Where local processing wins
Local recognition keeps recordings on the device, works after setup without a connection, and avoids a per-request provider dependency. It is a strong default for confidential writing, unreliable networks, and anyone who wants the processing path to remain inspectable.
The cost is local work: the model occupies disk space and recognition speed depends on the machine. Smaller models reduce that burden while larger models can favor stronger hardware.
Where cloud processing helps
A cloud provider can reduce the compute required on a laptop and may return results faster than a large local model on modest hardware. That convenience introduces a network dependency and sends the selected data across a provider boundary.
The important product behavior is consent at the layer level. A cloud speech provider needs audio; a cloud cleanup provider may need only transcript text. Those are different disclosures and should be different switches.
A route you can explain
Quill keeps recognition and Scribe cleanup independent. You can keep both local, use cloud recognition with local cleanup, use local recognition with cloud cleanup, or select cloud providers for both.
Before choosing, write down what may leave the device, what happens without internet access, where credentials live, and which mode types directly into your work. If the route cannot be explained in one sentence, it is too easy to misconfigure.
- Local + local: maximum control and offline capability
- Cloud recognition + local cleanup: spare speech-model compute
- Local recognition + cloud cleanup: audio stays local; transcript text may not
- Cloud + cloud: least local compute, largest provider dependency