrete In value and call the corresponding dictionary can to substantially reduce amount of calculation.

The beneficial effects of the invention are as follows：The method is based on microphone mirror image and simplifies dictionary, under the framework of sparse Bayesian, Realize that auditory localization and the synchronous of room reflections coefficient are estimated to solve.Dictionary is precalculated out using the inventive method, it is complete The efficiency estimated into auditory localization and reflectance factor can get a promotion, so in the case where amount of calculation is not obviously improved, Improve the precision of sound localization method.The present invention has very big potential practical valency to the auditory localization under complex indoor environment Value, the method that can also estimate as parameters,acoustic in closing space.

Different signal to noise ratios are done to the collection signal for obtaining respectively to process and position, select frequency range be 1kHz~ 2kHz.The calculating time under the inventive method and literature method is counted respectively, and table 1 shows that correlation method is calculated at same The time-consuming situation of one-time positioning task is completed, two kinds of the inventive method offer take, and one is not quantify reflectance factor condition Under be calculated on (left side), one kind be quantify reflectance factor under the conditions of be calculated on (right side).

The time-consuming contrast of the distinct methods of table 1

Table 1 shows while accurate estimation reflectance factor, to calculate time-consuming more time-consuming than literature algorithms few.The inventive method The calculating time is not influenceed by RMR room reverb degree, and the Grid dimension that it is only divided in room has direct relation, Grid dimension Mesh is more, and solution vector dimension is higher, expends the time more.

The present invention is elaborated with reference to the accompanying drawings and detailed description.

Brief description of the drawings

Fig. 1 is the flow of the sound localization method based on parametrization Bayes's dictionary learning under strong reverberant ambiance of the invention Figure.

Fig. 2 is the inventive method test system connection block diagram.

Fig. 3 is the sound source information vector x iterative convergent process under 10*10 grids.

Fig. 4 is the sound source information vector x iterative convergent process under 20*20 grids.

Fig. 5 is that reflectance factor iterative mean square error (MSE) when different signal to noise ratio lower sensor numbers are 10*10 is bent Line.

Fig. 6 is that reflectance factor iterative mean square error (MSE) when different signal to noise ratio lower sensor numbers are 20*20 is bent Line.

In figure, MSE_x, MSE_beta, MSE_x0X, β and literature method gained x are corresponded to respectively₀。

Specific embodiment

Reference picture 1-6.Sound localization method tool based on parametrization Bayes's dictionary learning under strong reverberant ambiance of the invention Body step is as follows：

Step one, localizing environment are chosen：This method is only available for rectangle or other can direct access virtual source according to shape Room.The indoor mesh generation for carrying out auditory localization in selected room in certain one-dimensional or two dimensional surface, it is assumed here that indoor net Lattice point number is G.

Step 2, structure signal model：It is sound source position in a room that the target information for obtaining is needed during auditory localization, The quantization of this information is discrete to turn to limited mesh point for being distributed in the interior space and outdoor mirror image space, each mesh point Possible sound source amplitude is x_g.Assuming that sound source is likely to be present in any one mesh point, the reception signal y of optional position (r, F) it is represented by formula (1)

F is current frequency in formula, and c is the velocity of sound in air, and G is the total number of mesh point, β_gIt is the reflection of correspondence virtual source Coefficient, x_gTo reflect the vector of sound source position.On all possible G mesh point, x_gSome only little non-zero entry numbers, The false reputation source position of actual sound source position and mirror image space corresponding to room space.When dictionary is constructed, only in house interior Setting mesh point, each dictionary element is superimposed for sound source in-house network lattice point that may be present with the passage between receiving point mirror image. Therefore can obtain

y_f=A_fx_f (4)

Formula (2) right-hand member is expressed as total superposition of the acoustic energy contribution of microphone and its mirror image, and formula (3) is dictionary atom computing Formula, N0 is receiving point mirror image series,It is n_s,iThe corresponding sound pressure reflection coefficient of rank mirror image.Arrangement formula (3) can obtain formula (4). In formula, y_f∈C^M×1,x_f∈C^G×1,A_f∈C^M×G。

Step 3, receiving point mirror image are solved：It is the rectangular room of lx × ly × lz for size, it is assumed that receiving point position is (x₀,y₀,z₀), certain grade of position of receiving point mirror image is (x_i,y_i,z_i), l, m, n are integer of the boundary between-∞ and+∞, then appoint The space coordinates of meaning receiving point mirror image is represented by

x_i=2llx ± x₀

y_i=2mly ± y₀ (5)

z_i=2nlz ± z₀

The reflection series of mirror image can be given by

N=| 2l- Δs_l|+|2m-Δ_m|+|2n-Δ_n| (6)

