# Local dictation

Nock can recognize speech on your computer using a downloaded model. This keeps the recording audio on the device for recognition. Open **Settings → Dictation → Where your voice is turned into text** to choose **On this computer**, or keep **Automatic** for local-first routing with an eligible cloud safety net.

## Download a speech model

1. Open **Settings → Dictation**.
2. Find **Where your voice is turned into text**.
3. Under **Speech models**, choose **Download** for a model that supports your language.
4. Wait for the download and verification to finish. You can pause or resume a download, and use **Delete** to remove a downloaded model.
5. If more than one model is installed, choose **Use** beside the one you want.

**Standard** is the smaller English model. **Higher accuracy** supports a broader set of European languages and requires more disk space and memory. Model availability and size appear in the card. Models are separate downloads, not part of the installer.

On Mac, **Recogniser on this Mac** also offers **Nock’s model** or **Apple’s** recognizer. Apple's recognizer stays on-device and can be used when its local language support is installed. On Windows, the local choice is Nock's downloaded model.

## Make local-only a hard rule

Turn on **Never send my voice off this computer**. Nock then uses local recognition only. If the model is missing, the language is unsupported, or the computer cannot run it, dictation stops and shows the reason rather than sending audio to cloud recognition.

This setting covers speech-to-text audio and also blocks **Polished** cleanup, which needs an online model. With local-only enabled, choose **Light** or **Off** instead. When local-only is off, **Polished** can send recognized text for cleanup. Hosted agent requests, cloud spoken replies, account features, updates, downloads, and connected services remain separate network operations; local dictation does not make those features offline. See [privacy controls](/nock/privacy-controls).

## Choose the right route

| Setting                                   | Recognition behavior                                                                                          |
| ----------------------------------------- | ------------------------------------------------------------------------------------------------------------- |
| **Automatic**                             | Local first; an eligible cloud route can be used when local recognition is unavailable or measured as slow.   |
| **On this computer**                      | Requests local recognition and does not choose cloud as its fallback.                                         |
| **Cloud**                                 | Requests the configured cloud speech route; availability depends on account access or your own supported key. |
| **Never send my voice off this computer** | Overrides cloud selection for dictation; if local recognition cannot run, the turn refuses.                   |

An in-progress recording stays on the route it started with. A failed cloud connection does not guarantee a mid-recording switch to local recognition.

## Platform notes

Nock's downloaded dictation model requires macOS 14 or later on Mac. On macOS 13, the model is unavailable; Apple's on-device recognizer may still be available. Windows uses Nock's model. The Mac's **Your words** list is not passed to Apple's recognizer. On the local Nock model, terms may be corrected after recognition; on cloud speech they may be sent as hints.

If the model cannot download, check the card's status and connection, then use **Try again** if shown. If the downloaded model cannot run or does not know your language, read the status beneath the recognition choices before choosing another route. For guided checks, see [dictation troubleshooting](/nock/dictation-troubleshooting).