Training/evaluating single speaker?
75 views
Skip to first unread message
Michal Kolodziejczyk
Sep 25, 2014, 7:33:15 AM9/25/14
to bob-...@googlegroups.com
Hello,
given trained background model, what is the simplest way to train additional single speaker, and then to verify this single speaker with another utterance? Any tools to use?
given trained background model, what is the simplest way to train additional single speaker, and then to verify this single speaker with another utterance? Any tools to use?
Laurent El Shafey
Sep 25, 2014, 7:43:36 AM9/25/14
to bob-...@googlegroups.com
Hello,
One possibility is to use the xbob.db.verification.filelist package.
Spear runs a complete evaluation on a specific dataset.
This xbob.db.verification.filelist package allows you to use your own customized lists for training, speaker enrollment and probing.
Inside the spear package, you have examples on how to define such lists, e.g.:
https://github.com/bioidiap/spear/tree/master/protocols/timit/2
In your case, you will have to:
1. Put the trained background model at the location spear is supposed to generate it. This way it won't be regenerated and only used for the next processing steps.
2. Define your customized lists: in the 'dev' (cf. https://github.com/bioidiap/spear/tree/master/protocols/timit/2/dev) directory,
2.a. create the 'for_models.lst' with the file(s) you want to use for speaker enrollment and
2.b. create the 'for_scores.lst' with the file(s) you want to use for probing the speaker model(s)
3. Create a configuration file for this database (e.g.: https://github.com/bioidiap/spear/blob/master/config/database/timit.py)
4. Run spear using this database configuration file and selecting the output directories, such as the pre-trained background model can be successfully read.
Cheers,
Laurent
One possibility is to use the xbob.db.verification.filelist package.
Spear runs a complete evaluation on a specific dataset.
This xbob.db.verification.filelist package allows you to use your own customized lists for training, speaker enrollment and probing.
Inside the spear package, you have examples on how to define such lists, e.g.:
https://github.com/bioidiap/spear/tree/master/protocols/timit/2
In your case, you will have to:
1. Put the trained background model at the location spear is supposed to generate it. This way it won't be regenerated and only used for the next processing steps.
2. Define your customized lists: in the 'dev' (cf. https://github.com/bioidiap/spear/tree/master/protocols/timit/2/dev) directory,
2.a. create the 'for_models.lst' with the file(s) you want to use for speaker enrollment and
2.b. create the 'for_scores.lst' with the file(s) you want to use for probing the speaker model(s)
3. Create a configuration file for this database (e.g.: https://github.com/bioidiap/spear/blob/master/config/database/timit.py)
4. Run spear using this database configuration file and selecting the output directories, such as the pre-trained background model can be successfully read.
Cheers,
Laurent
On 25/09/14 13:33, Michal Kolodziejczyk wrote:
Hello,
given trained background model, what is the simplest way to train additional single speaker, and then to verify this single speaker with another utterance? Any tools to use?
--
-- 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.
-- Dr. Laurent El Shafey Idiap Research Institute Centre du Parc - Rue Marconi 19 CH-1920 Martigny, Suisse Phone: +41 27 721 77 95 Fax: +41 27 721 77 12 http://www.idiap.ch/~lelshafey
Reply all
Reply to author
Forward
0 new messages