CN104503963A - Method for processing head-related impulse response data set - Google Patents

Abstract

The embodiment of the invention provides a method for processing a head-related impulse response data set. The method comprises the following steps: a processor groups the head-related impulse response (HRIR) data set according to an angle of elevation and/or a horizontal angle; the processor selects the HRIR sample point with a minimum reconstruction error value in the grouped HRIR data set as a high-dimensional HRIR sample point through locally linear embedding (LLE); the processor reduces the dimension of the HRIR sample point to a low-dimensional space to obtain low-dimension HRIR sample points; the processor adopts sectional straight line fitting to select characteristic HRIR sample points in the low-dimension HRIR sample points; and the processor combines the characteristic HRIR sample points into a three-dimensional matrix according to an angle-of-elevation sequence and/or a horizontal angle sequence. The embodiment of the invention solves the problem of excessive errors of a characteristic HRIR selection method.

Description

Head coherent pulse response data sets disposal route

Technical field

The embodiment of the present invention relates to computer science, particularly relates to a kind of head coherent pulse response data sets disposal route.

Background technology

Human auditory can not only tone, loudness of a sound, the tone color of perceived sounds, can also differentiate Sounnd source direction and distance.In order to make not audience at the scene also can obtain " telepresenc " true to nature, people have invented three dimensional audio technology.Three dimensional audio technology can reproduce the acoustic information of original sound field truly, makes hearer produce the listening experience of " on the spot in person ".Three-dimensional audio realizes by " binaural sound " technology, namely head related transfer function (Head Related Transfer function is adopted, hereinafter referred to as: HRTF) filtering is carried out to single channel sound-source signal, by the sound with sense of direction obtained, then reset with earphone or a pair loudspeaker.HRTF time-domain representation be called the response of coherent pulse (Head Related Impulse Responses, hereinafter referred to as: HRIR).Binaural sound learns a skill and is widely used in blind person's sense of hearing navigation, flight training, virtual reality.HRTF describes the process that sound propagates into from sound source in ear, and the trunk, head, ear etc. of people have diffraction to voice signal, reflect, the effect such as to block, and can be equivalent to filter action.Usually the HRTF database obtained through accurate experiment measurement is adopted in reality.As shown in Figure 1, this database is a three-dimensional matrice, stores the HRTF of thousands of sound source locus, and the shock response duration of each HRTF is longer, and Equivalent FIR Filter exponent number is very high, and everyone HRTF is not quite similar.Therefore, the HRTF data of the person that store multiple-position measurement, need to consume huge storage space.When adopting embedded chip synthesis three-dimensional sound, the more aobvious scarcity of its storage space, becomes the bottleneck of three dimensional audio technology application.Now, if can with auditory localization performance loss inconspicuous for cost, exchange the very large storage space reduction of HRTF database for, just better can solve HRTF database and expend the large problem of memory space, reduce power consumption and the cost of the chip used in practical application.

In prior art, adopt manifold learning LLE algorithm to carry out dimensionality reduction to one group of HRIR, extract feature HRIR and represent all HRIR of this group, and other non-feature HRIR carries out interpolation by the feature HRIR of this group and obtains.The all HRIR being about to be positioned at same horizontal angle are divided into one group, and to often organizing HRIR, by dimensionality reduction module, by higher-dimension HRIR sample point dimensionality reduction to lower dimensional space, obtain the HRIR sample point of low-dimensional, this dimensionality reduction module uses LLE algorithm; Then, choose module by unique point, carrying out cluster to often organizing low-dimensional sample point, obtaining the low-dimensional sample point corresponding to cluster centre, this higher-dimension HRIR corresponding to low-dimensional sample point is the feature HRIR selected by this group, and what unique point chose that module adopts is K-means clustering algorithm.When reconstructing all HRIR according to feature HRIR, adopt two step interpolation: the first, adopt linear interpolation, according to the feature HRIR of this group, reconstruct all non-feature HRIR of this group; The second, the Neighborhood matrix neighbor obtained in LLE module in basis and weight matrix W, again carries out interpolation to non-feature HRIR, obtains the HRIR after all reconstruct of this group.

