In this line:
https://github.com/NVIDIA/mellotron/blob/d5362ccae23984f323e3cb024a01ec1de0493aff/text/__init__.py#L102
you are filtering "_" and "~" symbols.
Also, one of the main advice to improve alignment map convergence is to add special symbols to start and end of every sentence. Usually these symbols are exactly "_" and "~" (_ex: _What is your name?~_). But you filter out exactly these symbols and do not add them anywhere in the code.
Hi @rafaelvalle
In this line: https://github.com/NVIDIA/mellotron/blob/d5362ccae23984f323e3cb024a01ec1de0493aff/text/__init__.py#L102 you are filtering "
_
" and "~
" symbols. Also, one of the main advice to improve alignment map convergence is to add special symbols to start and end of every sentence. Usually these symbols are exactly "_
" and "~
" (_ex:_What is your name?~
_). But you filter out exactly these symbols and do not add them anywhere in the code.It is interesting, that you've included the "
_
" symbol here: https://github.com/NVIDIA/mellotron/blob/d5362ccae23984f323e3cb024a01ec1de0493aff/text/symbols.py#L11 but anyway filter it out next in sentence preprocessing.So the questions are: 1) What is the reason of filtering out "
_
" and "~
" symbols? 2) Why don't you use them as start and end symbols in sentences?