Perfect text appears at your cursor, in any app.
Pick any modifier — Left/Right Option or ⌘. Hold to push-to-talk, or flip on Toggle mode to tap-start / tap-stop.
Talk normally. Fillers and "um"s are trimmed, sentences get capitalised and punctuated, custom replacements applied. 99 languages.
Text appears at your cursor in any app — Slack, Gmail, Notion, Xcode, Terminal. Saved to searchable history automatically.
Ten compact overlays — all 180×42, all voice-reactive, all optional. Flip between them in Settings → General whenever you want a different look.
Position the overlay, hide it entirely, or add subtle audio cues. Everything off by default — turn on what suits you.
Bottom-center, top-center, or bottom-right. Drag-to-snap onto any display.
Hide the overlay entirely. Status lives quietly in the menu bar. Perfect for screen-shares.
Subtle ticks on start and paste if you want them. Off by default.
Paravoice runs Whisper directly on your machine. No API keys. No accounts. No uploads. Airplane mode works fine.
WhisperKit + CoreML on Apple Neural Engine. Turbo is 8× faster than Large at 99% of the accuracy. 30 seconds of audio in under 500ms on M2.
Auto-detect on, mid-sentence switching, code-switching handled.
helloTiny · Base · Small · Medium · Large v3 Turbo. Swap anytime.
Works on flights, in tunnels, at conferences. Always.
Apple's voice-processing pipeline removes keyboard clicks, HVAC hum, and café chatter before Whisper even sees it.
Built-in VAD auto-stops after a configurable silence window. Stop talking, text appears — no key-release needed.
Paravoice cleans up your voice automatically — fillers trimmed, sentences capitalised, punctuation inferred, and your personal dictionary applied. Every toggle is optional.
"um hey sarah like quick update we're uh on track to ship thursday qa finishes the regression suite tomorrow morning i'll send the uh gh pr link by end of day"
"Hey Sarah, quick update— we're on track to ship Thursday. QA finishes the regression suite tomorrow morning. I'll send the GitHub PR link by end of day."
"Um", "uh", "like" — gone. Mid-sentence restarts collapsed.
First letter of every sentence, proper nouns, and "I" — always correct.
Commas, periods, and question marks inferred from cadence and syntax.
"gh pr" → "GitHub PR". "k8s" → "Kubernetes". Your dictionary, your rules.
Every dictation saved locally via SwiftData. Re-paste from the menu bar.
All processing happens on your Mac in under 10ms. Never slows the insert.
The defaults are opinionated and good. But every knob is there if you want to fiddle.
Four modifier options out of the box. Toggle mode turns any of them into tap-to-start / tap-to-stop.
Pick the Whisper variant that fits your Mac. Turbo is our default on M-series. Smaller models trade accuracy for memory.
Force English, Spanish, Japanese — or leave it on auto-detect.
Temperature, beam size — for people who know what those mean.
Auto-pause after 800ms. Or 2s. Or never.
CGEventTap native. Works in any app, any window, any Space.
Inserts at cursor via Accessibility API — not the clipboard.
Live indicator. Idle → listening → transcribing → paused.
Built for M-series. Neural Engine accelerated. ~145 MB + model.
One toggle via SMAppService. Always ready from the menu bar.
Every transcript stored locally. Search, re-paste, export, delete.
Five steps. Permissions, model download, hotkey pick, a quick try.
Signed, notarised, sandbox-aware. TCC permissions survive updates.
"I write 3,000 words of documentation a day. Paravoice cut that to under two hours."
"Finally a voice-to-text tool that doesn't ship my voice to someone's cloud. And it's faster."
"I dictate all my PRs now. Shipping tech specs 4× faster. Never going back."
Yes. Paravoice ships WhisperKit and runs it locally — no network code in the audio path. The one-time model download on first launch is the only outbound request, and even that can be pre-seeded. Verify with Little Snitch if you want.
Large v3 Turbo (default) — fastest and most accurate on any M-series Mac. Small if you want lower memory on older M1s. Tiny for the smallest download if you don't need peak accuracy.
Yes — Left Option, Right Option, Left Command, or Right Command. Set it in Settings → General. Flip on "Toggle mode" to use it as tap-to-start / tap-to-stop instead of push-to-talk. A min-hold threshold prevents accidental triggers from short taps.
Pill (default), Orb, Robot, Neural mesh, Iridescent sphere, Minimal pill, Ring, Vinyl, Soundwave, and Halo. All 180×42. All voice-reactive (they pulse with your RMS level in real time). Pick any of them in Settings → General, or hide the overlay entirely with Stealth mode.
99 languages via Whisper multilingual. Auto-detect on, mid-sentence switching works, code-switching handled. You can also lock a specific language if auto-detect guesses wrong.
Locally, in a SwiftData store inside ~/Library/Application Support/Paravoice/. Browse and re-paste from the menu bar dropdown. Delete one, delete all, or turn off history entirely in Settings → Privacy.
Yes. Settings → Text Processing → Custom replacements. Add pairs like gh pr → GitHub PR or k8s → Kubernetes. Applied after transcription, before paste.
No — Paravoice is Apple Silicon only (M1 or later). WhisperKit leans on the Neural Engine for real-time performance, so Intel support isn't planned.
Paravoice for Teams is coming. Single sign-on, MDM profile, admin policies. Email teams@paravoice.ai.