# Voice dictation

Scient desktop transcribes a short voice recording on the local computer. You
can optionally ask the selected AI provider to correct the resulting transcript
text before Scient inserts or sends it.

1. Select the microphone in the chat composer.
2. On first use, choose **Multilingual Small** for faster setup,
   **Multilingual Medium** for higher expected accuracy, or
   **Multilingual Turbo** for the most demanding dictation on a powerful
   computer. Scient downloads and verifies the selected local model. You can
   cancel setup and resume the download later.
3. Allow microphone access when macOS, Windows, or Linux asks.
4. While recording, choose cancel, insert the transcript, or transcribe and
   send. `Esc` cancels and `Enter` inserts.

If microphone access was denied on macOS, select **Open Settings** beside the
voice error, enable Scient under **Privacy & Security → Microphone**, then quit
and reopen Scient. A downloaded model does not itself grant microphone access.
If the Mac restricts microphone access through parental controls or device
management, those restrictions must be resolved before recording can start.

Dictation continues the current draft without forcing a new line. Existing
spaces and paragraph breaks you entered are preserved. Speech-recognition
segment boundaries do not introduce paragraph breaks.

Audio and the initial transcript are processed by a bundled local speech model.
Audio remains on the local computer and is never sent to Scient or an AI
provider.

**Correct transcripts with an LLM** in **Settings → Voice** is off by default.
When it remains off, the transcript stays local. When you turn it on, Scient
sends the transcript text—not the recording—to the selected, authenticated AI
provider, together with the selected language when one is set. While correction
is pending, you can use the original local transcript immediately. If the
provider is unavailable, unsupported, not authenticated, times out, or returns
an invalid correction, Scient falls back to the exact local transcript.

The model download and an enabled provider correction are therefore the two
possible network paths for voice dictation. Provider correction follows that
provider's own service and data-handling terms.

Voice dictation is available in the desktop app. The macOS voice runtime
requires macOS 12 or newer. A recording is limited to three minutes. If no
speech is detected, Scient leaves the draft unchanged.

To change models later, open **Settings → Voice**. The page shows which model
is recommended for the current machine, which model is in use, download
progress, and removal controls.