How to create model using newly arrived speaker data and compute its score with previous models

[Quazi Marufur Rahman]

Hi,

I have successfully executed ./bin/verify.py and EER for DEV is 1.89%. (I am not using EVAL). Now, I have collected some audio

from a speaker (spk1, spk2, spk3 and spk4).wav and want to 

1. create a model for that particular speaker using (spk1, spk2 and spk3).wav 
   
2. use spk4.wav to calculate similarity with all generated models.

I tried to use --execute-only flag, keeping spk1-3 in for_models and spk4 in for_probes to achieve this but the sequence of instructions (preprocessing, extraction, projection, enrollment, score-computation) are not clear to me. In a short, how can I use this system using newly arrived speaker and session?

Thanks in advance.

[Laurent El Shafey]

Hi,

You can define your own custom evaluation protocol using filelists with this package:
https://github.com/bioidiap/bob.db.verification.filelist

VoxForge itself is built on top of this package. Just look at the way this is done here:

https://github.com/bioidiap/bob.db.voxforge/tree/master/bob/db/voxforge/lists

Best,

Laurent

--
-- 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.

[Quazi Marufur Rahman]

Thanks for the reply. I

've placed files in appropriate location.

Is there any way to use --execute-only flag for a particular user?

Example: enrollment and compute-score for a single user.

You received this message because you are subscribed to a topic in the Google Groups "bob-devel" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/bob-devel/kxwJU9WEplc/unsubscribe.
To unsubscribe from this group and all its topics, send an email to bob-devel+...@googlegroups.com.

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

--

Quazi Marufur Rahman

Department of Computer Science and Engineering

University of Dhaka

Bangladesh

[Laurent El Shafey]

As far as I remember, --execute-only only refers to a particular task / step and not to a particular user.

But nothing prevents you to define a specific directory dev/ (or eval/) in your custom filelist database with a single user to enrol and scores against this specific user model.

[Manuel Günther]

As Laurent pointed out, the --execute-only flag executed only parts of the evaluation tool chain, but for all images of the database.
The scripts that you want to use are:
./bin/preprocess.py
./bin/extract.py
./bin/enroll.py
./bin/score.py

These scripts were generated to preprocess one file, extract features from one file, enroll a model from one (or more) files, and finally, compute the score between one or more models and one or more probes. It seems that these tools haven't yet been documented, but the --help option on the scripts should show some more details on how to use them.

Enjoy!
Manuel

[Quazi Marufur Rahman]

Thanks for the response. I've figured it out already.
One more thing, Is it possible to increase the accuracy of the ubm-gmm based system? I want to keep EER under 1%. Could you please suggest me a starting point?

Thanks in advance.

--

[Manuel Günther]

The easiest way to keep you EER below 1% is to throw out people from our database that are not recognized correctly ;-)

More seriously, you can try to optimize the parameters of your experiment (such as the number of Gaussians in your GMM). One easy way of trying out (combinations of) different parameters is to use the ./bin/grid_search.py script, which is documented here:
http://pythonhosted.org/bob.bio.base/more.html#finding-the-optimal-configuration

Let me know your experience with that script. It is rather new and I am not 100% sure that it will work correctly.

Enjoy!
Manuel

[André Anjos]

On Tue, Dec 1, 2015 at 6:03 PM, Manuel Günther <siebe...@googlemail.com> wrote:

The easiest way to keep you EER below 1% is to throw out people from our database that are not recognized correctly ;-)

There is always some light emanating from this mailing list! ;-) A

--

Dr. André Anjos
Idiap Research Institute
Centre du Parc - rue Marconi 19
CH-1920 Martigny, Suisse
Phone: +41 27 721 7763
Fax: +41 27 721 7712
http://andreanjos.org

[Quazi Marufur Rahman]

Could you please emanate some more light? :) 

I need the name of the paper/algorithm on which gmm-voxforge is working on.

--

-- 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 a topic in the Google Groups "bob-devel" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/bob-devel/kxwJU9WEplc/unsubscribe.
To unsubscribe from this group and all its topics, send an email to bob-devel+...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.