MODULE A
CAPTURE
TRIGGERHOTKEYHOLD / TOGGLE
GATEVAD
Configurable hotkeys for dictation and voice-to-AI. A small pill shows your live voice level — nothing else gets in the way.
TYPE: CONTROLOUT ●
Hold a key, speak, release — your words are already text, pasted where your cursor lives. Transcription happens while you talk, not after. Every backend is yours to choose.
Three modules. Voice goes in the left, text comes out the right. Everything between is swappable.
Configurable hotkeys for dictation and voice-to-AI. A small pill shows your live voice level — nothing else gets in the way.
Speech becomes text through any backend — hosted API, self-hosted server, or fully on-device. Same workflow either way.
Your own directive prompts shape the output. A vocabulary dictionary keeps names and jargon spelled the way you spell them.
Most open-source tools record first and transcribe after — speak five minutes, then watch a progress bar for two more. Grain transcribes while you speak, using rolling windows, overlap and voice-activity detection. Release the key: it pastes. Instantly.
SIMULATION — 5 MIN OF SPEECH ON AN ORDINARY MACHINE. RECORD-FIRST: ~2 MIN OF WAITING. GRAIN: PASTED ON RELEASE — REAL-WORLD DELAY IS RARELY EVEN HALF A SECOND. WINDOWS ARE SIZED SO ACCURACY DOESN'T DROP.
Idle, the entire application sits at about 100 MB. Press the hotkey and Parakeet 0.6B wake-loads instantly — it lives in RAM only while you speak, then hands the memory back on your schedule.
The ~1 GB exists only while you're actually speaking. The rest of the time, Grain is a ~100 MB background process — the UI is fully decoupled from the backend, so closing the console changes nothing.
Unload the model the moment you finish speaking, after a timer, or never. Wake-loading means even “instant” feels instant. // CLICK THE KNOB
Add every key you have — even multiple keys from the same provider. Grain rotates across them on daily quota, latency and context length, and falls through when one is down. You just keep speaking.
Every request is scored against your whole rack — keys you stacked, limits you've hit, how long your text is — and routed to whichever backend wins. When one fails, the next is already armed.
Select text anywhere on your machine, hit a shortcut, and ask: summarize it, restructure it, turn it into an email — or just ask a question about it. It runs on your configured processing layer, and the router picks the smartest model you've got.
NO ACCOUNT · BRING YOUR OWN KEYS · MIT LICENSE