Not able to get results for ivector

[abhipo]

Hi,

I am running the following command.
I am taking around 10 hours of data of 80 speakers. I am taking 20 speakers for development and 20 for evaluation. I am not getting desired desired results. The scores of the evaluation are not able to distinguish them from imposters.
Should i increase the data or dimension of ubm or tv matrix. How can i debug. Actually same amount of data is working fine for ubm gmm algorithm
Thanks

./bin/verify.py -p energy-2gauss -e mfcc20 -a ivec-cosine-voxforge -s ivector --groups {dev,eval} -vvv -d /home/abhipo/miniconda2/envs/
bob_py27/lib/python2.7/site-packages/bob/bio/spear/config/database/my_custom_db.py -r4 >log_ivector_seq

[Amir Mohammadi]

Hi,

Could you also please post the contents of the ivec-cosine-voxforge configuration file?

Could you please also confirm that you have 40 speakers in the training set? Do you mind sharing more information about your database?

Also, you can put your database configuration anywhere. You don't have to put in bob.bio.spear source directory.

Thanks,
Amir

--
-- You received this message because you are subscribed to the Google Groups bob-devel group. To post to this group, send email to bob-...@googlegroups.com. To unsubscribe from this group, send email to bob-devel+...@googlegroups.com. For more options, visit this group at https://groups.google.com/d/forum/bob-devel or directly the project website at http://idiap.github.com/bob/
---
You received this message because you are subscribed to the Google Groups "bob-devel" group.
To unsubscribe from this group and stop receiving emails from it, send an email to bob-devel+...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[abhipo]

Hi

By config file, do you mean "pkgs/name-VERSION-py27_1/lib/python2.7/site-packages/bob/bio/spear/config/algorithm/ivec_cosine_voxforge.py" ?

Its contents are
algorithm = bob.bio.gmm.algorithm.IVector(
    # IVector parameters
    subspace_dimension_of_t = 100,
    update_sigma = True,
    tv_training_iterations = 25,  # Number of EM iterations for the TV training
    # GMM parameters
    number_of_gaussians = 256,
    training_threshold = 0.0,
    use_lda = True,
    use_wccn = True,
    use_plda = False,
    lda_dim = 50,
    plda_dim_F  = 10,
    plda_dim_G = 50,
    plda_training_iterations = 200,
)

The files I am using for data consists male and female speaker speech in telephone and office environments. The total speech is around ten hours. I rechecked the data , the data consists of 59 hours of speech of more than 50 people.

Do you think I should increase the number of gaussian components ?


Best regards,

[Amir Mohammadi]

Hi,

I think the best way would be to try to tune these parameters:

algorithm = bob.bio.gmm.algorithm.IVector(
    # IVector parameters
    subspace_dimension_of_t = 100,
    update_sigma = True,
    tv_training_iterations = 25,  # Number of EM iterations for the TV training
    # GMM parameters
    number_of_gaussians = 256,
    training_threshold = 0.0,
    use_lda = True,
    use_wccn = True,
    use_plda = False,
    lda_dim = 50,
    plda_dim_F  = 10,
    plda_dim_G = 50,
    plda_training_iterations = 200,
)

using your development data if you think you do not have any bugs.

Otherwise, you have look at the intermediate steps and see where it goes wrong.

[Pavel Korshunov]

Hi,

Try to make plda_dim_F value the same as the number of the speakers in the training set (10 is probably too low). What is the number of Gaussians you are using in UBM that you said works fine for you? You can set it to the same number for i-vectors as well in number_of_gaussians parameter, or use pre-trained UBM for iVector instead of re-training it again.

Also, i-vectors usually require quite a lot of data for each speaker (more than GMM or ISV-based algorithms) to lead to a good performance.

In any case, as Amir said, you need to tune the parameters for your data. To do this properly, you need to have a good understanding of how the training algorithms work.

-- You received this message because you are subscribed to the Google Groups bob-devel group. To post to this group, send email to bob-...@googlegroups.com. To unsubscribe from this group, send email to bob-devel+unsubscribe@googlegroups.com. For more options, visit this group at https://groups.google.com/d/forum/bob-devel or directly the project website at http://idiap.github.com/bob/

---
You received this message because you are subscribed to the Google Groups "bob-devel" group.

To unsubscribe from this group and stop receiving emails from it, send an email to bob-devel+unsubscribe@googlegroups.com.

For more options, visit https://groups.google.com/d/optout.

--
-- You received this message because you are subscribed to the Google Groups bob-devel group. To post to this group, send email to bob-...@googlegroups.com. To unsubscribe from this group, send email to bob-devel+unsubscribe@googlegroups.com. For more options, visit this group at https://groups.google.com/d/forum/bob-devel or directly the project website at http://idiap.github.com/bob/

---
You received this message because you are subscribed to the Google Groups "bob-devel" group.

To unsubscribe from this group and stop receiving emails from it, send an email to bob-devel+unsubscribe@googlegroups.com.

For more options, visit https://groups.google.com/d/optout.

--

Dr. Pavel Korshunov

Biometric group

Idiap Research Institute

Rue Marconi 19

CH - 1920 Martigny

Switzerland

Room: 207

Tel:  +41 27 721 77 56

Fax: +41 27 721 77 12