Due in feature HRIR leaching process, adopt K-mean cluster module extract the feature HRIR obtained and be weighted reconstruct, reconstructed error is bigger than normal.And reconstructing in all HRIR processes according to feature HRIR, the reconstructed error that the two step interpolation adopted in weighting reconstructed module obtain is bigger than normal, when this is owing to carrying out second step interpolation, non-feature HRIR is through linear interpolation, itself is with error, if use these HRIR to be reconstructed other orientation HRIR, can cause deviation accumulation.

Summary of the invention

The embodiment of the present invention provides a kind of head coherent pulse response data sets disposal route, to overcome technical matters large for the reconstructed error of HRIR in prior art.

Embodiments provide a kind of head coherent pulse response data sets disposal route, comprising:

Head coherent pulse response HRIR data set divides into groups according to the elevation angle and/or horizontal angle by processor;

Described processor is by choosing the HRIR of reconstructed error minimum value as higher-dimension HRIR sample point in Local Liner Prediction LLE after grouping HRIR;

Described HRIR sample point dimensionality reduction to lower dimensional space, is obtained low-dimensional HRIR by described processor;

Described processor adopts sectional straight line fitting selected characteristic HRIR sample point in described low-dimensional HRIR sample point;

Described feature HRIR sample is combined into three-dimensional matrice according to elevation angle order and/or horizontal angle order by described processor.

Further, described processor, by choosing the HRIR of reconstructed error minimum value in Local Liner Prediction LLE after grouping HRIR as HRIR sample point, comprising:

Processor chooses a HRIR sample point in HRIR after grouping, calculates the Euclidean distance between described HRIR sample point and other HRIR sample points, and chooses the point of proximity of described HRIR sample point according to described Euclidean distance;

Construct the local covariance matrix of described HRIR sample point, determine the conditional value of reconstructed error

$\underset{j=1}{\overset{K}{\mathrm{\Σ}}}{w}_j^i=1---\left(1\right)$

Wherein, described K is and HRIR sample point x _ineighbor point number, described j is the subscript of a jth neighbor point, described in for using a jth neighbor point x _jreconstruct x _iweights;

By described local covariance matrix Q ⁱand described conditional value, adopt formula

$w_j^i=\frac{\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{\left(Q^i\right)}_{\mathrm{jk}}^{-1}}{\underset{j=1}{\overset{K}{\mathrm{\Σ}}}\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{\left(Q^i\right)}_{\mathrm{jk}}^{-1}}---\left(2\right)$

Calculate the weights reconstruction matrix w of described HRIR sample point _ij, wherein, described K is local covariance matrix Q ⁱdimension, described i is the subscript of i-th sample point, and described k is local covariance matrix Q ⁱkth dimension, described j is the subscript of a jth neighbor point;

By described w _ijsubstitute into reconstructed error function formula

$\min \mathrm{\ϵ}\left(W\right)=\underset{i=1}{\overset{I}{\mathrm{\Σ}}}{|x_i-\underset{j=1}{\overset{K}{\mathrm{\Σ}}}{w}_j^i{x}_j|}^2---\left(3\right)$

Determine that the HRIR sample point corresponding to reconstructed error minimum value is the higher-dimension HRIR sample point that can carry out dimension-reduction treatment, wherein, described I is that sample is always counted, and described i is the subscript of i-th sample, and described K is setting and described HRIR sample point x _ithe number of the sample point that Euclidean distance is minimum, described x _ifor the described HRIR sample point chosen, described x _jfor described x _ia jth neighbor point, described min ε (W) is minimal reconstruction error amount;

Described HRIR sample point dimensionality reduction to lower dimensional space, is obtained low-dimensional HRIR, comprises by described processor:

Formula will be met

$\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{y}_i=0,\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{y}_i{y_i}^T=I---\left(4\right)$

Described HRIR sample point substitute into formula

$\min \mathrm{\ϵ}\left(Y\right)=\underset{i=1}{\overset{I}{\mathrm{\Σ}}}{|y_i-\underset{j=1}{\overset{K}{\mathrm{\Σ}}}{w}_j^i{y}_j|}^2---\left(5\right)$

