Does dictation software record your voice? It depends on the tool. Cloud-based apps transmit your audio to remote servers, where it may be stored, reviewed by humans for quality control, or used to retrain models. On-device engines process speech locally and discard the audio the instant text appears. Check the retention policy first.
That single-sentence answer covers the mechanics, but the real question people are asking is more specific: who actually listens to my dictation, how long is it kept, and can I stop it? The honest answer touches three separate issues that make up voice data privacy: where processing happens, what a company’s retention policy says, and whether a human review programme exists behind the scenes. Apple, Google and Microsoft have all run human review programmes on voice data at some point, and each has changed its practices after public scrutiny. This article walks through what’s verified, what’s changed, and how to check any dictation tool — including ours — before you trust it with sensitive dictation.
Does Dictation Software Record Your Voice? On-Device vs Cloud Explained
Every dictation tool needs to turn sound waves into text, and there are only two places that conversion can happen: on your device, or on a server somewhere else. The difference sounds technical, but it decides who has access to your voice.
Cloud-based dictation records your audio, compresses it, and sends it over the internet to a data centre for processing. The transcript comes back a moment later. This is how many consumer voice assistants and dictation apps still work today, because cloud servers can run larger models than a phone or laptop can handle alone. The trade-off is that your voice — an audio file, not just text — now exists on someone else’s infrastructure, subject to their retention policy, security practices, and any human review programme they run for quality control.
On-device processing keeps the entire journey local. The audio never leaves your Windows PC, Mac, phone, or browser; the speech model runs on your own hardware, converts speech to text, and — in a well-built implementation — discards the audio the moment it has done its job. Nothing is uploaded, so there’s nothing sitting on a server for a hacker, a subpoena, or an employee to find later.
| Factor | Cloud Processing | On-Device Processing |
|---|---|---|
| Audio leaves your device | Yes, transmitted to servers | No |
| Typical retention | Days to months, depending on policy | None — discarded after transcription |
| Human review risk | Possible, if a QA programme exists | Not possible — no one holds the file |
| Works offline | No | Yes |
| Raw accuracy ceiling | Can be higher with large models | Improving fast, hardware-dependent |
The Human Review Programmes Apple, Google and Microsoft Ran on Voice Data
This isn’t a hypothetical risk — it happened, and it happened to three of the biggest technology companies at once. Human review of voice recordings has a real, documented history: in 2019, reporting revealed that Apple, Google and Microsoft had all been running programmes where contractors listened to snippets of voice assistant recordings to grade accuracy and improve their models.
Apple’s programme, called “Siri grading,” had contractors review a small sample — under 0.2% of Siri requests — including recordings accidentally triggered by background noise, some of which captured private medical conversations. Apple suspended the programme, later limited human review to its own employees, and made grading opt-in. Apple went on to pay a $95 million settlement in a related class action.
Google’s version came to light when a contracted language reviewer leaked over 1,000 Dutch-language Assistant recordings to a Belgian broadcaster, some containing addresses, medical details and a recording of a person in apparent distress. Google confirmed roughly 0.2% of audio snippets were reviewed for transcription quality and suspended EU reviews while it investigated. More recently, in October 2025, Google agreed to a record $1.375 billion settlement with the Texas Attorney General covering several privacy claims, including the collection of biometric identifiers — voiceprints among them — through Google Assistant.
Microsoft ran a similar programme for Cortana voice commands and Skype Translator calls, with contractors transcribing clips that sometimes included addresses and personal conversations, and has since updated its privacy disclosures to state explicitly that human review of voice data can occur.
None of this means these companies are uniquely careless — it means voice data is genuinely hard to govern once it leaves your device, and every major platform has had to learn that lesson in public.
What GDPR Voice Data Rules Actually Require
Under UK and EU data protection law, your voice isn’t automatically a “special category” of personal data — but it can become one quickly. The UK’s Information Commissioner’s Office confirmed in its 2024 biometric data guidance that a voice recording qualifies as biometric data, and therefore Article 9 special category data, once it is technically processed in a way that can identify or verify who is speaking.
That distinction matters in practice. Once voice data crosses into special category territory, a company generally needs explicit consent or another specific legal basis to process it, plus a documented Data Protection Impact Assessment if the processing is large-scale. A privacy policy that vaguely mentions “improving our services” doesn’t clear that bar.
For anyone dictating client notes, patient information, or confidential business communication, the practical takeaway is this: a tool that keeps your voice on-device sidesteps most of this exposure entirely, because there’s no biometric data being transmitted or stored for regulators to examine in the first place. A tool that uploads audio to the cloud is playing by a stricter rulebook — and you’re entitled to ask exactly how it complies.
