How does this work with novels in progress?

[thos-grol]

Say we want to add a chapter to an epub, can the tool keep the same context and everything, skip translating all the previous chapters and just translate the latest chapter using context etc?

[hydropix]

Hi,
Based on my tests, the context from the preceding chapters isn't necessary and is actually quite counterproductive. When TBL translates, it breaks the entire content into chunks of 200 lines (i.e., short paragraphs), and the preceding chunk is used as context, but that's as far as it goes. If you absolutely need to guide the translation with specific guidelines, there is a way to add custom instructions by adding a file to the "Custom_Instructions" folder and then selecting it from the options menu.

If you want more flexibility in the translation, you can also use plain text or a DOCX file and do manual copy-paste?

What’s your use case? Do you not want to have to retranslate the entire book from scratch? I’m curious.

[thos-grol]

Do you not want to have to retranslate the entire book from scratch? I’m curious.
This.

Use case is scrape -> translate -> tts.

Thanks for letting me know this information. To test this, I just joined all the untranslated text files and then used this tool to translate and then just ran a script to segment and then return translated chapters to proper locations. That should fix the problem.

[hydropix]

I'm working on another project (undisclosed) that makes heavy use of TTS, and I've noticed that the text needs to be cleaned up to remove certain characters that otherwise produce strange sounds... If you run into this issue, I recommend filtering the text a bit.

[thos-grol]

I'm working on another project (undisclosed) that makes heavy use of TTS, and I've noticed that the text needs to be cleaned up to remove certain characters that otherwise produce strange sounds... If you run into this issue, I recommend filtering the text a bit.

Quick question is the 200 limit due to gpu limitations? What if you used larger line limits or character limits to fit token windows? ie. for a 5090.

[thos-grol]

I've tried building an automated pipeline:

scrape -> chunk -> translate with context -> stitch -> refine -> tts
with all intermediary files left on disk. If we delete one file, the process will restart for that file upon re-run, preserving state and allowing for updates.

System Prompt:

You are a master literary editor. Your task is to refine the provided translation into a work of high literature.

Core Directives:

Omit Needless Words: (Strunk & White). If a word does not add meaning, delete it.

Active Voice: Favor active verbs over passive ones.

The "Iceberg" Technique (Hemingway): Use simple, forceful language for action. Subtext should live beneath the surface.

Precise Diction (Nabokov): When describing internal states or scenery, use the most specific noun or adjective possible. Avoid "very," "really," or "suddenly."

Rhythmic Variance: Vary sentence length to create a "pulse" in the prose.

Think I will try something like this prompt for refinement.