Android
On-device mode (offline)
Run speech recognition and AI text actions locally on your Android phone, without sending audio to the cloud.
Updated 2026-10-06
On-device mode runs speech recognition on your phone. After a one-time model download, dictation works without an internet connection and your audio is not sent to Dictāta's cloud service.
On-device mode is available on Android only.
Before you start
- You need an internet connection once, to download the model.
- Models are stored on the phone. Check that you have enough free storage (see the sizes below).
- On-device mode does not use your Dictāta cloud minutes.
- Speed and accuracy depend on your phone. Older or low-memory phones work best with the Fast model.
Turn on on-device transcription
- Open the Dictāta app and tap Account in the bottom menu.
- Tap Settings in the top right corner.
- In the Provider section, tap Transcription service.
- Change the setting from Managed to On-device.
- When the On-device model section appears, tap Selected model and choose a model.
- Tap Download model and wait until the status says Ready.
Now dictate with the keyboard as usual. If you start dictating before the model is downloaded, the keyboard shows "Download the on-device model in Dictāta settings to enable local dictation".
Choose a speech model
| Model | Download size | Best for |
|---|---|---|
| Fast | about 104 MB | Older phones, short dictation, lowest delay |
| Balanced (recommended) | about 161 MB | Good speed with better accuracy than Fast |
| Accurate | about 315 MB | Higher accuracy, a bigger download and more processing time |
| Experimental AI | about 2.6 GB | Large on-device audio model. For testing, not recommended as the default |
You can download several models and switch between them. Use Delete model to free space.
On-device AI model for rewrite and prompts
Speech recognition and AI text actions use separate models. To use prompts, AI Enhancement, rewrite and Context Actions in on-device mode, download the on-device AI model as well:
- In the same settings screen, find the On-device AI model section.
- Tap Download AI model. The model, Gemma 4 E2B, is about 2.6 GB, so use Wi-Fi.
- Wait until the AI model status says Ready.
Without this model, dictation works but AI text actions do not.
Switch back to cloud
Return to Account -> Settings -> Provider -> Transcription service and choose Managed. Downloaded models stay on your phone until you delete them.
Offline fallback
If your connection drops while you use the cloud service, the keyboard shows "No connection" with a Switch to Local button. This works once you have a model downloaded, so it is worth downloading one in advance.
Troubleshooting
Download fails or stops
Check that you have enough free storage and a stable connection, then tap Download model again. Failed downloads show a "Model download failed" message.
The keyboard says to download the model
The provider is set to On-device but the selected model is not downloaded. Open Account -> Settings and tap Download model for the model shown in Selected model.
The first dictation is slow
Local AI models need time to start the first time they are used. Keep the keyboard open for a moment before speaking, and try a smaller model if the delay stays long.
Quality is lower than cloud mode
On-device models are smaller than the cloud models. Try Balanced or Accurate, speak clearly, and add difficult names to your Glossary. If you need the highest accuracy, use Managed mode.
Free up space
Open Account -> Settings and tap Delete model or Delete AI model.