Question mark and others in the dataset

[JRMeyer]

>>> vcjobacc
[August 30, 2019, 1:50pm]

Hello everyone! slash
I prepare ds for the TTS. I want the model to pronounce question
sentences clearly question-like, I mean, I english we can realize that
we've just been asked because of at least a special order of words,
while in some other languages there are no such restrictions, the recipe
is just add '?' at the end; when speaking the intonation will make clear
that the question has just been ascked. slash
So, I want to have around 25k sentences with various length ending
sometimes with '.', sometimes with '?', sometimes with '!'. slash
Will the TTS learn to pronounce it right? How many ?-sentences do I need
then? slash
Now I have around 315, which seems to me too little. Should it be
something around 2k or so? slash
One more question. slash
Let's assume we haven't added one quite hard to pronounce word in the
dataset, so the TTS makes misstakes on that one. To fix it, clearly,
that word should be added to the dataset, right? But how many times
should it occur (roughly)? slash
Thank you all for the answers in advance!

[This is an archived TTS discussion thread from discourse.mozilla.org/t/question-mark-and-others-in-the-dataset]

[JRMeyer]

>>> alchemi5t
[August 30, 2019, 3:46pm]

To answer your question about special characters(like ? And ,)
influencing the TTS to help with intonations, Yes it works exactly like
that. You will get the expected prosody for questions and pauses for
commas if they are in the dataset and the dataset is consistent with
their audio.

The other questions about how many is very hard to answer. You'd have to
find what works empirically, or add enough to be proportionately to not
worry about it.

Also

{.avatar

> Let's assume we haven't added one quite hard to pronounce word in the
> dataset, so the TTS makes misstakes on that one. To fix it, clearly,
> that word should be added to the dataset, right? But how many times
> should it occur (roughly)? slash
> Thank you all for the answers in advance!

If you try fixing problems like this, you are more likely to add bias
than fix the problem. Like I said, empirically find it, or add a ton of
data hope it works.

[Archived Post]