Offline dictation for Mac, from setup to insertion

The current MumbleFlow release requires Apple Silicon and macOS 26 or later. It is a native Swift app built with AppKit and SwiftUI, not a browser shell. Intel Macs and older macOS releases cannot run this build.

Dictation needs the Microphone permission. Global Fn handling and text insertion need Accessibility access. Oat requests Screen & System Audio Recording only when you choose to record a meeting. These are distinct permissions, so basic dictation does not require meeting-audio access.

MumbleFlow uses FluidAudio's Silero VAD and the English Parakeet TDT 0.6B v2 Core ML model. Downloaded model assets are versioned and verified before use. The current speech recognizer is English-only.

1. Install MumbleFlow. Open the signed DMG, drag the app to Applications, and launch it so onboarding is visible.
2. Allow the microphone. Speak while watching the input indicator to confirm that the selected device is producing a signal.
3. Allow Accessibility. If macOS has just granted access, return to MumbleFlow and let the status refresh before testing insertion.
4. Download the speech assets. Wait for the English model status to show ready before disconnecting.
5. Choose polishing behavior. Use Apple Foundation Models when available, install the optional local Qwen3 fallback, or turn polishing off for raw transcription with deterministic whitespace normalization only.
6. Add important vocabulary. Put names, acronyms, and uncommon terminology in the local personal dictionary before a real test.

For the architecture behind these steps, read how local speech to text works.

• Hold Fn, speak, then release: the capture begins on the first key-down and ends as soon as you let go.
• Double-tap Fn: recording locks for a longer passage. Press Fn once more to finish and insert.
• Press Escape: cancel the current capture without inserting or saving it.
• Use the fallback shortcut: the default is Option-Space when Fn is reserved by a keyboard or system feature.

The floating pill shows an audio-reactive waveform while listening and a processing state afterward. MumbleFlow waits for the final transcript rather than typing unstable partial words into your document.

Start in a local editor. Dictate a short sentence, a correction, and a numbered list while online. Confirm that the final text appears at the cursor, reaches the clipboard, and appears in Dictation History. Then turn Wi-Fi off and repeat the same samples.

Test the two processing paths separately. Raw transcription should work as long as the Parakeet assets are present. Polishing should also work offline when Apple Foundation Models are available on the system or the optional Qwen3 fallback is installed. A very short phrase may intentionally skip language-model analysis to reduce waiting time.

Next, repeat the insertion test in the browser or desktop apps you use every day. A cloud editor may still need internet to load or sync even though MumbleFlow's transcription is local. For a strict privacy test, use an unsynced local document and follow the private dictation checklist.

• No pill appears: verify that MumbleFlow is running, Accessibility is enabled, and Fn does not conflict with macOS Dictation or another utility. Try Option-Space.
• The waveform does not move: check the microphone permission, input device, mute state, and System Settings input meter. Reconnect a Bluetooth microphone if its route changed.
• The pill turns red after silence: no speech was detected. Try again closer to the microphone; there is no error dialog to dismiss for a silent capture.
• The model is unavailable: reconnect, reopen onboarding or Settings, and finish the verified speech-model download.
• Text is copied but not inserted: refocus the editor, confirm Accessibility access, and test another field. The clipboard preserves the result when an application rejects automated insertion.
• Polishing is disabled: check the on-device intelligence status. Raw dictation still works; enable Apple Intelligence or install the local Qwen3 fallback for semantic cleanup.

If the same stage continues to fail, contact the MumbleFlow team with the Mac model, macOS version, app version, and processing mode. Do not attach confidential recordings.