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

[hts-users:02644] Re: Question about the unseen models


model searching procedure is done in HTS engine.


2010/12/8 那兴宇 <nxy-yzqs@xxxxxxx>


Thanks. Do you mean the similar work has been done in the model conversion convertion procedure?

If so, the 2nd question still confuses:
2. What happens when a "real unseen label" is generated? I mean the label which has never occured in the training data or the test data.

When 'make unseen', the unseen model items are from test labels pre-defined in the training stage.
Then when we test a label that is not included in the  pre-defined test labels, the OOV problem still exist, especially when the training data is limited.

>At 2010-12-07 14:34:47,"Wu ZZ" <nkwuzz@xxxxxxxxx> wrote:
>hts-engine format models are for HTS-engine. similar work as 'make unseen' is done is HTS-engine.



>2010/12/7 那兴宇 <nxy-yzqs@xxxxxxx>
>Hi, everyone

>There are two questions about the unseen model issue:
>1. In the demo script, there is not make_unseen_model step before the convertion to hts-engine format models, why?
>2. What happens when a "real unseen label" is generated? I mean the label which has never occured in the training >data or the test data.

>I tested the "real unseen label" and got really wierd synthesized speeches.

>It has confused me for a while, please help!
>Thanks in advance.

Na Xingyu
Beijing Institute of Technology, China










References
[hts-users:02641] Question about the unseen models, 那兴宇
[hts-users:02642] Re: Question about the unseen models, Wu ZZ
[hts-users:02643] Re: Question about the unseen models, 那兴宇