Re: [Kaldi-developers] GMM based model statistics

[Daniel P. <dp...@gm...>]

The number of leaves should decrease as you decrease the data, but
less than proportionally (e.g. 1/10 the data -> 1/4 the number of
leaves).
The number of Gaussians per leaf, which is the ratio of tot-num-gauss
to num-leaves, should decrease as you decrease the data, e.g. 1/10 the
data -> maybe 1/2 the average number of Gaussians per leaf.  The total
number of Gaussians should decrease in a
slightly-less-than-proportional way as you decrease the amount of
data.

Dan


On Tue, Oct 29, 2013 at 12:04 PM, Arif Khan <ife...@gm...> wrote:
> Thanks Daniel.
>
> Do we have any fixed rule, or it is just trying different combinations and
> see the results.
> Are the No. of leaves and No. of Gaussian dependent on each other OR they
> are independent means one can change either of them in any proportions.
>
>
>
> Thank you very much.
>
> Best regards,
> Arif
>
>
> On Tue, Oct 29, 2013 at 4:39 PM, Daniel Povey <dp...@gm...> wrote:
>>
>> If you have only 1 hour of data, you will have to modify the arguments
>> to the triphone-training script, i.e. the number of tree leaves and
>> number of Gaussians should be reduced.
>> Dan
>>
>>
>> On Tue, Oct 29, 2013 at 11:37 AM, Arif Khan <ife...@gm...> wrote:
>> > Thanks Daniel for you quick answer.
>> >
>> > I have another question. I train a model using the WSJ S1 and S5 recipe,
>> > and
>> > got good results on monophone system than triphone. I have about 1000
>> > utterances, with length of ~1 hr, and about 300 vocabulary size.
>> >
>> > In theory,  the triphone system should perform better than monophone. So
>> > I
>> > don't know if something is wrong in the tree construction or any thing
>> > else
>> > could be fixed.
>> >
>> > Best regards,
>> > Arif
>> >
>> >
>> > On Tue, Oct 29, 2013 at 4:17 PM, Daniel Povey <dp...@gm...> wrote:
>> >>
>> >> Sorry, there is no code per se to get those kind of stats, but you
>> >> could perhaps convert the alignments into phone sequences and get the
>> >> stats by hand (see ali-to-phones).
>> >>
>> >> Regarding the tree construction process, there is probably some
>> >> documentation on kaldi.sf.net; if there is a particular aspect of that
>> >> that is unclear, please let us know; but  otherwise, I doubt anyone
>> >> has time to respond to your question right now.
>> >>
>> >> Dan
>> >>
>> >>
>> >> On Tue, Oct 29, 2013 at 10:42 AM, Arif Khan <ife...@gm...>
>> >> wrote:
>> >> > Dear Kaldi authors,
>> >> >
>> >> > I want to do some analysis of the training data. Basically I want to
>> >> > find
>> >> > out the number of phones (mono phone and tri phone) that appeared in
>> >> > the
>> >> > training set. (relative frequency). Is there any module/script
>> >> > available
>> >> > for
>> >> > it.
>> >> >
>> >> > Also, I want to find out the tree construction process. I know the
>> >> > basics
>> >> > from wsj/s5 recipe. But, some more details will be helpful.
>> >> >
>> >> >
>> >> > Best regards,
>> >> > Arif
>> >> >
>> >> >
>> >> >
>> >> > ------------------------------------------------------------------------------
>> >> > Android is increasing in popularity, but the open development
>> >> > platform
>> >> > that
>> >> > developers love is also attractive to malware creators. Download this
>> >> > white
>> >> > paper to learn more about secure code signing practices that can help
>> >> > keep
>> >> > Android apps secure.
>> >> >
>> >> >
>> >> > http://pubads.g.doubleclick.net/gampad/clk?id=65839951&iu=/4140/ostg.clktrk
>> >> > _______________________________________________
>> >> > Kaldi-developers mailing list
>> >> > Kal...@li...
>> >> > https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>> >> >
>> >
>> >
>
>