Voice Dictation
Speak your message instead of typing it — the transcript is inserted at the cursor in the message box.
Voice dictation turns speech into text in the message box. Press the mic button (or the shortcut), speak, and stop; the transcript is inserted where your cursor is. Nothing is sent to the agent until you send the message yourself, so you can edit the text first.
New in 0.14.2
Voice dictation is available from Fabric Agents 0.14.2.
It uses your own OpenAI API key, so it needs a one-time setup.
Turning it on
The mic button is always in the message box, next to the Prompt Enhancer. Until dictation is set up it is dimmed; click it (or open Settings → Input → Voice dictation) to set it up:
- Switch on Enable voice dictation.
- Under API key, either:
- choose Voice key and paste an OpenAI API key — it is checked against OpenAI before it is saved, and stored encrypted on this computer, or
- pick an existing OpenAI API-key connection to reuse its key. ChatGPT sign-in connections can't be used: their tokens can't call the transcription API.
- Optionally pick a Model and a Language.
From then on the mic button is active.
Dictating
| Action | macOS | Windows / Linux |
|---|---|---|
| Start | Mic button or ⌘ ⇧ M | Mic button or Ctrl + Shift + M |
| Finish | ✓, Enter, the mic button or the shortcut | ✓, Enter, the mic button or the shortcut |
| Cancel | ✕ or Esc | ✕ or Esc |
While you speak, the message box shows a live waveform of your voice, a timer and the ✕ / ✓ buttons. When you finish, it shows Transcribing… until the text is inserted at your cursor; anything you had already typed stays as it is, and pressing Enter to finish never sends the message. You can still cancel while it transcribes.
- Shortcut behavior in settings switches between press to start and stop and hold to talk (records while the shortcut is held).
- Taps shorter than a third of a second are ignored, and a recording stops on its own after 5 minutes (the timer turns red in the last 30 seconds).
- The shortcut only applies to the focused chat panel.
Is my key working?
Under API key, Settings shows the key in use — only its first characters and last four, like sk-proj-…a1b2 — with when it was added and whether it works:
| Status | Meaning |
|---|---|
| Working | OpenAI accepted the key. |
| Rejected by OpenAI | The key is wrong, revoked, or has no access to the chosen model. Paste a new key. |
| Couldn't check | OpenAI couldn't be reached (for example, offline). The key may still be fine. |
The key is checked when you open Settings (if the last check is more than 10 minutes old), when you press Check again, and every time you dictate.
Models
| Model | Notes |
|---|---|
gpt-4o-transcribe | Default. Most accurate. |
gpt-4o-mini-transcribe | Cheaper, slightly less accurate. |
whisper-1 | The original Whisper model. |
Language takes an ISO code such as en or de. Leave it empty and the language is detected automatically; setting it helps with short recordings.
Microphone access
- macOS asks for microphone access the first time you dictate. If you declined, allow Fabric Agents under System Settings → Privacy & Security → Microphone.
- Windows and Linux use the system's default input device; make sure it is enabled and not muted.
Privacy
- Audio is recorded only while dictation is active, and the microphone is released as soon as you stop.
- The recording is sent from the app straight to OpenAI for transcription and is not stored by Fabric Agents.
- Your API key stays on this computer; it is never shown again in settings and is not sent to the agent or to any server other than OpenAI.