<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[MumbleFlow]]></title><description><![CDATA[MumbleFlow]]></description><link>https://mumbleflowapp.hashnode.dev</link><generator>RSS for Node</generator><lastBuildDate>Wed, 18 Nov 2026 15:06:15 GMT</lastBuildDate><atom:link href="https://mumbleflowapp.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[Private dictation on Mac in 2026: what “local” really means and how to check it]]></title><description><![CDATA[What “local” dictation on a Mac actually means in 2026, how to check it yourself, and an honest look at MumbleFlow’s limits.]]></description><link>https://mumbleflowapp.hashnode.dev/private-dictation-mac-2026-what-local-means</link><guid isPermaLink="true">https://mumbleflowapp.hashnode.dev/private-dictation-mac-2026-what-local-means</guid><category><![CDATA[macOS]]></category><category><![CDATA[privacy]]></category><category><![CDATA[dictation]]></category><category><![CDATA[Productivity]]></category><category><![CDATA[AI]]></category><dc:creator><![CDATA[MumbleFlow]]></dc:creator><pubDate>Wed, 18 Nov 2026 01:22:12 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6afbc9df962ee4fc1ea07ba5/f1f7af51-a505-47f3-8d18-890e487e7fa1.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Disclosure: I build <a href="https://mumble.helix-co.com">MumbleFlow</a>, a local dictation and meeting-notes app for Mac. This is a checklist you can run yourself, not a benchmark.</em></p>
<p>“Local” and “private” appear on almost every dictation page now, and they do not all mean the same thing. Some apps recognize speech on your Mac after you download a model. Some upload every utterance, then delete the audio. Some default to the cloud and offer an on-device option in settings. Some keep recognition local, then send the transcript elsewhere for cleanup.</p>
<p>If you care where your voice goes, the adjective is not enough.</p>
<h2>What to verify</h2>
<p>Five questions matter more than the landing page:</p>
<ol>
<li><strong>Where does the audio go?</strong> Is recognition happening on the Mac, or is the recording uploaded — every time, or only for optional features?</li>
<li><strong>What is retained, and by whom?</strong> “Discarded after transcription” is not the same as “never left the device.” Training, cloud history, and “improve the model” toggles are separate questions.</li>
<li><strong>Do you need an account?</strong> An account is not automatically a problem, but it usually means a server-side identity comes with the product.</li>
<li><strong>What is the pricing model?</strong> A subscription can be a fair price for a maintained product, and often comes with cloud features. A one-time price does not prove the app is local either.</li>
<li><strong>Where does the finished text go?</strong> Even a perfectly local recognizer hands text to whatever app is focused. A browser editor, syncing notes app, or clipboard manager can move that text afterward.</li>
</ol>
<p>If the product also records meetings, add two more: does recording start only when you ask, and does a bot join the call?</p>
<h2>How to check on your own Mac</h2>
<p><strong>Start with the offline test.</strong> Install the app, finish any required model download, then turn off Wi-Fi and Ethernet. Open TextEdit, start a dictation, and speak a short sentence. If the app errors, spins, or refuses to start, it needed the network for that recognition. If a finished sentence appears, recognition for that session did not depend on a live connection.</p>
<p>That is the most useful test I know, and it needs no extra software. It is not complete: an app can still connect later for updates, license checks, analytics, or a cloud history you did not notice, and it will not catch a cleanup feature that uploads text once you are back online.</p>
<p><strong>Watch the network while you speak.</strong> <a href="https://www.obdev.at/products/littlesnitch/index.html">Little Snitch</a> is a macOS firewall that can show or block outgoing connections per application. Start a dictation and see whether the app opens a connection as you speak or when the transcript appears. Activity Monitor’s Network tab is coarser, but a burst of bytes sent during a short dictation is worth a closer look. Do not rely on the built-in macOS firewall alone: it is primarily an incoming-connection filter, not a log of uploads.</p>
<p><strong>Read the settings already on the Mac.</strong> For Apple’s built-in Dictation, open System Settings → Keyboard, enable Dictation, and read the text directly under it. <a href="https://support.apple.com/guide/mac-help/use-dictation-mh40584/mac">Apple’s guide</a> says that text tells you whether voice inputs and transcripts for general text Dictation — composing messages and notes, but not dictating in a search box — are processed on your device and not sent to Siri servers.</p>