Obtain loss function

$\min \mathrm{\ϵ}\left(Y\right)=\underset{j=1}{\overset{N}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{m}_{\mathrm{ij}}{y_i}^T{y}_j---\left(6\right)$

Choose formula

M＝(I-W) ^T(I-W) (7)

Minimum d the nonzero eigenvalue of middle M is updated to described loss function, obtains with the proper vector of Y as the corresponding sample point mapping to lower dimensional space, wherein, and described y _ifor described x _ilow-dimensional map vector, described y _jfor described y _ia jth neighbor point, described min ε (Y) is loss function, and described i is the subscript of i-th sample, and described j is the subscript of a jth neighbor point, and described I is that sample is always counted, and described k is and y _icontiguous counts.

Further, described processor adopts sectional straight line fitting selected characteristic HRIR sample point in the HRIR sample point of described low-dimensional, comprising:

Any two starting points for the one-dimensional manifold of low-dimensional HRIR group adopt formula

$\left(\begin{array}{c}{y}_{1d}\\ y_{2d}\\ \·\\ \·\\ \·\\ y_{\mathrm{Nd}}\end{array}\right)=\left(\begin{array}{cccc}{y}_{11}& y_{12}& \·\·\·& y_{1,d-1}\\ y_{21}& y_{22}& \·\·\·& y_{2,d-1}\\ \·& \·& & \·\\ \·& \·& & \·\\ \·& \·& & \·\\ y_{N1}& y_{N2}& \·\·\·& y_{N,d-1}\end{array}\right)\left(\begin{array}{c}{a}_0\\ a_1\\ \·\\ \·\\ \·\\ a_{d-1}\end{array}\right)---\left(8\right)$

Calculated line coefficient (a ₀, a ₁..., a _d-1), wherein, described y is the point of the one-dimensional manifold of low-dimensional HRIR group, and described d is space dimensionality, and described a is straight line coefficient;

According to the bearing of trend of one-dimensional manifold, add N point, adopt formula

$e_N=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\left[{\left|\right|y_{\mathrm{id}}-{\hat{y}}_{\mathrm{id}}\left|\right|}_2\right]}^2---\left(9\right)$

Calculate N point error of fitting e _n, wherein, described N counts for described any two starting point institute structures are straight, and described i is that i-th section of straight line is counted, and described y is the point of the one-dimensional manifold of low-dimensional HRIR group, and described d is space dimensionality, described in for the actual point of the one-dimensional manifold of low-dimensional HRIR group;

If the described error of fitting e of N point _nbe less than maximum error e _mAX, then N point is joined in the point range between two starting points; If the described error of fitting e of N point _nbe greater than maximum error e _mAXthen using the end points of N-1 point as this section of straight line, then the higher-dimension HRIR corresponding to described end points is feature HRIR.

Further, described e _mAX=ENER × ratio, wherein, described ratio is error of fitting ratio, and described ENER is the normalized energy that low-dimensional embeds stream shape.

The embodiment of the present invention, HRIR data set divides into groups according to the elevation angle and/or horizontal angle by processor, described processor is by choosing the HRIR of reconstructed error minimum value as higher-dimension HRIR sample point in Local Liner Prediction LLE after grouping HRIR, described processor by described HRIR sample point dimensionality reduction to lower dimensional space, obtain low-dimensional HRIR, adopt sectional straight line fitting selected characteristic HRIR sample point in described low-dimensional HRIR sample point, according to elevation angle order and/or horizontal angle order, described feature HRIR sample is combined into three-dimensional matrice, solve the problem that in prior art, extraordinary HRIR choosing method error is excessive.

Accompanying drawing explanation

In order to be illustrated more clearly in the embodiment of the present invention or technical scheme of the prior art, be briefly described to the accompanying drawing used required in embodiment or description of the prior art below, apparently, accompanying drawing in the following describes is some embodiments of the present invention, for those of ordinary skill in the art, under the prerequisite not paying creative work, other accompanying drawing can also be obtained according to these accompanying drawings.

Fig. 1 is HRTF three-dimensional data base schematic diagram in prior art;

Fig. 2 is HRIR data set process flow figure of the present invention;

Fig. 3 is HRIR 3-D data set process flow figure of the present invention;

Fig. 4 is Local Liner Prediction LLE schematic diagram of the present invention;

Fig. 5 is the present invention 2 dimensional feature HRIR position view;

Fig. 6 is head relative coordinate system schematic diagram of the present invention.

