speaker recognition using Voxceleb V1 recipe

[Zhor BENHAFID]

hello, 

i am trying to do a speaker recognition system then i  follow  voxcele v1 recipe  using the same data all the stage up to 5  was done but when i run stage  6

 to get scoring   something goes wrong  with the trails

i get this in the  voxceleb1_test_scoring.log

[David Snyder]

Recently, the VoxCeleb corpus was updated, and the speaker IDs were changed. My guess is that you're using this newer version of the corpus. Updating the recipe to handle both older and newer versions of the corpus is on our list of things to do. 

Fortunately, fixing this on your own is quite straightforward. You have to write a script that converts the speaker IDs in the trials to the new format, or, equivalently, write a script that converts the speaker IDs in the data directory into the old format.

This file will probably be helpful for you: http://www.openslr.org/resources/49/vox1_meta.csv. The first column is the newer speaker ID format, and is probably what you're using. The second column is the old speaker ID format that the trials file expects. 

[David Snyder]

Another option, that might be even easier, is to download the newest version of the trials file from VGG: http://www.robots.ox.ac.uk/~vgg/data/voxceleb/meta/veri_test.txt.

You'll need to write a script that converts that into the Kaldi format. Look at data/voxceleb1_test/trials to see what that should look like. You could even pull out the code from https://github.com/kaldi-asr/kaldi/blob/master/egs/voxceleb/v1/local/make_voxceleb1.pl#L53 to https://github.com/kaldi-asr/kaldi/blob/master/egs/voxceleb/v1/local/make_voxceleb1.pl#L77 and apply it to http://www.robots.ox.ac.uk/~vgg/data/voxceleb/meta/veri_test.txt. That should give you a Kaldi formatted trials file. 

[David Snyder]

Also, since you're using a version of VoxCeleb1 which is not yet supported by the Kaldi recipe, there may be some unexpected behavior when preparing the corpus and dividing it into a train and evaluation set using the older scripts. Be careful that you didn't train your system on the evaluation set!

I suggest fully understanding how to prepare the corpus yourself, and ensure that data is getting split properly into train and eval sets. You may need to retrain this system if the split was not performed correctly. 

[Daniel Povey]

I wonder whether we could cache a version-numbered form of VoxCeleb on OpenSLR, since they don't seem to be doing that themselves?  Is it against the terms?
Dan

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
To post to this group, send email to kaldi...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/b6aef8a5-c6ce-43f5-87c9-0a9dcf3ac278%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[entn-at]

In principle it should be allowed, as the data is published under CC BY-SA 4.0, which allows redistribution: http://www.robots.ox.ac.uk/~vgg/data/voxceleb/files/license.txt