Local speech to text built for your Mac

MumbleFlow is written in Swift, with AppKit handling global keyboard events, microphone capture, Accessibility-based insertion, and the floating dictation pill. SwiftUI powers onboarding, settings, dictation history, search, and Oat's meeting views. It is not an Electron or Tauri wrapper.

The speech pipeline uses FluidAudio's Silero VAD to find speech and silence, then runs the English Parakeet TDT 0.6B v2 Core ML model on the Mac. VAD prevents long stretches of silence from becoming transcript noise, while Core ML lets the recognizer use Apple's on-device hardware efficiently.

The app and speech assets are separate: install the signed Mac app first, then download and verify the versioned model files during onboarding. Once those assets are present, dictation does not need a remote transcription service. For a model-by-model comparison, see MumbleFlow versus whisper.cpp.

1. Hold Fn for a quick thought. Recording starts on key-down so the opening word is not lost. Release Fn to stop immediately and process the utterance.
2. Double-tap Fn for longer dictation. The capture locks on, so you can speak without holding a key. Press Fn again to finish.
3. Press Escape to cancel. The active capture is discarded rather than transcribed or inserted.
4. Watch the compact pill. Its waveform responds to your input level, then changes to a processing indicator while recognition and optional polishing run.
5. Receive one stable result. MumbleFlow inserts the completed text at the active cursor and also makes it available on the clipboard and in local dictation history.

A configurable fallback shortcut is available when Fn conflicts with a keyboard or macOS setting. The offline Mac dictation guide covers setup and permission checks.

Speech recognition answers “what words were spoken?” Polishing answers “how should those words read?” Keeping those jobs separate lets you disable semantic rewriting when you want a close transcript or enable it for messages, paragraphs, corrections, and spoken lists.

When polishing is enabled, MumbleFlow can remove fillers and abandoned starts, resolve a correction such as “five—actually six,” add punctuation and paragraph breaks, and turn an explicitly spoken sequence into a numbered list. Short utterances can skip language-model polishing to reduce delay. Names, numbers, links, technical terms, tone, and substantive detail are meant to be preserved.

MumbleFlow uses Apple's on-device Foundation Models when they are available. If Apple Intelligence is unavailable, an optional local Qwen3 model can provide on-device polishing instead. Raw dictation remains available without either language model, and deterministic whitespace cleanup is used when polishing is off.

A local personal dictionary helps MumbleFlow recognize the uncommon names, acronyms, and vocabulary you use. Its entries and your encrypted dictation history stay on this Mac. Learn more about those boundaries in the private voice-to-text guide.

Local processing is a meaningful privacy and offline-use choice, not a guarantee that every transcript will be faster or more accurate. A noisy room can hurt either approach. The destination also matters: text inserted into a cloud document is governed by that document provider after insertion.

Use the same microphone and three repeatable samples: a short message, a paragraph with a self-correction, and a structured request such as a three-item to-do list. Add a name, number, URL, or technical term that matters in your work. Then compare the raw transcript, polished result, insertion reliability, and time from release to usable text.

Also test the apps you write in most often. MumbleFlow first targets the focused editable element through macOS Accessibility and can fall back to a paste workflow when an app exposes its editor differently. Browser editors, native apps, and Electron apps do not all present text fields the same way.

If you also need long-form meeting capture, MumbleFlow includes Oat: an explicit meeting recorder that combines microphone and ScreenCaptureKit system audio, builds a speaker-attributed transcript, and creates cited summaries. Explore how that differs from quick dictation on the MumbleFlow product page.