<p>When you enable Dictation, macOS may ask whether you want to improve Siri and Dictation. “Share Audio Recordings” allows Apple to store audio of those interactions, and Apple may review a sample. You can decline, and change it later under System Settings → Privacy &amp; Security → Analytics &amp; Improvements. <a href="https://www.apple.com/legal/privacy/data/en/ask-siri-dictation/">Apple’s Siri, Dictation &amp; Privacy page</a> is the companion document.</p>
<p>Also check Privacy &amp; Security for Microphone and Accessibility, which almost every system-wide dictation app needs, and Screen &amp; System Audio Recording if a meeting recorder captures computer audio. Grant only what the feature you are using requires.</p>
<p><strong>Read the in-app data controls, not just the homepage.</strong> Look for separate switches for model training, cloud history, and features that read the text around the cursor. Check the privacy policy for usage statistics, support reports that attach audio, and third-party AI providers. Then try one dictation without creating an account. If you cannot, the product has a server-side identity before you have said a word.</p>
<h2>What the official pages actually say</h2>
<p>Here is what the makers publish on the pages I checked, without inventing a feature matrix.</p>
<p><a href="https://wisprflow.ai/privacy">Wispr Flow</a> is explicit that transcription happens in the cloud. Audio is sent over TLS, transcribed on their servers, and deleted after transcription unless you turn on Dictation cloud storage. You can turn off “Improve the model for everyone,” which they say stops your audio, transcripts, and edits being used to train their models. That is a real control, not on-device recognition. <a href="https://wisprflow.ai/data-controls">Their data controls</a> state this plainly: “Transcription always occurs on the cloud.” They also note that they may collect usage statistics such as word counts regardless of those controls, and that Notetaker transcripts are stored in their cloud for cross-device sync.</p>
<p><a href="https://superwhisper.com/is-superwhisper-safe">Superwhisper</a> can go either way, and the default matters. Their safety page says that with a local model, transcription runs on the device and audio never leaves it; with a cloud model, audio is sent to a server and is not stored there. It also says the app uses their cloud models by default, because they work with no download, and that you can switch to a local model. Recordings are saved on the device by default, with automatic retention as short as one day or as long as “Forever.” Their <a href="https://superwhisper.com/offline-transcription">offline transcription page</a> describes the local-model path, including use beyond the Mac.</p>
<p><a href="https://www.macwhisper.com/">MacWhisper</a> takes the opposite default. <a href="https://docs.macwhisper.com/article/52-keeping-transcriptions-private">Their documentation</a> says transcription and speaker identification are performed locally by default, after you download a model. If you use a cloud transcription provider, DeepL translation, or a remote AI prompt, that content leaves the Mac. Local tools such as Ollama or LM Studio are the exception they document for prompts. The app also covers file transcription and meetings, so check which features you actually enabled.</p>
<p>None of this is a reason to avoid a product on its own. It is a reason to match the configuration to what you thought you were getting.</p>
<h2>Where MumbleFlow fits — including the limits</h2>
<p>I built <a href="https://mumble.helix-co.com">MumbleFlow</a> for the case I wanted myself: a native Mac app that does polished dictation and meeting notes on the computer, without a product account.</p>
<p>After you download the app and the verified English speech models, recognition, writing cleanup, meeting notes, and search run on the Mac. There is no product account, analytics SDK, cloud content server, or cloud sync. Dictation audio is discarded after processing, unless you explicitly turn on the optional latest-test diagnostic. Encrypted dictation history stays for seven days by default if you leave it on; you can pin entries, clear them, or turn it off. Meeting audio expires after 30 days; transcripts and notes stay until you delete them. Oat, the meeting workspace, starts only when you choose Record Meeting. There is no meeting bot and no ambient buffer.</p>
<p>It is not invisible to the network. You still need internet access to buy the app, to download the app and models, and to get software updates. “Works offline after setup” is more honest than “never talks to a network.” Run the offline test, and a connection monitor if you use one.</p>