Δ l=1 is represented and is sought x_iX in formula₀Before take it is negative；Δ l=0 is represented and is sought x_iX in formula₀Before take just；Δ m=1 is represented and is sought y_iFormula Middle y₀Before take it is negative；Δ m=0 is represented and is sought y_iY in formula₀Before take just；Δ n=1 is represented and is sought z_iZ in formula₀Before take it is negative；Δ n=0 is represented and is sought z_i Z in formula₀Before take just.

Step 4, construction solve dictionary：By the dictionary A under the different quantized values of formula (3) construction reflectance factor β_fAnd its To the first derivative A' of β_fWith matrix of second derivatives A "_f。

Step 5, position vector x_fSolve：When indoor acoustic fix ranging is carried out, the mesh point number G in room is not only much More than microphone number M, and more than sound source number.That is x_fWith spatial sparsity, estimate to contain by sparse Bayesian method The x of sound source location information.Observation signal is obtained by formula (7) sampling.

y_f|x_f,α₀~CN (A_fx_f,α₀ ^-1I_M) (7)

I_MIt is the unit matrix of M dimensions, x_fObtained by following formula：

x_f| α~CN (0, Λ^-1I) (8)

Wherein CN tables in above formula Give instructions in reply Gaussian Profile, Gamma represents that gamma is distributed.Above-mentioned probabilistic model can be asked by variation Bayes-expectation-maximization algorithm Solution.Wherein x_fPosterior probability can be obtained by formula (9)：

Represent the estimate to dependent variable.

The method for providing each parameter Estimation directly below, whereinObeying average is Gamma distribution.α₀Obeying average isHeight This distribution, the mark of Tr representing matrixs.

The estimation of step 6, parameter beta：Reflectance factor β parameters are included in above-mentioned each dictionary, estimates that the parameter can be equivalent to Minimize formula (10)：

The optimization problem can be solved by Newton methods, and iterative formula is as follows：

A' in formula_fRepresent to corresponding parameter beta derivation, A "_fRepresent second dervative.By in sparse Bayesian framework Dictionary introduces reflectance factor as parameter, and solves the problem by Newton method iteration, is solved by substep, may be implemented in and estimates Meter sound source position x_fWhile, estimate reflectance factor β.On the one hand the estimation for changing parameter is room acoustics parameter Estimation, together When also can obtain more accurately room acoustics information, can reach more preferable auditory localization effect.

Step 7, parameter beta quantify：In the case where room information is totally unknown, the initial value of β is traditionally arranged to be 0.5, passes through Bayes's Sparse methods with parameter are utilized respectively formula (9) and formula (10) to sound source position x_fEstimated with reflectance factor β, instead The estimation of coefficient estimation and sound source position is penetrated while carrying out, i.e., is repeated between (9) and (10), often walk iteration all Need to estimate reflectance factor and by quantifying to find immediate dictionary and carry out next step iteration, β differences will cause dictionary many It is secondary to recalculate, bring huge amount of calculation.

In order to avoid computing repeatedly dictionary, β is quantified as some discrete data points in its codomain scope (0~1) first (present invention takes 1000 uniform quantization methods of point), then calculates the corresponding dictionary of the value of these discrete points before iteration Come, the β that will be calculated quantifies onto corresponding centrifugal pump and calls the corresponding dictionary can to substantially reduce amount of calculation.

Application Example：

Room is an enclosing square space.Its length, width and height is respectively lx=4m, ly=5m, lz=3m.

Step 1：Sustained height is selected inside closed room, the linear microphone array of 15 array elements, array position is arranged It is x=1.44m, at intervals of 0.2m, initiating terminal microphone position is y=0.1m in z=1.5m, y direction.

Step 2：According to room-sized, the plane for selecting z=1.5m is main reference zone, and x, y direction respectively divide 10 sections The uniform mesh point for carrying out mesh generation, i.e. house interior is 100 (or 20 sections totally 400 points), sizing grid be 40cm × 50cm(20cm×25cm).Mesh coordinate takes grid element center position coordinates.

