--- title: 'Where Your Voice Goes: Dictation Privacy, Honestly' excerpt: 'Learn where dictation audio can travel, what privacy questions to ask, and how Keebye keeps speech-to-text and cleanup on your Mac.' date: '2026-07-21' lastModified: '2026-07-21' author: 'Teodor Deleanu' tags: - dictation - macos - privacy keywords: - is dictation private - does dictation send audio to the cloud - private dictation app mac ctaTitle: 'Keep your prompts on your Mac' ctaBody: 'Start your free trial, download a model, then turn Wi-Fi off and check for yourself that dictation keeps working.' --- If your prompts contain unreleased code, client details, or the rough version of a security decision, "is dictation private?" is not a settings question. It is a data-flow question: where does the recording go, what processes it, and what remains after the text reaches your cursor? The answer depends on architecture. Some dictation tools send recorded audio to a server. Others run speech-to-text on your device. Either design can be explained clearly. The useful question is not whether a product uses the word _private_, but whether you can draw its full path from microphone to text. ## Where audio can travel in a dictation pipeline A cloud-transcription path usually looks like this: `microphone -> recorded audio -> vendor server -> speech-to-text -> optional AI cleanup -> text back to your Mac` Sending audio to a server is not sinister. It is how a cloud architecture produces a transcript, and the same server path often handles AI cleanup before returning the text. The trade-off is concrete: your voice and everything you said must transit third-party infrastructure. The vendor's policies determine how that data is processed, whether anything is stored, and for how long. Cleanup deserves its own line in the diagram. A tool can run speech-to-text locally, then send the transcript to a server for AI rewriting. In that case, the audio may stay on your computer while the contents of your speech still leave it as text. An on-device path can be shorter: `microphone -> local speech-to-text -> local cleanup -> text field` Local transcription removes the server from the speech-to-text step. It does not answer every privacy question. The app could still send telemetry, sync transcripts, call a cloud model for cleanup, or contact a backend for account checks. "On-device" tells you where one part runs. You still need the rest of the map. ## Five questions to ask any dictation app 1. **Where does speech-to-text run?** Ask whether recorded audio goes to a server for transcription. "Encrypted in transit" answers how it travels, not whether it travels. 2. **Where does cleanup run?** If the app fixes grammar, removes filler words, or rewrites a transcript, find out whether that work happens locally or through a cloud model. 3. **Does the app send telemetry?** Crash reports, usage analytics, and diagnostics are separate from transcription. Ask what each report contains and whether you can disable it. 4. **What is stored, and for how long?** Separate audio from transcripts. Check the device, the vendor's servers, backups, and any history or sync feature. 5. **What happens in password fields?** A dictation app should recognize secure fields and refuse to insert text there. Otherwise spoken secrets can pass through the normal dictation path or appear in history. These questions work for any vendor. If the cloud-versus-local question is driving your buying decision, [the Keebye vs Wispr Flow guide](/compare/keebye-vs-wispr-flow) frames the trade-off without asserting what you should value most. ## Where Keebye sends your voice Keebye's dictation path stays on your Mac. All three speech-to-text options run on-device: Parakeet is the default English engine, Canary is the opt-in 25-language engine, and Apple native is also available. Rule-based cleanup and the optional local-LLM polish stay on the Mac too. The dictation path makes no network calls. Keebye connects for model downloads, app update checks, and sign-in plus trial or plan checks -- never to send audio or transcripts. Once you have downloaded a model, transcription and cleanup keep working offline. [The offline dictation guide](/blog/dictation-that-works-offline) explains that dependency boundary in more detail. Keebye has no in-app telemetry. Dictation history stays on your Mac, stores text rather than audio, and automatically deletes entries after 30 days. Keebye also refuses to dictate into secure or password fields by design. That boundary matters when your voice is carrying more than a grocery list. If you use dictation to write agent instructions, code-review notes, or prompts about unreleased work, the [dictation for AI prompts guide](/dictation-for/ai-prompts) covers the workflow around those sensitive inputs. ## Test the offline claim yourself You can observe whether dictation keeps working without a network connection. Start the trial and sign in, download a speech model, then turn Wi-Fi off. Dictate into an ordinary text field. The transcript still appears because speech-to-text and cleanup use the local components already on your Mac. This test demonstrates offline operation after the model download. It does not verify backend requests, storage, telemetry, or the app's full privacy posture. It is not a substitute for a full security audit, but it is a useful check that a privacy claim changes observable behavior. ## Limitations Keebye is early access and macOS-only. If you need a production-hardened release or another operating system today, it is not the right fit yet. Keebye connects to our backend for sign-in and trial or plan checks, never for audio or transcripts. That connection supports account operations even though dictation itself works offline after the model download. The keebye.com website uses standard web analytics. The Keebye app itself has no telemetry. Website measurement and app behavior are separate systems, and the distinction matters. On-device transcription can still make mistakes. Local processing changes where your data goes; it does not make speech recognition infallible. Review sensitive prompts, commands, and messages before you send them. Privacy claims should be specific enough to inspect. For Keebye, offline operation is one observable claim: download the model, disconnect the network, and keep dictating.