Long-form translation usually forces a trade-off between fidelity (format, timestamps) and model context limits. This project removes that trade-off by splitting documents into managed segments that keep contextual continuity and reassemble outputs so the original EPUB/SRT/DOCX layout remains intact — even for multi-hundred-page books.
What Sets It Apart
- Document-first preservation: keeps EPUB structure, styles, and SRT timecodes synchronized so translated output mirrors the source layout — meaning less manual reformatting after translation.
- Unlimited-length processing: an intelligent chunking and checkpoint system lets you translate arbitrarily large files and resume interrupted runs without redoing previous work, which is essential for long novels or large subtitle sets.
- Provider-agnostic workflow: works with local runtimes (Ollama, llama-based servers) and many cloud providers (OpenAI-compatible endpoints, OpenRouter, Poe, Mistral, Gemini), so teams can choose trade-offs between cost, latency, and privacy.
- Multiple output-focused options: supports a refine/pass-two mode for literary polishing and optional TTS generation, letting you tune for raw accuracy or final-readability depending on the target use.
Who It's For and Trade‑offs
Great fit if you need to translate long-form content (books, transcripts, subtitle batches) while keeping native file structure and timestamps intact — useful for publishers, translators, localization teams, and archiving projects. Look elsewhere if you need real-time or streaming translation of short chat-like messages: the system is optimized for batch, checkpointed translation, not low-latency conversational translation.
Trade-offs: translation quality depends heavily on chosen model and prompt/refinement strategy (some models produce better literal fidelity, others better literary style). Using cloud providers incurs API costs and potential privacy considerations; running locally requires model and runtime setup and more local compute. Also note the repository is AGPL-3.0, which affects redistribution and integration choices.
Where It Fits
Positions itself between file-format-aware localization tools and generic LLM wrappers: it focuses on end-to-end document fidelity (format + timestamps) and operational reliability for very large inputs, rather than being a lightweight chat client or a pure MT engine.
