The short answer
Voice transcription is private when your audio and transcript never leave your machine. Every other privacy promise (encryption in transit, retention policies, compliance badges) is a mitigation for the risk created by sending your voice to someone else's server. This guide explains where cloud transcription sends your data, who should care, what to ask any vendor, and how to verify a privacy claim yourself instead of taking it on trust.
Where cloud transcription actually sends your voice
Most dictation and transcription tools work the same way: the app records audio, uploads it to the vendor's servers, a model transcribes it there, and the text comes back. That round trip is fast and invisible, and it means your spoken words exist, at least briefly, on infrastructure you do not control. Four things follow from that:
- Retention you do not fully control. Vendors keep audio and transcripts for varying periods, under policies that can change. Deleting your copy does not necessarily delete theirs.
- A larger breach surface. Your words are only as safe as the vendor's security, and a transcription service is an attractive target precisely because of what people dictate into it.
- Possible use in training. Some services reserve the right to use customer audio to improve their models. The settings that opt you out are not always on by default.
- Third-party access. Data held by a vendor can be reachable by that vendor's staff, subprocessors, and legal process, none of whom you chose.
None of this means cloud services are careless. It means the privacy question is structural: once your voice leaves your machine, you are trusting someone else's controls.
Who should care the most
Everyone benefits from keeping their voice to themselves, but for some professions this is a duty rather than a preference:
- Legal: dictated memos, case notes, and correspondence are often privileged. Sending privileged material through an outside transcription vendor adds a party your client never agreed to. See our guide to dictation for lawyers.
- Healthcare and therapy: patient and session notes are among the most sensitive data there is. Keeping them on-device means there is no transcription vendor to vet, audit, or contract with for that step.
- Finance and business: deal notes, client details, and internal plans dictated into a cloud tool sit outside your company's security perimeter.
- Journalism and research: source conversations and unpublished work deserve the same protection as the people behind them.
The questions to ask any transcription vendor
If you are evaluating a dictation or transcription tool, these five questions separate marketing from architecture:
- Is my audio processed on my device, or uploaded to your servers?
- If it is uploaded: how long do you keep audio and transcripts, and can I verify deletion?
- Is my audio or text ever used to train models? Is that opt-in or opt-out?
- Which subprocessors touch my data?
- Can I verify any of this myself, from the outside?
A local-first tool answers the first question with "neither audio nor transcripts leave your machine", which makes most of the remaining questions moot for the transcription step itself.
Local processing: the structural answer
Modern speech recognition no longer requires a data center. Open models like Whisper run well on an ordinary Windows PC, which means the transcription step can happen entirely on your own hardware. When transcription is local, privacy stops being a policy and becomes a property of the architecture: there is no vendor server holding your words, nothing to retain, nothing to breach upstream, and nothing to train on. If you want the technical detail, read how local speech-to-text works.
How to verify a privacy claim instead of trusting it
"We don't upload your audio" is a testable claim, and you should test it:
- The airplane-mode test. Disconnect from the network entirely and dictate. A genuinely local tool keeps transcribing; a cloud tool stops working.
- Watch the wire. Tools like Wireshark or GlassWire show every connection an app makes. A local-first app should show no traffic during dictation.
- Read the vendor's own disclosure. A vendor confident in its architecture will document every network call its app makes. MurmurTone publishes exactly that on its network transparency page: update checks, license validation, optional model downloads, and the in-app purchase flow, and nothing else. Dictation itself makes no network calls.
A practical privacy checklist for dictation
- Prefer tools that transcribe on-device; verify with the airplane-mode test.
- Check what the app itself phones home for (updates and licensing are normal; audio upload is not).
- Confirm there is no analytics or telemetry you did not agree to.
- Treat transcripts with the same care as any other sensitive file on your machine: disk encryption and good account hygiene still matter.
- For regulated work, confirm your organization's own policies before adopting any new tool, local or not.
Where MurmurTone fits
MurmurTone was built so the privacy answer is structural. It is a Windows voice-to-text app that runs 100% offline: your voice is transcribed on your own machine by local Whisper models and is never uploaded. The app contains no analytics and no telemetry, and every network call it does make (update checks, license validation, optional model downloads, the purchase flow) is documented publicly. It supports 18+ languages, can translate speech to English, and uses your NVIDIA GPU for speed if you have one, all without a connection.
Try it yourself. Download MurmurTone with a 14-day free trial (no credit card), run the airplane-mode test, and watch it keep working.