What happens to your audio
Per Wispr Flow’s security and compliance FAQ (reviewed September 2026), dictation audio is sent to its cloud for transcription. An account and a network connection are part of the product’s architecture.
Its current data-controls documentation separates Improve the model for everyone, Dictation Cloud Storage, local history and Context Awareness. Turning one off does not change the others. The security overview notes that Improve the model for everyone was previously called Privacy Mode.
The context boundary
Wispr Flow’s Context Awareness documentation says the feature is on by default. Depending on the platform and the active app, context can include app information, text around the cursor, on-screen text, variable and file names, conversation content and a screenshot. It can be turned off in Data & Privacy, and Enterprise administrators can turn it off organization-wide.
Context Awareness is separate from the model-improvement, cloud-storage and local-history settings. A security review should inspect each control, and the organization policy actually applied to the account.
What “local” means, and how to verify it yourself
A privacy label is not an observable boundary. Read the vendor’s documentation, then watch what the exact build does on the network, including ours.
- Terminal. Run
nettop, built into macOS, and dictate for 30 seconds. Sustained upstream traffic during dictation needs investigating; timing and byte counts alone don’t identify the payload. - Connection monitor. Little Snitch, the Objective Development firewall, shows every connection in real time, per app and per host. Dictate and watch.
- Blocked network. Disconnect once the documented prerequisites are in place. A failure can reveal a dependency, but first rule out missing local models, activation, permissions and unrelated bugs.
LockedIn Flow runs Parakeet TDT v3 on the Apple Neural Engine, and cleanup uses local rules. Contextual cleanup uses Apple’s on-device model on macOS 26 with Apple Intelligence, off until you turn it on. Once the speech model is installed, dictation needs no network connection, and no account is required. The full method is in how to test an offline dictation claim.
Side by side
| Area | LockedIn Flow | Wispr Flow |
|---|---|---|
| Audio processing | On the Mac; never saved, never sent | Cloud |
| Transcript history | In memory until you quit, by default; encrypted on the Mac if kept longer | Dictation Cloud Storage and local-history controls are separate |
| App and text context | The app in front picks the formatting profile; field text is re-read only for opt-in edit learning | Context Awareness is on by default and can be turned off |
| Screenshot context | Not used | Can be included while Context Awareness is on |
| Account | Not required | Required |
| Model-improvement control | None needed: audio and transcripts are never uploaded | Improve the model for everyone |
| Offline dictation | Yes, once the speech model is installed | No |
Stronger on this rowWispr Flow facts from its documentation, Sept 26, 2026
The honest trade-off
Wispr Flow’s cloud buys real things: Windows, iOS and Android support and 100+ languages. LockedIn Flow is for Apple silicon Macs and focuses on 25 European languages. If cross-platform matters more than keeping audio on the device, Wispr Flow is the honest pick. If you live on a Mac and want dictation where “private” is something you can check rather than something you’re asked to believe, that’s why we exist.