What stays on your device when you speak?
A practical explanation of local transcription, model downloads, optional text processing, and the questions to ask about voice privacy.
Follow the audio
A voice app has several distinct jobs: capture sound, turn it into text, save that text, and help you use it. Calling the whole product ‘local’ obscures the boundaries between those jobs. The useful question is what crosses each boundary.
SaidClear performs speech recognition on your device. A local model turns recorded audio into a transcript. Downloading that model requires a connection; using an installed model does not require sending the recording to a transcription service. The first download and an ordinary recording therefore have different network behavior.
A transcript deserves the same care as a recording
Text can carry the same confidential details as sound: a customer name, an unreleased result, or a difficult conversation. Local transcription alone does not make subsequent cloud processing private.
In SaidClear, optional sync sends text and metadata to the workspace you pair with. Hosted summaries, questions, and translations are a separate choice: approved text goes to DeepSeek. Your original remains available. Before enabling a provider, review its processing terms and your organization’s policy; do not assume that ‘AI’ implies a particular retention window.
Check the whole lifecycle
An export you share, an operating-system backup, and a deleted account are separate parts of the lifecycle. Look for a local erase control as well as account deletion, and check whether deletion covers transcript checkpoints and queued sync entries. Removing a visible row is not the same as removing its stored copies.
A practical evaluation is to download the model, disconnect, record a short sample, and check that the transcript remains available after restarting. Then enable each network feature separately with synthetic text. A product should be able to explain every transfer without asking you to trust a slogan.