[hts-users:00437] Re: training error

[Alexis Moinet <alexis.moinet@xxxxxxxxxxxxxxx>]

Heiga ZEN (Byung Ha CHUN) wrote:
> Alexis Moinet wrote:
> >  Processing Data: sentence_0016_00.cmp; Label sentence_0016_00.lab
> >  Unable to traverse 530 states in 482 frames
> >  WARNING [-7324]  StepBack: Bad data or over pruning
> >  in /home/user/work/HTS-2.0/htk-3.4/HTKTools/HERest
> "Unable to traverse 530 states in 482 frames" means that the sentence 
> HMM corresponding to sentence_0016_00.lab contains 530 states but 
> sentence_0016_00.cmp consists of 482 frames.
> Therefore, it is unable to perform forward-backward algorithm.
> Please check whether sentence_0016_00.lab (context-dependent) is 
> correct or not.

Well, I understand this, but the lab file seems correct.
There are 63 phonemes in it (including silences), so how does HERest 
chose the number of states to be 530 ?
I guess it depends on the duration model trained in previous steps of 
training (which are "monophone" and "tie transition probabilities" ) ?

The point is : if we use only sentence_0001 until sentence_0020, we have 
no warning[-7324] (--> the sentence_0016 is ok) and we can synthesize 
(very) poor speech
If we use sentence_0001 until sentence_0300, we got warning[-7324] for 
sentence_0016 (and some other sentences as well) and the training 
eventually stops because of further error most probably caused by the 
warning.

so it might be that the HMM models computed during the monophone 
training and used to initialize the fullcontext HMMs cause the "530 
states" problem ?

by the way, we use 22kHz sound files, does it matter ? (since you seem 
to use 16kHz for HTS-demo)

> Hi Alexis,
>
> # How are you doing :-)

# Fine, thanks. What about you ? :-)

Best regards,

Alexis