LocalFlow streams a draft transcript while you speak, so releasing the key usually pastes instantly. The final on-device pass runs in ~61 ms — cloud dictation is still waiting on the network.
Audio goes from the microphone to the Neural Engine to your cursor — a closed loop on your own silicon. There is no server to trust, because there is no server.
The full dictation pipeline, reimplemented with local components in native Swift.
Release the key and the draft pastes in 0–5 ms via simulated ⌘V — clipboard restored afterwards.
Parakeet detects the language natively, mid-sentence. No toggle, no settings dive.
The audio-so-far is re-transcribed several times per second, so you watch your words arrive as you speak.
Right Option by default; remap to ⌘, ⌃ or F13–F20. Taps under 0.3 s are ignored.
Rule-based cleanup strips the ums and ähs offline — or hand polishing to a local LLM via Ollama.
Slack, Xcode, your browser, a terminal — text lands wherever your cursor is.
Press and hold the hotkey. A pop sound confirms recording.
16 kHz capture; the glass pill shows your level and a live draft.
Parakeet TDT runs on the Apple Neural Engine — CoreML, no GPU spin-up.
Clean text appears at your cursor. Bottle sound. Done.
macOS 14+ on Apple Silicon. Downloads the on-device model once, then works fully offline.
Download for macOSUnsigned for now — the release page has a one-time first-launch step, or build from source.