Sentence endings are often missing

[theslipperyCarrot]

When I have an audio generated from a longer text file, individual words at the end of the sentence or entire sentences are often missing.
Especially at the beginning of a file, it seems to me. Sometimes complete lines are not created at all in the middle of it.

Is it perhaps a matter of attitude? I've already played around with fade-out.

But there are also often three to five words that are missing. Especially after a punctuation mark such as a period or comma, the rest is often missing.
Best Regards and Thanks for ideas!

[lukaszliniewicz]

This looks like a problem with the model or the voice sample you're using. What model is it? The default one or a fine-tuned one? Can you try different voice clips of the same speaker to see if it helps?

[theslipperyCarrot]

Thanks for the tip!
So I should try other reference files under "Speaker Voice"?
It is a finetuned XTTS model with a self-rained RVC model behind it.

Unfortunately I can't train a new XTTS model so easily, because my XTTS-Finetune-WebUI no longer works.

[theslipperyCarrot]

I tried a first test with another file as a reference, which is a bit longer (27s).
That seems to make a big difference! Thank you very much!

[lukaszliniewicz]

I don't have much time now, but when I have a moment, I will update the training workflow. The version I'm running locally produces much cleaner results.

[theslipperyCarrot]

That would be really great!

Not all of Pandrator's functions are currently running for me. XTTS training and evaluation are not yet working. And i don´t have tried everything I guess. But I think that's my problem. I'm thinking about formatting my computer and completely resetting it in order to install and use Pandrator in the admin account afterwards.
I guess with evaluation, the audio generated is read out and checked again with Whisper?
This should actually fix the problem with missing words.

Pandrator is imo the absolute best application for AI text-to-speech!
Thank you so much for this cool program!

With a different reference audio, the problem has definitely gotten much better.
Maybe my experience will help someone. I thought I would make the best use of the reference audio that the XTTS-Finetune-WebUI had worked out. That was a mistake! That was short and had clipping at the end.

Here, a well-chosen clean audio clip improved my overall quality!
It is also imo the case with the training of XTTS and RVC that quality is more important than quantity of the source-audio!

[lukaszliniewicz]

I've updated the training workflow. Please try a clean installation and let me know if it works now.

[theslipperyCarrot]

Thank you very much!
I will try a new, clean installation.