Dataset shift compensation in machine learning

1. A method for inter-dataset variability compensation in a speaker recognition system, the method comprising:

operating at least one hardware processor of a speaker recognition system to compensate for inter-dataset variability, by:

receiving a heterogeneous development dataset comprising: multiple samples embodied as digital files, and digital metadata associated with at least some of the multiple samples;

dividing the multiple samples into multiple homogenous subsets, based on the digital metadata;

computing a statistical measure of high-level features of each of the multiple homogenous subsets, to produce multiple central high-level features for the multiple homogenous subsets, respectively;

computing an inter-dataset variability subspace spanned by the multiple central high-level features;

removing the inter-dataset variability subspace from the high-level features of the multiple homogenous subsets, to produce denoised samples;

training the speaker recognition system using the denoised samples; and

performing speaker recognition using the trained speaker recognition system.

2. The method according to claim 1 , wherein the high-level features are selected from the group consisting of:

i-vectors,

GMM (Gaussian Mixture Model) supervectors,

HMM (Hidden Markov Model) supervectors,

d-vectors,

JFA (Joint Factor Analysis) supervectors,

LBP (Local Binary Patterns),

HOG (Histograms of Oriented Gradients), and

EBIF (Early Biologically-Inspired Features).

3. The method according to claim 2 , wherein the speaker recognition system is selected from the group consisting of:

a PLDA (Probabilistic Linear Discriminant Analysis)-based system,

an SVM (Support Vector Machine)-based system,

a neural network-based system,

a NAP (Nuisance Attribute Projection)-based system,

a WCCN (Within-Speaker Covariance Matrix)-based system, and

an LDA (Linear Discriminant Analysis)-based system.

4. The method according to claim 3 , wherein the multiple samples are speech samples.

5. The method according to claim 4 , wherein the heterogeneous development dataset is devoid of speech samples from a target domain of the speaker recognition.

6. The method according to claim 4 , wherein the metadata comprises at least one parameter selected from the group consisting of: speaker gender, spoken language and recordation setting.

7. The method according to claim 4 , wherein the computing of the inter-dataset variability subspace comprises PCA (Principal Component Analysis).

8. A method comprising:

operating at least one hardware processor of a PLDA (Probabilistic Linear Discriminant Analysis) speaker recognition system to compensate for inter-dataset variability, by:

receiving a heterogeneous development dataset comprising multiple speech samples;

dividing the multiple speech samples into multiple homogenous subsets;

for each subset i of the multiple homogenous subsets: (a) estimating PLDA hyper-parameters {μ i , B i , W i }, wherein μ denotes a center of an i-vector space, B denotes a between-speaker covariance matrix and W denotes a within-speaker covariance matrix, and (b) computing an i-vector subspace S μ corresponding to {μ i }, an i-vector subspace S W corresponding to {W i }, and an i-vector subspace S B corresponding to {B i };

joining i-vector subspaces S μ , S W and S B into a single subspace S;

removing subspace S from i-vectors of the multiple speech samples, to produce denoised speech samples;

training the PLDA speaker recognition system using the denoised speech samples; and

performing speaker recognition using the trained PLDA speaker recognition system.

9. The method according to claim 8 , further comprising smoothing B by linear interpolation using an estimated diagonal of B.

10. The method according to claim 8 , wherein:

the heterogeneous development dataset further comprises metadata associated with at least some of the multiple speech samples; and

the dividing is based on the metadata.

11. The method according to claim 8 , wherein the computing of each of the i-vector subspaces S μ , S W and S B comprises PCA (Principal Component Analysis).

12. The method according to claim 8 , further comprising computing an average of squared {W i }, and finding a k number of largest eigenvalues of the squared {W i }, wherein the k largest eigenvalues span the i-vector subspace S W .

13. The method according to claim 12 , further comprising whitening the i-vector subspace S W with respect to W.

14. The method according to claim 8 , further comprising computing an average of squared {B i }, and finding an m number of largest eigenvalues of the squared {B i }, wherein the k largest eigenvalues span the i-vector subspace S B .

15. The method according to claim 14 , further comprising whitening the i-vector subspace S B with respect to B.

16. A computer program product for inter-dataset variability compensation for speaker recognition, the computer program product comprising a non-transitory computer-readable storage medium having program code embodied therewith, the program code executable by at least one hardware processor of a PLDA (Probabilistic Linear Discriminant Analysis) speaker recognition system to:

compensate for inter-dataset variability in the PLDA speaker recognition system, by:

receiving a heterogeneous development dataset comprising multiple speech samples;

dividing the multiple speech samples into multiple homogenous subsets;

for each subset i of the multiple homogenous subsets: (a) estimating PLDA (Probabilistic Linear Discriminant Analysis) hyper-parameters {μ i , B i , W i }, wherein μ denotes a center of an i-vector space, B denotes a between-speaker covariance matrix and W denotes a within-speaker covariance matrix, and (b) computing an i-vector subspace S μ corresponding to {μ i }, an i-vector subspace S W corresponding to {W i }, and an i-vector subspace S B corresponding to {B i };

joining i-vector subspaces S μ , S W and S B into a single subspace S;

removing subspace S from i-vectors of the multiple speech samples, to produce denoised speech samples;

training the PLDA speaker recognition system using the denoised speech samples; and

performing speaker recognition using the trained PLDA speaker recognition system.

17. The computer program product according to claim 16 , wherein the program code is further executable by the at least one hardware processor to smooth B by linear interpolation using an estimated diagonal of B.

18. The computer program product according to claim 16 , wherein:

the heterogeneous development dataset further comprises metadata associated with at least some of the multiple speech samples; and

the dividing is based on the metadata.

19. The computer program product according to claim 16 , wherein the computing of each of the i-vector subspaces S μ , S W and S B comprises PCA (Principal Component Analysis).

20. The computer program product according to claim 16 , wherein the program code is further executable by the at least one hardware processor to:

compute an average of squared {W i };

find a k number of largest eigenvalues of the squared {W i }, wherein the k largest eigenvalues span the i-vector subspace S W ;

whiten the i-vector subspace S W with respect to W;

compute an average of squared {B i };

find an m number of largest eigenvalues of the squared {B i }, wherein the m largest eigenvalues span the i-vector subspace S b ; and

whiten the i-vector subspace S B with respect to B.