Embodiment

For making the object of the embodiment of the present invention, technical scheme and advantage clearly, below in conjunction with the accompanying drawing in the embodiment of the present invention, technical scheme in the embodiment of the present invention is clearly and completely described, obviously, described embodiment is the present invention's part embodiment, instead of whole embodiments.Based on the embodiment in the present invention, those of ordinary skill in the art, not making the every other embodiment obtained under creative work prerequisite, belong to the scope of protection of the invention.

Fig. 2 is head coherent pulse response data sets process flow figure of the present invention, and as shown in Figure 2, the method for the present embodiment can comprise:

Head coherent pulse response HRIR data set divides into groups according to the elevation angle and/or horizontal angle by step 101, processor;

Step 102, described processor are by choosing the HRIR of reconstructed error minimum value as higher-dimension HRIR sample point in Local Liner Prediction LLE after grouping HRIR;

Described HRIR sample point dimensionality reduction to lower dimensional space, is obtained low-dimensional HRIR by step 103, described processor;

Step 104, described processor adopt sectional straight line fitting selected characteristic HRIR sample point in described low-dimensional HRIR sample point;

Described feature HRIR sample is combined into three-dimensional matrice according to elevation angle order and/or horizontal angle order by step 105, described processor.

Specifically, in HRIR database, each orientation for the HRIR of a plane should be equally distributed.Illustrate, in the HRIR database of CIPIC, the measurement orientation in horizontal angle is respectively-80 ° ,-65 ° ,-55 ° ,-45 ° ~ 45 °, 55 °, 65 °, 80 °.Wherein, between [-45 °, 45 °], uniform sampling is carried out with 5 ° of intervals.Therefore, the HRIR between [-45 °, 45 °] of each horizontal angle plane should be compiled is one group.And-80 of two ends ° ,-65 ° ,-55 ° ,-45 ° ~ 45 °, 55 °, 65 °, the HRIR in the 80 ° of orientation feature HRIR that will be directly brought into concentrates.Fig. 3 a is each elevation plane of HRIR three-dimensional data after grouping, and Fig. 3 b is each horizontal angle plane of HRIR three-dimensional data after grouping.For the HRIR after grouping by Local Liner Prediction (Locally Linear Embedding, hereinafter referred to as: LLE) calculate reconstructed error, and the HRIR choosing reconstructed error minimum value is as higher-dimension HRIR sample point, by this higher-dimension HRIR sample point dimensionality reduction to lower dimensional space, corresponding low-dimensional HRIR is obtained in lower dimensional space, adopt sectional straight line fitting selected characteristic HRIR sample point from low-dimensional HRIR sample point, finally according to the elevation angle and/or horizontal angle order, described feature HRIR sample is combined into three-dimensional matrice.

The embodiment of the present invention, HRIR data set divides into groups according to the elevation angle and/or horizontal angle by processor, described processor is by choosing the HRIR of reconstructed error minimum value as higher-dimension HRIR sample point in Local Liner Prediction LLE after grouping HRIR, described processor by described HRIR sample point dimensionality reduction to lower dimensional space, obtain low-dimensional HRIR, adopt sectional straight line fitting selected characteristic HRIR sample point in described low-dimensional HRIR sample point, according to elevation angle order and/or horizontal angle order, described feature HRIR sample is combined into three-dimensional matrice, solve the problem that in prior art, extraordinary HRIR choosing method error is excessive.

