Speech to text for Mac that turns talking into finished writing

Mac voice typing should not require opening a recorder, exporting an audio file, or moving text between tools. MumbleFlow lives in the menu bar and gives you two Fn gestures for short and long-form dictation. A compact frosted pill confirms that capture is active, its waveform responds to microphone level, and a processing indicator replaces the waveform after you stop.

1. 01

   Focus your destination

   Place the cursor in the message, document, browser editor, or other editable field where the finished writing should go. MumbleFlow targets the field that is focused at insertion time.

2. 02

   Speak in your own words

   Hold Fn for a quick thought or double-tap Fn to record hands-free. Ramble, pause, introduce a list, or correct yourself aloud instead of mentally composing every sentence first.

3. 03

   Finish or cancel

   Release Fn to finish a held recording, press Fn once to finish a locked recording, or press Escape to cancel. MumbleFlow processes the capture only after you finish.

4. 04

   Use one stable result

   The completed text is inserted at the cursor, copied to the clipboard, and optionally retained in encrypted Dictation History. Unstable partial words are not typed into the destination as you speak.

Quiet sound cues mark the start, locked-recording, and finish states. If a keyboard reserves Fn, the fallback shortcut provides the same capture flow. The offline dictation setup guide walks through the microphone, Accessibility, model, and shortcut checks.

Built-in dictation often treats every spoken fragment as part of the final document. MumbleFlow separates speech recognition from writing polish. The recognizer first produces a transcript; an optional on-device language model then edits that transcript conservatively.

Polishing is instructed to preserve meaning, tone, names, numbers, links, and substantive detail. It is useful for messages, emails, working notes, task lists, and first drafts—but it is not a guarantee that every correction or uncommon term will be interpreted perfectly. Review consequential names, dates, numbers, and instructions before sending.

After processing, MumbleFlow first attempts direct selected-text replacement through macOS Accessibility. It can fall back to a standard paste event when a browser or cross-platform editor exposes focus differently. Because applications control their own editors, the result is also left on the clipboard for recovery. See dictation and insertion troubleshooting for practical checks.

Dictation History keeps recent results available when it is enabled. Unpinned entries expire after seven days by default; a useful entry can be pinned or copied, and unpinned entries can be cleared. For the complete data path, read how private voice to text works.

MumbleFlow uses FluidAudio’s Silero voice-activity detector to find spoken sections and an English Parakeet TDT 0.6B v2 Core ML model for recognition. The model assets download during setup, are verified, and then run on the Apple Silicon Mac rather than calling a hosted transcription API for each utterance.

Longer text can be polished with Apple’s on-device Foundation Models when they are available. If Apple Intelligence is unavailable, an optional local Qwen3 model can provide semantic cleanup. Settings show which processing path is active, and polishing can be turned off without disabling raw transcription.

Temporary dictation audio is deleted after processing unless you enable opt-in diagnostics to retain the latest test recording. The transcript, clipboard, destination app, and optional history each have different privacy implications, so local recognition does not make text private after you insert it into a third-party service. The privacy and retention page explains those boundaries.

MumbleFlow’s current release is for Apple Silicon Macs running macOS 26 or later, with English recognition. It requires Microphone access for capture and Accessibility access for global Fn handling and text insertion. The native Swift, SwiftUI, and AppKit application is distributed outside the Mac App Store as a signed and notarized DMG.

Choose it when you want local-first voice typing, system-wide controls, optional writing polish, a personal dictionary, recoverable clipboard output, and encrypted local history. It is not currently a Windows, Intel Mac, mobile, multilingual, or cloud-sync product.

If your primary need is long-form calls rather than quick writing, explore Oat AI meeting notes and speaker-aware meeting transcription. For the whole product, visit the complete MumbleFlow feature list.