The local path
A local dictation path records a short audio segment, runs it through a speech model stored on the device, and inserts the resulting text. Once the model is installed, recognition does not depend on a network connection.
Quill uses local speech recognition by default. Its cleanup layer is a separate decision, so choosing local recognition does not silently enable a cloud rewriting service.
What your hardware changes
Larger speech models generally require more disk space and more compute. A smaller model can be the better everyday choice on a laptop because it returns text quickly without demanding a dedicated GPU.
Quill offers a range of local model sizes and falls back to CPU processing when a compatible accelerator is unavailable. Optional cloud recognition is an escape hatch for users who prefer to spare local hardware; it is not the default route.
- Disk space for the selected speech model
- CPU or GPU time while recognition runs
- A one-time model download before offline use
- A visible choice before any cloud provider is used
How to verify an offline claim
Install the model, disconnect the machine from the network, and dictate a new sentence into an ordinary text field. The test should complete without a provider login or API key.
Also inspect where provider keys are stored and whether saving a key activates it. Quill keeps supported cloud keys in Windows Credential Manager or macOS Keychain, and a saved key remains inactive until its provider is selected.