Fig. 4 is Local Liner Prediction LLE schematic diagram of the present invention, and as shown in Figure 4, the method for the present embodiment can comprise:

Step 201, processor choose a HRIR sample point in HRIR after grouping, calculate the Euclidean distance between described HRIR sample point and other HRIR sample points, and choose the point of proximity of described HRIR sample point according to described Euclidean distance;

Step 202, construct the local covariance matrix of described HRIR sample point, determine the conditional value of reconstructed error

$\underset{j=1}{\overset{K}{\mathrm{\Σ}}}{w}_j^i=1---\left(1\right)$

Wherein, described K is and HRIR sample point x _ineighbor point number, described j is the subscript of a jth neighbor point, described in for using a jth neighbor point x _jreconstruct x _iweights;

Step 203, by described local covariance matrix Q ⁱand described conditional value, adopt formula

$w_j^i=\frac{\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{\left(Q^i\right)}_{\mathrm{jk}}^{-1}}{\underset{j=1}{\overset{K}{\mathrm{\Σ}}}\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{\left(Q^i\right)}_{\mathrm{jk}}^{-1}}---\left(2\right)$

Calculate the weights reconstruction matrix w of described HRIR sample point _ij, wherein, described K is local covariance matrix Q ⁱdimension, described i is the subscript of i-th sample point, and described k is local covariance matrix Q ⁱkth dimension, described in, j is the subscript of a jth neighbor point;

Step 204, by described w _ijsubstitute into reconstructed error function formula

$\min \mathrm{\ϵ}\left(W\right)=\underset{i=1}{\overset{I}{\mathrm{\Σ}}}{|x_i-\underset{j=1}{\overset{K}{\mathrm{\Σ}}}{w}_j^i{x}_j|}^2---\left(3\right)$

Determine that the HRIR sample point corresponding to reconstructed error minimum value is the higher-dimension HRIR sample point that can carry out dimension-reduction treatment, wherein, described I is that sample is always counted, and described i is the subscript of i-th sample, and described K is setting and described HRIR sample point x _ithe number of the sample point that Euclidean distance is minimum, described x _ifor the described HRIR sample point chosen, described x _jfor described x _ia jth neighbor point, described min ε (W) is minimal reconstruction error amount;

Specifically, in LLE algorithm, each HRIR can think a sample point in higher dimensional space, is designated as x _i, i=1,2 ..., I.By calculating each sample point x _iand Euclidean distance between all the other sample points, and using apart from a minimum K sample point as x _ineighbor Points.K value can preset, and is taken as 12 in the present embodiment.The calculating of following calculating through type (1) rebuilds weight matrix W.

$\min \mathrm{\ϵ}\left(W\right)=\underset{i=1}{\overset{I}{\mathrm{\Σ}}}{|x_i-\underset{j=1}{\overset{K}{\mathrm{\Σ}}}{w}_j^i{x}_j|}^2---\left(4\right)$

Wherein, x _jfor x _ia jth Neighbor Points; for using x _jreconstruct x _iweights, need satisfy condition construct the local covariance matrix Q of this sample point ⁱ, its element calculated by publicity (2),

${q^i}_{\mathrm{jk}}={(x_i-x_j)}^T(x_i-x_k)---\left(5\right)$

Wherein, described T is matrix transpose computing;

By formula (5) with in conjunction with, use formula (6) to calculate w _ij

$w_j^i=\frac{\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{\left(Q^i\right)}_{\mathrm{jk}}^{-1}}{\underset{j=1}{\overset{K}{\mathrm{\Σ}}}\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{\left(Q^i\right)}_{\mathrm{jk}}^{-1}}---\left(6\right)$

Wherein, Q ⁱ∈ R ^{k × K}.For preventing due to Q ⁱbe singular matrix and cannot calculate, must to Q ⁱcarry out regularization,

Q ⁱ＝Q ⁱ+rI (7)

Wherein, r is regularization coefficient; I ∈ R ^{k × K}for unit matrix.

Higher-dimension HRIR sample point dimensionality reduction to the concrete steps of lower dimensional space, comprises by described processor:

Higher-dimension HRIR sample point need meet

$\min \mathrm{\ϵ}\left(Y\right)=\underset{i=1}{\overset{I}{\mathrm{\Σ}}}{|y_i-\underset{j=1}{\overset{K}{\mathrm{\Σ}}}{w}_j^i{y}_j|}^2---\left(8\right)$

Wherein, ε (Y) is loss function, y _ix _ilow-dimensional map vector, y _jy _ja jth Neighbor Points, y _ifollowing condition need be met,

$\underset{i=1}{\overset{I}{\mathrm{\Σ}}}{y}_i=0,\frac{1}{N}\underset{i=1}{\overset{I}{\mathrm{\Σ}}}{y}_i{y}_i^T=I---\left(9\right)$

