The dictation pipeline
When you dictate, Kalam runs a short pipeline: record your voice, transcribe it, optionally polish the text with AI using context from the app you are in, then type the result at your cursor.
This is the same four-step shape you see in the mode editor (Dictation → a mode → the pipeline rail), plus an early voice actions branch. Two of the remaining steps are optional:
- Hear your voice — always on. Speech-to-text turns audio into a raw transcript.
- Read your surroundings — optional. Captures context (app content, clipboard, selection, date/time) to inform polishing. Pro and above.
- Improve the text — optional. An AI model rewrites the transcript for tone, structure, and formatting. Pro and above.
What happens on each dictation
- Press and hold the hotkey. Kalam remembers which app and text field you are in, optionally snapshots your clipboard and selected text, and starts recording. The floating pill shows it is listening.
- Release. Recording stops and audio goes to your speech engine — on-device or cloud, per the active mode. The pill shows processing.
- Raw transcript. The speech engine returns text. At most one transcription call runs per dictation.
- Optional context + AI polish. If the active mode has Improve or surroundings on, Kalam makes one AI call that rewrites the transcript using your writing instructions and any captured context. If the AI declines or fails, the raw transcript is used instead.
- Insert. The final text is typed into your app, or copied to the clipboard as a fallback if direct typing is blocked.
Two calls, maximum
One dictation costs at most one transcription plus one AI polish call. There is no hidden chattiness.
Local vs cloud at each step
| Step | On your device | Leaves your device |
|---|---|---|
| Record + clean up audio | Always | Never |
| Speech-to-text | Local engines | Groq / OpenAI / Kalam Cloud if a cloud provider is selected |
| Surroundings (context) | Captured locally | Sent only inside the AI polish request, if polish runs |
| AI polish | Local/custom endpoint if configured | Groq / OpenAI / etc., or Kalam-hosted on Max |
| Snippets, voice commands, dictionary | Always | Never |
| Typing the text | Always | Never |
So "does my audio leave my device" is decided by the speech engine, and "does my text leave my device" is decided by AI polish. Speech-to-text modes and AI polish cover each in detail.
Three other ways dictation can end
- Voice actions. Checked on the raw transcript before Improve. A prefix (
new note,search for,go to,copy this, …) skips typing. Optional press enter at the end types then sends Return (off by default). See Voice actions. - Voice editing. A separate hotkey lets you highlight text and speak an instruction to rewrite it. Pro and above, Windows.
- Sensitive apps. If the focused app matches your sensitive-app list, Kalam forces on-device speech-to-text and skips context and AI for that dictation. See Sensitive apps and privacy.