AI polish
Improve with AI is the optional rewrite stage after speech recognition. On Pro and Pro trial, it strips verbal filler, fixes punctuation and capitalization, and formats output according to How you want it written.
Plain dictation runs with AI polish off for fast verbatim capture, while shipped profiles (Email, Message, Notes, Coding, Document) turn it on by default.
Before and after examples
| Spoken transcript | Active profile | Polished output |
|---|---|---|
| "uh hi sarah yeah I'll send the revised Q3 deck over by um Thursday afternoon thanks" | Hi Sarah,\n\nI'll send the revised Q3 deck over by Thursday afternoon.\n\nThanks, | |
| "hey let's push standup fifteen minutes later wait actually make it ten thirty" | Message | Hey, let's push standup 15 minutes later—actually, let's make it 10:30. |
| "for the api handler um check if the token expired first then return four oh one" | Coding | In the API handler, check if the token has expired first, then return 401. |
If you leave How you want it written blank on a profile with Improve enabled, Kalam runs a lightweight cleanup baseline—removing filler words and fixing grammar without inventing extra sentences.
When AI polish runs
On Pro or Pro trial, Improve runs once per take when:
- Improve with AI is switched on for the active profile (or context was captured via Read surroundings), and
- You did not speak a Commands prefix (such as "New note…" or "Search for…"), and
- A language provider (Groq, OpenAI, Google AI Studio, Anthropic, OpenRouter, or a custom OpenAI-compatible endpoint) is connected.
If the language model errors or times out, Kalam automatically falls back to inserting your raw spoken transcript so nothing you said is lost.
Snippets, Dictionary replacements, and inline Voice commands (period, new line, scratch that) apply only on raw transcripts when AI polish did not run. When AI polish runs, the language model handles formatting naturally from your speech.
What the model receives
When Improve runs, Kalam sends:
- Your How you want it written prompt (or the default cleanup prompt).
- The spoken transcript.
- Permitted Context awareness details (if Read surroundings is enabled on that profile).
Audio files, dictation history, dictionary lists, and snippets are never included.
When the focused window matches your Sensitive apps list, Kalam forces on-device speech recognition, blocks all Read surroundings context capture, and blocks Commands. If Improve with AI is still switched on for that profile, it processes only the raw transcript with zero screen, window, or clipboard context attached. To prevent any cloud call inside sensitive windows, turn Improve off on that profile or switch to Plain dictation.
Sense safeguards for polish
Configure Sense under Dictation → Models:
- Clean take: Evaluates short takes (up to 45 English words outside sensitive apps) to skip unnecessary AI calls when a clean spoken take needs no rewrite.
- Polish check: Compares the original transcript against the polished output afterward and flags a notice in History if content was dropped or meaning shifted.
While Sense is turned on, Clean take and Polish check can send short non-sensitive transcripts to Kalam's cloud check. Turn Sense off on Dictation → Models to disable those checks completely.