Re: [Kaldi-developers] Initializing DiagGmm models

[Xavier A. <xan...@gm...>]

Hi,
Thanks for your answers,
I need to take a closer look at sgmmbin/init-ubm-flat.cc. I did look
at the other files you propose and did not see exactly what I meant. I
see that gmmbin/gmm-init-model-flat.c is able to compute a single
Gaussian GMM given all data and that gmmbin/gmm-global-acc-stats and
gmmbin/gmm-global-est can retrain the model (i.e. EM reestimation),
but I believe I saw that those only allow me to grow the model by
splitting the N Gaussians with highest weight. This is not the same as
splitting uniformly all Gaussians (regardless of weights) or
performing Kmeans+splitting of all Gaussians.
I will try using the functions you propose and if the results are not
good enough I will try to implement the splitting myself (After
looking at it more I see that it should not be too difficult anyway).

Thanks

Xavier Anguera

On Mon, Aug 20, 2012 at 10:53 AM, Arnab Ghoshal <ar...@gm...> wrote:
> Or you could create something like sgmmbin/init-ubm-flat.cc with the
> option for creating both diag and full GMMs (you can look at
> gmmbin/gmm-init-model-flat.cc for how an HMM/GMM is initialized). You
> can then train it with gmmbin/gmm-global-acc-stats and
> gmmbin/gmm-global-est in the usual fashion.
>
> On Mon, Aug 20, 2012 at 9:47 AM, Daniel Povey <dp...@gm...> wrote:
>> That's interesting.
>>
>> In the past I've done this type of thing by clustering Gaussians of a
>> speech-reco system, but if you want to start from scratch, you could write a
>> program called gmm-global-init that would initialize a model, say  with a
>> single Gaussian, and train it, maybe on a small amount of data at first, and
>> keep mixing up.  The program gmm-global-est has an option to mix up the
>> #Gaussians.
>>
>> Dan
>>
>> On Sun, Aug 19, 2012 at 12:10 PM, Xavier Anguera <xan...@gm...> wrote:
>>>
>>> Hi again,
>>> I am now trying to train some models given my extracted features. I
>>> want to use these models for speaker-ID experiments. For this reason I
>>> was looking for some simple method to initialize the models given some
>>> training data (something like std-perturbed Gaussian splitting, or
>>> K-means equivalents) but I do not find anything straightforward in the
>>> main code or the examples.
>>> Does anyone have an example I can work with, or any suggestion on how
>>> to implement it?
>>>
>>> Thanks
>>>
>>> Xavier Anguera
>>>
>>>
>>> ------------------------------------------------------------------------------
>>> Live Security Virtual Conference
>>> Exclusive live event will cover all the ways today's security and
>>> threat landscape has changed and how IT managers can respond. Discussions
>>> will include endpoint security, mobile security and the latest in malware
>>> threats. http://www.accelacomm.com/jaw/sfrnl04242012/114/50122263/
>>> _______________________________________________
>>> Kaldi-developers mailing list
>>> Kal...@li...
>>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>>
>>
>>
>> ------------------------------------------------------------------------------
>> Live Security Virtual Conference
>> Exclusive live event will cover all the ways today's security and
>> threat landscape has changed and how IT managers can respond. Discussions
>> will include endpoint security, mobile security and the latest in malware
>> threats. http://www.accelacomm.com/jaw/sfrnl04242012/114/50122263/
>> _______________________________________________
>> Kaldi-developers mailing list
>> Kal...@li...
>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>>