Use sparse matrix W ∈ R ^{i × I}store work as x _jbelong to x _iduring Neighbor Points, the element in W otherwise, w _ij=0.Loss function is written as

$\min \mathrm{\ϵ}\left(Y\right)=\underset{i=1}{\overset{I}{\mathrm{\Σ}}}\underset{j=1}{\overset{I}{\mathrm{\Σ}}}{m}_{\mathrm{ij}}{y}_i^T{y}_j---\left(10\right)$

Wherein, m _ijit is the element of matrix M.Symmetric matrix M ∈ R ^{i × I}calculate by formula (11),

M＝(I-W) ^T(I-W) (11)

For making formula (10) medial error minimum, then choose the proper vector corresponding to minimum d nonzero eigenvalue of M as output Y.Because minimum eigenwert is close to zero, generally get the individual minimum eigenwert characteristic of correspondence vector of front 2 ~ d+1 as Y.Wherein, d is the embedding dimension, is set to 3 in the present embodiment.

In the present embodiment, can adopt other manifold learnings by higher-dimension HRIR sample point dimensionality reduction to lower dimensional space as Isometric Maps algorithm (Isomap), local retaining projection (Locality PreservingProjection, hereinafter referred to as LPP) etc.

The detailed process of sectional straight line fitting is:

Get initial two points of one-dimensional manifold, use formula

$\left(\begin{array}{c}{y}_{1d}\\ y_{2d}\\ \·\\ \·\\ \·\\ y_{\mathrm{Nd}}\end{array}\right)=\left(\begin{array}{cccc}{y}_{11}& y_{12}& \·\·\·& y_{1,d-1}\\ y_{21}& y_{22}& \·\·\·& y_{2,d-1}\\ \·& \·& & \·\\ \·& \·& & \·\\ \·& \·& & \·\\ y_{N1}& y_{N2}& \·\·\·& y_{N,d-1}\end{array}\right)\left(\begin{array}{c}{a}_0\\ a_1\\ \·\\ \·\\ \·\\ a_{d-1}\end{array}\right)---\left(12\right)$

Least square fitting is utilized to obtain straight line coefficient (a ₀, a ₁..., a _d-1);

According to the bearing of trend of one-dimensional manifold, often add a bit, use formula (12) and formula

$e_N=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\left[{\left|\right|y_{\mathrm{id}}-{\hat{y}}_{\mathrm{id}}\left|\right|}_2\right]}^2---\left(13\right)$

Calculate (a ₀, a ₁..., a _d-1) and e _n, by e _nwith maximum error e given in advance _mAXrelatively, if e _n<e _mAX, then this point is joined in point range above, until e _n>=e _mAX, using the end points of N-1 point as this section of straight line.For the error of fitting threshold value e in sectional straight line fitting _mAX, adopt formula

e _MAX＝ENER×ratio (14)

Calculate, wherein ENER is the normalized energy that low-dimensional embeds stream shape, shown in (15); Ratio is error of fitting ratio, manually need set, is set to 5 × 10 in the present embodiment ^-10.Wherein,

$\mathrm{ENER}=\underset{i=1}{\overset{I}{\mathrm{\&Sigma;}}}{\left[{\left|\right|h_i\left|\right|}_2\right]}^2/\max \left|h_i\right|---\left(15\right)$

Using N point and N+1 point as initial 2 points of next section of fitting a straight line, go to step 1., until all sample point matching is complete.

After obtaining the end points of all segmented linear, using the higher-dimension HRIR corresponding to these cut-points as feature HRIR.

This programme feature HRIR collection chooses result as shown in Figure 5.In the present embodiment, feature samples clicks delivery block can adopt other subsection curve drafting, as secondary, trilinear matching.

In order to verify validity of the present invention, carry out computer simulation experiment.The database adopted during experiment is the left ear HRIR data measuring individual subject_003 in the HRIR database in CIPIC laboratory, and three-dimensional HRIR matrix size is D1 × D2 × L, and numerical value is 50 × 25 × 200.

Adopt subjective listen test:

Test HRIR collection: choose the HRIR collection data that two groups are respectively the subject_003 of CIPIC, and compression reconfiguration data set.

