Private voice to text: follow every data boundary

1. Capture begins only when you activate it. Hold Fn for momentary dictation or double-tap Fn to lock recording. Press Escape to cancel.
2. Recognition stays on the Mac. FluidAudio's Silero VAD segments speech, and the English Parakeet TDT 0.6B v2 Core ML model creates the raw transcript.
3. Polishing can also stay local. Apple Foundation Models or the optional Qwen3 fallback can clean fillers, corrections, punctuation, paragraphs, and lists. Polishing can be switched off.
4. The final result goes to your chosen field. MumbleFlow uses macOS Accessibility when possible and a clipboard-paste fallback when necessary.
5. Dictation audio is normally deleted. The temporary microphone recording is removed after processing, except for the latest test recording when you enable opt-in diagnostics; the resulting text follows your Dictation History settings.

That last insertion step crosses an important boundary. If you dictate into a synced note, cloud document, browser form, message, or email, the destination app controls the text after it receives it. Local transcription cannot override another product's storage, sharing, or training settings.

Encryption reduces the risk of readable files at rest, but it is not a substitute for Mac login security, FileVault, software updates, backups, and physical control of the device. Another user or process with access to your unlocked account may still reach information the app can open.

Oat starts only when you select Record Meeting. ScreenCaptureKit captures the selected Mac system or meeting-app audio while the microphone is recorded as a separate timestamped stream. Video is neither captured nor retained. The microphone channel identifies the local user; remote audio is diarized to separate other speakers.

During recording, Oat transcribes completed speech chunks and can answer “What did I miss?” from the preceding five minutes. At the end it organizes a topic summary, timestamped timeline, decisions, action items, and open questions. Generated claims link back to transcript segments so you can inspect the source and, while retained, play the audio at that timestamp.

Meeting audio is encrypted locally and expires after 30 days by default; the transcript and notes can remain after audio deletion. Immediate deletion and storage controls are available. Confirmed speaker labels and voice profiles stay local so Oat can suggest a known name in later meetings rather than confidently guessing.

The native dictation and meeting inference pipelines run locally once their assets are installed. Network access is still used to download MumbleFlow, retrieve verified model assets, and receive software updates. The website and purchase flow are separate from the desktop inference pipeline.

The current download cannot connect Gmail, Google Calendar, or Slack. Read-only connector infrastructure exists in the codebase, but production credentials and provider reviews are unfinished. Until the product explicitly announces availability, MumbleFlow search covers meetings captured and stored inside the app.

Oat has no meeting bot and does not automatically join or start a call. Read the offline dictation guide to test the speech workflow separately from network-dependent downloads and destination apps.

• Check the active microphone and destination before starting, especially when a headset reconnects or several editors are open.
• Use an unsynced local document when the final text must not leave the Mac.
• Review clipboard tools and Universal Clipboard because MumbleFlow copies the final result and may use paste as an insertion fallback.
• Disable history or clear unpinned entries when your retention needs are shorter than the seven-day default.
• Delete meeting audio early when citations no longer need playback, instead of waiting for the 30-day cleanup.
• Verify source availability before relying on it. The current release searches local MumbleFlow meetings; Gmail, Calendar, and Slack connections are not enabled.
• Review names, numbers, and generated summaries before acting on them; local AI is still capable of mistakes.

Local processing is a clear architectural property, not a compliance certification or a promise against every threat. For recognition details, continue to the local STT guide.