raw in · polished out
Say it messy.
It lands clean.
Fillers dropped. Grammar fixed. Meaning kept — by your local model.
01The flow
Press, speak, release.
No app to open, no window to switch to. Outspoken lives behind a hotkey and types where your cursor already is.
Hold the hotkey
Anywhere, in any app. A small overlay appears at the bottom of your screen.
Speak naturally
Ums, false starts, mid-sentence pivots — say it the way you'd say it out loud.
Release
whisper.cpp transcribes on-device. Your local model tidies grammar, punctuation, and formatting.
Text lands in place
Polished text lands at your cursor in one atomic insert — and your clipboard is restored right after.
See it live
Talking beats typing.
02Privacy model
Local by design, not by settings toggle.
There is no Outspoken cloud. Audio is captured and transcribed in-process. If you enable rewriting, the transcript goes to a model server on localhost — anything that isn't 127.0.0.1 is refused outright, redirects included, unless you explicitly approve your own server: one host, spelled out by you, off by default. The only other network access: Whisper model downloads you start yourself, checksum-verified, from huggingface.co. Every connection is listed in Settings → Network activity.
Where your data actually lives
- audio
- Processed in memory. Never written to disk, never uploaded.
- transcripts
- Saved only if you turn on History — plain local files under
~/Library/Application Support/Outspoken/. - dictionary & snippets
- One local JSON file. Edit it by hand, version it, back it up.
- rewrite prompts
- Sent only to the endpoint you configure. Default: nowhere.
- account
- None. Download, drag to Applications, talk.
03Every app
If you can type there, you can speak there.
Outspoken types wherever your cursor is — native apps, Electron apps, browser text fields — and matches the register of the room: full sentences in email, loose in chat, a tidy comment in your editor.
Hi Priya — the revised SOW is attached, with the timeline moved up a week. Could you review section 3 before Friday?
shipping the fix now — should be live in ~10 min. will post in #incidents when it lands
Decision: consolidate both onboarding flows before the October release. Owner: Dana. Revisit at the next planning review.
// Debounce resize events so the chart doesn't re-render on every frame
Date picker resets to UTC when the browser locale is set to JST
coffee roasters near bernal heights open before 7am
- Mac
- iPhonein the works
- Windowson the roadmap
Write faster in every app, on any device.
Seamless speech-to-text wherever you type — on your Mac today, on your iPhone next. And the rule travels with it: your voice never leaves your hardware.
04Bring your own models
Plug in your own AI server. Or none at all.
Speech-to-text is built in and runs in-process. For cleanup — grammar, punctuation, formatting, tone — point Outspoken at any OpenAI-compatible endpoint on this Mac: Ollama, a llama.cpp server, LM Studio. Prefer verbatim? Switch rewriting off and get exactly what you said.
Transcription
Rewriting
Network
Works fully offline
Unplug the router mid-sentence. Nothing changes.
Models are files, not subscriptions
Whisper model files (GGML) on your disk. Swap or delete them anytime — downloads happen in-app, SHA-256-verified.
Raw mode
Turn rewriting off for verbatim transcripts — quotes, minutes, legal.
05The details
Built for people who dictate all day.
Custom dictionary
Teach it “Kubernetes”, “Priya”, “PostgREST” once. Lives with your snippets in one local JSON file you can edit by hand — changes apply instantly.
Voice snippets
Say “insert my address” and the whole block lands — sign-offs, links, canned replies. Snippet contents never reach the AI server.
Formatting commands
“New paragraph.” “Bullet point.” “Question mark.” Structure and punctuation are spoken, not typed — in English and German.
App-aware tone
Mail comes out professional, Slack stays casual, your editor gets a code comment — picked from the app in front. One toggle turns it off.
Raw transcript mode
One toggle for verbatim output — every word as said, no model in between.
History, memory-only
Recent dictations stay in memory — nothing touches disk unless you flip “Save history”. The menu keeps score: words dictated · 0 sent to the cloud.
Live translation
Dictate in German, English lands — translated on-device by Whisper. Other target languages (French, Spanish, Italian, …) run through your local AI server.
Secure fields stay off-limits
When macOS flags secure input — password fields, login prompts — nothing is typed or pasted. Ever.
Clipboard-safe insertion
Text lands in one atomic insert — nothing can be dropped or doubled on the way. Your clipboard is restored automatically right after, verified by a change check.
06Get the beta
Download Outspoken.
Outspoken 0.3.1 is a free beta for macOS 14 or later on Apple silicon (M1 and newer). The disk image ships with the tiny multilingual Whisper model bundled, so it dictates the moment you drag it into Applications — bigger models are one click away in Settings.
Outspoken-0.3.1.dmg · ~87 MB · no account · no installer
Signed & notarized by Apple: every download is Developer-ID-signed and has passed Apple's notarization scan — it opens like any Mac app, no warnings, no workarounds. macOS will simply ask once for microphone and accessibility permission, which is what makes dictation work.
07Questions
Fair questions.
How is this different from Wispr Flow?
Cloud dictation tools like Wispr Flow send your audio to their servers for transcription and cleanup. Outspoken does the same job — hotkey, polish, every app — with models that run on your machine. No audio, text, or context ever reaches us, because there's no “us” in the loop at all.
Does it work offline?
Entirely. Transcription is in-process and rewriting talks to your own endpoint, so airplane mode is a normal Tuesday. The only thing that ever needs a connection is downloading a model file once.
What does the rewrite step send, and where?
The transcript plus your style rules, to the endpoint you configured — typically http://localhost:11434, and never anything non-local. Snippet contents are swapped for placeholders before the request, so they never leave the app. Turn rewriting off and no request is made at all.
Which models does it use?
Speech-to-text runs on the whisper.cpp family, from tiny to large-v3 — including large-v3-turbo and a quantized q5 build — pick your speed/accuracy trade-off. Downloads happen in the app, SHA-256-verified. Rewriting works with any chat model behind an OpenAI-compatible API on this Mac: Ollama, llama.cpp, LM Studio, vLLM.
Can it translate while I dictate?
Yes. Speak German and English comes out — Whisper translates to English on-device, no server involved. For other target languages (French, Spanish, Italian, …) the transcript goes through your local AI server instead — still never off this Mac.
What are the hardware requirements?
Any Apple silicon Mac running macOS 14 or later. The default small model (~490 MB) decodes well over 30× realtime in our M-series measurements; lighter and larger models are one click away in Settings.
Windows or Linux?
macOS first. Windows and Linux are on the roadmap — join the waitlist and say which one you need.
08Early access
Be first in line.
Outspoken is in free beta on macOS (Apple silicon). Leave an email — it's the only thing we'll ever ask you for.
You're on the list. One email when it's ready — that's it.
one launch email · no tracking pixels · address never shared