Test subject: 4 normal testers of the sense of hearing: 2 women, 2 male sex;

Test angle (φ, θ): 7 orientation are tested, be respectively (0 °,-35 °), (0 °, 0 °), (0 °, 35 °), (-22.5 °, 0 °), (33.8 °, 0 °), (61.9 °, 0 °), (90 °, 0 °).There are front and back owing to using general HRIR and obscure this possibility, these 7 orientation therefore chosen all are positioned at front half-plane, and head relative coordinate system as shown in Figure 6.

Testing audio: be two groups of HRIR convolution (rebuilding HRIR after original HRIR, compression) respectively that 0.4 second white noise of 44.1kHz is identical with orientation by sampling rate, obtain the binaural sound of 0.4 second.Repeated 10 times, do interval with 0.1 second quiet, obtain the 5 seconds testing audios in two groups, this orientation.Perception angular distance option table is as shown in table 1,

Table 1

Difference angle	Describe	Affinity score
			0°	Identical	5
0°～5°	Just perceptible difference	3
			5°～15°	Significant difference	2
15°～30°	Very different	1
			>30°	Complete difference	0

Two groups, each orientation testing audio is played to 4 testers respectively, carries out angle difference between audition alternative two groups.Then search reserved portion according to difference angle in table 2.The score in 7 orientation is done on average, obtains the listening test results shown in table 3.Listening test results as known from Table 2, three kinds of scheme HRIR after overcompression, reconstruct distinguish with original HRIR in audition.

Table 2

Scenario Name	YXH	LZL	SJW	LD
					This programme	5.00	4.71	5.00	5.00

Calculated amount: when synthesizing binaural sound, the HRIR in required orientation is divided into three kinds of situations:

The first situation, when required HRIR belongs to feature HRIR collection, can directly use, and calculated amount is 0;

The second situation, when the both sides of required HRIR are feature HRIR, as horizontal angle index in Fig. 5 be 1, elevation angle index be 2 time, in elevation plane, the HRIR of these both sides, position belongs to feature HRIR collection, the HRIR of its both sides is used to carry out linear interpolation, reconstruct this orientation and need 200 real multiplications, 200 real additions.

The third situation, when required HRIR both sides not existing characteristics HRIR time, if horizontal angle index in Fig. 5 is 5, when elevation angle index is 2, the HRIR of this position needs by the feature HRIR linear interpolation of 4 around it, and now reconstructing this orientation HRIR needs 200 real multiplications, 600 real additions.

Comprehensive three kinds of situations, that average computational load of the method is 133.4 real multiplications, 202.6 real additions.

The present embodiment, head coherent pulse response HRIR data set divides into groups according to the elevation angle and/or horizontal angle by processor, by choosing the HRIR of reconstructed error minimum value in Local Liner Prediction LLE after grouping HRIR as higher-dimension HRIR sample point, by described HRIR sample point dimensionality reduction to lower dimensional space to low-dimensional HRIR, adopt sectional straight line fitting selected characteristic HRIR sample point in described low-dimensional HRIR sample point, according to elevation angle order and/or horizontal angle order, described feature HRIR sample is combined into three-dimensional matrice.Solve the technical matters that reconstructed error is little in HRIR selected characteristic HRIR process.The feature HRIR quantity of carrying out dividing into groups obtained according to the elevation angle and/or horizontal angle is less, thus make feature HRIR have more representativeness, adopt sectional straight line fitting selected characteristic HRIR sample point in described low-dimensional HRIR sample point, selected feature HRIR is more accurate.

Last it is noted that above each embodiment is only in order to illustrate technical scheme of the present invention, be not intended to limit; Although with reference to foregoing embodiments to invention has been detailed description, those of ordinary skill in the art is to be understood that: it still can be modified to the technical scheme described in foregoing embodiments, or carries out equivalent replacement to wherein some or all of technical characteristic; And these amendments or replacement, do not make the essence of appropriate technical solution depart from the scope of various embodiments of the present invention technical scheme.

