Re: [Kaldi-developers] hmm topology and gmm per state

[Arif K. <ife...@gm...>]

Hi,

Thanks you very much for your reply. I have already found the scripts where
one can introduce more than two topologies for hmm.

Can we some how control or specify the #Gaussian per state, or How to find
out the #Gaussian per state used in a given model file.

Can you give me some brief explanation of the following option

po.Register("power", &power, "If mixing up, power to allocate Gaussians
to"" states.");

which is found in the following two modules.

1. gmmbin/gmm-est.cc File Reference
http://kaldi.sourceforge.net/gmm-est_8cc.html
2. gmmbin/gmm-mixup.cc File Reference
http://kaldi.sourceforge.net/gmm-mixup_8cc.html

Note:

You sense very right, I have to redo the Aurora 2.0 experiments with Kaldi,
Unfortunately the default implementation of Kaldi gave me greater WER, than
reported in the paper. So that's why I have to come closer to the technique
which they have adopted in Aurora 2.0 experiments with HTK.

Their technique:
Hmm Model: Word based hmm model not phone based.
No of emitting states in HMM:  16 states for each digit, 3 states for
Silence and 1 state for short pause (sp).
# Gaussian per state: From 6 to 16 GM per state for digits, silence and sp.

Any advice or suggestion will be welcomed and acknowledged.

Regards,
Arif


On Fri, May 10, 2013 at 5:57 PM, Daniel Povey <dp...@gm...> wrote:

> To use more than two HMM topology entries you would have to change the
> scripts.
> Kaldi automatically allocates the #Gaussians in each state.
> I sense that you are trying to exactly duplicate an HTK system, which is
> not a very good idea.
> Dan
>
>
>
> On Fri, May 10, 2013 at 11:25 AM, Arif Khan <ife...@gm...> wrote:
>
>> Hi,
>>
>> I have question two question:
>>
>> 1. Can we use more than two hmm topology entries in Kaldi. For
>> example "Silence phones", "Nonsilence phones" and "Short pause" if each one
>> of them has different
>> no of hmm states.
>>
>> 2. How can we define/control the no. of Gaussians per state for different
>> topology entries. For example if I want to use
>> for "Silence phones", "Nonsilence phones" and "Short pause" different no
>> of Gaussians per state during the training.
>>
>> Any help will be highly appreciated.
>>
>> Regards,
>> Arif
>>
>>
>> ------------------------------------------------------------------------------
>> Learn Graph Databases - Download FREE O'Reilly Book
>> "Graph Databases" is the definitive new guide to graph databases and
>> their applications. This 200-page book is written by three acclaimed
>> leaders in the field. The early access version is available now.
>> Download your free book today! http://p.sf.net/sfu/neotech_d2d_may
>> _______________________________________________
>> Kaldi-developers mailing list
>> Kal...@li...
>> https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>>
>>
>