Long Russian technical book (11h expected) outputs only 35min audio + poor Russian TTS quality

[Flexible78]

https://prnt.sc/WGFR0sCeLTM9 screen here --------
Hi Drew!

Love the project, works great with English books. But have two serious issues with Russian technical content:

1. Massive content truncation

• Uploaded EPUB: ~600MB, ~426 blocks detected
• Progress reaches 100%, but final MP3 is only 35 minutes instead of expected ~11 hours
• Console shows "End of Block X" messages, but clearly not all content is processed
• No errors in final output, just... stops

2. Poor Russian TTS quality (XTTSv2)

• Heavy OCR artifacts ("іі плюс а [ ноль К ноль") cause synthesis failures
• Frequent combine_audio_sentences() crashes: TypeError: expected str, bytes or os.PathLike object, not NoneType
• Russian pronunciation is robotic/unclear compared to English

Environment:

• Windows 11, RTX 3060 6GB CUDA
• Settings: Temp 0.3, RepPen 1.5, Top-k 40, Top-p 0.85, Speed 1.4
• Language: Russian (ru)

Question: Is there a way to:

1. Force process ALL blocks (disable truncation)?
2. Improve Russian OCR handling or recommend preprocessing?
3. Alternative TTS engine for Russian technical books?

Thanks for great tool!

[ROBERT-MCDOWELL]

without the full terminal log, nothing can be debugged.
also russian models is not ours, it's the one you select from XTTSv2, Fairseq, Vits or other TTS engine projects, it's to their repo you should complain to. you are free to create your own fine-tuned model with perfect russian and share it to us so we will add to the russian fine tuned XTTSv2 list. you can also make an offer to us to train a new XTTSv2 (or else? which one you chosed?) with perfect russian.
OCR is not perfect and won't be. did E2A installed the right OCR dictionnary (tesseract-ocr-rus) when you installed E2A or run it?
reason why it needs the FULL install and first conversion log.
Also provide the ebook you try to convert.

[Flexible78]

where i can find log? in tmp 1.5GB flac files but in result only 35min

[ROBERT-MCDOWELL]

the log is the TERMINAL where you run E2A

[Flexible78]

1.log
yes pls

[ROBERT-MCDOWELL]

first of all, when you run (apparently in gradio mode) E2A, you should every time refresh your browser, gradio does not reconnect to a new python launch automatically, so the conversion will start badly without to refresh.

[Flexible78]

i did it. but maj be forget in this case/ but i try 6 atemps and nothing result more than 35minutes

[Flexible78]

I can provide it, but it takes up 600 MB. Should I upload it to the cloud?

[ROBERT-MCDOWELL]

600MB your ebook? are you kidding?

[Flexible78]

unfortnly nope cause i rec0gnize it with OCR by myself (pdf element software)

[ROBERT-MCDOWELL]

follow up on issues section.

[Flexible78]

ok/ and i am fix issue in code - sorry - cause nothing possivility to change to CUDA mode)) sorry i little improve your code)) and after it i can use CUda and Nvidia video))

[ROBERT-MCDOWELL]

sorry I don't understand what you said... go here to continue the discussion
#1653