AI polish and presets
AI polish tidies the recognised text before it is inserted. What it does depends on the preset in use. The default preset, Proofread, fixes punctuation and typos and removes filler words and repetitions; it does not change what you said, translate it or add to it.
Turn it on or off with the AI Polish switch at the top right of the window. The two menus next to it choose the preset and the model that polishes the text; the Recognition and polish card on the home page has the same switch and menus. A scene can turn it on or off, and choose a preset, for a particular app.
Services
AI polish uses the service and model chosen on the AI models page, or from the model menu next to the AI Polish switch: the built-in service, OpenAI, Groq, Google AI Studio, SiliconFlow, Alibaba Cloud Model Studio, DeepSeek, Kimi, Zhipu GLM, Volcengine Ark, Tencent Hunyuan, Xiaomi MiMo, OpenRouter, a local Ollama, or any OpenAI-compatible endpoint. The same menu is on the home page; it lists the models of the built-in service and of each provider that is set up. See Cloud services and Provider setup for setting one up.
If the service does not answer, is busy, or its answer is cut off at the length limit, the recognised text is inserted without polish, and the dictation is not lost; once it is in, the overlay says it was not polished and why, for example service busy. The history's details say why too, including what the service answered, and let you polish it again. For the notice the built-in service shows when it is busy, see The built-in service. When the selected model's quota is used up, AI polish can move on to other models in order: see When a model's quota runs out in Cloud services.
Presets
A preset tells AI polish what to do with the recognised text. Choose the one in use from the preset menu on the home page or next to the AI Polish switch, from AI Polish in the tray menu, or under Presets on the AI models page. While AI polish runs, the overlay shows the name of the preset, and each history entry records it.
| Preset | What it does |
|---|---|
| Proofread (default) | Fixes typos, homophones and punctuation, splits paragraphs, removes filler words and stutters. When you correct yourself ("no, I mean…"), only the corrected wording is kept. Symbols you say in file names, paths and commands are written as symbols, for example "dash dash" as "--" |
| Prompt optimizer | Rewrites a spoken request into a clear prompt for another AI: one precise instruction for a simple request, a numbered list when there are several points, with the background, requirements, conditions, limits and exceptions all included. A question stays a question and is rewritten, never answered; a statement is only proofread; English stays English |
| Clarify intent | Keeps the corrections, removes repetition and turns several points into a list; the conditions and exceptions of each point stay |
| Casual chat | Short, conversational sentences without a closing full stop; every time, number and name you mention stays |
| Chinese ⇄ English | Fixes recognition errors and keeps only the corrected wording, then translates between Chinese and English, leaving nothing out |
| Key points | Key points and to-dos as two lists, without adding facts; no owner is named unless you named one, and numbers are written as you said them |
| Punctuation only | Adds punctuation and sentence breaks and changes no words |
| Formal | Rewrites spoken phrasing into complete, formal written language, keeping every listed item, condition and exception |
Your own presets
Under Presets on the AI models page, New preset creates a preset of your own, and Copy to custom starts one from a built-in preset, which cannot be changed itself. A preset has a name of up to 24 characters and a prompt of up to 4,000 characters that says how to handle the text. Voltip adds the output format, so the prompt does not need to describe it. You can keep up to 30 presets of your own.
Trial run processes a sample text with the current AI service before you save; the result is not kept.
When you delete a preset, the settings and scenes that used it switch to Proofread.
What the AI service receives
- The recognised text, after the dictionary corrections. Never audio.
- The instructions of the preset in use.
- The name of the app you are dictating into. You can turn this off.
- The window title, only if you turn it on.
- The extra instructions of the scene that applies, if any.
- The correct spellings from your dictionary, and the specialist terms of a built-in scene that applies, so that the service keeps them as written.
The two switches are under Settings → Scenes, in Context sent to AI polish. They apply only while AI polish is on; the recognition service never receives this context.
With the Type as you speak output mode, sentences are inserted as they are recognised, so AI polish does not run.