Offline dictation for macOS. Your audio never leaves your Mac. No subscription.
Wispr Flow was caught screenshotting your screen every few seconds. VOICE does the opposite: Parakeet ASR runs fully on-device, polish runs on-device via Qwen3, and audio is never transmitted anywhere unless you explicitly route a long input to cloud. Zero telemetry.
You: "add this to agenda for tomorrow standup"
VOICE: "Add this to the agenda for tomorrow's standup." (200ms)
| VOICE | Wispr Flow | SuperWhisper | |
|---|---|---|---|
| Offline transcription | ✅ On-device Parakeet | ❌ Cloud only | ✅ |
| Offline polish | ✅ On-device Qwen3 | ❌ Cloud only | ✅ (setup required) |
| Meeting notes | ✅ No bot required | ❌ | ❌ |
| Polish in-place | ✅ | ✅ | ❌ |
| Screenshots your screen | ❌ Never | ✅ Every few seconds¹ | ❌ |
| Pricing | Free / open source | $15/mo | $249 one-time |
| App Store | ❌ | ✅ | ✅ |
¹ Wispr Flow's CTO publicly acknowledged the behavior after users discovered it. The app screenshots your screen every few seconds for "context awareness."
Hotkey dictation Hold to push-to-talk. Double-tap to lock hands-free. Floating pill overlay shows recording state.
On-device ASR Parakeet TDT 0.6B handles transcription locally. Whisper-class accuracy, no network hop.
On-device polish Qwen3 1.7B or 4B runs via MLX on Apple Silicon. Complex or long inputs route automatically to Cerebras (Qwen3-235B) in the cloud. You control this threshold.
Cleanup levels None / Light / Medium / High. Medium is the default.
Personality modes Neutral / Formal / Casual / Excited. Neutral matches your tone. Formal decontracts. Casual keeps slang. Excited adds energy without adding fluff.
My Style Paste writing samples. VOICE extracts your stylistic patterns and injects them into every polish call.
Polish in-place Select any text in any app, press ⌥1. Cerebras rewrites it with three presets: Fix (grammar only), Smooth (flow), Elevate (stronger word choice).
Meeting notes Captures Google Meet, Zoom, and Teams audio via ScreenCaptureKit. No bot. No browser extension. No calendar integration required. When the call ends, VOICE transcribes all participants and auto-summarizes with key decisions and action items.
- Parakeet TDT (on-device ASR, under 1s) — speech recognition
- TextFormatter (regex layer) — fixes common homophones before the LLM sees the text
- Qwen3 LLM (1.7B local or 235B cloud) — grammar, tone, formatting
- PostProcessor (deterministic rules) — formatting cleanup for lists, decimals, punctuation
- Clipboard + Paste — inserts into whatever app is active
git clone https://github.com/fortun8te/voice.git
cd Voice
make installPress Cmd+Right, speak, release. Done.
For cloud API setup, stable code signing, or model swapping: SETUP.md.
# Reset permissions if stuck
RESET_TCC=1 make install- Privacy: audio never leaves your Mac for standard dictation. Cloud only fires on explicit routing or Polish in-place.
- Latency: 200ms on M-series. Cloud-only tools add 2-3s round trips.
- No subscription: Qwen3 1.7B is 530MB. It runs on the hardware you already own.
- No account: nothing to log into, nothing to revoke.
- Offline: works on a plane, in a basement, anywhere.
- Icon flickers -- Icon Troubleshooting
- TCC re-prompts -- Stable Code Signing
- Quiet input -- 5x gain boost is applied automatically for low-volume microphones
MIT. Use it, fork it, sell it.
Made by fortun8te.