Why On-Device Processing Is Becoming the Standard
The direction of travel across the industry is unmistakable. Apple has been shifting more of its speech processing onto its Neural Engine, running newer dictation features locally rather than routing audio to a server by default. Google has done the same with on-device models built for phones, aimed at handling short commands and dictation without a round trip to the cloud.
Two forces are pushing this shift. The first is hardware — phones and laptops now carry enough dedicated processing power to run genuinely capable speech models locally, which wasn’t true a decade ago. The second is regulation and reputation: GDPR, UK data protection law, and a growing list of US state privacy laws all make cloud audio processing more expensive and legally riskier than it used to be, and the settlements above turned that risk from theoretical into very real numbers.
None of this means every on-device claim is equally trustworthy. “Processed on-device” and “zero data retention” are different promises, and some tools blend both approaches depending on the feature, so it’s worth checking exactly which claim a given product is making rather than assuming.
How to Check Whether Your Dictation Software Is Safe
If you’ve wondered whether dictation software is safe enough for confidential work, you don’t need a law degree to find out — a few minutes with the privacy policy and settings menu tells you most of what you need:
- Search the privacy policy for “retain,” “store,” and “human review.” If audio retention isn’t mentioned at all, that’s a gap, not reassurance.
- Look for an explicit on-device or “local processing” claim. Vague language like “we take your privacy seriously” isn’t a technical commitment.
- Check whether there’s an opt-out for quality review or model training, and whether it’s on or off by default.
- Look for independent compliance claims — GDPR compliance, HIPAA readiness, or similar — rather than taking marketing copy at face value.
- Test it offline. If dictation still works with your Wi-Fi turned off, the processing is happening on your device, not a server.
That last test is the simplest one, and it works on almost anything with a microphone. If you handle client calls, notes, or confidential communication for a living, running it is worth thirty seconds. If you handle patient records specifically, it’s worth reading a dedicated breakdown of medical dictation software and what compliance actually requires there, since the stakes and the rules are higher than for general writing.
Where Genie 007 Fits Into This
We built Genie 007 around the on-device end of that comparison, for the reasons above. Genie 007 processes your dictation and commands locally — your boss can’t see it, and we can’t see it either. There’s no audio storage and no logging: nothing about what you say is routed through our servers to be read, graded, or used to train anything. It’s AES-256 encrypted, GDPR compliant, and built to HIPAA-ready standards for anyone dictating in a regulated field.
That privacy foundation is only half the story, though. Plain dictation — even private, on-device dictation — still hands you a literal transcript of every “um,” false start, and clunky sentence you spoke: private, but not necessarily useful yet. Genie Mode goes a step further. Instead of transcribing word-for-word, you speak your intent — “draft a polite decline to this meeting request” — and Genie 007 writes the finished message in the right tone for wherever you’re sending it, without the audio ever leaving your device. If a previous cloud-based tool is what worried you in the first place, our guide to offline dictation software compares local-processing options in more depth. If privacy is a business-wide concern rather than a personal one, dictation software for small business covers what to look for when a whole team dictates client-facing work. And for the fuller idea behind turning spoken intent into finished text rather than a raw transcript, what is think-to-text explains the concept end to end.
Does Siri record everything you say?
No — Siri only records and processes audio after it detects its wake phrase or a manual trigger, not continuously. Apple handles many requests using on-device models where possible, though some are still sent to Apple’s servers, and a small sample was previously reviewed by human contractors before Apple changed that policy in 2019.
Do dictation apps store your voice recordings?
It depends entirely on the app’s architecture and retention policy. Cloud-based dictation tools commonly store audio for some period, often for model improvement or quality review, unless they explicitly state a no-retention policy, while on-device tools discard the audio locally the moment it’s converted to text, so nothing is ever stored.
Is on-device dictation more private than cloud-based dictation?
Generally, yes, because the audio never leaves your device, so there’s no file sitting on a remote server that could be breached, subpoenaed, or reviewed by a human. Cloud dictation can still be reasonably secure, but that depends entirely on the provider’s encryption, retention limits, and whether a human review programme exists.
Can dictation software be used for confidential medical or legal notes?
Only if the tool’s privacy architecture actually supports it — look for on-device processing, zero data retention, and GDPR or HIPAA-ready compliance rather than assuming general dictation software is suitable by default. For clinical work specifically, a dedicated medical dictation workflow is safer than a consumer app with vague privacy language.
Try Genie 007 Free
If keeping your voice off someone else’s server matters to you, Genie 007 processes every word on-device, whether you’re dictating notes, drafting an email, or refining a whole document by voice.
Download Genie 007 for Windows and try it free, or compare plans on the pricing page first. No credit card required.
Written by Bill Kiani, founder of Genie 007.
