SPEAR GMM-UBM - Determining cluster arrangement

[Neve Stagy‏]

Hello again everyone,

I want to dissect the resulting Gaussians to determine how a model sample's individual feature-vectors are clustered. 

I've read the implementation details, however I'm still confused on several technicalities:

1. Projector.hdf5 appears identical to all the others in /models/USERNAME.hdf5 - why is this? What are the user models meant to show?

2. /projected/USERNAME/SAMPLENAME.hdf5/n contains scalars for each Gaussian. Are these weights? Can I use this to infer which Gaussian each sample "leans" toward?

I'm familiar with GMM theory on how specific phonemes and fricatives tend to be clustered together - is there any systematic way of proving this based on the data produced by SPEAR? 

Best,

-Stefan

[Amir Mohammadi‏]

Hi Stefan,

I'm not sure if I understood all of your questions but regarding your last question about phonemes, I think you'll need an acoustic model from speech recognition to show that. An acoustic model can tell you to which phoneme a sample belongs to and then you can see where does that sample gets clustered to in your speaker verification GMM.
Does that make sense?

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.

[Neve Stagy‏]

So, I basically want to dig deeper into how individual feature vectors are clustered - i.e. if it behaves as predicted (clustering by phonemes/fricatives) or unpredictably. 

I'm not totally sure I need an acoustic model... is there no way to examine how a sample's feature vectors are clustered?

If I need an acoustic model, how could I go about building one? Does SPEAR support any of this?

Best,

-Stefan

[Tiago Freitas Pereira‏]

Hi Stephan

So, regarding your questions

1. The user models are a result of MAP adaption using the UBM (Projector.hdf5) as a prior. Both, user models and UBM persist the same information (means, diagonal covariance and weights). What do you mean by "identical"? Are they really identical?

Please, have a look and play with the adaptation factor (keyword `relevance_factor` in bob.bio.gmm.algorithm.GMM).

2. Those are the GMM Statistics.  Have a look here for more information (https://www.idiap.ch/software/bob/docs/bob/bob.learn.em/master/guide.html#gmm-statistics) and (https://www.idiap.ch/software/bob/docs/bob/bob.learn.em/master/py_api.html#bob.learn.em.GMMStats)

Cheers

--
-- 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.

--

Tiago