Re: [Kaldi-users] kaldi lattice

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

They can be arbitrarily off in principle, but it just won't matter for
any purpose that we use the lattice.
The lattice-boost-ali.cc looks at the frame-level phone identities (I
think), so all that matters is whether those frame-level things are
aligned between the num and den lats, which they are--it's just that
they are not aligned with the words, which that program anyway
ignores.
Dan


On Thu, Oct 24, 2013 at 11:22 AM, Hao Tang <lar...@gm...> wrote:
> Thanks for the clarification. I am sorry it's getting a little bit off
> topic. I just want to confirm it is safe to assume that the times
> before aligning are not too off in the context of rescoring a dense
> lattice.
>
> I've looked at the MMI and MPE script. latbin/lattice-boost-ali.cc
> boosts the not time-aligned denominator with the time-aligned
> numerator, so I guess the answer to the above question is yes?
>
> Hao
>
> On Wed, Oct 23, 2013 at 8:13 PM, Daniel Povey <dp...@gm...> wrote:
>> We implement something slightly different called MFPE which does not
>> require the start/end information.  Anyway most recipes use BMMI which
>> also does not require it.
>> Dan
>>
>>
>> On Wed, Oct 23, 2013 at 8:58 PM, Hao Tang <lar...@gm...> wrote:
>>> I am confused. Maybe I misunderstood something in the following paper.
>>>
>>> https://wiki.inf.ed.ac.uk/twiki/pub/CSTR/ListenSemester1_2007_8/povey_mpe.pdf
>>>
>>> Won't you need to know the start and end of frame indices for each arc
>>> in the lattice in order to run forward-backward on each phone arc?
>>>
>>> Hao
>>>
>>> On Wed, Oct 23, 2013 at 7:32 PM, Daniel Povey <dp...@gm...> wrote:
>>>> BTW, the time information does not matter at all for MMI training.
>>>>
>>>> Dan
>>>>
>>>>
>>>> On Wed, Oct 23, 2013 at 8:30 PM, Daniel Povey <dp...@gm...> wrote:
>>>>> I just updated that paragraph to the below.
>>>>>
>>>>>  If you want the time information of the lattices explicitly (rather than having
>>>>>  to count transition-ids), there is a function LatticeStateTimes (for Lattice),
>>>>>  and CompactLatticeStateTimes (for CompactLattice), which will give you the time
>>>>>  where each state is located (a number from 0 to the number of frames in the
>>>>>  file).  Be careful that in general, the words are not synchronized with to the
>>>>>  transition-ids, meaning that the transition-ids on an arc won't necessarily all
>>>>>  belong to the word whose label is on that arc.  This means that the times you
>>>>>  get from the lattice will (as far as the word labels are concerned) be inexact.
>>>>>  The same is also true of the weights; these are also not synchronized with
>>>>>  either the words or the transition-ids on a particular arc.  If you want exact
>>>>>  times (e.g. for conversion to HTK lattices, or for sclite scoring), then you
>>>>>  should run the program lattice-align-words.  This program only works if you
>>>>>  built your system with word-position-dependent phones, and it requires certain
>>>>>  command-line options to tell it which phones are in which position in the word.
>>>>>  See egs/wsj/s3/run.sh for an example (search for align).  There is an
>>>>>  alternative program, lattice-align-words-lexicon, that you can use if your
>>>>>  system does not have word-position-dependent phones.
>>>>>
>>>>> Dan
>>>>>
>>>>>
>>>>>
>>>>>>
>>>>>> I'm trying to understand the last paragraph in the lattice page.
>>>>>>
>>>>>> http://kaldi.sourceforge.net/lattices.html
>>>>>>
>>>>>> What does it mean that the labels are pushed relative to the
>>>>>> transition-ids and why does Kaldi choose to do this?
>>>>>>
>>>>>> In addition, what is latbin/lattice-align-words.cc doing exactly?
>>>>>> After aligning, I see plenty of silence phones appear in the lattice,
>>>>>> and the weights seem to be pushed. I assume the weights after aligning
>>>>>> are not meaningful for individual edges anymore and are only
>>>>>> meaningful for the whole path?
>>>>>>
>>>>>> Thanks,
>>>>>> Hao
>>>>>>
>>>>>> ------------------------------------------------------------------------------
>>>>>> October Webinars: Code for Performance
>>>>>> Free Intel webinars can help you accelerate application performance.
>>>>>> Explore tips for MPI, OpenMP, advanced profiling, and more. Get the most from
>>>>>> the latest Intel processors and coprocessors. See abstracts and register >
>>>>>> http://pubads.g.doubleclick.net/gampad/clk?id=60135991&iu=/4140/ostg.clktrk
>>>>>> _______________________________________________
>>>>>> Kaldi-users mailing list
>>>>>> Kal...@li...
>>>>>> https://lists.sourceforge.net/lists/listinfo/kaldi-users