# On-device dictation

> Download a recognition model and type speech into the field you are using.

Dictation is an experimental feature, disabled by default. It records when you use its shortcut, transcribes on your Mac and inserts text at the cursor.

## Set it up

1. Open **Settings → Dictation** and enable dictation.
2. Grant Microphone and Accessibility when requested.
3. Select a microphone and an on-device model.
4. Download the model from Settings; model weights are not bundled with the app.
5. Record a shortcut and choose toggle or hold-to-talk behavior.
6. Try a short sentence in a disposable text field.

## Local models

| Family | Variants | Language coverage documented by the app |
| --- | --- | --- |
| NVIDIA Parakeet | Redux, the default lightweight option; Ultra for quality | 25 languages |
| Qwen3-ASR | 0.6B and 1.7B | 30 languages and 22 Chinese dialects |

Redcast uses Core ML and Accelerate in its bundled helper. The first use can take longer while model components are prepared. Download and disk requirements are shown by the model settings; keep enough space for the selected model.

## Record and insert

In toggle mode, use the shortcut to start and finish. Hold-to-talk supports a key chord or a single physical modifier, including Globe. The recording capsule shows the active recording. Keep the destination field focused while speaking.

The finished transcript is inserted through the same text-insertion path as snippets. [Dictation History](/docs/voice/history) keeps the final text when retention is enabled, even if delivery to the destination fails.

## Local versus cloud

On-device mode does not send your recording to an AI provider. [Cloud transcription](/docs/voice/cloud) is a separate opt-in and uses an explicitly approved model. Switching between them does not change the meaning of the dictation shortcut: both produce text at the cursor.

## Troubleshooting

- No recording: check the microphone selection, macOS Microphone permission and shortcut conflict.
- Words recognized but not inserted: check Accessibility and try a simple editable field.
- Model not ready: finish its download before dictating; first initialization can take time.
- Wrong language or names: select a suitable model or use cloud transcription's language and spelling hints.
- Transcription failed: open Dictation History and retry if its recording is still retained.

Cancel a recording you do not want to submit. Silence and cancelled recordings do not create history entries.