[mary-users] TOBI ignored in US English HSMM voices?
Joe Crumpton
crumpton at cse.msstate.edu
Fri Jul 12 18:08:51 CEST 2013
Marcela,
Thank you for the information. I looked at the training text for the CMU voices and I think some of the sentences are questions, they just don't have a question mark at the end.
Examples:
( arctic_a0500 "Who the devil gave it to you to be judge and jury." )
( arctic_a0071 "What was the object of your little sensation." )
( arctic_a0042 "How could he explain his possession of the sketch." )
I am planning to change the punctuation for the obvious questions and train a new voice. I will let you know how it works.
Thank you,
Joe Crumpton
>>> On 7/11/2013 at 6:32 AM, Marcela Charfuelan <Marcela.Charfuelan at dfki.de>
wrote:
> Dear Joe Crumpton,
>
> The commented lines in the config file:
>
> # voice.cmu-slt-hsmm.F0.data not used for hmm models
>
> are specific for unit selection voices, therefore not used in hmm voices.
>
> Regarding other parameters in the voice building process to make pitch
> changes more pronounced, as far as I remember we used all the context
> features available in MARY to train the models, but it is important to
> remember that the text processing modules for English are not the same
> as for German, so I guess the components for German are better, but I am
> not so sure about the details here, maybe Ingmar can help/explain a bit
> here.
>
> The data is also different, in the bits data there are several examples
> of sentences that have question marks and exclamations, in the cmu-slt
> there are not; so for the cmu-slt there are no context models trained
> with ? or !, so the generation process uses the ones more close to that.
>
> Regards,
> Marcela.
More information about the Mary-users
mailing list