AI Polish and Translate

Sona's polish and translation features are optional. Local transcription works without them, but the AI steps need a configured provider in Settings > LLM Service. This page stays focused on LLM Polish and Translate; if you mainly want a transcript-side recap, jump straight to AI Summary.

Best for

• Cleaning up transcript wording after the local transcription pass
• Producing translated segments while keeping the source transcript visible
• Teams that want polishing and translation to happen inside the same editor flow

Before you start

• You already have transcript segments from Live Record, Batch Import, or Projects and Inbox.
• You are ready to configure the feature you need in Settings > LLM Service.

Connect the provider first

1. Open Settings > LLM Service.
2. Choose between Local Offline LLM (built-in llama.cpp engine with downloadable Qwen presets or custom GGUF imports for 100% offline polish and translation), Local ONNX NMT (lightweight local translation engine), or Cloud Providers (unified aimux multi-provider integration supporting OpenAI, Anthropic Claude, Google Gemini, Groq, Mistral, Ollama, DeepSeek, etc.).
3. In Feature Models, assign the model for Polish Model and Translation Model.
4. For reasoning models, Sona automatically parses and isolates thinking traces (<think> blocks), preventing internal chain-of-thought tokens from polluting the final transcript.
5. If using a cloud provider, enter required credentials and click Test Connection.
6. Return to the main editor after the required feature model is assigned.

Run polish when you want a cleaner draft

1. Make sure Polish Model is assigned in Settings > LLM Service.
2. In the app, open the LLM Polish action.
3. Choose what you need: LLM Polish, Re-transcribe, Undo, Redo, or Advanced Settings.
4. Open Advanced Settings if you want to manage Auto-Polish, Auto-Polish Frequency, Keywords, directive Polish Modes (Clean & Readable, Verbatim Strict, Formal Written), and Custom Context. The fuller tuning view is collected in Vocabulary and Advanced Settings.

Run translation when you need bilingual output

1. Make sure Translation Model is assigned in Settings > LLM Service.
2. Click the Translate button (redesigned with a compact popover card).
3. Choose the target language.
4. Click Start Translation or Retranslate. In live recording and batch queue modes, you can also enable non-blocking real-time bilingual translation and the streaming translation scheduler.
5. Bilingual translation display is enabled by default once translated. Use Show Translations or Hide Translations to toggle the view, and double-click any translated segment directly in the editor to make independent inline adjustments.

What the AI steps change

• LLM Polish updates transcript text in place.
• Translate stores translation text per segment and can display it under the original text.

Useful notes

• Polish Model and Translation Model are configured separately. One provider can serve both, or you can split them.
• Translation can use dedicated translation providers such as Google Translate (Free) or Google Translate (API), but polish needs an LLM-capable provider and model.
• Translation target languages currently include Chinese (Simplified), English, Japanese, Korean, French, German, and Spanish.
• Re-transcribe is only available when the current transcript came from a saved project item.
• If what you need next is a transcript-side recap instead of rewritten transcript text, continue to AI Summary.
• If you are really here for Auto-Polish, keywords, or custom context, continue to Vocabulary and Advanced Settings.
• Once the text looks right, move to Export and Settings.