Dysgraphia is a neurological condition that affects written expression. People with dysgraphia often know exactly what they want to say, but the physical act of forming letters, organizing words on a page, or sustaining the motor planning needed for typing can be slow, painful, or unreliable. The International Dyslexia Association and the U.S. National Institutes of Health both publish general information on dysgraphia as part of broader learning-difference resources.

Voice dictation bypasses the handwriting bottleneck entirely. Instead of fighting a pen or a keyboard, you speak and the computer types. For many people with dysgraphia, that's not a convenience — it's the difference between writing and not writing.

But not all dictation tools are equal for this job. Cloud services can lag, drop connections, or charge by the minute. Built-in options like Apple Dictation are free but limited in accuracy and require an internet connection for full quality. A local AI dictation app on Mac runs the speech model directly on your machine, so the same kind of voice-to-text that big cloud services offer is available without the network trip.

For accessibility use cases, "local" isn't a marketing bullet — it changes how the tool fits into your life. Four things matter most.

Privacy. When audio stays on your Mac, you don't have to wonder who hears it. For therapy notes, medical dictation, schoolwork, or anything personal, that's not a small thing.

Reliability. Cloud APIs go down, rate-limit you, change pricing, and require API keys that expire. Local models don't. They work when your wifi doesn't.

Latency. The model is already loaded in memory. There's no round trip to a server, so the time between finishing a sentence and seeing it on screen is shorter. That makes the flow feel like talking, not submitting jobs.

Cost. MetaWhisp's local mode is free and unlimited. There's no per-minute meter, no subscription required to actually use the core feature, and no account to create. For an accessibility tool, that pricing model fits the use case better than consumption pricing ever could.

The setup is short on purpose. If you can install a Mac app and grant a permission, you can finish this.

1. Download MetaWhisp from the download page and drag it to Applications.
2. Grant microphone permission when prompted. macOS will block audio capture until you allow it in System Settings → Privacy & Security → Microphone.
3. Pick a global hotkey. The default is Right Option (⌥). Hold it, talk, release.
4. Test it. Open any text field — Notes, Messages, a browser address bar — and hold the hotkey while saying a sentence. Release. The text should appear.
5. Turn on auto-paste in settings if you want the transcript to drop into whichever app is in front of you.

That's the whole installation. The model is ~950 MB and downloads the first time you run the app, then lives locally forever.

Friction matters more for accessibility tools than for any other category. If the tool requires four clicks to start, it won't get used when the writing block hits.

The global hotkey lets you trigger dictation from anywhere — Pages, Word, Google Docs, Slack, a code editor, a web form. You don't switch apps, you don't click a record button, you don't navigate a menu. You hold the key, talk, release. The whole interaction is hands-on-keyboard and voice, which matters when the alternative is a hand that's already tired.

Auto-paste takes it one step further: when you release the hotkey, the transcript lands directly in whatever text field currently has focus. No copy, no paste, no clicking into a separate window to grab the text. The transcript just shows up where you were already working.

Combined, those two features turn dictation into something you can interrupt and resume without losing your place — which is exactly how dictation is supposed to feel.

Speech is messier than writing. People pause mid-sentence, restart, use filler words, talk around a point before getting to it. For someone with dysgraphia, that unstructured flow can feel like the same problem they have with written output — only now it's spoken.

MetaWhisp's processing modes address that with four settings.

• Raw outputs exactly what you said, including "um"s and false starts.
• Correct fixes grammar and punctuation, removes filler words, and produces clean prose.
• Structured reorganizes rambling speech into a more readable paragraph format.
• Rewrite produces a polished final draft from rough notes.

Raw mode runs entirely on-device for free. Correct, Structured, and Rewrite work on the free tier when you add your own OpenAI or Cerebras API key — only the transcript text, never the audio, goes to your own API. On Pro ($30/year or $7.77/month), the same modes are built-in with no key needed.

The only first-party accuracy number we publish is from our own LibriSpeech test-clean run: 2.76% word error rate, which works out to roughly 97% accuracy on that standard benchmark. That's the same Whisper large-v3-turbo model OpenAI released and documented on the Whisper GitHub repo and the Hugging Face model card.

In my own seven-app head-to-head test on the same audio file, MetaWhisp landed around 3.7% WER — comparable to SuperWhisper, Wispr Flow, and MacWhisper, and well ahead of Apple Dictation (11–14% on the same clip). I'm not an ML researcher; that's one founder running one test, not an industry benchmark. Take it as a data point, not a citation.

Honest limit: we haven't benchmarked legal or medical domain accuracy. If you need medical-grade claims, look elsewhere until those numbers exist publicly.

Yes and yes. Local mode runs the entire speech recognition pipeline on the Apple Silicon Neural Engine using WhisperKit. The app needs no network connection to transcribe.

That matters more than it sounds.

• Hospital and clinic wifi often blocks third-party cloud services.
• School networks throttle or filter unfamiliar APIs.
• Airplane mode is, by definition, offline.
• Home internet goes down at exactly the wrong time.

An accessibility tool that fails when the network fails isn't really an accessibility tool — it's a dependency on someone else's infrastructure. Local mode removes that dependency.