Quality of the ctc decoder limit inference words to known words only

[JRMeyer]

>>> elpimous_robot
[December 1, 2017, 8:22pm]

Hello all,

I'm using a homemade model, for french language, files are ok.

My mic array records in continuous, and inference is done when a vader
function cut.

le a eaefanke eethe slash
It doesn't correspond to any word in my vocabulary.txt

my question :

How could I restrict inference to known words ? (and forget others)

Thanks all.

[This is an archived TTS discussion thread from discourse.mozilla.org/t/quality-of-the-ctc-decoder-limit-inference-words-to-known-words-only]

[JRMeyer]

>>> francob
[December 1, 2017, 8:34pm]

Hi, if you check out the hyperparameters and increase LM_WEIGHT and
VALID_WORD_COUNT_WEIGHT then you should get better results.

[Archived Post]

[JRMeyer]

>>> elpimous_robot
[December 1, 2017, 8:55pm]

Hi Francob. slash
Thanks, I'll try it... slash
PS : what hyperparams do you use ?

[Archived Post]

[JRMeyer]

>>> francob
[December 1, 2017, 9:24pm]

5 for LM_WEIGHT and 3 for VALID_WORD_COUNT_WEIGHT, haven't optimized
them yet though

[Archived Post]

[JRMeyer]

>>> dbanka
[December 27, 2017, 7:41am]



> NT_WEIGHT, haven't optimized them yet though

Hey Francob, How can we optimize LM_WEIGHT , WORD_COUNT_WEIGHT and
VALID_WORD_COUNT_WEIGHT

[Archived Post]

[JRMeyer]

>>> elpimous_robot
[December 28, 2017, 7:17pm]

Hi. They are params ! slash
set it like this :

> -- WORD_COUNT_WEIGHT = 5 slash
> ...

[Archived Post]