Dictation engines

Click the mic on the command bar and your words become the command text. Which engine listens is up to you (Settings → Voice):

The four engines

1. Browser — the WebView’s built-in speech recognition. Zero setup, decent accuracy, needs no key.
2. Parakeet (offline) — NVIDIA’s Parakeet model running locally. Free, private, works with no internet. Details.
3. OpenAI — recorded audio transcribed via your OpenAI key. Strong accuracy, network + key required.
4. Realtime — a live conversational session (OpenAI realtime or Grok) that hears you continuously instead of per-clip. Details.

Dictation cleanup

Whatever the engine hears passes through a cleanup layer before it lands (Settings → Voice → Dictation cleanup, on by default):

• Fillers and stutters out — pure interjections (um, uh, er, hmm…) and stutter doubles (“the the”) are removed, and the punctuation they leave behind is tidied. Words that can carry meaning (“like”, “actually”, “you know”) are never touched by the rules.
• Names spelled right — your open agent windows (Alex, Remy…) and the app’s own vocabulary (Claude Code, Kanban, Sprint Board…) are case-corrected, and close mishearings (“jemini”, “cloud code”) are fixed within a strict edit-distance guard.
• AI refine (optional, off by default) — long dictations get a one-shot pass through your local agent that also removes false starts and keeps only the corrected version when you self-correct mid-sentence (“at 2… actually 3”). It preserves meaning, names, numbers, and technical detail, and swaps in silently only if you haven’t edited the text meanwhile.

Cleanup means shorter prompts — fewer tokens on every dictated command — with the ask intact.

Good to know

• Dictation types into whichever text box has focus — an agent prompt, a note, a card title — and defaults to the command bar when nothing is focused. Clicking the mic doesn’t steal your field’s focus.
• In the command bar, every voice command works exactly like typing.
• Spoken replies (TTS) are configured separately — OS voices, OpenAI, or Grok — in voice settings.
• Keys never leave your machine; they live in the local .env / settings store.