AI dictation · meeting notes · local search
Every feature. One private Mac workspace.
MumbleFlow combines system-wide speech to text for Mac with Oat meeting notes. Dictate into the app you are already using, turn explicitly recorded meetings into cited knowledge, and search what was said—all with the core AI pipeline running on your Mac.
Apple Silicon · macOS 26+ · English · Models download during setup
Voice typing across macOS
Speak naturally. Insert finished writing.
MumbleFlow is a native speech-to-text app for Mac, designed for messages, documents, browser editors, task lists, and any other focused text field that accepts normal Mac input. It listens only after your shortcut, transcribes on device, applies conservative formatting, and inserts one final result.
Hold or lock Fn
Hold Fn for a short thought, or double-tap Fn to keep recording hands-free. Press Fn once to finish a locked recording; Escape cancels without inserting.
On-device recognition
Silero voice detection and the English Parakeet TDT v2 speech model turn microphone audio into text locally on Apple Silicon.
Conservative cleanup
MumbleFlow removes fillers and abandoned starts, follows explicit spoken corrections, and adds punctuation without treating rewriting as permission to change your point.
Document structure
Longer thoughts can become readable paragraphs, while clearly introduced tasks can become numbered lists instead of a single run-on block.
Meaning checks
Before insertion, validation checks help protect names, numbers, links, negation, concrete details, and the meaning present in the recognized speech.
Final text only
MumbleFlow waits for a stable result before writing to the focused field, avoiding a trail of partial recognition that changes underneath your cursor.
Short dictations still receive deterministic punctuation and cleanup but skip unnecessary semantic-model work. Longer passages can use Apple Intelligence or the optional local Qwen model. Output varies with audio quality, pronunciation, context, and the destination editor; MumbleFlow does not promise perfect recognition or a fixed processing time.
- 01
Activate
Focus the destination, then hold Fn for push-to-talk or double-tap Fn to lock recording.
- 02
Speak
Talk in your normal cadence. You can say a correction aloud, pause between paragraphs, or clearly introduce a list of tasks.
- 03
Review locally
Parakeet recognizes the speech, cleanup formats it, and meaning checks compare the proposed result with the recognized content.
- 04
Insert
The finished text goes to the currently focused field and the clipboard; encrypted history is available if you keep it enabled.
Controls and recovery
Adapt the workflow without giving up control
Voice typing is personal: a surname, internal acronym, or keyboard conflict can matter more than a long generic feature list. MumbleFlow keeps those everyday controls in the Mac app and stores personalization locally.
Personal dictionary
Save uncommon names, acronyms, product terms, and specialist vocabulary in an encrypted local dictionary. Eligible terms can use an optional acoustic rescoring model, and every entry remains editable or removable.
Dictation History
Keep encrypted final text for seven days by default, pin entries that should remain, disable history entirely, or clear every unpinned result.
Clipboard fallback
Every completed result is copied. If a particular editor rejects automatic insertion, your words are still available to paste manually.
Shortcut choices
When Fn conflicts with another tool or keyboard setting, MumbleFlow provides a configurable fallback shortcut that defaults to Option-Space.
Visible processing state
A compact frosted HUD shows listening, waveform, lock, processing, model, and no-speech states, with restrained sound cues for key transitions.
Optional local model
Eligible Macs can use Apple Intelligence for semantic work; Settings offers a removable local Qwen3 fallback when Apple Intelligence is unavailable.
For the full data path—from microphone capture through clipboard and destination app—read the private voice-to-text guide. If a shortcut, permission, model, or editor is not behaving as expected, use the setup and troubleshooting checklist.
Oat meeting capture
Record the conversation you chose to remember
Oat is MumbleFlow's local-first AI meeting-notes workspace. Start it deliberately, select the computer audio you want, and keep a readable record of the call without retaining video. Microphone and computer-audio permissions remain visible macOS controls.
Recording by choice
Oat begins only after you select Record Meeting. Choose all computer audio or one running app; there is no calendar-triggered or ambient recording buffer.
Two audio paths
Your microphone and selected computer audio are captured separately, helping Oat distinguish you from voices coming through the call.
No retained video
The computer-audio workflow does not retain a screen or camera video stream. It records the selected sound needed for the meeting transcript.
Live transcript
Follow timestamped speech while the meeting is underway, then reopen the transcript later from the local meeting workspace.
Speaker labels
Your microphone is labeled You. Oat separates remote voices, lets you correct a label, and can reuse conservatively confirmed speaker profiles on this Mac.
Five-minute catch-up
What did I miss? summarizes the preceding five minutes into topics, decisions, questions, and next steps while the recording continues.
Resume a meeting
Continue an existing meeting in the same encrypted audio and transcript record while retained audio exists, preserving its timestamp offset.
Speaker separation is an aid to reviewing a conversation, not identity verification. You can correct a remote label, and Oat stores a confirmed local speaker profile for cautious reuse. The underlying timestamped transcript remains available when a label or generated note needs checking.
Grounded meeting intelligence
From transcript to outcomes—with a way back
A transcript records the words; useful meeting notes organize what happened. Oat builds structured views while retaining source IDs, so summaries are navigational layers over the meeting rather than detached claims.
Topics
Review the subjects the conversation covered without replacing the underlying transcript.
Timeline
Scan important moments in time order and return to their supporting transcript timestamps.
Decisions
Keep decisions attached to one or more source moments instead of presenting them as unsupported model recollections.
Next steps
Surface follow-up work, including owners or due dates only when those details are present in the conversation.
Open questions
Preserve unresolved issues as questions so ambiguity is not quietly rewritten as agreement.
Source citations
Select a citation to open the supporting transcript moment and, while retained meeting audio exists, play the corresponding passage.