Step 3：According to room wall surface material, when obtaining each wall acoustic absorptivity or direct measurement room reverberation by tabling look-up Between determine room average sound pressure reflectance factor β₀。

Step 4：A pre- mesh point for dividing is randomly selected as sound source position, computer, Yi Jixiang are connected according to accompanying drawing 2 The collecting device answered.

Step 5：Open all devices and drive sound source, persistently play predetermined white noise signal and set gathered time domain The sample frequency of signal is simultaneously recorded, and preserves data.Short time discrete Fourier transform is done to each preservation data, each microphone letter is obtained Number frequency domain representation y_f。

Step 6：Selected coordinate reference points, measure relative position of each microphone in room, according to house interior grid Point is divided, and solves the positional information r of each microphone mirror image_iAnd the reflection series n of each mirror image_s,i。

Step 7：According to each microphone mirror position r_i, mirror reflection series n_s,iAnd each wall reflectance factor initial value β₀Carry out Quantify to solve, according to the corresponding dictionary atom of each mesh point in formula (5) solution room, form the dictionary A under each Frequency point_fAnd Its Jacobian matrix A'_f、A”_f。

Step 8：Based on y_fAnd A_f, the position vector x under each independent frequency is obtained according to formula (9)_fAnd combine each frequency The position vector x that point data is solved_t, room reflections coefficient estimation is carried out according to formula (10), the step is typically repeated repeatedly straight To position vector x_tAnd reflectance factor β is not in change.

Step 9：According to solving result x_f、x_tThe corresponding mesh point of middle maximum is legal using we under the Frequency point The sound source position that position obtains.

The situation reference picture 3, Fig. 4 of reflectance factor iteration convergence in the case of different initial values and mesh generation in this method. Wherein thick dashed line is actual value, it can be seen that under different initial values in the case of 10*10,20*20 mesh generation, using the inventive method The actual value of reflectance factor can finally be converged to.

Claims (1)

1. under a kind of strong reverberant ambiance based on parametrization Bayes's dictionary learning sound localization method, it is characterised in that including with Lower step：

Step one, localizing environment are chosen and can be used in rectangle or the room according to shape direct access virtual source；Select certain in room One-dimensional or two dimensional surface carries out indoor mesh generation, it is assumed that indoor mesh point number is G；

Step 2, structure signal model：It is sound source position in a room that the target information for obtaining is needed during auditory localization, this The quantization of information is discrete to turn to limited mesh point for being distributed in the interior space and outdoor mirror image space, the sound of each mesh point Source amplitude is x_g；Assuming that sound source is present in any one mesh point, reception signal y (r, f) of optional position is expressed as formula (1)

$y(r,f)=\underset{g=0}{\overset{G}{\mathrm{\Σ}}}{x}_g{\mathrm{\β}}_g\exp (j2\mathrm{\π}f/c|r-r_{m,g}\left|\right)/\left(4\mathrm{\π}\right|r-r_{m,g}\left|\right)---\left(1\right)$

In formula, f is current frequency, and c is the velocity of sound in air, and G is indoor mesh point number, β_gIt is the reflectance factor of correspondence virtual source, x_gTo reflect the vector of sound source position；On all of G mesh point, x_gSome only little non-zero entry numbers, corresponding to room The actual sound source position in space and the false reputation source position in mirror image space；When dictionary is constructed, mesh point only is set in house interior, Each dictionary element is that the in-house network lattice point that sound source is present is superimposed with the passage between receiving point mirror image；Thus obtain

$H(r_i,g_j,f)=\underset{n_{s,i}=0}{\overset{N_0}{\mathrm{\Σ}}}{\mathrm{\β}}^{n_{s,i}}\exp (j2\mathrm{\π}f/c|r_{n_{s,i}}-g_j\left|\right)/\left(4\mathrm{\π}\right|r_{n_{s,i}}-g_j\left|\right)---\left(3\right)$

y_f=A_fx_f (4)

Formula (2) right-hand member represents total superposition of microphone and its contribution of mirror image acoustic energy, and formula (3) is dictionary atom computing formula, N₀For Receiving point mirror image series,It is n_s,iThe corresponding sound pressure reflection coefficient of rank mirror image；Arrangement formula (3) obtains formula (4)；In formula, y_f∈C^{M ×1},x_f∈C^G×1,A_f∈C^M×G；

