For the local speech and default cleanup path, the reviewed SuprFlow app uses a downloaded model on your Mac to turn speech into text. It does not show a step that sends your recording or transcript to SuprFlow to train a model. That is the narrow technical answer. It should not be stretched into “nothing ever leaves the device” or “every AI provider is contractually barred from training on my words.”
If you switch on hosted AI formatting, SuprFlow sends transcript text and formatting instructions to the provider or custom endpoint you chose. You then need that provider's current terms for the exact API account, model, region, and plan. A bring-your-own-key setting makes the request direct from your Mac; it does not remove the provider from the data path.
First, identify which AI operation you mean
| Operation | What happens in the reviewed app | The question to ask |
|---|---|---|
| Speech recognition | A downloaded WhisperKit model runs on the Mac. | Is this the model and signed release installed on our machines? |
| Default text cleanup | Local rules process the recognized words. | Is Cloud API formatting disabled in the active mode? |
| Ollama formatting | Transcript text is sent to a configured Ollama endpoint, localhost by default. | Is the endpoint actually an approved local server? |
| Cloud API formatting | Transcript and prompt are sent to the selected provider's API. | What do that provider's current retention, training, logging, and subprocessor terms say for our account? |
| Usage telemetry | SuprFlow 1.0.0 sends events through TelemetryDeck. Its event schema includes a mode-name field, though the app does not expose a custom-mode editor. An API error response can be forwarded in truncated form, and some model identifiers are editable. | Have free-form fields been removed or sanitized and a signed release tested for content leakage? |
| Local history | Recent transcriptions are written to a file on the Mac. | Does our device, backup, and retention policy allow this copy? |
“No model training,” “no cloud storage,” “no content transfer,” and “no local history” are four different requirements. Turning one off does not prove the others. This distinction also appears in Wispr Flow's own data-preferences documentation, which separates model improvement, cloud storage, and local history rather than treating them as a single privacy switch.
How to set up a local-only content path
Use a Mac that meets SuprFlow's current requirements and install the approved signed build. Download the speech model before the offline test. In the app, choose the default local cleanup, and inspect any mode that can override the formatter. Do not select Cloud API. If you use Ollama for additional cleanup, verify the actual endpoint is a local service controlled by your organization; a custom URL can be remote. Dictate synthetic sample text and verify insertion in the intended application. Record the app build, model, settings, and test result so the decision can be repeated after updates.
This configuration is about dictation content. It does not imply that an account/license check, update check, model download, or release telemetry never makes a network request. Review those separately. Also decide how much transcript history and clipboard history the Mac should retain.
What a firm can and cannot promise
A product team can describe and test the software's observed behavior. It cannot make a blanket contractual promise on behalf of every hosted model provider a user might configure. If a bank, law firm, audit practice, or investment team requires no training or no external processing, write that requirement into its approved configuration, provider contract where applicable, device policy, and staff instructions. Recheck it when the app, provider, endpoint, or plan changes.
Do not equate “on-device” with automatic legal compliance. A locally generated note can still be exposed through a cloud document, a shared clipboard, a lost laptop, or a mistaken recipient. A person should verify confidential names, numbers, and claims before the result becomes a client record.
Does SuprFlow itself train a model on my recordings?
The local recognition and cleanup path in SuprFlow 1.0.0 performs inference on the Mac; it does not show an upload of dictated content to SuprFlow for model training. Confirm the installed release and published data policy before treating that observation as a production guarantee.
Can a hosted formatter use my transcript for training?
Its provider terms decide. SuprFlow sends the transcript text and prompt directly to the selected endpoint in the Cloud API path. Check the provider's current contract and settings; do not assume the answer from SuprFlow's marketing.
Is telemetry the same as training data?
No. SuprFlow's telemetry is an operational event pipeline, not an intended model-training feed. But that does not make it content-free: it can send a truncated provider error response that might echo dictated text, and some model identifiers are editable. Its mode-name field would be another user-content risk if custom modes become reachable; the app does not currently show that editor. Ask for the telemetry fields of the signed release you install. An organization may prohibit telemetry regardless of training use.