Claims (4)

1. a head coherent pulse response data sets disposal route, is characterized in that, comprising:

Head coherent pulse response HRIR data set divides into groups according to the elevation angle and/or horizontal angle by processor;

Described processor is by choosing the HRIR of reconstructed error minimum value as higher-dimension HRIR sample point in Local Liner Prediction LLE after grouping HRIR;

Described HRIR sample point dimensionality reduction to lower dimensional space, is obtained low-dimensional HRIR by described processor;

Described processor adopts sectional straight line fitting selected characteristic HRIR sample point in described low-dimensional HRIR sample point;

Described feature HRIR sample is combined into three-dimensional matrice according to elevation angle order and/or horizontal angle order by described processor.

2. method according to claim 1, is characterized in that, described processor, by choosing the HRIR of reconstructed error minimum value in Local Liner Prediction LLE after grouping HRIR as HRIR sample point, comprising:

Processor chooses a HRIR sample point in HRIR after grouping, calculates the Euclidean distance between described HRIR sample point and other HRIR sample points, and chooses the point of proximity of described HRIR sample point according to described Euclidean distance;

Construct the local covariance matrix of described HRIR sample point, determine the conditional value of reconstructed error

Wherein, described K is and HRIR sample point x _ineighbor point number, described j is the subscript of a jth neighbor point, described in for using a jth neighbor point x _jreconstruct x _iweights;

By described local covariance matrix Q ⁱand described conditional value, adopt formula

Calculate the weights reconstruction matrix w of described HRIR sample point _ij, wherein, described K is local covariance matrix Q ⁱdimension, described i is the subscript of i-th sample point, and described k is local covariance matrix Q ⁱkth dimension, described j is the subscript of a jth neighbor point;

By described w _ijsubstitute into reconstructed error function formula

Determine that the HRIR sample point corresponding to reconstructed error minimum value is the higher-dimension HRIR sample point that can carry out dimension-reduction treatment, wherein, described I is that sample is always counted, and described i is the subscript of i-th sample, and described K is setting and described HRIR sample point x _ithe number of the sample point that Euclidean distance is minimum, described x _ifor the described HRIR sample point chosen, described x _jfor described x _ia jth neighbor point, described min ε (W) is minimal reconstruction error amount;

Described HRIR sample point dimensionality reduction to lower dimensional space, is obtained low-dimensional HRIR, comprises by described processor:

Formula will be met

Described HRIR sample point substitute into formula

Obtain loss function

Choose formula

M＝(I-W) ^T(I-W) (7)

Minimum d the nonzero eigenvalue of middle M is updated to described loss function, obtains with the proper vector of Y as the corresponding sample point mapping to lower dimensional space, wherein, and described y _ifor described x _ilow-dimensional map vector, described y _jfor described y _ia jth neighbor point, described min ε (Y) is loss function, and described i is the subscript of i-th sample, and described j is the subscript of a jth neighbor point, and described I is that sample is always counted, and described k is and y _icontiguous counts.

3. method according to claim 1 and 2, is characterized in that, described processor adopts sectional straight line fitting selected characteristic HRIR sample point in the HRIR sample point of described low-dimensional, comprising:

Any two starting points for the one-dimensional manifold of low-dimensional HRIR group adopt formula

Calculated line coefficient (a ₀, a ₁..., a _d-1), wherein, described y is the point of the one-dimensional manifold of low-dimensional HRIR group, and described d is space dimensionality, and described a is straight line coefficient;

Add N point according to the bearing of trend of one-dimensional manifold, adopt formula

Calculate N point error of fitting e _n, wherein, described N counts for described any two starting point institute structures are straight, and described i is that i-th section of straight line is counted, and described y is the point of the one-dimensional manifold of low-dimensional HRIR group, and described d is space dimensionality, described in for the actual point of the one-dimensional manifold of low-dimensional HRIR group;

If the described error of fitting e of N point _nbe less than maximum error e _mAX, then N point is joined in the point range between two starting points; If the described error of fitting e of N point _nbe greater than maximum error e _mAXthen using the end points of N-1 point as this section of straight line, then the higher-dimension HRIR corresponding to described end points is feature HRIR.

4. method according to claim 3, is characterized in that, described e _mAX=ENER × ratio, wherein, described ratio is error of fitting ratio, and described ENER is the normalized energy that low-dimensional embeds stream shape.