Voice dictation

Dictation transcribes microphone audio into the composer. Configure it under Settings → Voice.

Choose an engine

Dictation engine offers:

• OpenAI — uses gpt-transcribe and the saved OpenAI key.
• OpenRouter — uses gpt-transcribe through OpenRouter and its saved key.
• Qwen3-ASR (on-device) — runs Qwen3-ASR inference on this machine after model download.
• Local — runs English Whisper inference on this machine.
• Qwen3-ASR (local server) — optional local ASR HTTP server

The engine is read each time the microphone stops, so changing it does not require a restart. The mic does not inherently require an OpenAI key; its requirement follows the selected engine.

Waveform chooses the listening visualizer in the composer: Bars, Dots, Line, or Mirror. It does not change transcription.

Use Local

Local remains disabled until at least one model is installed:

• Whisper Tiny — Fast, English, about 41 MB.
• Whisper Small — Recommended, English, about 249 MB.

Use Install Whisper Tiny or Install Whisper Small. Installation does not switch Dictation engine to Local by itself.

When a model is loaded, Unload Whisper Tiny/Small releases its runtime memory. Delete Whisper Tiny/Small cache removes the downloaded files. The UI can mark a model as recommended for the current machine, but never installs it automatically.

Dictate

Use the composer microphone or Ctrl+M / ⌘M. Start recording, speak, then stop. The transcript is inserted into the composer for review before send.

Always review names, code identifiers, paths, and commands. Transcription output is user prompt text; it is not executed until you send it and the selected mode permits the resulting work.

Privacy boundary

OpenAI and OpenRouter receive the recorded audio for transcription. Local keeps inference on this machine after model download. Local model files are cached under app user data.

For microphone, missing-key, download, load, or model errors, use Indexing and dictation issues.