[hts-users:02495] Re: using of HVite

[John Dines <dines@xxxxxxxx>]

Are you are trying to align with standard HVite using HSMM from HTS?

You will find the 'dummy' transition matrix in the models will not 
usually permit meaningful alignment - you first need to convert HSMM to 
HMM (search previous HTS posts for this - Simon King posted a 
solution).  Forced alignment of HSMM (including duration distribution) 
with HVite is not supported in 2.1.

If this is not the problem, maybe you have some inf/nan in your feature 
files?

John

Sébastien Le Maguer wrote:
> Hi,
>
> I've got a problem with HVite. It always returns "No tokens survived 
> to final node of network". I would like to find how monophones are 
> aligned.
>
> I used this command :
>
> ./../tools/bin/HVite -A -B -C configs/trn.cnf -D -y lab -T 1  -o SW -f 
> -a -H models/qst001/ver1/cmp/monophone.mmf -I data/labels/mono.mlf -i 
> test/gen_align.mlf -S data/scp/train.scp data/dict/mono.dict 
> data/lists/mono.list
>
> mono.dict contains the same things that mono.list and all other files 
> are generated from the HTS 2.1 demo (For the moment, I'm using HTS 2.1).
>
> Here are samples of the messages I have:
>
> Aligning File: <...>/demo/base/data/cmp/cmu_us_arctic_slt_b0538.cmp
> Created lattice with 31 nodes / 30 arcs from label file
> No tokens survived to final node of network
> Aligning File: <...>/demo/base/data/cmp/cmu_us_arctic_slt_b0539.cmp
> Created lattice with 32 nodes / 31 arcs from label file
> No tokens survived to final node of network
>
> I think I'm misunderstanding something.
>
> Thanks
> Sebastien