AI Review
Proofread the subtitle text with a local (or remote) large language model - typos, spelling, grammar, punctuation and casing. Nothing is changed until you apply the suggestions you agree with.
- Menu: Tools → AI review…
- Shortcut: Ctrl+Alt+R (Cmd+Alt+R on macOS)
Engines
- llama.cpp — a managed local server. Pick a model from the curated list (Qwen 3.5, Gemma 3, Llama 3.1, EuroLLM, Phi-4 mini); Subtitle Edit downloads the engine and model on first use. A green dot marks models that are already downloaded. Custom
*.gguffiles placed in the llama.cpp models folder also appear. - Ollama — uses a running Ollama instance; type a model name or pick one from the server.
- OpenAI-compatible — any endpoint that speaks the OpenAI chat API: LM Studio, KoboldCpp, vLLM, a llama.cpp server on another machine, or cloud APIs (OpenAI, Groq, OpenRouter, DeepSeek, Mistral, Gemini). Enter the URL, model name, and an API key if the service needs one.
Using a cloud model
If a local model is too slow on your machine, select the OpenAI-compatible engine and point it at a cloud provider. Only the URL, model name and API key differ:
| Provider | URL | Example model |
|---|---|---|
| Google Gemini | https://generativelanguage.googleapis.com/v1beta/openai/chat/completions |
gemini-3.1-flash-lite |
| OpenAI | https://api.openai.com/v1/chat/completions |
gpt-4o-mini |
| DeepSeek | https://api.deepseek.com/v1/chat/completions |
deepseek-chat |
| Groq | https://api.groq.com/openai/v1/chat/completions |
llama-3.3-70b-versatile |
| Mistral | https://api.mistral.ai/v1/chat/completions |
mistral-small-latest |
| OpenRouter | https://openrouter.ai/api/v1/chat/completions |
google/gemini-3.1-flash-lite |
| LM Studio (local) | http://localhost:1234/v1/chat/completions |
name shown in LM Studio |
The API key goes in the API key field - Subtitle Edit sends it as a bearer token.
The subtitle is sent in batches, and free tiers usually limit how many requests you may send per minute. The timer field after the API key sets a delay in seconds between requests - raise it if the provider starts rejecting requests. It is off (0) by default, and the first request of a review never waits.
Reviewing only part of a subtitle is also possible: select the lines in the subtitle grid, then right-click → Selected lines → AI review….
Reviewing
Press Review to start. The subtitle is sent to the model in small batches, and suggestions appear in the grid while the review runs - you can inspect, check and uncheck rows before it finishes, or press Stop to keep what was found so far.
Each suggestion shows:
- Apply — checkbox deciding whether the fix is applied
- Line number and a category tag (spelling, grammar, punctuation, casing, other)
- Before / After — with word-level differences highlighted
- Selecting a row shows the model’s short reason below the grid
Filter the grid with the category chips above it. Press Apply N fixes to apply the checked suggestions - this is a single undo step (Ctrl+Z reverts everything).
Apply does not close the window. The applied rows disappear from the grid and everything else stays, so a review can be worked through in passes - press Select none, tick the suggestions you agree with, apply, then carry on with the rest. Each pass is its own undo step.
The buttons next to it are the usual pair: OK applies the checked suggestions and closes, so the last pass is a single click, and Cancel closes without applying what is still checked. Passes you already applied stay applied - use Ctrl+Z to undo them.
Listening to a line
When a video is loaded, a Play current button appears at the bottom left. It plays the selected suggestion’s line in the video player and pauses at the end of the line, so you can hear what was actually said before deciding on a fix. Double-clicking a suggestion does the same, as does F5 (or Ctrl/Cmd+Space) - F5 follows your Play selected lines shortcut. Space is not used for playback here: it toggles the Apply checkbox of the selected row.
Sentences across multiple lines
Lines are grouped into sentence units, so a sentence that continues over several subtitles is always reviewed as a whole, and the model sees a couple of surrounding lines as read-only context. Corrections never move words between lines, so timing and reading speed are unaffected. Suggestions belonging to the same sentence are checked and unchecked together.
Choosing a model
Any model in the list can proofread, but quality depends on how well it knows your language:
- English - Llama 3.1 8B is the strongest proofreader of its size; Qwen 3.5 9B and Gemma 4 12B are close.
- European languages (including Danish, Swedish, Norwegian, Finnish and Dutch) - the EuroLLM 2512 models are trained on all 24 official EU languages and are the best pick here. Take EuroLLM 22B (IQ4_XS) if it fits, EuroLLM 9B otherwise.
- Low-end machines - Gemma 4 E2B or Phi-4 mini run on almost anything, at a noticeable quality cost.
Rule of thumb: pick the largest model whose download size leaves about 2 GB of VRAM headroom for context. On a 16 GB card that means the 12.3 GB EuroLLM 22B or the 7.6 GB Gemma 4 12B; on 8 GB, a 4-6 GB model such as EuroLLM 9B or Qwen 3.5 4B (Q8_0). A model that does not fit still runs - llama.cpp keeps the rest in system RAM - but much more slowly.
Reviewing is not translating, so a bigger model mostly buys fewer false suggestions, not different categories of fix. If the review produces a lot of noise in your language, try a EuroLLM model before trying a larger one of the same family.
The prompt
The Edit prompt… button opens the review instructions sent to the model. {language} is replaced with the auto-detected subtitle language. The strict data-exchange contract is appended by Subtitle Edit and cannot be broken by prompt edits, so feel free to tailor the instructions - e.g. “also flag anachronisms” or “never touch song lyrics”.
Safety rails
- Nothing is applied automatically - you decide per suggestion.
- ASSA override blocks at the start or end of a line (
{\pos(...)\fs54...}) are never sent to the model - it only sees the text, and the blocks are restored unchanged around its correction. Lines that are pure drawings or tags are skipped. The ASSA actor and style name are sent as read-only context so the model knows who is speaking and whether a line is dialogue or a sign. - Suggestions that would add or remove the remaining formatting tags (
<i>, inline blocks) are discarded. - Suggestions that change a line’s length a lot are flagged with a warning and start unchecked, since they are usually rewrites rather than corrections.
- Replies from the model that do not follow the expected format are retried once and then skipped.