Frequently Asked Questions¶

General¶

Does it work offline?¶

Mostly yes. Specifically:

Translation needs an LLM. The free Gemini API is online; local Ollama / LM Studio through the Custom Provider settings is fully offline.
OCR with Tesseract or EasyOCR is offline.
STT with Whisper (default) is offline.
TTS with Edge TTS (default) is online; ElevenLabs / Google Cloud TTS / Gemini TTS are online (free or paid); Piper TTS is fully offline neural TTS — no key, no network calls once you've downloaded the per-language voice (~25–60 MB ONNX file) via Settings → Voice → Piper TTS → Download voices now.

For a fully air-gapped setup: Custom Provider → local LLM, Tesseract or EasyOCR for OCR, Whisper for STT, and Piper TTS for voice output.

Where are my translated files saved?¶

Next to the original by default, with a _translated_<src>_<tgt> suffix (e.g. report_translated_en_fr.docx). Override per-feature in Settings → General → Translation storage path.

Where are my settings stored?¶

INI file at:

OS	Path
Linux	`~/.config/ai-translate/settings.ini`
macOS	`~/Library/Preferences/ai-translate/settings.ini`
Windows	`%APPDATA%\ai-translate\settings.ini`

API keys live in the OS keychain (not in the INI). Translation history lives in a SQLite DB in the data directory.

How is my data handled?¶

Local-first — text never leaves your machine unless you're calling a cloud LLM / OCR / STT / TTS service.
No telemetry — the app doesn't phone home. The only outbound request the app itself makes is the optional GitHub-Releases update check (toggle in Settings → General); cloud backends only call their respective vendors.
API keys — stored in your OS keychain. The desktop app's keychain fallback is a plaintext INI when no keychain daemon is available.

Can I translate a Google Doc / Notion page?¶

Not directly. Export to .docx first, translate, then import the translated file back. Same for Notion (export as Markdown / HTML), Confluence (export as .docx), etc.

Choosing models / engines¶

Which LLM model should I use?¶

For most users:

Any Gemini Flash variant — free tier, fast, surprisingly good. Use for everyday translations. Names look like gemini-2.5-flash, gemini-3-flash-preview, etc., depending on what's currently available.
Any Gemini Pro variant — pay-per-token, higher quality. Use for important documents (legal, technical, customer-facing).
Local Ollama with a 7B-13B model — when you need offline / privacy.

The per-feature model picker means you can use a fast model for chat-style translation and reserve the expensive one for documents.

Which OCR engine should I use?¶

Tesseract for clean printed text in major scripts. Free, offline, fast.
EasyOCR for non-Latin scripts (CJK especially) and noisier images.
Google Cloud Vision for handwriting, mixed scripts, and the highest accuracy when you can pay.

Which STT method should I use?¶

Whisper local for offline / privacy.
Soniox for multi-speaker recordings — speaker labels round-trip into your SRT.
Google Cloud STT for telephony / medical audio (their domain models are good).
Gemini Live for real-time speech-to-speech translation.

Which TTS backend?¶

Edge TTS for free, high-quality voices.
ElevenLabs for premium / branded / cloned voices.
Google Cloud TTS for WaveNet voices in long-tail languages where Edge has thin coverage.
Gemini TTS for free natural prebuilt voices reusing your existing Gemini API key.
Piper TTS when you need offline / air-gapped voice output. Trade-off: each language needs a one-time ~25–60 MB voice download via Settings → Voice → Piper TTS → Download voices now, and 13 of the app's 45 languages have no Piper voice (those silently fall back to Edge TTS).

Workflow¶

How do I translate a whole folder?¶

Drop the folder into the Translate Document drop zone. Supported files inside (recursively) get queued; everything else is silently skipped. There's a 100-file drop cap; bigger batches → split into multiple drops.

Can I pause and resume translations?¶

Yes. Quit the app any time — Pending / Translating tasks resume on next launch. Per-task checkpointing means PDF page 47 of 100 isn't redone when you resume.

Can I edit a translation by hand?¶

For Translate Text — yes, click the right pane and type. The edit auto-saves to the entry's history record.

For Translate Document — open the translated file in your usual editor (Word, LibreOffice, etc.) and edit there. The app doesn't roundtrip the edits back into history.

Can I bulk-translate a list of strings?¶

Use the CLI:

ait *.txt --target French

Or for in-process strings (e.g. UI strings extracted from code), call the translate_text MCP tool with a list, or use the Python API directly:

from src.core.llm_engine import translate_text
out = translate_text(texts=["Hello", "World"], target_lang="French")

Glossary¶

Why isn't the LLM using my glossary?¶

Three things to check:

The set is active (checkbox checked).
The source term in your glossary actually appears in the source text (the per-call compression only sends the LLM entries that match the batch text — saves tokens, but means a typo'd source term is invisible).
The model is strong enough — flash-lite sometimes ignores hints that flash and pro honour.

Glossary terms are matched accent-insensitively?¶

Yes. Both glossary lookup and the search box in the Glossary page use a normalisation function that strips accents and case. So cafe, Café, and CAFE all match an entry whose source is Café.

Privacy¶

Do you collect any usage data?¶

No. The app has no analytics SDK. The optional update check polls a single GitHub Releases endpoint at startup; it's toggleable in Settings → General.

Are my API keys safe?¶

They're stored in your OS keychain (Keychain on macOS, Credential Manager on Windows, Secret Service on Linux). Other processes can't read them without your explicit permission. The fallback (when no keychain daemon is available — typically headless Linux servers) is a plaintext INI under your user's config directory; in that mode the keys are file-permission-protected but not cryptographically encrypted.