<p>The limits belong in the first paragraph of any fair review. This release is Apple Silicon only, macOS 26 or later, and English recognition only. Intel Macs, Windows, Linux, and phones are not supported. Cleanup can use Apple Intelligence or an optional local Qwen model, or you can keep raw transcription. External Gmail, Calendar, and Slack connections are not enabled in the current download. The price on the site is a $5 one-time purchase, with no recurring app subscription.</p>
<p>The destination caveat still applies. Text inserted into a browser editor or a syncing notes app follows that app’s rules, and clipboard managers can move a copy too.</p>
<p>If that shape matches what you need, the product and setup steps are at <a href="https://mumble.helix-co.com">mumble.helix-co.com</a>. If you need another language, an Intel Mac, or a cloud workspace that follows you across devices, the checklist above is still how I would evaluate the next app.</p>
]]></content:encoded></item><item><title><![CDATA[How I built a fully local dictation and meeting-notes app on macOS]]></title><description><![CDATA[Disclosure: I build MumbleFlow. This is a walk through the architecture, not a benchmark report. Most “AI voice” tools are a microphone attached to a…]]></description><link>https://mumbleflowapp.hashnode.dev/how-i-built-a-fully-local-dictation-and-meeting-notes-app-on-macos</link><guid isPermaLink="true">https://mumbleflowapp.hashnode.dev/how-i-built-a-fully-local-dictation-and-meeting-notes-app-on-macos</guid><category><![CDATA[Productivity]]></category><category><![CDATA[macOS]]></category><category><![CDATA[AI]]></category><category><![CDATA[SwiftUI]]></category><category><![CDATA[privacy]]></category><dc:creator><![CDATA[MumbleFlow]]></dc:creator><pubDate>Tue, 17 Nov 2026 03:32:43 GMT</pubDate><content:encoded><![CDATA[<p><em>Disclosure: I build <a href="https://mumble.helix-co.com">MumbleFlow</a>. This is a walk through the architecture, not a benchmark report.</em></p>
<p>Most “AI voice” tools are a microphone attached to a hosted API — a poor fit if you are dictating a product discussion or a meeting you do not want transcribed elsewhere.</p>
<p>I wanted a different shape: hold a key, speak the rough version, and have finished writing appear in the app I am already using, with recognition, cleanup, meeting notes, and search all running on the Mac after setup. That product is <a href="https://mumble.helix-co.com">MumbleFlow</a>.</p>
<h2>Recognition and cleanup are different jobs</h2>
<p>Recognition answers “what words were spoken?” Cleanup answers “how should those words read?” If you collapse those into one step, you cannot tell whether a bad result came from the speech model, the rewriter, or the insertion path. MumbleFlow keeps them separate.</p>
<p>The speech stack comes from FluidAudio: Silero VAD to find speech and silence, then the English Parakeet TDT 0.6B v2 Core ML model for recognition. Without VAD, long pauses become transcript noise. Core ML is why Parakeet TDT belongs on Apple Silicon: the model can use the Mac’s on-device hardware instead of a generic CPU-only loop.</p>
<p>This is not a whisper.cpp wrapper. Whisper is an excellent engine for local file transcription and for experiments where you want to swap model variants yourself. A system-wide dictation app also needs key-down capture, a status UI, model verification, and reliable insertion when a text field rejects input. We built that as a native Swift 6 app — SwiftUI for the workspace, AppKit for global events, Accessibility insertion, and floating panels — and used Parakeet rather than bundling Whisper weights.</p>
<p>Cleanup is optional and conservative. Apple’s on-device Foundation Models are the preferred semantic engine: they remove fillers and abandoned restarts, follow an explicit correction such as “five — actually six,” add punctuation and paragraphs, and turn a clearly introduced sequence into a numbered list, while trying not to rewrite names, numbers, links, or meaning. If Apple Intelligence is unavailable, an optional local Qwen3 1.7B model can do the same job through MLX. You can also turn polishing off and keep raw recognition with deterministic whitespace cleanup. Short utterances can skip the language model entirely.</p>
<p>Test those two outputs separately; if you only score the final sentence, you will mis-attribute errors.</p>
<h2>The unglamorous problem is insertion</h2>
<p>Hold Fn to start on key-down so the first word is not clipped, speak, and release to process. Double-tap Fn to lock recording for a longer passage, then press Fn once to finish. Escape cancels. If Fn is claimed by a keyboard or by macOS, a configurable fallback defaults to Option-Space.</p>
