[Subject Prev][Subject Next][Thread Prev][Thread Next][Date Index][Thread Index]

[hts-users:03933] Re: Full Context TTS Phones


If you do not have any context features in synthesize time, you'd
better also remove them in train time. You can just alter your
question set file and remove the related questions, no alteration to
the labels is needed. Then you can use partial label to synthesize.

The question set only does pattern match, so as long as the pattern
matches, whether you've got all the features does not matter. However,
if you trained the system with some feature which cannot be provided
in synthesis time, it may cause trouble when that feature is used (no
pattern matches, the "false" route is always chosen).

On Sun, Nov 17, 2013 at 4:59 PM, Ibrahim Sobh <im_sobh@xxxxxxxxxxx> wrote:
> Hi
>
> In training phase I use full context phones such as:
>
>
> S^i-f+i=r@1_3/A:1_0_2/B:0-0-3@2-1&11-3#5-2$1-2!2-2;11-2|i/C:1+1+3/D:content_4/E:content+2@4+2&4+2#1+1/F:content_2/G:0_0/H:13=5@1=1|NONE/I:0_0/J:13+5-1
>
>
> However in synthesizing phase, I have only the phone context information
> (p1p2p3p4p5p) and I do not have full information about other context factors
> (syllable, stress, pos, ...  ).
>
> In synthesis; I tried to use only part of the phone :
>
> i^f-i+r=l@2_2
>
> It worked!
>
> At synthesize time,   can we use parts of the full context phones if we do
> not have full information (or not sure of the context information)? any
> parts of the phone are optional?
>
> Or is it better to guess other context factors?!!
>
>
> Best Regards

Follow-Ups
[hts-users:03934] Re: Full Context TTS Phones, Ibrahim Sobh
[hts-users:03936] MDL vs ML, Tóth Bálint
References
[hts-users:03927] HMM extension (Multi-process supporting) link broken, Karthik Krishnan
[hts-users:03932] Full Context TTS Phones, Ibrahim Sobh