Generated notes depend on what participants actually said and on the available local model. They should be reviewed for consequential work. Learn how the separate audio channels, timestamps, labels, playback, and retention fit together on the meeting-transcription page.
Find it later
Search exact words, related meaning, and the source
MumbleFlow indexes meetings stored in the app with local lexical and semantic retrieval. Exact-word matching helps with names and specific phrases; meaning-aware retrieval can find a passage when your question uses different wording. The two result lists are combined before the strongest passages are shown.
- 01
Search meetings
Ask about transcript passages stored on this Mac. Ranked results work without a generative model.
- 02
Generate when available
Apple Intelligence or installed Qwen can synthesize an answer from only the highest-ranked source passages.
- 03
Validate citations
Returned source IDs must match records supplied to the model; unknown citation IDs are rejected.
- 04
Open the evidence
Follow a valid citation into the meeting transcript and supporting timestamp. Unsupported questions receive an explicit no-evidence response.
The current app searches MumbleFlow meetings only. Gmail, Google Calendar, and Slack are not connectable in this release, even though connector infrastructure exists in the codebase. See the precise current scope on the local AI search page.
Local-first by design
Keep speech, notes, and search on your Mac
Speech recognition, writing cleanup, speaker analysis, meeting intelligence, embeddings, retrieval, and generated search answers run locally after their models are installed. The desktop app has no MumbleFlow account, analytics SDK, cloud content server, cloud sync, or automatic ambient meeting recorder.
Dictation audio
Ordinary dictation audio is discarded after processing rather than becoming a permanent voice archive.
Dictation text
Encrypted history expires after seven days by default unless an entry is pinned; history can be disabled, or unpinned entries cleared.
Meeting audio
Separate microphone and computer-audio recordings are encrypted during capture and expire after 30 days unless deleted sooner.
Meeting knowledge
Transcripts, notes, citations, and confirmed speaker profiles remain locally until you delete the associated data.
Search protection
Customer fields are sealed with AES-GCM, lexical index terms are opaque, and the serialized semantic index is encrypted separately.
Deletion controls
Clear unpinned history, remove vocabulary, delete meeting audio or a whole meeting, and remove the optional Qwen model in Settings.
Local processing cannot determine what happens after text is inserted into another service. A cloud document, chat app, or shared editor applies its own storage and sharing rules. The privacy and retention page documents these boundaries, downloads, encryption details, and deletion controls.
Current-release scope
Built for a specific Mac setup
Supported
Apple Silicon · macOS 26+ · English
Required speech assets download during setup. Core voice typing and meeting workflows can then run without an active internet connection; purchases, downloads, and updates still need network access.
Not in this release
Intel, Windows, Linux, and mobile
Additional speech languages and live Gmail, Google Calendar, or Slack connections are also outside the current downloadable build. This page describes shipping behavior, not a promised roadmap.
For a model-by-model explanation, see how local speech to text works. For installation, permissions, and an offline test, follow the offline dictation setup guide.
Questions
Frequently asked questions
What features are included in MumbleFlow?
MumbleFlow includes system-wide voice dictation, on-device English speech recognition, writing cleanup, a personal dictionary, encrypted dictation history, and Oat meeting notes. Oat adds explicit microphone and computer-audio recording, live transcripts, speaker labels, five-minute catch-up, cited meeting notes, timestamp navigation, and local meeting search.
Can MumbleFlow type into other Mac apps?
MumbleFlow is designed to insert finished text into the field focused when processing ends. It supports native Mac text fields and browser editors, also copies the result to the clipboard, and can retain it in encrypted Dictation History. Individual apps and websites can still impose their own input behavior.
Does MumbleFlow work offline?
Core dictation, transcription, polishing, meeting intelligence, and meeting search run on the Mac after the required models are installed. An internet connection is still needed to buy and download the app, fetch model files and updates, and use any external source connection that may become available later.
Can MumbleFlow record and summarize meetings?
Yes. When you explicitly start a recording, Oat captures your microphone and selected computer audio without retaining video, builds a timestamped transcript, separates remote speakers, and creates cited topics, timelines, decisions, next steps, and open questions. What did I miss? summarizes the preceding five minutes while recording continues.
How does MumbleFlow protect dictation and meeting data?
Customer speech and text are processed locally. Dictation audio is discarded after processing, encrypted dictation text expires after seven days by default unless pinned, and encrypted meeting audio expires after 30 days. Transcripts and notes remain until you delete them. The Mac app has no product account, analytics SDK, cloud content server, or cloud sync.
Which Macs and languages does MumbleFlow support?
The current MumbleFlow release supports Apple Silicon Macs running macOS 26 or later and recognizes English speech. Intel Macs, Windows, Linux, iPhone, Android, and additional dictation languages are not supported in this release.
Can MumbleFlow search Gmail, Google Calendar, or Slack?
Not in the current download. MumbleFlow can search meetings captured and stored in the app. Gmail, Google Calendar, and Slack connector infrastructure exists in the codebase, but those external sources are not enabled while production credentials and provider reviews remain unfinished.
Keep exploring
The rest of MumbleFlow
AI dictation for Mac
Natural voice typing that removes fillers, catches corrections, formats lists, and inserts finished text.
Oat AI meeting notes
Reviewable, actionable meeting notes with live catch-up, speaker memory, timelines, and source citations.
Meeting transcription
Separate microphone and computer audio, readable speaker turns, timestamp playback, and grounded notes.
Local AI search
Private hybrid search and grounded answers that take you directly to the supporting source.
Privacy and security
A plain-language account of local processing, encryption, retention, network access, and deletion.
About MumbleFlow
Why MumbleFlow exists, how it is built, and the product principles behind it.
Looking for technical detail? Read the local speech-to-text guide.