<p>The result is one stable insertion into the focused field, a copy on the clipboard, and optional encrypted Dictation History — seven days by default, with pin, clear, and off. We wait for the finished text instead of streaming unstable partials under the cursor.</p>
<p>Insertion is where a lot of local STT tools feel unfinished. Native Mac apps, browser editors, and Electron apps do not expose text fields the same way. The first path is macOS Accessibility into the focused editable element; the fallback is clipboard paste. Every completed result is copied, so a stubborn editor does not mean the sentence is gone. Test that path in the apps you actually write in. A local personal dictionary holds names, acronyms, and jargon on that Mac.</p>
<h2>Meetings need a visible start, and every claim needs a source</h2>
<p>Oat, the meeting workspace inside MumbleFlow, starts only when you choose Record Meeting. There is no meeting bot, no calendar auto-start, and no ambient buffer. The microphone is your side of the conversation; selected computer audio comes from ScreenCaptureKit, without registering a video output and without retaining video. Keeping the channels separate means the microphone can be labeled “You” while remote audio is diarized into speaker turns you can rename.</p>
<p>While a meeting is still going, “What did I miss?” summarizes only the preceding five minutes, with citations back to that window. Afterward, Oat organizes topics, a timeline, decisions, next steps, and open questions. Owners and due dates appear only when they were actually spoken, and generated notes keep source IDs so you can jump to the transcript — and play the passage while encrypted audio is retained. Meeting audio expires after 30 days by default; transcripts and notes stay until you delete them. Speaker labels are suggestions, not proof of identity, and you still have to tell others in the room that you are recording.</p>
<h2>Search is retrieval first, generation second</h2>
<p>Search has to work without shipping the meeting library to a hosted index. MumbleFlow fuses two local retrievers: a blind-token FTS5 index with BM25 ranking for exact words, and Apple NLEmbedding sentence vectors for phrasing that does not match literally (“schedule change” versus “move the launch”). Equal-weight reciprocal-rank fusion combines the lists.</p>
<p>Ranked results work without a generative model. If Apple Intelligence or the optional Qwen model is available, an answer is generated only from the top retrieved chunks, and every citation ID has to match a record that was actually retrieved. Unknown IDs are rejected.</p>
<p>Storage is more specific than “the database is encrypted.” Sensitive fields are sealed with AES-GCM, lexical tokens are HMAC’d before they enter FTS5, and the USearch vector archive is encrypted separately. Keys live in This-Device-Only Keychain items. The current SQLite container is not advertised as SQLCipher-encrypted.</p>
<h2>“Local” is not the same as “never talks to a network”</h2>
<p>After the app and verified English speech assets are installed, core dictation, polishing, meeting summaries, and meeting search run on the Mac. The Mac app has no product account, analytics SDK, cloud content server, or cloud sync. Dictation audio is discarded after processing, except for an optional latest-test diagnostic.</p>
<p>Internet access is still required to buy the app, download models, and get updates. The optional Qwen3 1.7B fallback is a user-initiated download of roughly 930 MB. Automatic update checks are not enabled in the current package, and external Gmail, Calendar, and Slack connections are not enabled in this download. Local recognition also does not control the destination: text inserted into a cloud doc or a syncing notes app follows that app’s rules.</p>
<p>The current release is a Developer ID-signed, Apple-notarized app for Apple Silicon, macOS 26 or later, and English only — not Intel Macs, Windows, Linux, or mobile. It is a $5 one-time purchase, with no recurring app subscription.</p>
<p>If you are choosing between a hosted dictation product, a whisper.cpp toolchain, and a native app, skip the unverifiable feature matrix. Use the same microphone and a few fixed samples: a short reply, a mid-sentence correction, a spoken list, and a sentence with a name and a number. Score leftover edits, whether meaning changed, insertion reliability, and what still works with Wi-Fi off. For meetings, check citations and catch-up, not just whether a summary looks tidy.</p>
<p>The product site, with the local speech-to-text write-up and the privacy page, is at <a href="https://mumble.helix-co.com">mumble.helix-co.com</a>. If a specific insertion path or model behavior is wrong, I want to know.</p>
]]></content:encoded></item></channel></rss>