Everything happens where your cursor is.

Hold, talk, release

Hold fn while you speak and let go to insert. Tap it once instead to talk hands-free, then tap again to finish. Prefer another key? Use right ⌥, right ⌘, right ⌃, or record any shortcut.

• Prefer the mouse? Use the middle button or a side button the same way: hold to talk, click for hands-free.
• Esc cancels a recording; nothing is typed.
• fn still works for fn shortcuts: pressing fn with another key (like fn+←) cancels the dictation instead of recording.
• Sayburst warns you if macOS is set to open the emoji picker on the 🌐 key and shows where to change it.
• A small overlay shows your words as you speak, elapsed time and a live level meter, with stop and cancel buttons. On MacBooks with a notch, it grows out of the notch. It never steals focus.
• Auto-send, per app: Sayburst presses Return after inserting, so a chat message goes out the moment you let go.
• Optionally mutes your speakers while you talk, so music doesn't end up in the transcript.
• Pick any microphone. If it's unplugged mid-recording, Sayburst stops, keeps what you said, and tells you.

Sayburst Writer: AI polish that never leaves your Mac

Sayburst Writer is a small AI model (Qwen3 4B) that runs on your Mac's GPU. After you speak, it turns talk into writing:

• Self-corrections: “Tuesday at 3… actually, no, Wednesday” becomes “Wednesday at 3”.
• Formatting: emails get a greeting, paragraphs and a sign-off; spoken lists become bullets.
• Your words, your language: it doesn't summarize, add content or translate. It treats what you say as text to clean, never as instructions to follow.
• A safety check: if a rewrite would change a number you said, Sayburst types your cleaned transcript instead.
• Never in your way: press Skip while it's polishing to insert your cleaned text right away.

An optional 2.5 GB download. No account, no API key, works offline. Best on Macs with 16 GB of memory.

Writing modes

Every dictation passes through a mode. Cleanup runs on your Mac in milliseconds; modes with AI polish then hand the text to your writer.

Clean

Removes fillers (“um”, “uh”) and accidental repeats, fixes capitals and punctuation, understands “new line”, “new paragraph” and “bullet point”, and resolves self-corrections.

Message

Short and casual, no forced period at the end — for chat.

Email

Greeting, short paragraphs, sign-off on its own lines.

Notes

Short lines and bullets when you list things.

Raw

Exactly what the model heard.

Choose a mode per app — Email for Mail, Message for Slack — and override it for a single dictation from the menu bar. Each mode's settings and instructions are editable, and you can add your own.

Rewrite and ask, or bring your own model

Sayburst Writer handles polishing, rewriting and questions on your Mac. Prefer another model? Use Ollama or LM Studio, or OpenAI, Anthropic, OpenRouter, Groq or any OpenAI-compatible service with your own API key.

• Polish: every built-in mode has AI polish on; turn it off, or edit its instructions, per mode.
• Rewrite selection: select text, hold the rewrite shortcut, say “make this friendlier” or “turn this into bullet points”. The selection is replaced.
• Ask: hold the ask shortcut and ask a question. The answer appears in a panel with Copy and Insert buttons; nothing is typed until you choose.
• Context, if you allow it: selected text, clipboard text, or the current window's text can be included. Each is off by default.

Before a cloud provider is turned on, Sayburst shows exactly what will be sent. Audio is never sent. Sayburst never runs commands that a model or your speech suggests.

Vocabulary and replacements

Add names, products and jargon so they're spelled your way. Whisper models are nudged toward them while transcribing; with Parakeet, Sayburst corrects near-misses like “Saberst” → “Sayburst” — but never changes real words.

It learns as you fix. Correct a name by hand after dictating and Sayburst offers to add it to your dictionary, so it's right next time. It looks only at the field you dictated into, never at password fields, and keeps only the words you add (and the ones you declined, so it doesn't ask twice).

Replacements turn exact phrases into anything: “my email” → your address, “sig” → your signature. Import and export as JSON or CSV.

History and recovery

Every dictation is saved to a searchable history on your Mac. See what the model heard next to what was inserted, play back the recording if you kept it, copy, export (Markdown, JSON, CSV), delete, or run it again with a different model. Your stats — words written, time saved, speaking speed — live there too.

Text is never lost: if it can't be typed — you switched windows, a password field is focused, Accessibility is off — Sayburst keeps it and offers it back. ⌃⌥⌘V pastes your last dictation anywhere.

Choose how long to keep history (forever down to one day) or turn it off. Recordings are deleted after transcription unless you choose to keep them.

Notes and files

Capture a note: hold the notes shortcut and speak. Sayburst adds a timestamped bullet under today's date in a Markdown file you choose — handy with Obsidian or any notes folder.

Transcribe a file: drop an audio or video file onto the History window to transcribe it with the same local model.

Local speech models

Parakeet v3 is the default: fast and accurate in English and 24 other European languages, running on the Neural Engine. Whisper models cover 99 languages, including Chinese, Japanese, Korean, Hindi and Arabic. Downloads are verified against published checksums, can be cancelled and resume where they stopped. Compare models.