Talk it through with an AI, out loud. It asks the right questions, looks things up on the web while it keeps talking, and writes the text you were after. Or just dictate into any app. All from a global hotkey — transcription fully offline if you want.
No window switching, no copy-paste gymnastics. Press a key, speak, done.
Speech in, speech out. Discuss what you need — no key to hold — and the assistant asks who it is for, what tone, what must be in. Say the word, and the whole conversation becomes one finished text, typed and copied. Select text first and it is the text to rework.
The same conversation, about what is on your screen — the error, the page, the text you selected. Say “write it” and the text lands at your cursor without the conversation stopping, so the next sentence can ask for a shorter one. Or type straight into the overlay; pin it on top with F9.
Need a fact mid-sentence — a price, a date, a version, a person? The assistant hands the question to a web-searching model, says it is looking it up, and keeps talking. The answer arrives between turns and lands in the final text.
Talk over the reply and it stops mid-sentence and listens — your first words are kept. Turns end on meaning, not on a stopwatch: a small on-device model reads your prosody and lets you hesitate.
Optionally, the focused window is captured and described once by a vision model, in parallel with your first sentence. You never have to read the stack trace out loud. Press S to look again.
Speak and watch your words appear at the cursor — in any application. Backed by Whisper large-v3 for state-of-the-art accuracy, with refinement levels and live translation.
Cascade: transcription → chat model → voice, every slot local or cloud, ~1.8 s. Realtime: one OpenAI speech-to-speech session, ~1.0 s, with the prosody of speech. Live: an OpenAI voice that delegates the thinking back to your own model. Presets pick a whole setup in one click.
Replies are read aloud by the voice you pick: OpenAI's multilingual voices, Groq's Orpheus, or Piper, fully on your machine — no key, no network, first sample in 0.2 s.
Dictation gives you what you said. Talk mode gives you what you meant — a spoken conversation that reasons, searches, and writes.
or, with the Realtime engine: one OpenAI speech-to-speech session holds the whole conversation — ~1.0 s and the prosody of speech, the server handles turn-taking and interruption.
or, with the Live engine: gpt-live-1 speaks and listens while delegating the thinking — the same ~1.0 s, and your own chat model answers again.
research() tool hands the question to a web-searching model on a thread; the reply is spoken between turns, never mid-sentence, and stays in the brief.Hold F3 and speak. You mumble, Loquivox polishes. Pick a refinement level and watch the same take land differently — try them:
Choose every engine — transcription, turn detection, voice. Go fully local, or plug in the cloud provider you trust.
Customizable down to the keystroke — combos supported, too.
Hold the floor. Your speech lands as text at the cursor, in whatever app has focus.
Click a glowing key — or press the real one on your keyboard. Go on, this page is listening too. 😉
While dictating: Esc cancels, Space pauses & resumes, and an optional Refine key lets you pick this dictation's polish level on the fly.
In a conversation: Space ends the turn, S looks at the screen again, T sends the selection, Enter writes the text.
Pick your flavor. Every path ends with you talking to your computer.
Builds a static whisper-cli for offline transcription out of the box.
Get the package on the releases page.
Works on Debian, Ubuntu and Arch — X11 and Wayland are auto-detected.