PLDA enroll under bob.kaldi

[Chang Pan]

Hi,

I found it really strange that if I enroll one utterance of multiple speakers, then the plda score would be very accurate on speaker identiftication.

However, if I enroll multiple utterances of multiple speakers, the scores would be meaningless. 

I'm wondering if the package only supports enroll one utterance for each speaker. But the comment of plda_enroll() function is:

 feats : numpy.ndarray

        A 2D numpy ndarray object containing iVectors (of a single speaker).

My code is:

     def enroll_helper(self, ivector):

        return bob.kaldi.plda_enroll(ivector, self.plda_parameter[1])

    def enroll_all(self, nProcess=8, save=True):

        pool = Pool(nProcess)

        self.enrolled = pool.map(self.enroll_helper, self.train_data)

        pool.close()

        pool.join()

Could anyone tell what goes wrong here?

Any help or suggestion would be helpful!

[Milos]

Hi Chang,

I guess that current implementation supports only one utterance per speaker. The function plda_enroll() is general, valid for multiple utterance as well, however, the function plda_score() assumes that only one utterance has been used for enrollment.

More specifically, bob.kaldi calls the ivector-plda-scoring (the Kaldi binary) with default parameters. To support variable number of utterances, probably this comment of ivector-plda-scoring is valid:

For training examples, the input is the iVectors averaged over speakers;

a separate archive containing the number of utterances per speaker may be

optionally supplied using the --num-utts option; this affects the PLDA

scoring (if not supplied, it defaults to 1 per speaker).

Best,

Milos

[Sebastien Marcel]

bob.kaldi is using the PLDA implementation from Kaldi I think but since you are in the bob ecosystem you can also use the native PLDA implementation from bob itself

from the bob.learn.em package https://www.idiap.ch/software/bob/docs/bob/bob.learn.em/stable/py_api.html 

https://pythonhosted.org/bob.learn.em/guide.html

we are using it for our face recognition experiments http://pythonhosted.org/bob.bio.face/baselines.html#the-algorithms where we have several examples per class.

similarly it is used for speaker recognition in bob.spear

https://www.idiap.ch/software/bob/docs/bob/bob.bio.spear/master/index.html

On Friday, April 13, 2018 at 6:15:09 AM UTC+2, Chang Pan wrote:

[Amir Mohammadi]

Hi,

Here is the up-to-date version of links that Sebastien sent earlier.
https://www.idiap.ch/software/bob/docs/bob/bob.learn.em/stable/py_api.html
https://www.idiap.ch/software/bob/docs/bob/bob.learn.em/stable/guide.html
https://www.idiap.ch/software/bob/docs/bob/bob.bio.face/stable/baselines.html#the-algorithms
https://www.idiap.ch/software/bob/docs/bob/bob.bio.spear/stable/

Our documentations on pythonhosted.org is old and deprecated.

If you want to dig around and learn about Bob, you can also see this link:
https://www.idiap.ch/software/bob/docs/bob/docs/master/index.html

Best,

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.

[Chang Pan]

Thanks Marcel!

At this moment I would try to fix the bob.kaldi since most of my code is completed.

I'll try bob.learn later!

[Chang Pan]

Thanks Amir!

[Chang Pan]

Hi Milos,

Thank you for your advice. I think you're totally correct! I'm dealing with it.

[Chang Pan]

Even if I save num_utts.ark by plda_enroll() and add '--num-utts=ark:ivector_pkl/num_utts.ark' within plda_score, the score was still really strange :(

Thanks anyway...

[Chang Pan]

Hi Marcel,

I just tried to use bob.learn.em for my code. Even if I only enroll one utterance, the result is strange already.... I'm pretty sure my i-vector are extractor correctly since I've checked them by cosine function.

Could you check my implementation? Any suggestion would be helpful.

Thank you

On Friday, April 13, 2018 at 1:56:23 AM UTC-7, Sebastien Marcel wrote:

[captain...@gmail.com]

Hi Chang,

     I am doing speaker recognition by using bob.kaldi as well. I am a little confused about ivector+PLDA training, Do you use wechat, Is it possible that we can add 

friends on wechat?

    If you are glad to add friends, my wechat ID is : whiskywithrocks

   

    Thank you.