Step 3, receiving point mirror image are solved：It is the rectangular room of lx × ly × lz for size, it is assumed that receiving point position is (x₀, y₀,z₀), certain grade of position of receiving point mirror image is (x_i,y_i,z_i), l, m, n are integer of the boundary between-∞ and+∞, then arbitrarily connect The space coordinates of sink mirror image is expressed as

x_i=2llx ± x₀

y_i=2mly ± y₀ (5)

z_i=2nlz ± z₀

The reflection series of mirror image is given by

N=| 2l- Δs_l|+|2m-Δ_m|+|2n-Δ_n| (6)

Δ l=1 is represented and is sought x_iX in formula₀Before take it is negative；Δ l=0 is represented and is sought x_iX in formula₀Before take just；Δ m=1 is represented and is sought y_iY in formula₀ Before take it is negative；Δ m=0 is represented and is sought y_iY in formula₀Before take just；Δ n=1 is represented and is sought z_iZ in formula₀Before take it is negative；Δ n=0 is represented and is sought z_iIn formula z₀Before take just；

Step 4, construction solve dictionary：By the dictionary A under the different quantized values of formula (3) construction reflectance factor β_fAnd its to β's First derivative A'_fWith matrix of second derivatives A "_f；

Step 5, position vector x_fSolve：When indoor acoustic fix ranging is carried out, the mesh point number G in room is not only far longer than biography Sound device number M, and more than sound source number；That is x_fWith spatial sparsity, estimate to contain sound source by sparse Bayesian method The x of positional information；Observation signal is obtained by formula (7) sampling；

I_MIt is the unit matrix of M dimensions, x_fObtained by following formula：

Wherein In above formulaRepresent multiple Gaussian Profile, Gamma represents that gamma is distributed；Above-mentioned probabilistic model is solved by variation Bayes-expectation-maximization algorithm；Wherein x_fPosterior probability obtained by formula (9)：

Represent the estimate to dependent variable；

The method for providing each parameter Estimation directly below, whereinObeying average is Gamma distribution；α₀Obeying average isHeight This distribution, the mark of Tr representing matrixs；

The estimation of step 6, parameter beta：Reflectance factor β parameters are included in above-mentioned each dictionary, estimates the parameters equivalent in making formula (10) Minimize：

$\mathrm{\β}=\underset{\mathrm{\β}}{\arg }{\mathrm{min\Σ}}_f\left|\right|y_f-A_f\left(\mathrm{\β}\right){\hat{x}}_f|{|}_2^2+Tr\left(A_f{\mathrm{\ΣA}}_f^H\right)---\left(10\right)$

The optimization problem is solved by Newton methods, and iterative formula is as follows：

${\mathrm{\β}}^T={\mathrm{\β}}^{T-1}-{\mathrm{\ω\Σ}}_f\frac{y_f^H{A}_f^{\′}{\hat{x}}_f-\left|\right|A_f{\hat{x}}_f|{|}_2^2}{{\left(A_f^{\′\′}{\hat{x}}_f\right)}^H(y_f-A_f{\hat{x}}_f)-\left|\right|A_f{\hat{x}}_f|{|}_2^2}{|}_{\mathrm{\β}={\mathrm{\β}}^{T-1}}---\left(11\right)$

In formula, A'_fRepresent to corresponding parameter beta derivation, A "_fRepresent second dervative；By to the dictionary in sparse Bayesian framework Reflectance factor is introduced as parameter, and the problem is solved by Newton method iteration, solved by substep, realize estimating sound source Position x_fWhile, estimate reflectance factor β；On the one hand the estimation for changing parameter is room acoustics parameter Estimation, while also To more accurately room acoustics information, more preferable auditory localization effect is reached；

Step 7, parameter beta quantify：In the case where room information is totally unknown, the initial value of β is set to 0.5, by with parameter Bayes's Sparse methods, are utilized respectively formula (9) and formula (10) to sound source position x_fEstimated with reflectance factor β, reflectance factor Estimate to be carried out simultaneously with the estimation of sound source position, i.e., be repeated between formula (9) and formula (10), often walk iteration and be required for Estimate reflectance factor and by quantifying to find immediate dictionary and carry out next step iteration, β differences will cause dictionary repeatedly to weigh It is new to calculate, bring huge amount of calculation；

In order to avoid computing repeatedly dictionary, β is quantified as some discrete data points in its codomain scope (0~1) first, then The corresponding dictionary of the value of these discrete points is calculated before iteration, the β that will be calculated quantifies onto corresponding centrifugal pump And call the corresponding dictionary can to substantially reduce amount of calculation.